paper_id
uint32
0
3.26k
title
stringlengths
15
150
paper_url
stringlengths
42
42
authors
listlengths
1
21
type
stringclasses
3 values
abstract
stringlengths
393
2.58k
keywords
stringlengths
5
409
TL;DR
stringlengths
7
250
submission_number
int64
1
16.4k
arxiv_id
stringlengths
10
10
embedding
listlengths
768
768
github
stringlengths
26
123
400
Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional
https://openreview.net/forum?id=QwoGfQzuMa
[ "Sanjeev Raja", "Martin Sipka", "Michael Psenka", "Tobias Kreiman", "Michal Pavelka", "Aditi S. Krishnapriyan" ]
Poster
Transition path sampling (TPS), which involves finding probable paths connecting two points on an energy landscape, remains a challenge due to the complexity of real-world atomistic systems. Current machine learning approaches rely on expensive training procedures and under-utilize growing quantities of atomistic data,...
Transition path sampling, molecular dynamics, generative models, diffusion models, flow matching models
We repurpose generative models for transition path sampling via minimization of the Onsager-Machlup action functional.
14,150
2504.18506
[ -0.007958891801536083, -0.014622140675783157, -0.019496385008096695, 0.062098849564790726, 0.040422793477773666, -0.019099362194538116, 0.03231311962008476, 0.012289047241210938, -0.001217914978042245, -0.05349108204245567, 0.03704293072223663, -0.026775255799293518, -0.057798970490694046, ...
https://github.com/ASK-Berkeley/OM-TPS
401
Long-Term TalkingFace Generation via Motion-Prior Conditional Diffusion Model
https://openreview.net/forum?id=aINERD9MzJ
[ "Fei Shen", "Cong Wang", "Junyao Gao", "Qin Guo", "Jisheng Dang", "Jinhui Tang", "Tat-Seng Chua" ]
Poster
Recent advances in conditional diffusion models have shown promise for generating realistic TalkingFace videos, yet challenges persist in achieving consistent head movement, synchronized facial expressions, and accurate lip synchronization over extended generations. To address these, we introduce the \textbf{M}otion-pr...
Diffusion Model, TalkingFace, Pose
null
14,128
2502.09533
[ -0.013207346200942993, -0.005758811254054308, -0.008088936097919941, 0.032984018325805664, 0.03781864419579506, 0.05976742506027222, 0.044120654463768005, 0.005999595858156681, -0.034153301268815994, -0.04953264445066452, -0.018427005037665367, -0.000046069370000623167, -0.038177233189344406...
null
402
Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching
https://openreview.net/forum?id=6Eg1OrHmg2
[ "Aaron J Havens", "Benjamin Kurt Miller", "Bing Yan", "Carles Domingo-Enrich", "Anuroop Sriram", "Daniel S. Levine", "Brandon M Wood", "Bin Hu", "Brandon Amos", "Brian Karrer", "Xiang Fu", "Guan-Horng Liu", "Ricky T. Q. Chen" ]
Poster
We introduce Adjoint Sampling, a highly scalable and efficient algorithm for learning diffusion processes that sample from unnormalized densities, or energy functions. It is the first on-policy approach that allows significantly more gradient updates than the number of energy evaluations and model samples, allowing us ...
Sampling, Stochastic Optimal Control
We introduce a highly scalable algorithm for learning to sample from only energy functions, the first of its kind in terms of efficiency, which we scale to new benchmarks on conformer generation.
14,121
2504.11713
[ -0.028185531497001648, -0.015682512894272804, -0.006377448793500662, 0.06137854605913162, 0.029825979843735695, 0.00621642405167222, 0.013217817060649395, -0.03527146205306053, 0.004652166739106178, -0.08269938826560974, 0.04277865216135979, -0.03408320993185043, -0.08752160519361496, -0.0...
https://github.com/facebookresearch/adjoint_sampling
403
Confounder-Free Continual Learning via Recursive Feature Normalization
https://openreview.net/forum?id=7zrS5hHlfY
[ "Yash Shah", "Camila Gonzalez", "Mohammad H. Abbasi", "Qingyu Zhao", "Kilian M. Pohl", "Ehsan Adeli" ]
Poster
Confounders are extraneous variables that affect both the input and the target, resulting in spurious correlations and biased predictions. There are recent advances in dealing with or removing confounders in traditional models, such as metadata normalization (MDN), where the distribution of the learned features is adju...
deep neural networks, confounders, continual learning, invariant representations, statistical regression
We introduce the Recursive Metadata Normalization (R-MDN) layer to learn confounder-invariant feature representations under changing distributions of the data during continual learning.
14,097
2507.09031
[ 0.020570527762174606, -0.02781512215733528, -0.014263024553656578, 0.03569818660616875, 0.06475479900836945, 0.030586495995521545, 0.04819322004914284, 0.0022060938645154238, -0.035652946680784225, -0.02538507618010044, -0.011019432917237282, -0.002951475325971842, -0.08722478896379471, 0....
https://github.com/stanfordtailab/RMDN
404
Reliable and Efficient Amortized Model-based Evaluation
https://openreview.net/forum?id=HDbWrsgkB9
[ "Sang T. Truong", "Yuheng Tu", "Percy Liang", "Bo Li", "Sanmi Koyejo" ]
Poster
Comprehensive evaluations of language models (LM) during both development and deployment phases are necessary because these models are thought to possess numerous capabilities as well as safety risks. The average score across a wide range of benchmarks provides a signal that helps guide the use of these LMs in practice...
Model Evaluation, Amortization, Adaptive Testing
We perform reliable and efficient model-based evaluation by introducing amortized calibration and question generator, validating across 22 NLP benchmarks.
14,091
2503.13335
[ -0.010252236388623714, -0.0024230023846030235, -0.020736094564199448, 0.057708051055669785, 0.04161246493458748, 0.03788422420620918, 0.031038284301757812, 0.02259407937526703, -0.026806844398379326, 0.000897026329766959, -0.006661286111921072, 0.05261359363794327, -0.03355974704027176, -0...
https://github.com/sangttruong/reeval
405
Hybrid Spiking Vision Transformer for Object Detection with Event Cameras
https://openreview.net/forum?id=WZKcJZWG3P
[ "Qi Xu", "Jie Deng", "Jiangrong Shen", "Biwu Chen", "Huajin Tang", "Gang Pan" ]
Poster
Event-based object detection has attracted increasing attention for its high temporal resolution, wide dynamic range, and asynchronous address-event representation. Leveraging these advantages, spiking neural networks (SNNs) have emerged as a promising approach, offering low energy consumption and rich spatiotemporal d...
Spiking Neural Networks, Fall Detection, Object Detection With Event Cameras
null
14,090
2505.07715
[ -0.008268560282886028, -0.012937219813466072, -0.008635599166154861, 0.02876235544681549, 0.018999138846993446, 0.03721171244978905, 0.031201161444187164, 0.016608089208602905, -0.05479470640420914, -0.023685680702328682, -0.0212668776512146, -0.0104482751339674, -0.05685483291745186, 0.00...
null
406
PDE-Controller: LLMs for Autoformalization and Reasoning of PDEs
https://openreview.net/forum?id=7epYTVsWEI
[ "Mauricio Soroco", "Jialin Song", "Mengzhou Xia", "Kye Emond", "Weiran Sun", "Wuyang Chen" ]
Poster
We present PDE-Controller, a framework that enables large language models (LLMs) to control systems governed by partial differential equations (PDEs). Traditional LLMs have excelled in commonsense reasoning but fall short in rigorous logical reasoning. While recent AI-for-math has made strides in pure mathematics, area...
AI-for-Math, Large Language Model, Partial Differential Equation
We build an LLM termed "PDE-Controller" that can achieve reasoning and planning on PDE (partial differential equation) control problems.
14,088
null
[ -0.054756809026002884, -0.0038629192858934402, -0.0032740861643105745, 0.03897342085838318, 0.04655412584543228, 0.03211120888590813, 0.015867579728364944, -0.005109860096126795, -0.015021626837551594, -0.035519540309906006, -0.023815905675292015, 0.018396776169538498, -0.04304471239447594, ...
https://github.com/delta-lab-ai/pde-controller
407
Making Hard Problems Easier with Custom Data Distributions and Loss Regularization: A Case Study in Modular Arithmetic
https://openreview.net/forum?id=le8hVvWi6Q
[ "Eshika Saxena", "Alberto Alfarano", "Emily Wenger", "Kristin E. Lauter" ]
Poster
Recent work showed that ML-based attacks on Learning with Errors (LWE), a hard problem used in post-quantum cryptography, outperform classical algebraic attacks in certain settings. Although promising, ML attacks struggle to scale to more complex LWE settings. Prior work connected this issue to the difficulty of traini...
transformers, modular arithmetic, math, cryptography
We introduce two techniques, varying the diversity of training data and introducing a regularized loss function, to improve transformer learning on modular arithmetic, cryptanalysis, and other synthetic tasks
14,087
null
[ 0.008796654641628265, 0.011465875431895256, -0.0051483772695064545, 0.04868736118078232, 0.06328457593917847, 0.010114770382642746, 0.03617896884679794, -0.0351545549929142, -0.012123322114348412, -0.021171975880861282, 0.007699353154748678, -0.01669454760849476, -0.05666843056678772, 0.01...
https://github.com/facebookresearch/arithmetic
408
Attention-Level Speculation
https://openreview.net/forum?id=4OszSYdsgO
[ "Jack Cai", "Ammar Vora", "Randolph Zhang", "Mark O'Connor", "Mark C. Jeffrey" ]
Poster
As Large Language Models (LLMs) grow in size and context length, efficient inference strategies are essential to maintain low-latency token generation. Unfortunately, conventional tensor and data parallelism face diminishing returns when scaling across multiple devices. We propose a novel form—attention-level speculati...
Inference, Speculation, Attention, Transformer, Large Language Model
We show that attention-level speculation reduces LLM decode latency when tensor parallelism fails to scale.
14,085
null
[ -0.016044680029153824, -0.007465647999197245, -0.014515039511024952, 0.03788303583860397, 0.016829216852784157, 0.04273849353194237, 0.01023442018777132, 0.04139985889196396, -0.03690311312675476, 0.005740851629525423, -0.008509851060807705, 0.011928712017834187, -0.03778616711497307, 0.01...
https://github.com/mcj-group/alspec
409
Prediction-Powered E-Values
https://openreview.net/forum?id=rkegUc8d0c
[ "Daniel Csillag", "Claudio Jose Struchiner", "Guilherme Tegoni Goedert" ]
Poster
Quality statistical inference requires a sufficient amount of data, which can be missing or hard to obtain. To this end, prediction-powered inference has risen as a promising methodology, but existing approaches are largely limited to Z-estimation problems such as inference of means and quantiles. In this paper, we app...
prediction-powered inference, e-values, statistical inference, distribution-free methods
We extend prediction-powered inference to e-values, vastly expanding the set of inference tasks achievable in a prediction-powered manner while also benefiting from the usual virtues of e-values.
14,084
null
[ 0.014635016210377216, -0.037094514816999435, -0.0008755829185247421, 0.05464622378349304, 0.060633569955825806, 0.025763867422938347, 0.011052580550312996, 0.014311491511762142, -0.00821756199002266, -0.04631008952856064, -0.004247939679771662, 0.0036060120910406113, -0.07080129534006119, ...
https://github.com/dccsillag/experiments-prediction-powered-evalues
410
RAGGED: Towards Informed Design of Scalable and Stable RAG Systems
https://openreview.net/forum?id=4ufjBV6S4I
[ "Jennifer Hsia", "Afreen Shaikh", "Zora Zhiruo Wang", "Graham Neubig" ]
Poster
Retrieval-augmented generation (RAG) enhances language models by integrating external knowledge, but its effectiveness is highly dependent on system configuration. Improper retrieval settings can degrade performance, making RAG less reliable than closed-book generation. In this work, we introduce RAGGED, a framework fo...
Retrieval-Augmented Generation, RAG, Evaluation, Information Retrieval, Question Answering
We show that retrieval-augmented generation systems depend more on the reader’s ability to handle noise than on retrieval quality, and introduce a framework (RAGGED) to systematically evaluate their stability and scalability.
14,057
null
[ -0.0097917839884758, -0.039386067539453506, -0.006842681672424078, 0.05438033863902092, 0.03127874806523323, -0.022257914766669273, 0.012389221228659153, -0.015469135716557503, -0.04053826630115509, -0.030473342165350914, -0.037002015858888626, 0.018936699256300926, -0.07551413029432297, -...
https://github.com/neulab/ragged
411
H-Tuning: Toward Low-Cost and Efficient ECG-based Cardiovascular Disease Detection with Pre-Trained Models
https://openreview.net/forum?id=RLu1QIPiVr
[ "Rushuang Zhou", "Yuanting Zhang", "Yining Dong" ]
Poster
Fine-tuning large-scale pre-trained models provides an effective solution to alleviate the label scarcity problem in cardiovascular diseases (CVDs) detection using electrocardiogram (ECG). However, as the pre-trained models scale up, the computational costs for fine-tuning and inference become unaffordable on low-leve...
Electrocardiograph, Pre-trained models, Fine-tuning, Cardiovascular diseases.
This study allows for the utilization of pre-trained models with high computation efficiency and robust performance, exploring a path toward low-cost and efficient CVDs detection.
