List your company

Agent Safety Reasoning on ASSEBench-Corrected

84.72AccuracyDRAFT
30.005644.210358.41572.6197Feb 11, 2026

Evaluation results

MethodLinks
DRAFTBackbone=Qwen2.5-7B-Instruct2026.0284.7287.8682.9293.43
DRAFTBackbone=Qwen3-4B2026.0284.1787.2593.3391.55
DRAFTBackbone=Llama-3.1-8B-Instruct2026.0283.8987.3481.6393.9
ExtractorBackbone=Qwen2.5-7B-Instruct2026.0283.6187.3180.5695.31
SFTBackbone=Qwen3-4B2026.0282.7884.7389.1280.75
DRAFTBackbone=Qwen2.5-3B-Instruct2026.0281.6786.0878.1695.77
SFTBackbone=Llama-3.1-8B-Instruct2026.0281.3984.0985.183.1
SFTBackbone=Qwen2.5-7B-Instruct2026.0281.1183.9684.3683.57
ExtractorBackbone=Qwen3-4B2026.0280.5685.0478.0493.43
DRAFTBackbone=Qwen2.5-1.5B-Instruct2026.0280.2884.6778.4292.02
ExtractorBackbone=Llama-3.1-8B-Instruct2026.0278.6184.1974.8296.24
ExtractorBackbone=Qwen2.5-3B-Instruct2026.0278.0683.6474.8194.84
SFTBackbone=Qwen2.5-1.5B-Instruct2026.0274.1780.2573.2688.73
ExtractorBackbone=Qwen2.5-1.5B-Instruct2026.0273.6181.1970.2196.24
SFTBackbone=Qwen2.5-3B-Instruct2026.0273.6175.284.7167.61
VanillaBackbone=Qwen2.5-3B-Instruct2026.0265.1474.666.8184.44
LoRABackbone=Qwen2.5-7B-Instruct2026.0264.7269.8370.6769.01
VanillaBackbone=Qwen2.5-7B-Instruct2026.0263.3869.3470.3768.34
LoRABackbone=Qwen2.5-3B-Instruct2026.0261.1170.8363.6779.81
VanillaBackbone=Qwen3-4B2026.0260.867.6567.6567.65
LoRABackbone=Qwen3-4B2026.0260.5666.5166.8266.2
LoRABackbone=Llama-3.1-8B-Instruct2026.0257.551.4379.4138.03
VanillaBackbone=Llama-3.1-8B-Instruct2026.0257.153.7977.4241.21
LoRABackbone=Qwen2.5-1.5B-Instruct2026.0235.0119.8636.7113.62
VanillaBackbone=Qwen2.5-1.5B-Instruct2026.0232.1118.0433.6512.33