Grounded Payoff Tracking on BookSum
69.8Detection AccuracyCFPG
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CFPGModel=GPT-4.1-mini2026.01 | 69.8 | 166 | 11 | 5.76 | 0.647 | |
| Foreshadow-Aware PromptModel=GPT-4.1-mini2026.01 | 58 | 235 | 17 | 8.85 | 0.453 | |
| Context Refresh (Sim.)Model=GPT-4.1-mini2026.01 | 48.6 | 306 | 7 | 13.05 | 0.382 | |
| CFPGModel=Llama3.1 8B-Inst.2026.01 | 27.3 | 439 | 8 | 19.91 | 0.226 | |
| Foreshadow-Aware PromptModel=Llama3.1 8B-Inst.2026.01 | 25.8 | 451 | 5 | 20.04 | 0.182 | |
| Context Refresh (Sim.)Model=Llama3.1 8B-Inst.2026.01 | 22.1 | 482 | 3 | 23.72 | 0.16 | |
| Context Refresh (Sim.)Model=Qwen2.5 7B-Inst.2026.01 | 19.6 | 493 | 7 | 23.79 | 0.177 | |
| CFPGModel=Qwen2.5 7B-Inst.2026.01 | 19.6 | 490 | 5 | 26.23 | 0.184 | |
| Foreshadow-Aware PromptModel=Qwen2.5 7B-Inst.2026.01 | 15.6 | 522 | 2 | 27 | 0.114 | |
| CFPGModel=Qwen2.5 3B-Inst.2026.01 | 10.8 | 550 | 5 | 31.63 | 0.08 | |
| Context Refresh (Sim.)Model=Qwen2.5 3B-Inst.2026.01 | 6.7 | 587 | 0 | 32.81 | 0.039 | |
| Foreshadow-Aware PromptModel=Qwen2.5 3B-Inst.2026.01 | 4.5 | 601 | 0 | 34.67 | 0.022 |