14,054
null
[ -0.01832669787108898, -0.0016948599368333817, -0.002330989809706807, 0.04064866527915001, 0.06472072005271912, 0.0002673986309673637, -0.00363456760533154, -0.009664341807365417, -0.00830842088907957, -0.03506023436784744, 0.014467224478721619, 0.01613164320588112, -0.049247756600379944, 0...
https://github.com/KAZABANA/H-Tuning
412
How Do Transformers Learn Variable Binding in Symbolic Programs?
https://openreview.net/forum?id=kVtyv7bpnw
[ "Yiwei Wu", "Atticus Geiger", "Raphaël Millière" ]
Poster
Variable binding---the ability to associate variables with values---is fundamental to symbolic computation and cognition. Although classical architectures typically implement variable binding via addressable memory, it is not well understood how modern neural networks lacking built-in binding operations may acquire thi...
variable binding, mechanistic interpretability, causal interventions, transformers, language models
Developmental interpretability study of a small transformer trained to perform variable binding on programs
14,043
2505.20896
[ -0.028956279158592224, 0.016133051365613937, -0.03518116846680641, 0.02018851786851883, 0.031318046152591705, 0.02850199304521084, 0.057912860065698624, 0.0026894118636846542, -0.04489115625619888, -0.01382196694612503, -0.019710684195160866, 0.04686165228486061, -0.062478214502334595, 0.0...
null
413
One Stone, Two Birds: Enhancing Adversarial Defense Through the Lens of Distributional Discrepancy
https://openreview.net/forum?id=pb4om8rWRQ
[ "Jiacheng Zhang", "Benjamin I. P. Rubinstein", "Jingfeng Zhang", "Feng Liu" ]
Poster
*Statistical adversarial data detection* (SADD) detects whether an upcoming batch contains *adversarial examples* (AEs) by measuring the distributional discrepancies between *clean examples* (CEs) and AEs. In this paper, we explore the strength of SADD-based methods by theoretically showing that minimizing distribution...
adversarial defense, adversarial robustness, accuracy-robustness trade-off
null
14,032
2503.02169
[ -0.01122993417084217, 0.0007529081776738167, -0.032487791031599045, 0.04540696740150452, 0.03908493369817734, 0.02017449401319027, 0.04934120178222656, -0.02809862233698368, -0.02513570338487625, -0.0532168410718441, -0.001046887133270502, -0.020785028114914894, -0.04975874349474907, -0.01...
https://github.com/tmlr-group/DAD
414
Multimodal Medical Code Tokenizer
https://openreview.net/forum?id=UaTrcei5Ba
[ "Xiaorui Su", "Shvat Messica", "Yepeng Huang", "Ruth Johnson", "Lukas Fesser", "Shanghua Gao", "Faryad Sahneh", "Marinka Zitnik" ]
Poster
Foundation models trained on patient electronic health records (EHRs) require tokenizing medical data into sequences of discrete vocabulary items. Existing tokenizers treat medical codes from EHRs as isolated textual tokens. However, each medical code is defined by its textual description, its position in ontological h...
Medical Code, Tokenization, EHR
Multimodal Medical Code Tokenizer
14,031
2502.04397
[ 0.010968277230858803, -0.041827455163002014, -0.0024834908545017242, 0.0540037639439106, 0.04011806845664978, 0.010608633048832417, 0.020706063136458397, 0.038860052824020386, -0.015511480160057545, -0.03958813473582268, -0.01281748153269291, 0.018501976504921913, -0.06104500964283943, -0....
null
415
ELEMENTAL: Interactive Learning from Demonstrations and Vision-Language Models for Reward Design in Robotics
https://openreview.net/forum?id=grlezgVg4s
[ "Letian Chen", "Nina Marie Moorman", "Matthew Craig Gombolay" ]
Poster
Reinforcement learning (RL) has demonstrated compelling performance in robotic tasks, but its success often hinges on the design of complex, ad hoc reward functions. Researchers have explored how Large Language Models (LLMs) could enable non-expert users to specify reward functions more easily. However, LLMs struggle t...
Learning from Demonstration, Vision-Language Models, Inverse Reinforcement Learning
ELEMENTAL enables robots to learn user-aligned reward functions by combining user language instructions and demonstration through interactive self-reflection using vision-language models.
14,019
2411.18825
[ -0.002253294689580798, -0.0044607254676520824, -0.03155399486422539, 0.046704284846782684, 0.031107017770409584, 0.0023385847453027964, 0.01246605534106493, 0.018588507547974586, -0.015378800220787525, -0.018539266660809517, -0.05520618334412575, 0.04261702671647072, -0.0551457516849041, -...
null
416
Tight and Fast Bounds for Multi-Label Learning
https://openreview.net/forum?id=rcqVuXU2Gm
[ "Yi-Fan Zhang", "Min-Ling Zhang" ]
Poster
Commonly used evaluation metrics in multi-label learning all involve base loss functions, and the theoretical guarantees of multi-label learning often rely on the properties of base loss functions. Some recent theoretical works have used the Lipschitz continuity of base loss functions to prove the generalization bounds...
Multi-Label Learning, Generalization Bound
null
14,018
null
[ -0.02173779532313347, -0.00043350999476388097, 0.008425923064351082, 0.048852719366550446, 0.026471750810742378, 0.01796756684780121, 0.02281513251364231, -0.02480563335120678, -0.006712283007800579, -0.013785005547106266, -0.002946909051388502, 0.01673804223537445, -0.10023888945579529, -...
null
417
LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models
https://openreview.net/forum?id=SDjZtxDo35
[ "Dachuan Shi", "Yonggan Fu", "Xiangchi Yuan", "Zhongzhi Yu", "Haoran You", "Sixu Li", "Xin Dong", "Jan Kautz", "Pavlo Molchanov", "Yingyan Celine Lin" ]
Poster
Recent advancements in Large Language Models (LLMs) have spurred interest in numerous applications requiring robust long-range capabilities, essential for processing extensive input contexts and continuously generating extended outputs. As sequence lengths increase, the number of Key-Value (KV) pairs in LLMs escalates,...
KV Cache Storage, Long-Context Modeling, Large Language Models
We propose LaCache, a training-free KV cache optimization framework featuring a ladder-shaped storage pattern for accurate and efficient generative inference of LLMs.
14,017
null
[ -0.031162653118371964, -0.018474748358130455, -0.02944689802825451, 0.022233761847019196, 0.032670337706804276, 0.029773183166980743, 0.02735726721584797, 0.015422312542796135, -0.0029072524048388004, -0.013162532821297646, -0.014036851935088634, -0.01698424108326435, -0.03651227429509163, ...
https://github.com/GATECH-EIC/LaCache
418
CSG-ODE: ControlSynth Graph ODE For Modeling Complex Evolution of Dynamic Graphs
https://openreview.net/forum?id=7hEZd8Rtlh
[ "Zhiqiang Wang", "Xiaoyi Wang", "Jianqing Liang" ]
Poster
Graph Neural Ordinary Differential Equations (GODE) integrate the Variational Autoencoder (VAE) framework with differential equations, effectively modeling latent space uncertainty and continuous dynamics, excelling in graph data evolution and incompleteness. However, existing GODE face challenges in capturing time-var...
Graph Neural Network, Dynamic Graph, Graph Neural ODE, VAE
This work introduces ControlSynth Graph ODE (CSG-ODE), a novel approach enhancing dynamic graph modeling via inter-node importance weighting and nonlinear evolution, with its stable extension (SCSG-ODE) achieving superior performance and stability.
14,016
null
[ -0.03206136077642441, -0.01438787579536438, 0.022637629881501198, 0.039453037083148956, 0.044645801186561584, 0.04120391234755516, 0.036117732524871826, 0.04811836779117584, -0.026142532005906105, -0.06557892262935638, 0.03069395199418068, -0.031625159084796906, -0.05424904450774193, 0.027...
null
419
Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective
https://openreview.net/forum?id=p6nhzZ9ilZ
[ "Junze Deng", "Qinhang Wu", "Peizhong Ju", "Sen Lin", "Yingbin Liang", "Ness Shroff" ]
Poster
Rehearsal-based methods have shown superior performance in addressing catastrophic forgetting in continual learning (CL) by storing and training on a subset of past data alongside new data in current task. While such a concurrent rehearsal strategy is widely used, it remains unclear if this approach is always optimal. ...
Continual Learning, rehearsal-based methods, catastrophic forgetting
null
14,014
2506.00205
[ -0.022172851487994194, -0.024203836917877197, -0.025426756590604782, 0.04554365575313568, 0.04841611534357071, -0.007578390184789896, 0.0561918281018734, 0.04103715345263481, -0.052371203899383545, -0.03820706903934479, -0.003903187345713377, 0.01524582039564848, -0.042905330657958984, 0.0...
null
420
Weak-to-Strong Generalization Even in Random Feature Networks, Provably
https://openreview.net/forum?id=OUzDIhgiqr
[ "Marko Medvedev", "Kaifeng Lyu", "Dingli Yu", "Sanjeev Arora", "Zhiyuan Li", "Nathan Srebro" ]
Poster
Weak-to-Strong Generalization (Burns et al.,2024) is the phenomenon whereby a strong student, say GPT-4, learns a task from a weak teacher, say GPT-2, and ends up significantly outperforming the teacher. We show that this phenomenon does not require a complex and pretrained learner like GPT-4, can arise even in simple ...
weak-to-strong, generalization, random features, student-teacher, 2-layer NN, 2-layer networks
null
14,003
2503.02877
[ 0.002076071919873357, -0.04186156019568443, 0.0027037879917770624, 0.05363880842924118, 0.030963612720370293, 0.001960723428055644, 0.03447389975190163, 0.0053319246508181095, -0.006239231210201979, -0.011844987981021404, -0.016559353098273277, 0.006220129784196615, -0.07491976767778397, -...
null
421
Identification of Latent Confounders via Investigating the Tensor Ranks of the Nonlinear Observations
https://openreview.net/forum?id=WH3ZRH2jno
[ "Zhengming Chen", "Yewei Xia", "Feng Xie", "Jie Qiao", "Zhifeng Hao", "Ruichu Cai", "Kun Zhang" ]
Poster
We study the problem of learning discrete latent variable causal structures from mixed-type observational data. Traditional methods, such as those based on the tensor rank condition, are designed to identify discrete latent structure models and provide robust identification bounds for discrete causal models. However, w...
Causal discovery, discrete latent variables
We study the problem of learning causal structure of discrete latent variables
14,001
null
[ -0.020297950133681297, -0.019681056961417198, -0.023172184824943542, 0.03080523945391178, 0.0269092358648777, 0.0227718073874712, 0.056769926100969315, -0.006197032984346151, -0.020115232095122337, -0.04117939993739128, -0.000026596053430694155, -0.0023704811464995146, -0.052314069122076035,...
null
422
Disentangling and Integrating Relational and Sensory Information in Transformer Architectures
https://openreview.net/forum?id=lbrqeIipJr
[ "Awni Altabaa", "John Lafferty" ]
Poster
Relational reasoning is a central component of generally intelligent systems, enabling robust and data-efficient inductive generalization. Recent empirical evidence shows that many existing neural architectures, including Transformers, struggle with tasks requiring relational reasoning. In this work, we distinguish bet...
relational, reasoning, attention, transformers, inductive biases, sensory, relational, architecture, attention
We introduce an extension of the Transformer architecture with explicit relational computational mechanisms, integrating sensory and relational processing.
13,994
2405.16727
[ -0.029880044981837273, 0.013777694664895535, 0.0031817241106182337, 0.0412362664937973, 0.021667391061782837, 0.016170982271432877, 0.03152237832546234, 0.016088299453258514, -0.01757604442536831, -0.012762662023305893, -0.04305929318070412, 0.04812837764620781, -0.05427820235490799, -0.01...
https://github.com/awni00/dual-attention
423
Sample-specific Noise Injection for Diffusion-based Adversarial Purification
https://openreview.net/forum?id=6nbcwJVZNy
[ "Yuhao Sun", "Jiacheng Zhang", "Zesheng Ye", "Chaowei Xiao", "Feng Liu" ]
Poster
*Diffusion-based purification* (DBP) methods aim to remove adversarial noise from the input sample by first injecting Gaussian noise through a forward diffusion process, and then recovering the clean example through a reverse generative process. In the above process, how much Gaussian noise is injected to the input sam...
adversarial purification, adversarial robustness, diffusion-based adversarial purification, accuracy-robustness trade-of
null
13,990
2506.06027
[ 0.026283781975507736, -0.015035264194011688, 0.00728733791038394, 0.05295836925506592, 0.04248236119747162, 0.0271471980959177, 0.017398180440068245, -0.04024174064397812, 0.0016994228353723884, -0.06513915956020355, 0.01279362477362156, -0.024839259684085846, -0.07502604275941849, 0.00046...
https://github.com/tmlr-group/SSNI
424
NestQuant: nested lattice quantization for matrix products and LLMs
https://openreview.net/forum?id=4OWGON33HE
[ "Semyon Savkin", "Eitan Porat", "Or Ordentlich", "Yury Polyanskiy" ]
Poster
Post-training quantization (PTQ) has emerged as a critical technique for efficient deployment of large language models (LLMs). This work proposes NestQuant, a novel PTQ scheme for weights and activations that is based on self-similar nested lattices. Recent works have mathematically shown such quantizers to be informa...
large language models, quantization
Nested lattice codes for LLM quantization
13,986
2502.09720
[ -0.01367279514670372, -0.019062308594584465, -0.018962664529681206, 0.016696155071258545, 0.021277930587530136, 0.03340023383498192, 0.019266454502940178, -0.0014292850391939282, -0.02279704436659813, -0.010277844965457916, 0.017115531489253044, -0.020482441410422325, -0.06730596721172333, ...
https://github.com/cookiedoth/nestquant
425
Interpolating Neural Network-Tensor Decomposition (INN-TD): a scalable and interpretable approach for large-scale physics-based problems
https://openreview.net/forum?id=xfWYB81p5O
[ "Jiachen Guo", "Xiaoyu Xie", "Chanwook Park", "Hantao Zhang", "Matthew J. Politis", "Gino Domel", "Wing Kam Liu" ]
Poster
Deep learning has been extensively employed as a powerful function approximator for modeling physics-based problems described by partial differential equations (PDEs). Despite their popularity, standard deep learning models often demand prohibitively large computational resources and yield limited accuracy when scaling...
