Research map / Large language models / Alignment and preferences
Interpretable Reward Model via Sparse Autoencoder
Shuyi Zhang, Wei Shi, Sihang Li, Jiayi Liao, Tao Liang, Hengxing Cai, Xiang Wang
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 pretrained SAE encoder instead of random init (ablation) | SARM-4B · RewardBench 2 | Overall score: 68.4 -> 73.6 | Improved |
| pretrain SAE on sequence-level instead of token-level activations (ablation) | SARM-4B · RewardBench 2 | Overall score: 71.5 -> 73.6 | Improved |
| pretrain SAE on sequence-level instead of token-level activations (ablation) | SARM-4B · RewardBench 2 | Ties: 72.5 -> 79.6 | Improved |
| pretrain SAE on sequence-level instead of token-level activations (ablation) | SARM-4B · RewardBench 2 | Safety: 92.9 -> 91.3 | Hurt |
| move SAE from layer 7 to layer 14 (ablation) | SARM-2B · RewardBench 2 | Overall score: 36.21 -> 62.52 | Improved |
| move SAE from layer 14 to layer 21 (ablation) | SARM-2B · RewardBench 2 | Overall score: 62.52 -> 64.17 | Improved |
| move SAE from layer 21 to layer 28 (ablation) | SARM-2B · RewardBench 2 | Overall score: 64.17 -> 62.94 | Hurt |
| increase SAE feature dimension from 8x to 32x (ablation) | SARM-2B · RewardBench 2 | Overall score: 63.04 -> 62.94 | No clear effect |
| increase SAE sparsity k from 48 to 240 (ablation) | SARM-2B · RewardBench 2 | Overall score: 61.99 -> 62.89 | No clear effect |
| scale backbone from SARM-3B to SARM-4B | SARM · RewardBench 2 | Overall score: 64.2 -> 73.6 | Improved |
| multiply value-head weight of a safety feature by a constant | SARM · RewardBench 2 | reward distribution | Improved |
Closest papers by meaning
- GRAM: A Generative Foundation Reward Model for Reward Generalization (ICML 2025 · similarity 0.90)
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · similarity 0.90)
- RRM: Robust Reward Model Training Mitigates Reward Hacking (ICLR 2025 · similarity 0.89)
- Tool-Augmented Reward Modeling (ICLR 2024 · similarity 0.89)
- Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs (NeurIPS 2024 · similarity 0.89)
- Dense Reward for Free in Reinforcement Learning from Human Feedback (ICML 2024 · similarity 0.88)
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.
