ConflictVLA-Bench：评测VLA模型对前提冲突的行为反应
原标题：ConflictVLA-Bench: Benchmarking Behavioral Responses of Vision-Language-Action Models to Premise Conflicts
AI 摘要
研究者提出 ConflictVLA-Bench 基准，基于 LIBERO 构建 2826 个含前提冲突的任务，覆盖四类冲突家族与四种结构配置，并将冲突 rollout 与前提一致的成功参考 rollout 配对，同时评估结果与执行过程。对八个 VLA 模型的测试显示，无效前提使原目标完成率至少下降 17.3 个百分点，OpenVLA 下降达 56.2 个百分点；即便模型在一致任务上成功、在冲突任务上失败，它们仍常继续接近原目标、保留早期轨迹结构且动作幅度抑制有限，作者称之为 Failed Persistence。显式前提检查并未稳定产生选择性且协调的行为改变，说明仅凭终端失败无法判断模型是否已停止追求无效目标。
正文节选
ConflictVLA-Bench: Benchmarking Behavioral Responses of Vision-Language-Action Models to Premise Conflicts Abstract While Vision-Language-Action (VLA) models perform strongly on manipulation tasks, their responses to invalid task premises remain underexplored. Existing evaluations of premise conflicts often focus on terminal task outcomes, yet task failure alone cannot distinguish behavioral disengagement from continued pursuit followed by an execution error. We call the latter pattern Failed Pe