Scientific Machine Learning; Large-scale simulation; Partial Differential Equations (PDEs); Numerical Analysis; Tensor decomposition
Novel neural networks based on interpolation theory from numerical analysis that has faster training/solving speed and better accuracy for modeling high-dimensional large-scale physics-based PDEs
13,975
2503.02041
[ -0.0628504827618599, -0.008879337459802628, -0.03225196152925491, 0.024619871750473976, 0.023497125133872032, 0.03016706369817257, -0.02153804525732994, -0.03319978713989258, -0.03829273208975792, -0.05409856513142586, 0.012884043157100677, -0.02196122519671917, -0.025198882445693016, 0.02...
null
426
MetaOptimize: A Framework for Optimizing Step Sizes and Other Meta-parameters
https://openreview.net/forum?id=H78W6bTkuZ
[ "Arsalan Sharifnassab", "Saber Salehkaleybar", "Richard S. Sutton" ]
Poster
We address the challenge of optimizing meta-parameters (hyperparameters) in machine learning, a key factor for efficient training and high model performance. Rather than relying on expensive meta-parameter search methods, we introduce MetaOptimize: a dynamic approach that adjusts meta-parameters, particularly step size...
Continual optimization, meta-parameter optimization
null
13,965
2402.02342
[ -0.01704292744398117, 0.0006455992697738111, 0.0036813330370932817, 0.021179521456360817, 0.04363933578133583, 0.03132762387394905, 0.027731910347938538, -0.030453084036707878, -0.035084232687950134, -0.03105519525706768, -0.0070618218742311, 0.022217871621251106, -0.059746820479631424, -0...
https://github.com/sabersalehk/MetaOptimize
427
Calibrating Video Watch-time Predictions with Credible Prototype Alignment
https://openreview.net/forum?id=aYUjVw9zcO
[ "Chao Cui", "Shisong Tang", "Fan Li", "Jiechao Gao", "Hechang Chen" ]
Poster
Accurately predicting user watch-time is crucial for enhancing user stickiness and retention in video recommendation systems. Existing watch-time prediction approaches typically involve transformations of watch-time labels for prediction and subsequent reversal, ignoring both the natural distribution properties of labe...
Prototype learning, optimal transport, recommendation
null
13,964
null
[ 0.02090689167380333, -0.02041442319750786, 0.024398740381002426, 0.03286106511950493, 0.04680551216006279, 0.016128653660416603, 0.03325171768665314, -0.023670533671975136, 0.005993055645376444, -0.047492615878582, 0.02787826955318451, 0.01571653224527836, -0.04112515598535538, 0.015138657...
null
428
Auditing Prompt Caching in Language Model APIs
https://openreview.net/forum?id=gUj2fxQcLZ
[ "Chenchen Gu", "Xiang Lisa Li", "Rohith Kuditipudi", "Percy Liang", "Tatsunori Hashimoto" ]
Poster
Prompt caching in large language models (LLMs) results in data-dependent timing variations: cached prompts are processed faster than non-cached prompts. These timing differences introduce the risk of side-channel timing attacks. For example, if the cache is shared across users, an attacker could identify cached prompts...
audit, prompt caching, privacy, transparency, large language models
We develop and conduct statistical audits to detect prompt caching and cache sharing in real-world language model APIs, as prompt caching may cause privacy leakage.
13,956
2502.07776
[ -0.031765080988407135, -0.014459549449384212, -0.025754228234291077, 0.04163333773612976, 0.03783588111400604, -0.004910953342914581, 0.04502400383353233, 0.01888093166053295, -0.029601436108350754, 0.0069028837606310844, -0.0011922871926799417, 0.0031620520167052746, -0.040794119238853455, ...
https://github.com/chenchenygu/auditing-prompt-caching
429
Tensor Product Neural Networks for Functional ANOVA Model
https://openreview.net/forum?id=Ci3nWnys6T
[ "Seokhun Park", "Insung Kong", "yongchan Choi", "Chanmoo Park", "Yongdai Kim" ]
Poster
Interpretability for machine learning models is becoming more and more important as machine learning models become more complex. The functional ANOVA model, which decomposes a high-dimensional function into a sum of lower dimensional functions (commonly referred to as components), is one of the most popular tools for ...
Interpretability, Trustworthy AI, Functional ANOVA model, Generalized additive models, Tensor product neural network
In this paper, we propose a novel neural network which guarantees a unique functional ANOVA decomposition and thus is able to estimate each component stably and accurately.
13,951
2502.15215
[ -0.015201019123196602, -0.007200188934803009, -0.019511157646775246, 0.01926008239388466, 0.05023723840713501, 0.04208449646830559, 0.03802543878555298, -0.03767339885234833, -0.013860649429261684, -0.04929281771183014, 0.015493139624595642, 0.0071974932216107845, -0.06335272639989853, 0.0...
https://github.com/ParkSeokhun/ANOVA-TPNN
430
Instruction-Following Pruning for Large Language Models
https://openreview.net/forum?id=juARG7yu4P
[ "Bairu Hou", "Qibin Chen", "Jianyu Wang", "Guoli Yin", "Chong Wang", "Nan Du", "Ruoming Pang", "Shiyu Chang", "Tao Lei" ]
Poster
With the rapid scaling of large language models (LLMs), structured pruning has become a widely used technique to learn efficient, smaller models from larger ones, delivering superior performance compared to training similarly sized models from scratch. In this paper, we move beyond the traditional static pruning approa...
Large language model, model pruning, contextual sparsity, pre-training, fine-tuning
null
13,949
2501.02086
[ -0.010735591873526573, -0.015905236825346947, -0.025026269257068634, 0.01430116780102253, 0.029098574072122574, 0.03844838589429855, 0.02423861250281334, 0.0033150396775454283, -0.03827274218201637, -0.005829138215631247, -0.017614340409636497, 0.038972996175289154, -0.04246808588504791, -...
null
431
Audio Flamingo 2: An Audio-Language Model with Long-Audio Understanding and Expert Reasoning Abilities
https://openreview.net/forum?id=xWu5qpDK6U
[ "Sreyan Ghosh", "Zhifeng Kong", "Sonal Kumar", "S Sakshi", "Jaehyeon Kim", "Wei Ping", "Rafael Valle", "Dinesh Manocha", "Bryan Catanzaro" ]
Poster
Understanding and reasoning over non-speech sounds and music are crucial for both humans and AI agents to interact effectively with their environments. In this paper, we introduce Audio Flamingo 2 (AF2), an Audio-Language Model (ALM) with advanced audio understanding and reasoning capabilities. AF2 leverages (i) a cust...
audio, sound, large audio-language model
A parameter-efficient Audio Language Model that achieves sota on several audio understanding and reasoning benchmarks.
13,946
2503.03983
[ 0.008212975226342678, -0.008388071320950985, 0.0019365499028936028, 0.039116788655519485, 0.020609784871339798, 0.01471472717821598, 0.02291111648082733, 0.008905801922082901, -0.03355046734213829, -0.020314762368798256, -0.023489313200116158, 0.05392513796687126, -0.05030494183301926, -0....
null
432
It's My Data Too: Private ML for Datasets with Multi-User Training Examples
https://openreview.net/forum?id=8bGEHOTvmq
[ "Arun Ganesh", "Ryan McKenna", "Hugh Brendan McMahan", "Adam Smith", "Fan Wu" ]
Poster
We initiate a study of algorithms for model training with user-level differential privacy (DP), where each example may be attributed to multiple users, which we call the multi-attribution model. We first provide a carefully chosen definition of user-level DP under the multi-attribution model. Training in the multi-attr...
differential privacy, multi-attribution
We initiate a study of algorithms for model training with user-level differential privacy (DP), where each example is associated with multiple users
13,939
2503.03622
[ -0.010612680576741695, -0.009176193736493587, -0.003696841187775135, 0.057546038180589676, 0.0633062943816185, 0.009108968079090118, 0.03506677225232124, -0.03914681449532509, -0.020827241241931915, -0.016443779692053795, -0.013937640003859997, 0.005233838688582182, -0.06715412437915802, -...
null
433
Understanding the Kronecker Matrix-Vector Complexity of Linear Algebra
https://openreview.net/forum?id=2qTwKMDAsD
[ "Raphael A Meyer", "William Joseph Swartworth", "David Woodruff" ]
Poster
We study the computational model where we can access a matrix $\mathbf{A}$ only by computing matrix-vector products $\mathbf{A}\mathrm{x}$ for vectors of the form $\mathrm{x} = \mathrm{x}_1 \otimes \cdots \otimes \mathrm{x}_q$. We prove exponential lower bounds on the number of queries needed to estimate various pr...
Tensors, Kronecker measurements, Sketching, Matrix-Vector, Lower Bound, Query Complexity
First exponential lower bounds on the number of Kronecker matrix-vector products to solve Linear Algebra Problems
13,930
2502.08029
[ -0.030107716098427773, -0.01681511662900448, 0.026398515328764915, 0.045746028423309326, 0.03488593176007271, -0.0009833944495767355, 0.03245925530791283, 0.01981852948665619, -0.018884247168898582, -0.040273576974868774, -0.008160714991390705, -0.01081705279648304, -0.04143477603793144, 0...
null
434
Fair Clustering via Alignment
https://openreview.net/forum?id=jImlK83NmV
[ "Kunwoong Kim", "Jihu Lee", "Sangchul Park", "Yongdai Kim" ]
Poster
Algorithmic fairness in clustering aims to balance the proportions of instances assigned to each cluster with respect to a given sensitive attribute. While recently developed fair clustering algorithms optimize clustering objectives under specific fairness constraints, their inherent complexity or approximation often r...
Clustering, Fairness, Trustworthy AI
This paper proposes a new algorithm for fair clustering based on a novel decomposition of the fair clustering objective, achieving an approximately optimal trade-off between fairness level and clustering utility for any given fairness level.
13,919
2505.09131
[ -0.007444286253303289, -0.008922328241169453, 0.001785732340067625, 0.03544239699840546, 0.027602680027484894, 0.047415487468242645, -0.007455514743924141, 0.015034948475658894, -0.03499200940132141, -0.026246504858136177, 0.004871387034654617, -0.01967771165072918, -0.07770044356584549, -...
https://github.com/kwkimonline/FCA
435
Computing Voting Rules with Improvement Feedback
https://openreview.net/forum?id=VFI7HottBp
[ "Evi Micha", "Vasilis Varsamis" ]
Poster
Aggregating preferences under incomplete or constrained feedback is a fundamental problem in social choice and related domains. While prior work has established strong impossibility results for pairwise comparisons, this paper extends the inquiry to improvement feedback, where voters express incremental adjustments rat...
Preference Aggregation, Incomplete Preferences, Improvement Feedback, Positional Scoring Rules, Condorcet-Consistent Rules, Computational Social Choice
This paper studies the limits of learning social choice functions using t-improvement feedback.
13,914
2502.12542
[ -0.025266515091061592, -0.02219613827764988, 0.004712138324975967, 0.0447116456925869, 0.03514911234378815, 0.015729187056422234, 0.009559799917042255, -0.0018503495957702398, -0.03980805352330208, -0.03247055038809776, -0.016739772632718086, -0.005141374655067921, -0.06663225591182709, -0...
https://github.com/VasilisVar00/Computing-Voting-Rules-with-Improvement-Feedback
436
Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models
https://openreview.net/forum?id=19kqoNoc2N
[ "Yuan Li", "Zhengzhong Liu", "Eric P. Xing" ]
Poster
Optimizing data mixtures for supervised fine-tuning (SFT) of large language models (LLMs) is critical for developing general-purpose models, yet this area remains underexplored. In this paper, we frame data mixing as an optimization problem and introduce a novel method designed to minimize validation loss. Our approach...
Large language models, supervised fine-tuning, data mixing
null
13,907
null
[ -0.015483811497688293, -0.022720400243997574, 0.008185308426618576, 0.0412227138876915, 0.040814466774463654, 0.03045506402850151, 0.025665998458862305, -0.005758362356573343, -0.025161243975162506, -0.016777098178863525, -0.017485149204730988, 0.021212808787822723, -0.06429030746221542, 0...
null
437
Protriever: End-to-End Differentiable Protein Homology Search for Fitness Prediction
https://openreview.net/forum?id=GZ7gwOZ6Or
[ "Ruben Weitzman", "Peter Mørch Groth", "Lood Van Niekerk", "Aoi Otani", "Yarin Gal", "Debora Susan Marks", "Pascal Notin" ]
Poster
Retrieving homologous protein sequences is essential for a broad range of protein modeling tasks such as fitness prediction, protein design, structure modeling, and protein-protein interactions. Traditional workflows have relied on a two-step process: first retrieving homologs via Multiple Sequence Alignments (MSA), th...
