Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders

arXiv:2605.16339 · cs.LG · Submitted 2026-05-07 · Read on arXiv

cs.LG

Submitted: 2026-05-07

Updated: 2026-09-30

Code: https://github.com/shunchang-liu/pisa

Terminology

• Benchmark — 1× in this paper · explained in 2 episodes
• Hallucination — 1× in this paper · explained in 4 episodes
• Latent Space — 1× in this paper · explained in 7 episodes
• Sparse Autoencoders (SAEs) — 1× in this paper · explained in 3 episodes
• Steering — 1× in this paper · explained in 2 episodes

Sources

• Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
• Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
• Exploring and Addressing Reward Confusion in Offline Preference Learning
• Learning Safety Constraints for Large Language Models
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• Safe RLHF: Safe Reinforcement Learning from Human Feedback
• Toy Models of Superposition
• Detecting Strategic Deception Using Linear Probes
• RewardBench: Evaluating Reward Models for Language Modeling
• SAFER: Probing Safety in Reward Models with Sparse Autoencoder
• TruthfulQA: Measuring How Models Mimic Human Falsehoods
• Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy
• SAE-V: Interpreting Multimodal Models for Enhanced Alignment
• The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models
• Improving Dictionary Learning with Gated Sparse Autoencoders
• Universal Jailbreak Backdoors from Poisoned Human Feedback
• Polysemanticity and Capacity in Neural Networks
• Proximal Policy Optimization Algorithms
• Towards Understanding Sycophancy in Language Models
• The Trickle-down Impact of Reward (In-)consistency on RLHF

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks