Research map / Large language models / Alignment and preferences
CAMEL: Confidence-Gated Reflection for Reward Modeling
Zirui Zhu, Hailun Xu, Yang Luo, Yong Liu, Kanchan Sarkar, Kun Xu, Yang You
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reflection to all test inputs | CAMEL-Fast · RewardBench | accuracy: 90.5 -> 92.8 | Improved |
| add reflection to all test inputs | CAMEL-Fast · RM-Bench | accuracy: 74.8 -> 84.2 | Improved |
| add reflection to all test inputs | CAMEL-Fast · JudgeBench | accuracy: 65.2 -> 71.6 | Improved |
| add reflection to all inputs (ablation) | Qwen3-14B base · Avg of three benchmarks | accuracy: 71.9 -> 73.8 | Improved |
| add reflection to all inputs (ablation) | Qwen3-14B-SFT · RewardBench | accuracy: 90.6 -> 90.5 | No clear effect |
| add reflection to all inputs (ablation) | Qwen3-14B-GRPO · RewardBench | accuracy: 91.2 -> 90.0 | Hurt |
| add reflection to all inputs (ablation) | Qwen3-14B-GRPO · JudgeBench | accuracy: 62.9 -> 74.2 | Improved |
| add SFT on preference data (ablation) | Qwen3-14B · RewardBench | accuracy: 81.9 -> 90.6 | Improved |
| add counterfactual prefix augmentation and confidence gating to GRPO + reflection (ablation) | CAMEL · Avg of three benchmarks | accuracy: 82.7 -> 81.1 | Hurt |
| add counterfactual prefix augmentation to GRPO (ablation) | CAMEL · RewardBench | accuracy: 91.2 -> 92.4 | Improved |
| raise confidence threshold tau from 3 to 7 (ablation) | CAMEL · RM-Bench | accuracy: 81.3% -> 87.2% | Improved |
| raise confidence threshold tau from 5 to 7 (ablation) | CAMEL · JudgeBench | accuracy: 67.1% -> 66.0% | Hurt |
Closest papers by meaning
- Bayesian Preference Learning for Test-Time Steerable Reward Models (NeurIPS 2026 · similarity 0.90)
- RRM: Robust Reward Model Training Mitigates Reward Hacking (ICLR 2025 · similarity 0.90)
- Capturing Individual Human Preferences with Reward Features (NeurIPS 2025 · similarity 0.90)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.89)
- RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style (ICLR 2025 · similarity 0.89)
- WARM: On the Benefits of Weight Averaged Reward Models (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
