autotrust/JEV-27B-VL is a vision-capable extension of autotrust/JEV-27B, pairing the same calibrated System 1 typed-decision interface (yes/no, pick-one-of-2–256-options, or 0–5 rating, returned as a probability distribution in a single forward pass) with the unmodified Qwen3.8-27B as System 2, now extended to accept images and multimodal prompts up to 256K tokens via a POST /v1/decide endpoint. Text-only decisions match the original JEV-27B almost exactly (mean probability difference 0.010 across 1,000 checks), while image-based System 1 decisions prove strong across a wide range of zero-shot applications: 75% success on a MuJoCo robot-arm pick-and-place task and 95% on 60 multi-step browser computer-use tasks (both the best or tied-best in the JEV family), an AUC of 0.727 for zero-shot short-video recommendation from thumbnail covers alone (matching collaborative filtering trained on 59,045 users' logs), 73.2% on Plan-RewardBench (top of that paper's leaderboard) and higher precision than all 16 AgentRewardBench leaderboard judges at matched recall, and 78.3% on VL-RewardBench (above every model on its 2025 leaderboard, including GPT-4o and Claude 3.5 Sonnet). It also extends choice questions to up to 256 options with no retraining via two-letter labels beyond the trained 16, and the card documents concrete prompt-writing guidance — one-line "use when" descriptions for similar options measurably help, while wording, JSON vs. plain text, and extra instructions do not. It's served via a patched vLLM server (serve_decide.py) requiring --max-num-seqs 8 for correct multimodal LoRA behavior, with image-decision calibration not yet systematically measured, and is released under Apache-2.0, built from the unchanged Qwen3.8-27B weights plus the JEV System 1 adapter and decision head.