Computational Biology
null
13,905
2506.08954
[ -0.036998450756073, -0.018537422642111778, -0.0024318681098520756, 0.008180540055036545, 0.06874968111515045, -0.01452836487442255, 0.04290353134274483, 0.003001568838953972, 0.0006440570577979088, -0.03257494792342186, 0.034185025840997696, -0.011789404787123203, -0.047901369631290436, 0....
https://github.com/OATML-Markslab/Protriever
438
Settling the Maximin Share Fairness for Scheduling among Groups of Machines
https://openreview.net/forum?id=Z4I9cf7WY6
[ "Bo Li", "Fangxiao Wang", "Xing Shiji" ]
Poster
We study the fair scheduling of jobs among groups of (unrelated) machines and focus on the maximin share (MMS) fairness at the group level. The problem was first introduced by Li et al. [NeurIPS 2023], where each group consists of a number of identical machines (or identical up to different speeds), and the cost o...
Fair division, Maximin share, Job scheduling
The paper studies the fair scheduling of jobs among groups of (unrelated) machines and focus on the maximin share (MMS) fairness at the group level.
13,902
null
[ -0.0249528419226408, -0.0002374522591708228, -0.013236546888947487, 0.027318531647324562, 0.04609173908829689, 0.033057741820812225, 0.013601166196167469, 0.015146524645388126, -0.04482307285070419, -0.04821097105741501, -0.021742448210716248, -0.020389001816511154, -0.07779212296009064, -...
null
439
When can in-context learning generalize out of task distribution?
https://openreview.net/forum?id=YKyza9lrv4
[ "Chase Goddard", "Lindsay M. Smith", "Vudtiwat Ngampruetikorn", "David J. Schwab" ]
Poster
In-context learning (ICL) is a remarkable capability of pretrained transformers that allows models to generalize to unseen tasks after seeing only a few examples. We investigate empirically the conditions necessary on the pretraining distribution for ICL to emerge and generalize \emph{out-of-distribution}. Previous wor...
In-context learning, Generalization, OOD, out-of-distribution, Machine Learning, ICL
We observe a transition between transformers that learn a specialized solution and a generalizing solution when trained to do in-context learning of linear functions with varying task diversity.
13,899
2506.05574
[ 0.0024857630487531424, -0.014649646356701851, -0.009173437021672726, 0.0169659536331892, 0.0278532262891531, -0.004955803509801626, 0.014814704656600952, 0.02287266217172146, -0.031806331127882004, -0.0033861014526337385, -0.054052501916885376, 0.03567073866724968, -0.06260913610458374, -0...
null
440
Bayesian Inference for Correlated Human Experts and Classifiers
https://openreview.net/forum?id=sw2pUzbTf1
[ "Markelle Kelly", "Alex James Boyd", "Sam Showalter", "Mark Steyvers", "Padhraic Smyth" ]
Poster
Applications of machine learning often involve making predictions based on both model outputs and the opinions of human experts. In this context, we investigate the problem of querying experts for class label predictions, using as few human queries as possible, and leveraging the class probability estimates of pre-trai...
human-ai, consensus, bayesian
null
13,891
2506.05636
[ 0.02118143066763878, 0.008336908183991909, -0.020087046548724174, 0.03951999172568321, 0.028645768761634827, 0.01245509460568428, 0.02294027805328369, -0.016019679605960846, -0.006259565707296133, -0.04306544363498688, -0.03248000890016556, 0.04074990749359131, -0.06593785434961319, -0.012...
null
441
SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning
https://openreview.net/forum?id=teJdFzLnKh
[ "Jinpeng Chen", "Runmin Cong", "Yuzhi Zhao", "Hongzheng Yang", "Guangneng Hu", "Horace Ip", "Sam Kwong" ]
Poster
Multimodal Continual Instruction Tuning (MCIT) aims to enable Multimodal Large Language Models (MLLMs) to incrementally learn new tasks without catastrophic forgetting, thus adapting to evolving requirements. In this paper, we explore the forgetting caused by such incremental training, categorizing it into superficial ...
Multimodal Continual Instruction Tuning, Multimodal Large Language Model, Continual Learning
null
13,874
2505.02486
[ -0.011698722839355469, -0.037017572671175, -0.017734205350279808, 0.024917012080550194, 0.030525336042046547, 0.04520779848098755, 0.029655832797288895, 0.010318891145288944, -0.0389757864177227, -0.01138371042907238, -0.0019288333132863045, 0.04769055172801018, -0.04703140631318092, 0.009...
https://github.com/jinpeng0528/SEFE
442
Structured Preconditioners in Adaptive Optimization: A Unified Analysis
https://openreview.net/forum?id=GzS6b5Xvvu
[ "Shuo Xie", "Tianhao Wang", "Sashank J. Reddi", "Sanjiv Kumar", "Zhiyuan Li" ]
Poster
We present a novel unified analysis for a broad class of adaptive optimization algorithms with structured (e.g., layerwise, diagonal, and kronecker-factored) preconditioners for both online regret minimization and offline convex optimization. Our analysis not only provides matching rate to several important structured ...
Shampoo, adaptive optimization, layerwise
null
13,872
2503.10537
[ -0.04867805913090706, -0.04298247769474983, 0.019447611644864082, 0.04174795374274254, 0.044447656720876694, 0.03748294711112976, 0.01903682015836239, -0.02863682620227337, -0.007298668380826712, -0.05627640709280968, 0.00007000276673352346, -0.011458129622042179, -0.05060293525457382, -0....
null
443
Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs
https://openreview.net/forum?id=FCm4laCLiH
[ "Bowen Tan", "Zheng Xu", "Eric P. Xing", "Zhiting Hu", "Shanshan Wu" ]
Poster
Synthetic data offers a promising path to train models while preserving data privacy. Differentially private (DP) finetuning of large language models (LLMs) as data generator is effective, but is impractical when computation resources are limited. Meanwhile, prompt-based methods such as private evolution depend heavily...
synthetic data, language model, differential privacy
We propose a novel framework for generating privacy-preserving synthetic data without extensive prompt engineering or billion-scale LLM finetuning.
13,871
2503.12347
[ -0.00591674679890275, -0.02016601152718067, -0.008134075440466404, 0.07038088142871857, 0.06891611218452454, 0.000897215970326215, 0.02372289076447487, 0.002826985204592347, -0.010212494060397148, 0.00589191447943449, -0.013937540352344513, 0.01740363799035549, -0.04865775629878044, -0.003...
null
444
CEGA: A Cost-Effective Approach for Graph-Based Model Extraction and Acquisition
https://openreview.net/forum?id=HnXElKZdEh
[ "Zebin Wang", "Menghan Lin", "Bolin Shen", "Ken Anderson", "Molei Liu", "Tianxi Cai", "Yushun Dong" ]
Poster
Graph Neural Networks (GNNs) have demonstrated remarkable utility across diverse applications, and their growing complexity has made Machine Learning as a Service (MLaaS) a viable platform for scalable deployment. However, this accessibility also exposes GNN to serious security threats, most notably model extraction at...
GNN Model Extraction, Cost Efficiency, Budget Limitation
This paper introduces CEGA, a novel Graph Neural Networks model extraction framework to efficiently select query nodes under strict budget constraints, exhibiting superior performance over alternatives.
13,870
2506.17709
[ -0.0005252293776720762, -0.010569787584245205, -0.012839737348258495, 0.038627635687589645, 0.04515990614891052, 0.02229633927345276, 0.006035310681909323, 0.0023796837776899338, 0.006643637549132109, -0.03519803658127785, 0.02532191202044487, -0.00624509435147047, -0.07297728955745697, 0....
https://github.com/LabRAI/CEGA
445
Observation Interference in Partially Observable Assistance Games
https://openreview.net/forum?id=rjZ2SWjwwB
[ "Scott Emmons", "Caspar Oesterheld", "Vincent Conitzer", "Stuart Russell" ]
Poster
We study partially observable assistance games (POAGs), a model of the human-AI value alignment problem which allows the human and the AI assistant to have partial observations. Motivated by concerns of AI deception, we study a qualitatively new phenomenon made possible by partial observability: would an AI assistant e...
assistance games, AI alignment, observation interference, partial observability, partially observable assistance games
We study when an AI assistant would have an incentive to interfere with a human's observations.
13,868
2412.17797
[ -0.04640784114599228, -0.0019475645385682583, -0.012275127694010735, 0.02467004582285881, 0.01894998364150524, 0.014654220081865788, 0.018572364002466202, 0.01209898479282856, -0.012216725386679173, -0.039726004004478455, -0.04379367083311081, 0.03170838579535484, -0.0857035219669342, -0.0...
null
446
Feasible Action Search for Bandit Linear Programs via Thompson Sampling
https://openreview.net/forum?id=GrF14Q0DNW
[ "Aditya Gangrade", "Aldo Pacchiano", "Clayton Scott", "Venkatesh Saligrama" ]
Poster
We study the 'feasible action search' (FAS) problem for linear bandits, wherein a learner attempts to discover a feasible point for a set of linear constraints $\Phi_* a \ge 0,$ without knowledge of the matrix $\Phi_* \in \mathbb{R}^{m \times d}$. A FAS learner selects a sequence of actions $a_t,$ and uses observations...
Safe Bandits, Linear Bandits, Thompson Sampling
An efficient method based on Thompson Sampling to find feasible actions for LPs with bandit feedback
13,853
null
[ -0.0035892415326088667, -0.005757463164627552, -0.025591667741537094, 0.011224769055843353, 0.03701866790652275, 0.021829208359122276, 0.033376820385456085, 0.0005698736640624702, -0.008535006083548069, -0.04592425376176834, -0.017016321420669556, 0.003910592757165432, -0.0728319063782692, ...
null
447
SafetyAnalyst: Interpretable, Transparent, and Steerable Safety Moderation for AI Behavior
https://openreview.net/forum?id=WUGrleBcYP
[ "Jing-Jing Li", "Valentina Pyatkin", "Max Kleiman-Weiner", "Liwei Jiang", "Nouha Dziri", "Anne Collins", "Jana Schaich Borg", "Maarten Sap", "Yejin Choi", "Sydney Levine" ]
Poster
The ideal AI safety moderation system would be both structurally interpretable (so its decisions can be reliably explained) and steerable (to align to safety standards and reflect a community's values), which current systems fall short on. To address this gap, we present SafetyAnalyst, a novel AI safety moderation fram...
AI safety, large language model, interpretability, content moderation
We propose a novel framework for interpretable safety moderation of AI behavior.
13,822
2410.16665
[ -0.004071331582963467, -0.008224387653172016, -0.0345488078892231, 0.019686773419380188, 0.03139486908912659, 0.01608264073729515, 0.017207980155944824, -0.0229649618268013, 0.0010070133721455932, -0.02801172062754631, -0.017486587166786194, 0.03285308927297592, -0.051943205296993256, -0.0...
null
448
Introducing 3D Representation for Dense Volume-to-Volume Translation via Score Fusion
https://openreview.net/forum?id=UHVk08XFkX
[ "Xiyue Zhu", "Dou Hoon Kwark", "Ruike Zhu", "Kaiwen Hong", "Yiqi Tao", "Shirui Luo", "Yudu Li", "Zhi-Pei Liang", "Volodymyr Kindratenko" ]
Poster
In volume-to-volume translations in medical images, existing models often struggle to capture the inherent volumetric distribution using 3D voxel-space representations, due to high computational dataset demands. We present Score-Fusion, a novel volumetric translation model that effectively learns 3D representations by ...
Diffusion models, 3D medical image generation, video generation
We find it hard to train 3D volume translation models in inverse problems that require high accuracy, such as super-resolution. We effectively introduced 3D representation by ensmbling 2D models' results with a 3D model.
13,813
null
[ -0.020149778574705124, 0.006235831882804632, 0.017392214387655258, 0.03471752628684044, 0.046991147100925446, 0.029946254566311836, -0.001893604639917612, -0.019950874149799347, -0.01086127944290638, -0.06341978907585144, 0.02663031406700611, -0.009593321941792965, -0.02969140186905861, 0....
null
449
LowRA: Accurate and Efficient LoRA Fine-Tuning of LLMs under 2 Bits
https://openreview.net/forum?id=Fm0nDMKBwC
[ "Zikai Zhou", "Qizheng Zhang", "Hermann Kumbong", "Kunle Olukotun" ]
Poster
Fine-tuning large language models (LLMs) is increasingly costly as models scale to hundreds of billions of parameters, and even parameter-efficient fine-tuning (PEFT) methods like LoRA remain resource-intensive. We introduce LowRA, the first framework to enable LoRA fine-tuning below 2 bits per parameter with minimal p...
Quantization, LoRA, PEFT
We introduce LowRA to enable LoRA fine-tuning below 2 bits per parameter with minimal performance loss, cutting memory use by up to 50%.
13,809
2502.08141
[ -0.009331641718745232, -0.02735772915184498, 0.005783254746347666, 0.029394879937171936, 0.04680436849594116, 0.058111127465963364, 0.004014260601252317, 0.030242919921875, -0.026815291494131088, -0.0016107558039948344, -0.0059625087305903435, 0.022266490384936333, -0.05854763835668564, 0....
null
450
Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks
https://openreview.net/forum?id=SJkpCMeIxu
[ "MAYNARA DONATO DE SOUZA", "Cleber Zanchettin" ]
Poster
Data scarcity and quality issues remain significant barriers to developing robust predictive models in medical research. Traditional reliance on real-world data often leads to biased models with poor generalizability across diverse patient populations. Synthetic data generation has emerged as a promising solution, yet ...
