Trajectory Attribution on Who&When n=58 (Hand-Crafted)
73Agent-level AccuracyFALAT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FALATBackbone=DeepSeek V3.22026.05 | 73 | 29.1 | |
| FALATBackbone=MiniMax 2.52026.05 | 66.3 | 25.9 | |
| FALATBackbone=GPT-5.4-Mini2026.05 | 65.5 | 22.2 | |
| AgentTracerBackbone=AgentTracer-8B2026.05 | 63.8 | 20.7 | |
| DirectBackbone=Qwen3-Coder2026.05 | 60.4 | 13.8 | |
| CHIEFBackbone=GPT-5.4-Mini2026.05 | 55.2 | 13.8 | |
| DirectBackbone=DeepSeek-R12026.05 | 53.4 | 6.9 | |
| DirectBackbone=Gemini-2.5-Pro2026.05 | 51.7 | 6.9 | |
| CHIEFBackbone=MiniMax 2.52026.05 | 51 | 17 | |
| CHIEFBackbone=DeepSeek V3.22026.05 | 50 | 10.3 | |
| DirectBackbone=LLaMA-3.2-3B2026.05 | 50 | 3.5 | |
| DirectBackbone=Claude-Sonnet-42026.05 | 50 | 18.9 | |
| DirectBackbone=Qwen3-32B2026.05 | 44.8 | 1.7 | |
| DirectBackbone=Qwen3-8B2026.05 | 39.5 | 3.5 | |
| DirectBackbone=GPT-4.12026.05 | 37.9 | 3.4 |