Research map / Large language models / Alignment and preferences
Factored Causal Representation Learning for Robust Reward Modeling in RLHF
Yupei Yang, Lin Yang, Wanxi Deng, Lin Qu, Fan Feng, Biwei Huang, Shikui Tu, Lei Xu
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 |
|---|---|---|---|
| factorize latent into causal and non-causal parts (vs single bottleneck latent, InfoRM-style) (ablation) | CausalRM · Math reward model ID avg | pairwise accuracy (%): 66.1 -> 70.1 | Improved |
| add latent factorization (ablation) | CausalRM · Math reward model length sensitivity | sigma_len: 0.14 -> 0.03 | Improved |
| add adversarial head with gradient reversal (ablation) | CausalRM · Math reward model OOD avg | pairwise accuracy (%): 81.5 -> 85.6 | Improved |
| add adversarial head with gradient reversal (ablation) | CausalRM · Math length sensitivity | sigma_len: 0.13 -> 0.03 | Improved |
| add reconstruction decoder loss (ablation) | CausalRM · Math reward model OOD avg | pairwise accuracy (%): 83.1 -> 85.6 | Improved |
| add KL regularization on causal latent zc (ablation) | CausalRM · Math reward model OOD avg | pairwise accuracy (%): 81.1 -> 85.6 | Improved |
| add KL regularization on non-causal latent znc (ablation) | CausalRM · Math reward model ID avg | pairwise accuracy (%): 64.0 -> 70.1 | Improved |
| add KL regularization on both latents (ablation) | CausalRM · Sycophancy-hacked test ID avg | pairwise accuracy (%): 46.0 -> 70.6 | Improved |
| replace Standard RM with CausalRM reward in PPO | Qwen2.5-Math-7B RLHF · Math RLHF ID avg | final-answer accuracy (%): 67.8 -> 74.0 | Improved |
| use InfoRM reward in PPO instead of Standard RM | Qwen2.5-Math-7B RLHF · Math RLHF ID avg | final-answer accuracy (%): 67.8 -> 48.0 | Hurt |
| use CausalRM instead of Standard RM under sycophantic prefix artifact | reward model on hacked Anthropic-HH · Sycophancy-hacked ID avg | pairwise accuracy (%): 59.2 -> 70.6 | Improved |
| use CausalRM instead of artifact-specific ODIN/MMD for length bias | math reward model · Math ID avg | pairwise accuracy (%): 69.3 -> 70.1 (vs MMD); 68.8 -> 70.1 (vs ODIN) | Improved |
Closest papers by meaning
- RRM: Robust Reward Model Training Mitigates Reward Hacking (ICLR 2025 · similarity 0.90)
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · similarity 0.89)
- InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling (NeurIPS 2024 · similarity 0.89)
- Rectifying Shortcut Behaviors in Preference-based Reward Learning (NeurIPS 2025 · similarity 0.89)
- The Trickle-down Impact of Reward Inconsistency on RLHF (ICLR 2024 · similarity 0.89)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · 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.