Medical tasks, synthetic data, data-centric
null
13,807
null
[ 0.009491266682744026, -0.047781527042388916, -0.011218124069273472, 0.025736799463629723, 0.06425101310014725, 0.008647844195365906, 0.03626403585076332, -0.01965891569852829, 0.010990607552230358, -0.05126713588833809, 0.006683646701276302, 0.010985517874360085, -0.05474826321005821, 0.01...
https://github.com/szanara/profile2gen
451
Imitation Learning from a Single Temporally Misaligned Video
https://openreview.net/forum?id=YV05KZt7v2
[ "William Huey", "Huaxiaoyue Wang", "Anne Wu", "Yoav Artzi", "Sanjiban Choudhury" ]
Poster
We examine the problem of learning sequential tasks from a single visual demonstration. A key challenge arises when demonstrations are temporally misaligned due to variations in timing, differences in embodiment, or inconsistencies in execution. Existing approaches treat imitation as a distribution-matching problem, al...
Learning from Videos, Inverse Reinforcement Learning, Reward Formulation
Learning sequential tasks from a temporally misaligned video requires a reward function that measures subgoal ordering and coverage.
13,799
2502.05397
[ 0.01952126808464527, -0.008378371596336365, -0.01817709021270275, 0.051443904638290405, 0.028902567923069, 0.04431671276688576, 0.03524244949221611, 0.05622291564941406, -0.031156834214925766, -0.02897619642317295, -0.002515978179872036, -0.0030297806952148676, -0.07874194532632828, -0.023...
https://github.com/portal-cornell/orca
452
The Logical Implication Steering Method for Conditional Interventions on Transformer Generation
https://openreview.net/forum?id=E7c9Jf1KjV
[ "Damjan Kalajdzievski" ]
Poster
The field of mechanistic interpretability in pre-trained transformer models has demonstrated substantial evidence supporting the ''linear representation hypothesis'', which is the idea that high level concepts are encoded as vectors in the space of activations of a model. Studies also show that model generation behavio...
Machine Learning, ICML, Mechanistic Interpretability, Interpretability, Safety, LLM, Transformer, Neuro-Symbolic, Language Modelling, Transformer, Steering
We show how to linear concepts in the representation space of a transformer to build a form of logical implication into models.
13,790
2502.03618
[ -0.03045603632926941, -0.028162149712443352, -0.02392447181046009, 0.013802532106637955, 0.05281561240553856, 0.014074398204684258, 0.013487914577126503, -0.010758206248283386, -0.021289026364684105, -0.0018224038649350405, -0.04065375775098801, 0.028089871630072594, -0.0296926349401474, 0...
null
453
Variational Phylogenetic Inference with Products over Bipartitions
https://openreview.net/forum?id=s1WJSRaJuy
[ "Evan Sidrow", "Alexandre Bouchard-Cote", "Lloyd T Elliott" ]
Poster
Bayesian phylogenetics is vital for understanding evolutionary dynamics, and requires accurate and efficient approximation of posterior distributions over trees. In this work, we develop a variational Bayesian approach for ultrametric phylogenetic trees. We present a novel variational family based on coalescent times o...
Phylogenetic Inference, Variational Bayes, COVID-19 Genetics, Linkage Clustering, Reinforce Estimators
We develop a variational inference approach for ultrametric phylogenetic trees that is differentiable, doesn't restrict tree space, and doesn't rely on MCMC subroutines.
13,788
2502.15110
[ -0.00030978440190665424, 0.024700885638594627, -0.01897118054330349, 0.01589975878596306, 0.046253420412540436, 0.03319495543837547, 0.0600934736430645, 0.032920487225055695, 0.0015709118451923132, -0.03968615457415581, 0.017230559140443802, -0.01450984738767147, -0.07889055460691452, 0.02...
https://github.com/EvanSidrow/VIPR
454
Shifting Time: Time-series Forecasting with Khatri-Rao Neural Operators
https://openreview.net/forum?id=emkdmORaj4
[ "Srinath Dama", "Kevin Course", "Prasanth B. Nair" ]
Poster
We present an operator-theoretic framework for temporal and spatio-temporal forecasting based on learning a *continuous time-shift operator*. Our operator learning paradigm offers a continuous relaxation of the discrete lag factor used in traditional autoregressive models, enabling the history of a system up to a given...
spatio-temporal modeling, time-series modeling, time-shift operator, Khatri-Rao neural operator, neural operator, operator learning, time-series forecasting
We present a new operator-theoretic paradigm for temporal and spatio-temporal forecasting problems by learning a continuous time-shift operator.
13,773
null
[ -0.013743511401116848, -0.02759147435426712, -0.0033409097231924534, 0.009816419333219528, 0.06431645900011063, 0.053032491356134415, 0.025745315477252007, 0.019933007657527924, -0.031275589019060135, -0.04704369977116585, -0.014471275731921196, 0.03433876112103462, -0.04578157514333725, 0...
https://github.com/srinathdama/ShiftingTime
455
Constrained Online Convex Optimization with Polyak Feasibility Steps
https://openreview.net/forum?id=EAAjvpE7sp
[ "Spencer Hutchinson", "Mahnoosh Alizadeh" ]
Poster
In this work, we study online convex optimization with a fixed constraint function $g : \mathbb{R}^d \rightarrow \mathbb{R}$. Prior work on this problem has shown $O(\sqrt{T})$ regret and cumulative constraint satisfaction $\sum_{t=1}^{T} g(x_t) \leq 0$, while only accessing the constraint value and subgradient at the ...
online convex optimization, online learning, constraints
null
13,771
2502.13112
[ -0.04618968069553375, -0.027797110378742218, -0.010636155493557453, 0.04532570019364357, 0.03853803873062134, 0.05127977207303047, 0.028925873339176178, 0.01806456968188286, -0.015575355850160122, -0.03324193134903908, -0.030447181314229965, -0.021278662607073784, -0.04444359987974167, -0....
https://github.com/shutch1/OCO-Polyak-Feasibility-Steps
456
Graph-Based Algorithms for Diverse Similarity Search
https://openreview.net/forum?id=dmN2fQ3woH
[ "Piyush Anand", "Piotr Indyk", "Ravishankar Krishnaswamy", "Sepideh Mahabadi", "Vikas C. Raykar", "Kirankumar Shiragur", "Haike Xu" ]
Poster
Nearest neighbor search is a fundamental data structure problem with many applications. Although the main objective of the data structure is to quickly report data points that are closest to a given query, it has long been noted that without additional constraints the reported answers can be redundant and/or duplicativ...
nearest neighbor search; diversity; algorithms
This paper presents provably efficient algorithms for incorporating diversity into the results of graph-based similarity search algorithms.
13,770
2502.13336
[ -0.03451179340481758, -0.03459222614765167, 0.002182057360187173, 0.04753302037715912, 0.042947251349687576, 0.039942558854818344, 0.015152689069509506, -0.02007749117910862, -0.007211420685052872, -0.05630391091108322, -0.004780015908181667, -0.028884781524538994, -0.06370501220226288, 0....
https://github.com/microsoft/DiskANN/tree/diversity
457
Sketch to Adapt: Fine-Tunable Sketches for Efficient LLM Adaptation
https://openreview.net/forum?id=zZXOXhxO6I
[ "Tianyi Zhang", "Junda Su", "Aditya Desai", "Oscar Wu", "Zhaozhuo Xu", "Anshumali Shrivastava" ]
Poster
Adapting pre-trained large language models (LLMs) is crucial but challenging due to their enormous size. Parameter-efficient fine-tuning (PEFT) techniques typically employ additive adapters applied to frozen model weights. To further reduce memory usage, model weights are often compressed through quantization. However,...
large language models, fine-tuning, sketching
null
13,768
2410.06364
[ -0.015952011570334435, -0.04230578988790512, 0.0015534713165834546, 0.04027601704001427, 0.06733496487140656, 0.03706151992082596, 0.016563231125473976, 0.01494801975786686, -0.036720503121614456, -0.04778387397527695, -0.025628995150327682, 0.015165474265813828, -0.05839946120977402, -0.0...
https://github.com/LeanModels/SketchTune
458
RestoreGrad: Signal Restoration Using Conditional Denoising Diffusion Models with Jointly Learned Prior
https://openreview.net/forum?id=NbjrGgxLPi
[ "Ching-Hua Lee", "Chouchang Yang", "Jaejin Cho", "Yashas Malur Saidutta", "Rakshith Sharma Srinivasa", "Yilin Shen", "Hongxia Jin" ]
Poster
Denoising diffusion probabilistic models (DDPMs) can be utilized to recover a clean signal from its degraded observation(s) by conditioning the model on the degraded signal. The degraded signals are themselves contaminated versions of the clean signals; due to this correlation, they may encompass certain useful informa...
Denoising diffusion probabilistic model, prior distribution, posterior, speech enhancement, image restoration
This paper proposes an integration of conditional denoising diffusion probabilistic models (DDPMs) into the variational autoencoder (VAE) framework to jointly learn a more informative diffusion prior in signal restoration applications.
13,761
2502.13574
[ -0.01213102973997593, -0.009538650512695312, -0.015174854546785355, 0.0646454393863678, 0.0699085146188736, 0.05350829288363457, 0.04689251258969307, -0.017116863280534744, -0.035822562873363495, -0.07877766340970993, 0.013294566422700882, 0.019987905398011208, -0.007963919080793858, 0.010...
null
459
Adversarial Reasoning at Jailbreaking Time
https://openreview.net/forum?id=aWd7mL5U9Q
[ "Mahdi Sabbaghi", "Paul Kassianik", "George J. Pappas", "Amin Karbasi", "Hamed Hassani" ]
Poster
As large language models (LLMs) are becoming more capable and widespread, the study of their failure cases is becoming increasingly important. Recent advances in standardizing, measuring, and scaling test-time compute suggest new methodologies for optimizing models to achieve high performance on hard tasks. In this pa...
LLMs, Jailbreaking, Reasoning, Test-time compute.
We employ a reasoning-driven framework to formulate and solve the jailbreaking problem as an optimization task.
13,747
2502.01633
[ -0.02802276611328125, -0.023234976455569267, -0.03663010522723198, 0.058396413922309875, 0.044722966849803925, 0.035780273377895355, 0.05329218879342079, -0.01780679076910019, -0.019374554976820946, 0.001764354994520545, -0.0032514971680939198, 0.0017099124379456043, -0.08209720999002457, ...
https://github.com/Helloworld10011/Adversarial-Reasoning
460
Hierarchical Equivariant Policy via Frame Transfer
https://openreview.net/forum?id=nAv5ketrHq
[ "Haibo Zhao", "Dian Wang", "Yizhe Zhu", "Xupeng Zhu", "Owen Lewis Howell", "Linfeng Zhao", "Yaoyao Qian", "Robin Walters", "Robert Platt" ]
Poster
Recent advances in hierarchical policy learning highlight the advantages of decomposing systems into high-level and low-level agents, enabling efficient long-horizon reasoning and precise fine-grained control. However, the interface between these hierarchy levels remains underexplored, and existing hierarchical methods...
robot learning, imitation learning, robotic manipulation, equivariance
We propose a equivariant hierarchical policy learning framework for visuomotor policy learning
13,739
2502.05728
[ -0.010626133531332016, 0.006863413378596306, 0.022566476836800575, 0.04080866649746895, 0.02165324054658413, 0.02636881358921528, -0.008199640549719334, -0.022349782288074493, -0.034627679735422134, -0.04910535737872124, 0.005933390464633703, -0.018022283911705017, -0.07234571874141693, -0...
null
461
Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion
https://openreview.net/forum?id=GGPM0z3dhU
[ "Kulin Shah", "Alkis Kalavasis", "Adam Klivans", "Giannis Daras" ]
Poster
There is strong empirical evidence that the stateof-the-art diffusion modeling paradigm leads to models that memorize the training set, especially when the training set is small. Prior methods to mitigate the memorization problem often lead to decrease in image quality. Is it possible to obtain strong and creative gene...
diffusion, memorization, corrupted data, limited samples, generative models, ambient diffusion
We study to what extent is possible to train powerful generative models without memorizing the training set.
13,735
2502.21278
[ -0.01118030771613121, -0.007611551322042942, -0.008902691304683685, 0.07028385996818542, 0.06340610235929489, -0.0012272571912035346, 0.028420135378837585, 0.012134828604757786, -0.031838882714509964, -0.04358929395675659, -0.00157573027536273, 0.010386806912720203, -0.042046334594488144, ...
https://github.com/kulinshah98/memorization_noisy_data
462
Dialogue Without Limits: Constant-Sized KV Caches for Extended Response in LLMs
https://openreview.net/forum?id=SuYO70ZxZX
[ "Ravi Ghadia", "Avinash Kumar", "Gaurav Jain", "Prashant J. Nair", "Poulami Das" ]
Poster
Autoregressive Transformers rely on Key-Value (KV) caching to accelerate inference. However, the linear growth of the KV cache with context length leads to excessive memory consumption and bandwidth constraints. Existing methods drop distant tokens or compress states in a lossy manner, sacrificing accuracy by discardin...
