Mathematical Reasoning on MATH level 3
56AccuracyGradient Shift
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Gradient Shiftvariant=mult2026.01 | 56 | 96 | -0.12 | — | |
| Gradient Shiftvariant=hybrid2026.01 | 55 | 92 | -0.15 | — | |
| Benchmark LLM (BenchAcc)2026.01 | 52 | — | — | — | |
| ME-ICPOBackbone=Qwen2.5-Math-7B, ME-ICPO=true2026.03 | 51.9 | — | — | 58.87 | |
| Weak-to-Strongalpha=1.0, k=402026.01 | 51 | 80 | -0.23 | — | |
| Weak-to-Strongalpha=1.5, k=402026.01 | 50 | 80 | -0.2 | — | |
| ME-ICPOBackbone=Qwen2.5-Math-1.5B, ME-ICPO=true2026.03 | 47.81 | — | — | 52.08 | |
| Weak-to-Strongalpha=2.0, k=402026.01 | 46 | 77 | -0.15 | — | |
| Qwen2.5-Math-7BME-ICPO=false2026.03 | 43.71 | — | — | 49.88 | |
| Qwen2.5-Math-1.5BME-ICPO=false2026.03 | 39.77 | — | — | 45.42 | |
| Gradient Shiftvariant=add2026.01 | 38 | 66 | -0.11 | — | |
| Helper LLM2026.01 | 37 | — | — | — |