提出新方法缓解多模态大模型在评估任务中的盲目认同偏差，提升判断准确性。
Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification
- 通过生成独立于输入的通用行为先验，再基于先验评估任务轨迹。
- 在多个场景中使错误检测能力提升25个百分点，任务完成率最高提高20个百分点。
- 适合需要可靠自动评估的AI自改进、在线监督等系统开发者使用。
验证器——对智能体行为赋予奖励的函数——在数学、编程和游戏等领域推动了人工智能进展。然而，将这些成果扩展到缺乏明确成功标准的领域仍具挑战：尽管人类能识别期望结果，但将其转化为可扩展规则并不容易。多模态大模型（MLLMs）凭借其世界知识、与人类偏好对齐的能力及推理能力，提供了有前景的解决方案。我们评估了13+种模型、28+种设计下的MLLM验证器在网页导航、计算机操作和机器人任务中的表现，涵盖数千条智能体轨迹。发现一个关键局限：MLLM存在过度认可智能体行为的倾向，称为‘共识偏差’。该偏差普遍存在，且在测试时缩放下依然顽固，可能损害依赖于MLLM判断或奖励的应用（如自改进、引导、在线监督）。我们讨论了评估和设计MLLM验证器的若干考量，并引入轻量级方法SGV，通过调节（无）条件生成来更好利用其能力。首先，让MLLM生成与评估数据无关的广泛行为先验；然后，基于自生成的先验，对候选轨迹进行推理与评估。该方法显著提升了人类对齐度，使错误检测提升25个百分点，准确率提升14个百分点。在自改进和在线监督中，其使OSWorld中的GUI专家、robomimic中的扩散策略以及VisualWebArena中的ReAct智能体的任务完成率超过此前最优水平，最高提升20个百分点。作为副产品，我们发布了更新版VisualWebArena，包含更强的智能体基线、更符合人类判断的标注者、容器化并行与高保真重置机制、超过10倍的速度提升，以及VWA-Lite——一个仅含1/3数据但评估保真度相当的子集。
原文摘要 · Abstract (English)
Verifiers--functions assigning rewards to agent behavior--have been key to AI progress in math, code, and games. However, extending gains to domains without clear-cut success criteria remains a challenge: while humans can recognize desired outcomes, translating this intuition into scalable rules is nontrivial. Multimodal LLMs (MLLMs) offer a promising solution, given their world knowledge, human-preference alignment, and reasoning capabilities. We evaluate MLLM verifiers across web navigation, computer use, and robotics, spanning 13+ models, 28+ designs, and thousands of trajectories from diverse agents. We identify a critical limitation: a strong tendency for MLLMs to over-validate agent behavior--a phenomenon we term agreement bias. This bias is pervasive, resilient to test-time scaling, and can harm applications relying on MLLM judgments/rewards (e.g., self-improvement, steering, online supervision). We discuss several considerations for evaluating and designing MLLM verifiers, and introduce SGV, a lightweight method that better leverages their capabilities by modulating (un)conditional generation. First, an MLLM is elicited to generate broad priors about desired behavior, independent of the data under evaluation. Then, conditioned on self-generated priors, it reasons over and evaluates a candidate trajectory. Our methods yield more human-aligned verifiers, improving failure detection by 25pp and accuracy by 14pp. In self-improvement and online supervision, they boost task completion of a GUI specialist in OSWorld, a diffusion policy in robomimic, and a ReAct agent in VisualWebArena--surpassing the previous state of the art by 20pp. As a byproduct, we release an update of VisualWebArena featuring strong agent baselines, more human-aligned oracles, container parallelism with high fidelity and proper resets, >10x speedups, and VWA-Lite, a 1/3 subset with comparable evaluation fidelity.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。