Large Language Model, Key-Value Cache Compression
An effective KV compression technique for long-response benchmarks
13,734
2503.00979
[ -0.0038458933122456074, -0.0459868460893631, -0.007679125759750605, 0.05542551726102829, 0.011358639225363731, 0.05002159997820854, 0.032186057418584824, 0.008384227752685547, -0.03159762918949127, -0.0371846929192543, -0.015049689449369907, 0.0007775839767418802, -0.04980017617344856, -0....
https://github.com/ghadiaravi13/MorphKV
463
Test-Time Graph Neural Dataset Search With Generative Projection
https://openreview.net/forum?id=824TCt6CkE
[ "Xin Zheng", "Wei Huang", "Chuan Zhou", "Ming Li", "Shirui Pan" ]
Poster
In this work, we address the test-time adaptation challenge in graph neural networks (GNNs), focusing on overcoming the limitations in flexibility and generalization inherent in existing data-centric approaches. To this end, we propose a novel research problem, test-time graph neural dataset search, which seeks to lear...
Test-Time Adaption (TTA); Graph Neural Networks (GNNs); Distribution Shift; Generative Models
Generative test-time graph adaptation method with distribution learning
13,731
null
[ -0.025160986930131912, -0.031865205615758896, -0.011472489684820175, 0.09685149043798447, 0.05317207798361778, 0.049977682530879974, 0.010431842878460884, 0.008640228770673275, 0.006509268656373024, -0.04293638467788696, 0.02761746384203434, -0.022329797968268394, -0.06273938715457916, -0....
null
464
What Has a Foundation Model Found? Inductive Bias Reveals World Models
https://openreview.net/forum?id=i9npQatSev
[ "Keyon Vafa", "Peter G. Chang", "Ashesh Rambachan", "Sendhil Mullainathan" ]
Poster
Foundation models are premised on the idea that sequence prediction can uncover deeper domain understanding, much like how Kepler's predictions of planetary motion later led to the discovery of Newtonian mechanics. However, evaluating whether these models truly capture deeper structure remains a challenge. We develop a...
world models, foundation models, inductive bias, large language models
We develop inductive bias probes to test whether foundation models have inductive biases toward specific world models.
13,728
null
[ -0.02878384478390217, -0.025878634303808212, -0.01737578772008419, 0.045151930302381516, 0.038763780146837234, 0.012680497951805592, 0.028971048071980476, 0.05190500244498253, -0.0470823310315609, -0.04681519791483879, 0.019357677549123764, 0.013850321993231773, -0.042779628187417984, -0.0...
null
465
AlphaPO: Reward Shape Matters for LLM Alignment
https://openreview.net/forum?id=LmdZ0pSWtG
[ "Aman Gupta", "Shao Tang", "Qingquan Song", "Sirou Zhu", "Jiwoo Hong", "Ankan Saha", "Viral Gupta", "Noah Lee", "Eunki Kim", "Siyu Zhu", "Parag Agrawal", "Natesh S. Pillai", "Sathiya Keerthi" ]
Poster
Reinforcement Learning with Human Feedback (RLHF) and its variants have made huge strides toward the effective alignment of large language models (LLMs) to follow instructions and reflect human values. More recently, Direct Alignment Algorithms (DAAs) have emerged in which the reward modeling stage of RLHF is skipped b...
llm, large language models, deep learning, alignment, preference tuning, post training, reward shaping
Use a parameter alpha to control the shape of the reward function for alignment training, in order to improve performance
13,723
2501.03884
[ -0.03836820647120476, 0.006527873687446117, -0.014600925147533417, 0.04469896852970123, 0.04013490676879883, 0.03729848563671112, 0.00621178699657321, 0.011904357001185417, -0.02853861078619957, -0.02715335041284561, -0.01692437380552292, 0.03367310389876366, -0.08266659826040268, -0.03013...
null
466
Addressing Concept Mislabeling in Concept Bottleneck Models Through Preference Optimization
https://openreview.net/forum?id=sQ6lqdjGBX
[ "Emiliano Penaloza", "Tianyue H. Zhang", "Laurent Charlin", "Mateo Espinosa Zarlenga" ]
Poster
Concept Bottleneck Models (CBMs) propose to enhance the trustworthiness of AI systems by constraining their decisions on a set of human understandable concepts. However, CBMs typically rely on datasets with assumedly accurate concept labels—an assumption often violated in practice which we show can significantly degrad...
Concept Bottleneck Models, Interpretable AI, XAI
null
13,713
2504.18026
[ -0.016595160588622093, -0.0031564836390316486, -0.021716006100177765, 0.09175974130630493, 0.019217999652028084, 0.03194037079811096, 0.00474978843703866, 0.0013364475453272462, 0.008805683813989162, -0.031150730326771736, -0.043681323528289795, 0.04825809970498085, -0.04700982943177223, 0...
https://github.com/Emilianopp/ConceptPreferenceOptimization
467
Near-optimal Sketchy Natural Gradients for Physics-Informed Neural Networks
https://openreview.net/forum?id=bKsZomnmqn
[ "Maricela Best Mckay", "Avleen Kaur", "Chen Greif", "Brian Wetton" ]
Poster
Natural gradient methods for PINNs have achieved state-of-the-art performance with errors several orders of magnitude smaller than those achieved by standard optimizers such as ADAM or L-BFGS. However, computing natural gradients for PINNs is prohibitively computationally costly and memory-intensive for all but small n...
Sci-ML, Physics Informed Neural Networks, Natural Gradients, Sketching
null
13,711
null
[ -0.031263284385204315, -0.027862608432769775, 0.007115622516721487, 0.04862460121512413, 0.040537890046834946, 0.035870712250471115, 0.020556287840008736, -0.00718253618106246, -0.047124724835157394, -0.04744722321629524, -0.01766875386238098, -0.007009885739535093, -0.057920947670936584, ...
null
468
KinDEL: DNA-Encoded Library Dataset for Kinase Inhibitors
https://openreview.net/forum?id=WBN0Mz3VAC
[ "Benson Chen", "Tomasz Danel", "Gabriel H. S. Dreiman", "Patrick J. McEnaney", "Nikhil Jain", "Kirill Novikov", "Spurti Umesh Akki", "Joshua L. Turnbull", "Virja Atul Pandya", "Boris P. Belotserkovskii", "Jared Bryce Weaver", "Ankita Biswas", "Dat Nguyen", "Kent Gorday", "Mohammad Sultan...
Poster
DNA-Encoded Libraries (DELs) represent a transformative technology in drug discovery, facilitating the high-throughput exploration of vast chemical spaces. Despite their potential, the scarcity of publicly available DEL datasets presents a bottleneck for the advancement of machine learning methodologies in this domain....
DEL, small molecule, dataset, benchmark, drug discovery
Dataset and benchmark paper for a 81 million small molecule DNA-encoded library to find hits for drug discovery
13,707
2410.08938
[ 0.0012995785800740123, -0.04138577729463577, -0.031098131090402603, 0.036755289882421494, 0.05289248749613762, -0.02482357621192932, 0.012416776269674301, -0.020043816417455673, -0.019084667786955833, -0.027370410040020943, 0.012281146831810474, 0.010680370032787323, -0.050592806190252304, ...
https://github.com/insitro/kindel
469
UI-Vision: A Desktop-centric GUI Benchmark for Visual Perception and Interaction
https://openreview.net/forum?id=5Rtj4mYH1C
[ "Shravan Nayak", "Xiangru Jian", "Kevin Qinghong Lin", "Juan A. Rodriguez", "Montek Kalsi", "Nicolas Chapados", "M. Tamer Özsu", "Aishwarya Agrawal", "David Vazquez", "Christopher Pal", "Perouz Taslakian", "Spandana Gella", "Sai Rajeswar" ]
Poster
Autonomous agents that navigate Graphical User Interfaces (GUIs) to automate tasks like document editing and file management can greatly enhance computer workflows. While existing research focuses on online settings, desktop environments, critical for many professional and everyday tasks, remain underexplored due to da...
Large language models, Multimodal models, LLMs, VLMs, Autonomous agents, GUI
We introduce UI-Vision, a comprehensive desktop-centric and license-permissive GUI understanding benchmark
13,703
null
[ -0.028404271230101585, 0.020090293139219284, -0.014317471534013748, 0.015958625823259354, 0.017149923369288445, -0.003841925412416458, 0.02027815766632557, 0.03119618259370327, -0.011331940069794655, -0.04763633757829666, -0.03676484152674675, 0.0264283437281847, -0.06678637117147446, -0.0...
https://github.com/uivision/UI-Vision
470
Online Laplacian-Based Representation Learning in Reinforcement Learning
https://openreview.net/forum?id=NXtoNstR96
[ "Maheed H. Ahmed", "Jayanth Bhargav", "Mahsa Ghasemi" ]
Poster
Representation learning plays a crucial role in reinforcement learning, especially in complex environments with high-dimensional and unstructured states. Effective representations can enhance the efficiency of learning algorithms by improving sample efficiency and generalization across tasks. This paper considers the L...
Reinforcement Learning, Representation learning, Online Learning, Graph Laplacian
We propose an online method for learning the Laplacian representation in reinforcement learning, and show theoretically and empirically it converges.
13,699
null
[ -0.02280195988714695, -0.017375357449054718, 0.011748621240258217, 0.04532364755868912, 0.03336559236049652, 0.04171537980437279, -0.0029852224979549646, 0.0038401756901293993, -0.03749873489141464, -0.03772575035691261, -0.0035893686581403017, -0.015638776123523712, -0.08520758152008057, ...
https://github.com/MaheedHatem/online_laplacian_representation
471
Modulated Diffusion: Accelerating Generative Modeling with Modulated Quantization
https://openreview.net/forum?id=y8hMadAgrz
[ "Weizhi Gao", "Zhichao Hou", "Junqi Yin", "Feiyi Wang", "Linyu Peng", "Xiaorui Liu" ]
Poster
Diffusion models have emerged as powerful generative models, but their high computation cost in iterative sampling remains a significant bottleneck. In this work, we present an in-depth and insightful study of state-of-the-art acceleration techniques for diffusion models, including caching and quantization, revealing t...
Diffusion Models; Efficiency; Quantization; Caching Method
null
13,693
2506.22463
[ 0.004063531756401062, -0.05434221774339676, -0.02026345767080784, 0.04119330272078514, 0.06195610761642456, 0.04773736745119095, -0.0030957586131989956, 0.015496144071221352, -0.017313828691840172, -0.0845181941986084, 0.015691258013248444, -0.02943892776966095, -0.046678729355335236, 0.02...
https://github.com/WeizhiGao/MoDiff
472
Memorization Sinks: Isolating Memorization during LLM Training
https://openreview.net/forum?id=sRJrMPu5Uu
[ "Gaurav Rohit Ghosal", "Pratyush Maini", "Aditi Raghunathan" ]
Poster
Large language models are susceptible to memorizing repeated sequences, posing privacy and copyright concerns. A popular mitigation strategy is to remove memorized information from specific neurons post-hoc. However, such approaches have shown limited success so far. In a controlled setting, we show that the memorizati...
Memorization, Localization, Unlearning
We propose a method to disentangle sequence memorization and general language model capabilities during pretraining.
13,672
2507.09937
[ -0.013028996996581554, 0.008642475120723248, -0.011689024046063423, 0.03220732882618904, 0.045860886573791504, 0.004468932282179594, 0.0527007058262825, 0.02659759856760502, -0.04186040535569191, 0.00482853502035141, 0.0030343260150402784, 0.015923568978905678, -0.06589404493570328, -0.003...
https://github.com/grghosal/MemSinks
473
Skip the Equations: Learning Behavior of Personalized Dynamical Systems Directly From Data
https://openreview.net/forum?id=2gpjvMEAMm
[ "Krzysztof Kacprzyk", "Julianna Piskorz", "Mihaela van der Schaar" ]
Poster
While black-box approaches are commonly used for data-driven modeling of dynamical systems, they often obscure a system's underlying behavior and properties, limiting adoption in areas such as medicine and pharmacology. A two-step process of discovering ordinary differential equations (ODEs) and their subsequent mathem...
dynamical systems, differential equations, ODE discovery, interpretability, transparency, verification
An interpretable framework for trajectory prediction from static features: a decision tree identifies the curve’s global shape, while generalized additive models assign the key properties such as start, peak, or inflection.
13,671
null
[ -0.03801590949296951, -0.014649578370153904, -0.016741018742322922, 0.03452346846461296, 0.06064034625887871, 0.016152985394001007, 0.036748822778463364, -0.0017886360874399543, -0.002666833810508251, -0.028197122737765312, 0.03600001707673073, 0.016727419570088387, -0.053220778703689575, ...
https://github.com/krzysztof-kacprzyk/EPISODE
474
Return Capping: Sample Efficient CVaR Policy Gradient Optimisation
https://openreview.net/forum?id=ebf2IYBrZO
[ "Harry Mead", "Clarissa Costen", "Bruno Lacerda", "Nick Hawes" ]
Poster
When optimising for conditional value at risk (CVaR) using policy gradients (PG), current methods rely on discarding a large proportion of trajectories, resulting in poor sample efficiency. We propose a reformulation of the CVaR optimisation problem by capping the total return of trajectories used in training, rather ...
Reinforcement Learning, Machine Learning, CVaR, Risk-Averse
We present a sample-efficient method for policy gradient CVaR optimisation by capping trajectory returns, rather than discarding trajectories.
