Reward Aggregation on Average HealthBench, WritingBench, PLawBench
75.7Average ScoreRuscaRL + GEAR
Evaluation results
| Method | Links | ||
|---|---|---|---|
| RuscaRL + GEARPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 75.7 | 6.5 | |
| RuscaRL + GEARPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 74.7 | 6.1 | |
| GEARPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Aggregation-only2026.06 | 73.9 | 7.3 | |
| GEARPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Aggregation-only2026.06 | 72.6 | 6.2 | |
| RuscaRLPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 69.3 | — | |
| RuscaRLPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 68.6 | — | |
| Flat aggregationPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Aggregation-only2026.06 | 66.6 | — | |
| Flat aggregationPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Aggregation-only2026.06 | 66.4 | — | |
| RaR-Explicit + GEARPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 66.1 | 6 | |
| RaR-Explicit + GEARPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 65.1 | 5.5 | |
| Hard gatingPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Aggregation-only2026.06 | 64.1 | — | |
| Hard gatingPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Aggregation-only2026.06 | 64 | — | |
| RaR-ExplicitPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 60 | — | |
| RaR-ExplicitPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 59.6 | — | |
| MeRF + GEARPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 58.9 | 5.5 | |
| MeRF + GEARPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 58.2 | 5.6 | |
| MeRFPolicy Model=Llama-3.1-8B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 53.4 | — | |
| MeRFPolicy Model=Qwen2.5-7B-Instruct, Pipeline Category=Plug-in replacement in rubric-guided RL pipelines2026.06 | 52.6 | — |