Downstream Repair on GSM8K trace-level 60/20/20 (test)
88.8Repair AccuracyFull trace
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Full traceRepair model=Gemma 4 E4B IT, Alpha=0.05, Input Strategy=Full trace2026.05 | 88.8 | — | — | |
| Whole-trace abst.Repair model=Gemma 4 E4B IT, Alpha=0.05, Input Strategy=Whole-trace abst.2026.05 | 88.6 | — | — | |
| CROPRepair model=Gemma 4 E4B IT, Alpha=0.05, Input Strategy=CROP prefix2026.05 | 88.3 | -0.91 | -1.53 | |
| CROPRepair model=Qwen2.5-7B-Instruct, Alpha=0.05, Input Strategy=CROP prefix2026.05 | 87.6 | 0.95 | 0.09 | |
| CROPRepair model=DeepSeek-R1-0528-Qwen3-8B, Alpha=0.05, Input Strategy=CROP prefix2026.05 | 87.4 | 1.83 | 0.89 | |
| Whole-trace abst.Repair model=Qwen2.5-7B-Instruct, Alpha=0.05, Input Strategy=Whole-trace abst.2026.05 | 86.3 | — | — | |
| Full traceRepair model=Qwen2.5-7B-Instruct, Alpha=0.05, Input Strategy=Full trace2026.05 | 85.1 | — | — | |
| Full traceRepair model=DeepSeek-R1-0528-Qwen3-8B, Alpha=0.05, Input Strategy=Full trace2026.05 | 84.9 | — | — | |
| Whole-trace abst.Repair model=DeepSeek-R1-0528-Qwen3-8B, Alpha=0.05, Input Strategy=Whole-trace abst.2026.05 | 84.6 | — | — | |
| Question onlyRepair model=Gemma 4 E4B IT, Alpha=0.05, Input Strategy=Question only2026.05 | 82.3 | — | — | |
| Question onlyRepair model=Qwen2.5-7B-Instruct, Alpha=0.05, Input Strategy=Question only2026.05 | 80.5 | — | — | |
| Whole-trace abst.Repair model=Llama3.1-8B, Alpha=0.05, Input Strategy=Whole-trace abst.2026.05 | 74.6 | — | — | |
| CROPRepair model=Llama3.1-8B, Alpha=0.05, Input Strategy=CROP prefix2026.05 | 73.9 | -1.18 | -2.02 | |
| Full traceRepair model=Llama3.1-8B, Alpha=0.05, Input Strategy=Full trace2026.05 | 73.3 | — | — | |
| Question onlyRepair model=Llama3.1-8B, Alpha=0.05, Input Strategy=Question only2026.05 | 71.1 | — | — | |
| Question onlyRepair model=DeepSeek-R1-0528-Qwen3-8B, Alpha=0.05, Input Strategy=Question only2026.05 | 70 | — | — |