List your company

Misaligned action detection on MISACTBENCH

89.9PrecisionDEACTION
45.38856.94468.580.056Feb 9, 2026

Evaluation results

MethodLinks
DEACTIONBackbone=GPT-5.1 Thinking2026.0289.976.885.982.8
DEACTIONBackbone=Claude Sonnet 4.52026.0288.27484.180.4
DEACTIONBackbone=Qwen3-VL-32B2026.0280.163.376.970.7
DEACTIONBackbone=GPT-5.1 Instant2026.0273.486.480.279.4
Task ShieldBackbone=GPT-5.1 Thinking2026.0261.373.667.866.9
Task ShieldBackbone=Claude Sonnet 4.52026.0259.575.866.566.6
InferActBackbone=GPT-5.1 Thinking2026.025670.162.562.3
Task ShieldBackbone=GPT-5.1 Instant2026.0251.488.85865.1
InferActBackbone=GPT-5.1 Instant2026.0251.487.35864.7
Task ShieldBackbone=Qwen3-VL-32B2026.0250.66956.658.4
InferActBackbone=Claude Sonnet 4.52026.0248.49653.164.3
InferActBackbone=Qwen3-VL-32B2026.0247.18951.161.6