List your company

Real-world Harm Prediction on ATBench

39AccuracyTRACES
14.87221.13627.433.664May 26, 2026

Evaluation results

MethodLinks
TRACESBase Model=Qwen3-4B2026.0539
TRACESBase Model=Llama3.1-8B-Instruct2026.0536.4
AgentDoGBase Model=Qwen3-4B2026.0535.6
Gemini-3-FlashBase Model=Original2026.0534.8
QWQ-32BBase Model=Original2026.0534.8
AgentDoGBase Model=Llama3.1-8B-Instruct2026.0534.4
Gemini-3-ProBase Model=Original2026.0532
GPT-5.2Base Model=Original2026.0530.8
Qwen3-4BBase Model=Original2026.0516.4
Llama3.1-8BBase Model=Original2026.0515.8