13,664
2504.20887
[ 0.0030869552865624428, -0.031857892870903015, 0.010510789230465889, 0.039230722934007645, 0.030101247131824493, 0.030953416600823402, 0.012187053449451923, -0.020286627113819122, 0.0006859114510007203, -0.024655897170305252, -0.04167110472917557, 0.031186312437057495, -0.06557580828666687, ...
https://github.com/HarryMJMead/cvar-return-capping
475
LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection
https://openreview.net/forum?id=om0CcjvEQh
[ "Xinyue Zeng", "Haohui Wang", "Junhong Lin", "Jun Wu", "Tyler Cody", "Dawei Zhou" ]
Poster
The proliferation of open-sourced Large Language Models (LLMs) and diverse downstream tasks necessitates efficient model selection, given the impracticality of fine-tuning all candidates due to computational constraints. Despite the recent advances in LLM selection, a fundamental research question largely remains nasce...
LLM Selection, PAC-Bayesian Theory, Generalization Bound, Scaling Laws
We derive a PAC-Bayesian generalization bound for LLM fine-tuning dynamics and propose LENSLLM, a framework that enables accurate, efficient performance prediction across diverse tasks.
13,652
2505.03793
[ -0.0225197933614254, -0.013970600441098213, 0.010299494490027428, 0.0011956783710047603, 0.030171256512403488, 0.047263529151678085, 0.005290290806442499, 0.03450271859765053, -0.03890376165509224, 0.0045259492471814156, -0.017353767529129982, 0.060118842869997025, -0.06665252894163132, -0...
https://github.com/Susan571/LENSLLM
476
A New Concentration Inequality for Sampling Without Replacement and Its Application for Transductive Learning
https://openreview.net/forum?id=NRVdvg7VMn
[ "Yingzhen Yang" ]
Poster
We introduce a new tool, Transductive Local Complexity (TLC), to analyze the generalization performance of transductive learning methods and motivate new transductive learning algorithms. Our work extends the idea of the popular Local Rademacher Complexity (LRC) to the transductive setting with considerable and novel c...
Concentration Inequality, Sampling Without Replacement, Transductive Local Rademacher Complexity, Transductive Learning
We present a new concentration inequality for the supremum of the empirical process associated with sampling without replacement, and apply our new concentration inequality to prove sharper generalization bound for transductive kernel learning.
13,647
null
[ -0.0024449729826301336, -0.020751364529132843, -0.013199537992477417, 0.043371401727199554, 0.04148615896701813, 0.020006101578474045, 0.021615460515022278, -0.028765501454472542, -0.0006847872282378376, -0.02494855225086212, -0.002031799405813217, -0.004856583662331104, -0.07198254019021988...
null
477
Improving Diversity in Language Models: When Temperature Fails, Change the Loss
https://openreview.net/forum?id=RsyMfsqzeG
[ "Alexandre Verine", "Florian Le Bronnec", "Kunhao Zheng", "Alexandre Allauzen", "Yann Chevaleyre", "benjamin negrevergne" ]
Poster
Increasing diversity in language models is a challenging yet essential objective. A common approach is to raise the decoding temperature. In this work, we investigate this approach through a simplistic yet common case to provide insights into why decreasing temperature can improve quality (Precision), while increasing ...
Language Models, Diversity, Precision, Recall, Temperature
We propose losses to improve diversity in language models to compensate for temperature failing.
13,645
null
[ 0.0024832061026245356, -0.012443349696695805, -0.02167680114507675, 0.053886592388153076, 0.07406137883663177, 0.032040905207395554, 0.044793423265218735, 0.0025877049192786217, -0.030223160982131958, -0.046508532017469406, -0.019092509523034096, 0.052075162529945374, -0.05823585391044617, ...
null
478
PENCIL: Long Thoughts with Short Memory
https://openreview.net/forum?id=6wglsDXIei
[ "Chenxiao Yang", "Nathan Srebro", "David McAllester", "Zhiyuan Li" ]
Poster
While state-of-the-art LLMs have demonstrated great promise of using long Chains-of-Thought (CoT) to boost reasoning, scaling it up to more challenging problems is fundamentally limited by suboptimal memory usage — intermediate computations accumulate indefinitely in context even no longer needed for future thoughts. W...
Large language models, chain-of-thought
null
13,637
2503.14337
[ -0.04707913473248482, -0.01939186081290245, -0.02040785364806652, 0.052996665239334106, 0.04338064044713974, 0.026782672852277756, 0.01143634133040905, 0.04453406110405922, -0.04394161328673363, -0.02468084543943405, -0.013002155348658562, -0.01053122989833355, -0.07642894983291626, -0.003...
https://github.com/chr26195/PENCIL
479
Universal Neural Optimal Transport
https://openreview.net/forum?id=t10fde8tQ7
[ "Jonathan Geuter", "Gregor Kornhardt", "Ingimar Tomasson", "Vaios Laschos" ]
Poster
Optimal Transport (OT) problems are a cornerstone of many applications, but solving them is computationally expensive. To address this problem, we propose UNOT (Universal Neural Optimal Transport), a novel framework capable of accurately predicting (entropic) OT distances and plans between discrete measures of variable...
Optimal Transport, Neural Operators, Meta Learning, Adversarial Training
We train a neural network to accurately predict optimal transport distances across datasets and dimensions with neural operators.
13,626
2212.00133
[ -0.04331616684794426, -0.023619096726179123, 0.014463535510003567, 0.04971840977668762, 0.04239910840988159, 0.03159353882074356, 0.003161810804158449, 0.019065767526626587, -0.021441156044602394, -0.06565768271684647, 0.007429058663547039, -0.009907357394695282, -0.06464650481939316, -0.0...
https://github.com/GregorKornhardt/UNOT
480
Policy-Regret Minimization in Markov Games with Function Approximation
https://openreview.net/forum?id=eZ5QyZV7zi
[ "Thanh Nguyen-Tang", "Raman Arora" ]
Poster
We study policy-regret minimization problem in dynamically evolving environments, modeled as Markov games between a learner and a strategic, adaptive opponent. We propose a general algorithmic framework that achieves the optimal $\mathcal{O}(\sqrt{T})$ policy regret for a wide class of large-scale problems characterize...
policy regret, Markov games, strategic opponents, function approximation, online learning, reinforcement learning, adversarial learning, Eluder dimension, multi-agent learning
A general algorithm that provably minimizes policy regret in complex games with strategic opponents.
13,580
null
[ -0.055449265986680984, -0.02146640047430992, 0.005149246193468571, 0.03804114833474159, 0.043406710028648376, 0.04872686043381691, 0.02023063600063324, -0.008477573283016682, -0.0294394139200449, -0.04681555926799774, -0.008689370937645435, 0.020635103806853294, -0.06883735209703445, -0.02...
null
481
MATH-Perturb: Benchmarking LLMs' Math Reasoning Abilities against Hard Perturbations
https://openreview.net/forum?id=OZy70UggXr
[ "Kaixuan Huang", "Jiacheng Guo", "Zihao Li", "Xiang Ji", "Jiawei Ge", "Wenzhe Li", "Yingqing Guo", "Tianle Cai", "Hui Yuan", "Runzhe Wang", "Yue Wu", "Ming Yin", "Shange Tang", "Yangsibo Huang", "Chi Jin", "Xinyun Chen", "Chiyuan Zhang", "Mengdi Wang" ]
Poster
Large language models have demonstrated impressive performance on challenging mathematical reasoning tasks, which has triggered the discussion of whether the performance is achieved by true reasoning capability or memorization. To investigate this question, prior work has constructed mathematical benchmarks when questi...
mathematical reasoning, benchmark, robustness
We construct MATH-P-Simple and MATH-P-Hard to benchmark LLM's math reasoning against simple and hard perturbations, and examine memorization issues.
13,579
2502.06453
[ -0.010428689420223236, -0.011682896874845028, -0.008194036781787872, 0.025894032791256905, 0.06193822994828224, -0.014852865599095821, 0.028308387845754623, 0.02412228472530842, -0.03722960501909256, -0.004423552192747593, 0.007033397909253836, 0.03758436441421509, -0.05209638550877571, -0...
null
482
DeepCrossAttention: Supercharging Transformer Residual Connections
https://openreview.net/forum?id=j3JBfFnGYh
[ "Mike Heddes", "Adel Javanmard", "Kyriakos Axiotis", "Gang Fu", "Mohammadhossein Bateni", "Vahab Mirrokni" ]
Poster
Transformer networks have achieved remarkable success across diverse domains, leveraging a variety of architectural innovations, including residual connections. However, traditional residual connections, which simply sum the outputs of previous layers, can dilute crucial information. This work introduces DeepCrossAtten...
residual network, cross attention, resnet, transformer
We speed up transformer training by introducing learnable, input-dependent residual connections combined with depth-wise cross attention.
13,577
2502.06785
[ -0.007600944954901934, -0.04211225360631943, 0.009585116989910603, 0.02408326230943203, 0.022836240008473396, 0.0185686107724905, 0.022343214601278305, 0.005014708731323481, -0.021176161244511604, -0.02442770265042782, -0.010938028804957867, -0.0004678514669649303, -0.037841614335775375, -...
null
483
Steerable Transformers for Volumetric Data
https://openreview.net/forum?id=Ax550Vokon
[ "Soumyabrata Kundu", "Risi Kondor" ]
Poster
We introduce Steerable Transformers, an extension of the Vision Transformer mechanism that maintains equivariance to the special Euclidean group $\mathrm{SE}(d)$. We propose an equivariant attention mechanism that operates on features extracted by steerable convolutions. Operating in Fourier space, our network utilizes...
Equivariance, transformers, vision transformers, steerable
SE(d) equivariant transformfers for volumetric data
13,573
2405.15932
[ 0.0023389062844216824, -0.014492909424006939, 0.03425915166735649, 0.011451809667050838, 0.008238513953983784, 0.02331320010125637, 0.005868254695087671, 0.042327187955379486, -0.02433774620294571, -0.026680169627070427, -0.04267903417348862, 0.030843690037727356, -0.04197344556450844, 0.0...
https://github.com/SoumyabrataKundu/Steerable-Transformer
484
Value-Based Deep RL Scales Predictably
https://openreview.net/forum?id=FLPFPYJeVU
[ "Oleh Rybkin", "Michal Nauman", "Preston Fu", "Charlie Victor Snell", "Pieter Abbeel", "Sergey Levine", "Aviral Kumar" ]
Poster
Scaling data and compute is critical in modern machine learning. However, scaling also demands _predictability_: we want methods to not only perform well with more compute or data, but also have their performance be predictable from low compute or low data runs, without ever running the large-scale experiment. In this ...
scaling laws, online reinforcement learning, q-learning
We establish that value-based online RL can be scaled predictably to larger data, larger compute, or generally larger budget
13,569
2502.04327
[ -0.03920581564307213, -0.040470466017723083, 0.015398947522044182, 0.028537819162011147, 0.06583262979984283, 0.046154215931892395, 0.02927190065383911, -0.008899823762476444, -0.04786132648587227, -0.021412860602140427, -0.01045935321599245, -0.010817966423928738, -0.08326325565576553, -0...
null
485
Hardware and Software Platform Inference
https://openreview.net/forum?id=kdmjVF1iDO
[ "Cheng Zhang", "Hanna Foerster", "Robert D. Mullins", "Yiren Zhao", "Ilia Shumailov" ]
Poster
It is now a common business practice to buy access to large language model (LLM) inference rather than self-host, because of significant upfront hardware infrastructure and energy costs. However, as a buyer, there is no mechanism to verify the authenticity of the advertised service including the serving hardware platfo...
ML security, ML governance
We introduce a new problem and solutions to identifying the software and hardware platform of an ML model solely based on its input-output behaviour. Abstract:
13,561
2411.05197
[ -0.013694032095372677, 0.012052156962454319, -0.03770384564995766, 0.06155068799853325, 0.04084087163209915, 0.04634597525000572, 0.005398775916546583, 0.03093746490776539, -0.020273389294743538, 0.001880268449895084, 0.009145612828433514, 0.00852496549487114, -0.045156024396419525, 0.0203...
https://github.com/ChengZhang-98/HSPI
486
Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs
https://openreview.net/forum?id=kqj2Cn3Sxr
[ "Aryan Gulati", "Brando Miranda", "Eric Chen", "Emily Xia", "Kai Fronsdal", "Bruno de Moraes Dumont", "Sanmi Koyejo" ]
Poster
Current mathematical reasoning benchmarks for large language models (LLMs) are approaching saturation, with some achieving $>$ 90% accuracy, and are increasingly compromised by training-set contamination. We introduce Putnam-AXIOM, a benchmark of 522 university-level competition problems drawn from the prestigious Will...
Benchmarks, Large Language Models, Mathematical Reasoning, Mathematics, Reasoning, Machine Learning
Putnam-AXIOM is a challenging mathematical reasoning benchmark for LLMs, revealing significant reasoning performance gaps and the impact of data contamination.
13,558
null
[ -0.03571672365069389, -0.0028211823664605618, -0.0155959976837039, 0.04678397253155708, 0.052323367446660995, 0.022723346948623657, 0.020224280655384064, 0.011270658113062382, -0.037967316806316376, 0.0017899598460644484, -0.010408896021544933, 0.04685540869832039, -0.04960194230079651, -0...
https://github.com/brando90/putnam-axiom
487
Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow
https://openreview.net/forum?id=6uPcJtMgWN
[ "Zhonglin Cao", "Mario Geiger", "Allan Dos Santos Costa", "Danny Reidenbach", "Karsten Kreis", "Tomas Geffner", "Franco Pellegrini", "Guoqing Zhou", "Emine Kucukbenli" ]
Poster
Fast and accurate generation of molecular conformers is desired for downstream computational chemistry and drug discovery tasks. Currently, training and sampling state-of-the-art diffusion or flow-based models for conformer generation require significant computational resources. In this work, we build upon flow-matchin...
Flow-matching, few-shot generation, equivariance, small molecules
We present a novel flow-matching objective and combine it with reflow to accelerate training and inference of molecular conformer generation
13,550
null
[ 0.00812478642910719, 0.008464884012937546, 0.0010857937159016728, 0.03754768893122673, 0.02937394753098488, -0.008874492719769478, 0.020047741010785103, -0.00591305922716856, 0.0063763875514268875, -0.060284268110990524, 0.011453083716332912, -0.02508923038840294, -0.0787276104092598, 0.00...
null
488
Do Vision-Language Models Really Understand Visual Language?
https://openreview.net/forum?id=ZPQU4uGMBA
[ "Yifan Hou", "Buse Giledereli", "Yilei Tu", "Mrinmaya Sachan" ]
Poster
Visual language is a system of communication that conveys information through symbols, shapes, and spatial arrangements. Diagrams are a typical example of a visual language depicting complex concepts and their relationships in the form of an image. The symbolic nature of diagrams presents significant challenges for bui...
vision-language model, visual language, evaluation, interpretation, diagram
null
13,533
2410.00193
[ 0.004573287442326546, 0.02815539948642254, -0.012395966798067093, 0.04420078545808792, 0.04853762686252594, -0.008277708664536476, 0.02317141182720661, 0.03442547842860222, -0.0332125723361969, -0.008532044477760792, -0.04729308560490608, 0.05177806317806244, -0.06313177943229675, 0.022837...
https://github.com/buseg/diagram-understanding
489
What can large language models do for sustainable food?
https://openreview.net/forum?id=f6SFHNfuMu
[ "Anna Thomas", "Adam Yee", "Andrew Mayne", "Maya B. Mathur", "Dan Jurafsky", "Kristina Gligorić" ]
Poster
Food systems are responsible for a third of human-caused greenhouse gas emissions. We investigate what Large Language Models (LLMs) can contribute to reducing the environmental impacts of food production. We define a typology of design and prediction tasks based on the sustainable food literature and collaboration with...
large language models, sustainability, climate, food, health, optimization
Motivated by food systems' major role in climate change, we define a typology of sustainable food development tasks, demonstrate expert-level LLM performance, and propose a novel framework integrating LLMs with combinatorial optimization.
13,531
2503.04734
[ -0.0006404874147847295, 0.006611267104744911, -0.03672971948981285, 0.006403076462447643, 0.0743509903550148, 0.007618316914886236, 0.004731159191578627, 0.0668702945113182, -0.03521649166941643, -0.022729497402906418, -0.03862399235367775, 0.02152751199901104, -0.054458972066640854, 0.008...
https://github.com/thomasat/llms-sustainable-food
490
No Free Lunch from Random Feature Ensembles: Scaling Laws and Near-Optimality Conditions
https://openreview.net/forum?id=z9GgK3CK39
[ "Benjamin Samuel Ruben", "William Lingxiao Tong", "Hamza Tahir Chaudhry", "Cengiz Pehlevan" ]
Poster
Given a fixed budget for total model size, one must choose between training a single large model or combining the predictions of multiple smaller models. We investigate this trade-off for ensembles of random-feature ridge regression models in both the overparameterized and underparameterized regimes. Using determinis...
Ensemble Learning, Regression, Scaling Laws
We show that ensembles of random feature models never outperform a single model of the same total size, and identify conditions where near-optimal performance by ensembles is possible in both the overparameterized and underparameterized regimes.
13,527
null
[ -0.0397513285279274, -0.024716226384043694, 0.0007964945398271084, 0.023394731804728508, 0.055423349142074585, 0.033158011734485626, 0.030259788036346436, -0.011463990435004234, -0.026629194617271423, -0.033688295632600784, -0.007865712977945805, 0.0037459908053278923, -0.0721263587474823, ...
null
491
GSM-∞: How Do your LLMs Behave over Infinitely Increasing Reasoning Complexity and Context Length?
https://openreview.net/forum?id=n52yyvEwPa
[ "Yang Zhou", "Hongyi Liu", "Zhuoming Chen", "Yuandong Tian", "Beidi Chen" ]
Poster
Recently, long-context large language models (LLMs) have shown strong performance in information retrieval and long-document QA. However, to tackle the most challenging intellectual problems, LLMs must reason effectively in long and complex contexts (e.g., frontier mathematical research). Studying how LLMs handle incre...
Long Context, Reasoning, Understanding, Benchmarks
We study LLMs reasoning ability decay with respect to increasingly harder problems and with respect to increasing context length through synthesized dataset generator that generates fine-grainedly controlled GSM8K-like problems.
13,525
null
[ -0.037817906588315964, 0.0018050476210191846, -0.008251053281128407, 0.03719432279467583, 0.053856462240219116, 0.014249990694224834, 0.01653323881328106, 0.010164570063352585, -0.041188616305589676, 0.011251971125602722, -0.005414178594946861, 0.03437498211860657, -0.0522780567407608, 0.0...
https://github.com/Infini-AI-Lab/gsm_infinite
492
Improving Model Alignment Through Collective Intelligence of Open-Source Models
https://openreview.net/forum?id=K4N9UvsuNB
[ "Junlin Wang", "Roy Xie", "Shang Zhu", "Jue WANG", "Ben Athiwaratkun", "Bhuwan Dhingra", "Shuaiwen Leon Song", "Ce Zhang", "James Zou" ]
Poster
Building helpful and harmless large language models (LLMs) requires effective model alignment approach based on human instructions and feedback, which necessitates high-quality human-labeled data. Constructing such datasets is often expensive and hard to scale, and may face potential limitations on diversity and genera...
Alignment, Open-Source Model, Mixture of Agents
We show how to leverage mixture-of-agents to generate synthetic data and feedback to effectively align models.
13,514
2505.03059
[ -0.006048107985407114, -0.011507401242852211, -0.03370802104473114, 0.04260747507214546, 0.026163579896092415, 0.01701465994119644, 0.05026693642139435, 0.029345108196139336, -0.018521303310990334, -0.008666643872857094, -0.004438921809196472, 0.0654793530702591, -0.09164879471063614, -0.0...
null
493
Robust Reward Alignment via Hypothesis Space Batch Cutting
https://openreview.net/forum?id=pL87Z7YTJS
[ "Zhixian Xie", "Haode Zhang", "Yizhe Feng", "Wanxin Jin" ]
Poster
Reward design in reinforcement learning and optimal control is challenging. Preference-based alignment addresses this by enabling agents to learn rewards from ranked trajectory pairs provided by humans. However, existing methods often struggle from poor robustness to unknown false human preferences. In this work, we pr...
Learning from Human Feedback, Inverse Reinforcement Learning, Preference Based Reinforcement Learning, Robust Learning
null
13,507
2502.02921
[ -0.026919253170490265, 0.006908709648996592, -0.02853541076183319, 0.03346285596489906, 0.042642880231142044, 0.03302827104926109, 0.012429798021912575, -0.008013010956346989, -0.01843644306063652, -0.06872047483921051, -0.03398821875452995, 0.03754556179046631, -0.07627295702695847, -0.05...
https://github.com/asu-iris/HSBC-Robust-Learning
494
Convergence of Consistency Model with Multistep Sampling under General Data Assumptions
https://openreview.net/forum?id=vsJsR3ieCx
[ "Yiding Chen", "Yiyi Zhang", "Owen Oertell", "Wen Sun" ]
Poster
Diffusion models accomplish remarkable success in data generation tasks across various domains. However, the iterative sampling process is computationally expensive. Consistency models are proposed to learn consistency functions to map from noise to data directly, which allows one-step fast data generation and multiste...
Consistency models, diffusion models, learning theory
null
13,478
2505.03194
[ -0.02408606931567192, -0.036506932228803635, -0.011855894699692726, 0.0760422796010971, 0.06482550501823425, 0.03499634563922882, 0.029757319018244743, 0.0003463137545622885, -0.013820717111229897, -0.06250905245542526, 0.012382456101477146, 0.005096776410937309, -0.08267838507890701, 0.01...
null
495
Wasserstein Policy Optimization
https://openreview.net/forum?id=oAKe7MG9GM
[ "David Pfau", "Ian Davies", "Diana L Borsa", "João Guilherme Madeira Araújo", "Brendan Daniel Tracey", "Hado van Hasselt" ]
Poster
We introduce Wasserstein Policy Optimization (WPO), an actor-critic algorithm for reinforcement learning in continuous action spaces. WPO can be derived as an approximation to Wasserstein gradient flow over the space of all policies projected into a finite-dimensional parameter space (e.g., the weights of a neural netw...
Policy Optimization, Wasserstein metric, Optimal Transport, Gradient Flow, Deep Reinforcement Learning, Actor-Critic, Continuous Control
We derive a novel policy gradient algorithm from Wasserstein gradient flows and show that it is simple and effective at deep reinforcement learning for continuous control.
13,462
2505.00663
[ -0.03455110639333725, -0.03176869824528694, 0.0075096567161381245, 0.04627956077456474, 0.035068705677986145, 0.022605478763580322, 0.017695942893624306, 0.006154200062155724, -0.02285953052341938, -0.030721420422196388, -0.014275556430220604, 0.009279274381697178, -0.06495025008916855, -0...
https://github.com/google-deepmind/acme/blob/master/examples/baselines/rl_continuous/run_wpo.py
496
Simplifying DINO via Coding Rate Regularization
https://openreview.net/forum?id=shTZSlk0HQ
[ "Ziyang Wu", "Jingyuan Zhang", "Druv Pai", "XuDong Wang", "Chandan Singh", "Jianwei Yang", "Jianfeng Gao", "Yi Ma" ]
Poster
DINO and DINOv2 are two model families being widely used to learn representations from unlabeled imagery data at large scales. Their learned representations often enable state-of-the-art performance for downstream tasks, such as image classification and segmentation. However, they employ many empirically motivated desi...
Self-supervised learning, DINO
We propose a methodology that simplifies and improves DINO, a widely used self-supervised learning algorithm.
13,457
2502.10385
[ 0.006151373498141766, -0.031300678849220276, -0.020863933488726616, 0.04914258420467377, 0.03980693966150284, 0.01667175441980362, -0.010905186645686626, 0.015162946656346321, -0.03501645103096962, -0.060841962695121765, 0.014084858819842339, -0.005980493500828743, -0.0652066320180893, 0.0...
https://github.com/RobinWu218/SimDINO
497
Boosting Protein Graph Representations through Static-Dynamic Fusion
https://openreview.net/forum?id=QbtBIE36Fd
[ "Pengkang Guo", "Bruno Correia", "Pierre Vandergheynst", "Daniel Probst" ]
Poster
Machine learning for protein modeling faces significant challenges due to proteins' inherently dynamic nature, yet most graph-based machine learning methods rely solely on static structural information. Recently, the growing availability of molecular dynamics trajectories provides new opportunities for understanding th...
Graph Neural Networks, Protein Modeling, Molecular Dynamics, Heterogeneous Graph
null
13,452
null
[ -0.0376705601811409, -0.01833515241742134, -0.014472565613687038, 0.048457659780979156, 0.022291388362646103, 0.000027701571525540203, 0.02387520670890808, 0.007335018832236528, 0.0021136109717190266, -0.04225708171725273, 0.039532944560050964, 0.0012341312831267715, -0.07769273221492767, ...
https://github.com/PKGuo/protein-static-dynamic-fusion
498
Symmetric Reinforcement Learning Loss for Robust Learning on Diverse Tasks and Model Scales
https://openreview.net/forum?id=YjBrt82S3v
[ "Ju-Seung Byun", "Andrew Perrault" ]
Poster
Reinforcement learning (RL) training is inherently unstable due to factors such as moving targets and high gradient variance. Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning from AI Feedback (RLAIF) introduce additional challenges. For instance, diverse preferences complicate the alignment ...
Reinforcement learning, Cross entropy, Symmetric loss functions
null
13,418
2405.17618
[ -0.011913019232451916, -0.025824597105383873, -0.0012548189843073487, 0.03127527981996536, 0.0015644378727301955, 0.016783349215984344, 0.005193524993956089, -0.001490914961323142, -0.05204537510871887, -0.047783397138118744, 0.0032809872645884752, 0.007612362504005432, -0.07415799051523209,...
https://github.com/shashacks/symmetric_rl
499
Joint Localization and Activation Editing for Low-Resource Fine-Tuning
https://openreview.net/forum?id=Lllg9YjAFX
[ "Wen Lai", "Alexander Fraser", "Ivan Titov" ]
Poster
Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, are commonly used to adapt LLMs. However, the effectiveness of standard PEFT methods is limited in low-resource scenarios with only a few hundred examples. Recent advances in interpretability research have inspired the emergence of activation editing (or ste...
activation editing, low-resource, localization
Activation editing approach that jointly optimizes the selection of intervention components and the intervention strategy in low-resource setting
13,402
2502.01179
[ 0.001111673889681697, -0.00749852042645216, -0.011533604003489017, 0.025394706055521965, 0.04956702142953873, 0.04249400272965431, 0.028472768142819405, -0.013276400044560432, -0.02849578857421875, -0.009614236652851105, -0.015387783758342266, 0.019743753597140312, -0.06175951659679413, -0...
https://github.com/wenlai-lavine/jola