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

• URL: http://arxiv.org/abs/2605.16339v1
• Date: Thu, 07 May 2026 16:48:48 GMT
• Title: Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders
• Abstract summary: Preference learning in large language models relies on reward models as proxies for human judgment.<n>We analyze this instability at the representation level under three semantic-preserving perturbation types.<n>We propose two SAE-based instability mitigation strategies.
• Score: 20.13794189010598
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Preference learning in large language models relies on reward models as proxies for human judgment. However, these models frequently exhibit preference instability, producing contradictory preference assignments in response to subtle, meaning-preserving input variations. We analyze this instability at the representation level under three semantic-preserving perturbation types: paraphrasing, pattern injection, and backdoor triggers. We attribute this instability to over-reliance on predictive yet brittle features, which we term unstable features, and isolate them via Sparse Autoencoders (SAEs) in a sparse latent space where benign and perturbed inputs activate distinctly separable patterns. Building on this separability, we propose two SAE-based instability mitigation strategies: SAE Feature Steering, which identifies and suppresses anomalously activated features at inference, and SAE Residual Correction, which learns adaptive adjustments over SAE features to restore correct preferences. Our methods substantially reduce incorrect preference assignments on harmlessness and hallucination benchmarks while preserving benign performance and general utility on other tasks, without retraining the reward model. Our code and data are available in \url{https://github.com/shunchang-liu/pisa}.

Related papers

• C$^{2}$R: Cross-sample Consistency Regularization Mitigates Feature Splitting and Absorption in Sparse Autoencoders [56.67585330425431]
  Sparse Autoencoders (SAEs) are widely used to interpret large language models by decomposing activations into sparse, human-understandable features.<n>These issues stem from inconsistent latent assignment across samples.<n>We introduce C$2$R, which penalizes the co-activation of directionally similar latents.
  arXiv  Detail & Related papers  (2026-06-29T17:45:31Z)
• Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders [7.5916620727820225]
  Sparse autoencoders (SAEs) are widely used to interpret neural network representations.<n>We study this question through emphfeature stability: for each SAE feature, we estimate the probability that a similar feature reappears in an independently trained SAE.<n>We find a pronounced functional asymmetry: stable features carry most of the reconstruction- and prediction-relevant signal, while unstable features have weak marginal impact.
  arXiv  Detail & Related papers  (2026-06-10T14:32:57Z)
• FACTOR: Counterfactual Training-Free Test-Time Adaptation for Open-Vocabulary Object Detection [63.91351553178842]
  FACTOR is a training-free Test-time adaptation for Open-vocabulary object detection.<n>It quantifies attribute sensitivity, semantic relevance, and prediction variation to selectively suppress attribute-dependent predictions.<n>Experiments on PASCAL-C, COCO-C, and FoggyCityscapes show that FACTOR consistently outperforms prior TTA methods.
  arXiv  Detail & Related papers  (2026-05-05T02:31:18Z)
• Functional Subspace Watermarking for Large Language Models [10.606696096012291]
  We propose a framework that anchors ownership signals into a low-dimensional functional backbone.<n>Our method achieves superior detection accuracy and statistical verifiability under multiple model attacks.
  arXiv  Detail & Related papers  (2026-03-19T11:44:34Z)
• CORE: Context-Robust Remasking for Diffusion Language Models [51.59514489363897]
  We propose Context-Robust Remasking (CORE), a training-free framework for inference-time revision.<n>Rather than trusting static token probabilities, CORE identifies context-brittle tokens by probing their sensitivity to targeted masked-context perturbations.<n>On LLaDA-8B-Base, CORE delivers consistent improvements across reasoning and code benchmarks, outperforming compute-matched baselines and improving MBPP by up to 9.2 percentage points.
  arXiv  Detail & Related papers  (2026-02-04T00:12:30Z)
• Analysis of Variational Sparse Autoencoders [1.675385127117872]
  We investigate whether incorporating variational methods into SAE architectures can improve feature organization and interpretability.<n>We introduce the Variational Sparse Autoencoder (vSAE), which replaces deterministic ReLU gating with sampling from learned Gaussian posteriors.<n>Our findings suggest that naive application of variational methods to SAEs does not improve feature organization or interpretability.
  arXiv  Detail & Related papers  (2025-09-26T23:09:56Z)
• Advancing Reliable Test-Time Adaptation of Vision-Language Models under Visual Variations [67.35596444651037]
  Vision-language models (VLMs) exhibit remarkable zero-shot capabilities but struggle with distribution shifts in downstream tasks when labeled data is unavailable.<n>We propose a Reliable Test-time Adaptation (ReTA) method that enhances reliability from two perspectives.
  arXiv  Detail & Related papers  (2025-07-13T05:37:33Z)
• Principled Input-Output-Conditioned Post-Hoc Uncertainty Estimation for Regression Networks [1.4671424999873808]
  Uncertainty is critical in safety-sensitive applications but is often omitted from off-the-shelf neural networks due to adverse effects on predictive performance.<n>We propose a theoretically grounded framework for post-hoc uncertainty estimation in regression tasks by fitting an auxiliary model to both original inputs and frozen model outputs.
  arXiv  Detail & Related papers  (2025-06-01T09:13:27Z)
• AFD: Mitigating Feature Gap for Adversarial Robustness by Feature Disentanglement [56.90364259986057]
  Adversarial fine-tuning methods enhance adversarial robustness via fine-tuning the pre-trained model in an adversarial training manner.<n>We propose a disentanglement-based approach to explicitly model and remove the specific latent features.<n>Our approach surpasses existing adversarial fine-tuning methods and adversarial training baselines.
  arXiv  Detail & Related papers  (2024-01-26T08:38:57Z)
• Feature Separation and Recalibration for Adversarial Robustness [18.975320671203132]
  We propose a novel, easy-to- verify approach named Feature Separation and Recalibration. It recalibrates the malicious, non-robust activations for more robust feature maps through Separation and Recalibration. It improves the robustness of existing adversarial training methods by up to 8.57% with small computational overhead.
  arXiv  Detail & Related papers  (2023-03-24T07:43:57Z)
• Robustness and Accuracy Could Be Reconcilable by (Proper) Definition [109.62614226793833]
  The trade-off between robustness and accuracy has been widely studied in the adversarial literature. We find that it may stem from the improperly defined robust error, which imposes an inductive bias of local invariance. By definition, SCORE facilitates the reconciliation between robustness and accuracy, while still handling the worst-case uncertainty.
  arXiv  Detail & Related papers  (2022-02-21T10:36:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.