Reward Hacking Onset Localization on CHERRL VerInstruct and HealthBench runs
490VerInstruct SPCC-Qwen
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| CC-Qwenframework=Claude Code, backbone=Qwen2026.06 | 490 | 220 | 96 | 91 | 341 | 474 | 198 | 80 | 0 | |
| CC-Haikuframework=Claude Code, backbone=Haiku2026.06 | 490 | 150 | 100 | 101 | 331 | 158 | 420 | 329 | 0 | |
| RHDA-397Bbackbone=qwen3.5-397B-A17B2026.06 | 489 | 157 | 76 | 83 | 385 | 459 | 167 | 20 | 0 | |
| RHDA-Plusbackbone=Qwen3.5-plus2026.06 | 482 | 132 | 86 | 75 | 383 | 454 | 120 | 11 | 0 | |
| Reference2026.06 | 478 | 116 | 91 | 68 | 301 | 460 | — | — | — | |
| CC-Opusframework=Claude Code, backbone=Opus2026.06 | 470 | 151 | 110 | 90 | 121 | 450 | 274 | 224 | 0 | |
| CC-Sonnetframework=Claude Code, backbone=Sonnet2026.06 | 463 | 218 | 93 | 68 | 437 | 446 | 269 | 86 | 0 | |
| CoT Monitor2026.06 | 332 | 169 | — | — | 283 | — | 217 | 172 | 3 |