Improved Representation Steering for Language Models

• URL: http://arxiv.org/abs/2505.20809v1
• Date: Tue, 27 May 2025 07:16:40 GMT
• Title: Improved Representation Steering for Language Models
• Authors: Zhengxuan Wu, Qinan Yu, Aryaman Arora, Christopher D. Manning, Christopher Potts,
• Abstract summary: We show how to improve representation steering via our new Reference-free Preference Steering (RePS)<n>On Gemma models with sizes ranging from 2B to 27B, RePS outperforms all existing steering methods trained with a language modeling objective.<n>In suppression, RePS matches the language-modeling objective on Gemma-2 and outperforms it on the larger Gemma-3 variants.
• Score: 50.86411958644953
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Steering methods for language models (LMs) seek to provide fine-grained and interpretable control over model generations by variously changing model inputs, weights, or representations to adjust behavior. Recent work has shown that adjusting weights or representations is often less effective than steering by prompting, for instance when wanting to introduce or suppress a particular concept. We demonstrate how to improve representation steering via our new Reference-free Preference Steering (RePS), a bidirectional preference-optimization objective that jointly does concept steering and suppression. We train three parameterizations of RePS and evaluate them on AxBench, a large-scale model steering benchmark. On Gemma models with sizes ranging from 2B to 27B, RePS outperforms all existing steering methods trained with a language modeling objective and substantially narrows the gap with prompting -- while promoting interpretability and minimizing parameter count. In suppression, RePS matches the language-modeling objective on Gemma-2 and outperforms it on the larger Gemma-3 variants while remaining resilient to prompt-based jailbreaking attacks that defeat prompting. Overall, our results suggest that RePS provides an interpretable and robust alternative to prompting for both steering and suppression.

Related papers

• GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs [56.93583799109029]
  GrAInS is an inference-time steering approach that operates across both language-only and vision-language models and tasks.<n>During inference, GrAInS hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale.<n>It consistently outperforms both fine-tuning and existing steering baselines.
  arXiv  Detail & Related papers  (2025-07-24T02:34:13Z)
• ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding [71.654781631463]
  ReAgent-V is a novel agentic video understanding framework.<n>It integrates efficient frame selection with real-time reward generation during inference.<n>Extensive experiments on 12 datasets demonstrate significant gains in generalization and reasoning.
  arXiv  Detail & Related papers  (2025-06-02T04:23:21Z)
• AlphaMaze: Enhancing Large Language Models' Spatial Intelligence via GRPO [0.0]
  Large Language Models (LLMs) have demonstrated impressive capabilities in language processing, yet they often struggle with tasks requiring visual spatial reasoning.<n>We introduce a novel two-stage training framework designed to equip standard LLMs with visual reasoning abilities for maze navigation.
  arXiv  Detail & Related papers  (2025-02-20T16:05:18Z)
• AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders [73.37603699731329]
  We introduce AxBench, a large-scale benchmark for steering and concept detection.<n>For steering, we find that prompting outperforms all existing methods, followed by finetuning.<n>For concept detection, representation-based methods such as difference-in-means, perform the best.
  arXiv  Detail & Related papers  (2025-01-28T18:51:24Z)
• Steering Language Model Refusal with Sparse Autoencoders [16.304363931580273]
  This work uncovers a tension between SAE steering-based safety improvements and general model capabilities.<n>Our findings reveal important open questions about the nature of safety-relevant features in language models.
  arXiv  Detail & Related papers  (2024-11-18T05:47:02Z)
• Steering Without Side Effects: Improving Post-Deployment Control of Language Models [61.99293520621248]
  Language models (LMs) have been shown to behave unexpectedly post-deployment. We present KL-then-steer (KTS), a technique that decreases the side effects of steering while retaining its benefits. Our best method prevents 44% of jailbreak attacks compared to the original Llama-2-chat-7B model.
  arXiv  Detail & Related papers  (2024-06-21T01:37:39Z)
• EMR-Merging: Tuning-Free High-Performance Model Merging [55.03509900949149]
  We show that Elect, Mask & Rescale-Merging (EMR-Merging) shows outstanding performance compared to existing merging methods. EMR-Merging is tuning-free, thus requiring no data availability or any additional training while showing impressive performance.
  arXiv  Detail & Related papers  (2024-05-23T05:25:45Z)
• ReCoRe: Regularized Contrastive Representation Learning of World Model [21.29132219042405]
  We present a world model that learns invariant features using contrastive unsupervised learning and an intervention-invariant regularizer. Our method outperforms current state-of-the-art model-based and model-free RL methods and significantly improves on out-of-distribution point navigation tasks evaluated on the iGibson benchmark.
  arXiv  Detail & Related papers  (2023-12-14T15:53:07Z)
• RecExplainer: Aligning Large Language Models for Explaining Recommendation Models [50.74181089742969]
  Large language models (LLMs) have demonstrated remarkable intelligence in understanding, reasoning, and instruction following. This paper presents the initial exploration of using LLMs as surrogate models to explain black-box recommender models. To facilitate an effective alignment, we introduce three methods: behavior alignment, intention alignment, and hybrid alignment.
  arXiv  Detail & Related papers  (2023-11-18T03:05:43Z)
• ControlVAE: Tuning, Analytical Properties, and Performance Analysis [14.272917020105147]
  ControlVAE is a new variational autoencoder framework. It stabilizes the KL-divergence of VAE models to a specified value. It can achieve a good trade-off between reconstruction quality and KL-divergence.
  arXiv  Detail & Related papers  (2020-10-31T12:32:39Z)
• ControlVAE: Controllable Variational Autoencoder [16.83870832766681]
  Variational Autoencoders (VAE) have been widely used in a variety of applications, such as dialog generation, image generation and disentangled representation learning. ControlVAE combines a controller, inspired by automatic control theory, with the basic VAE to improve the performance of resulting generative models.
  arXiv  Detail & Related papers  (2020-04-13T15:04:56Z)

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.