Investigating Generalization of One-shot LLM Steering Vectors

• URL: http://arxiv.org/abs/2502.18862v1
• Date: Wed, 26 Feb 2025 06:13:01 GMT
• Title: Investigating Generalization of One-shot LLM Steering Vectors
• Authors: Jacob Dunefsky, Arman Cohan,
• Abstract summary: We propose optimizing steering vectors through gradient descent on a single training example.<n>We find that the resulting vectors effectively mediate safety-relevant behaviors in multiple models.
• Score: 21.2431937128876
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Steering vectors have emerged as a promising approach for interpreting and controlling LLMs, but current methods typically require large contrastive datasets that are often impractical to construct and may capture spurious correlations. We propose directly optimizing steering vectors through gradient descent on a single training example, and systematically investigate how these vectors generalize. We consider several steering optimization techniques, including multiple novel ones, and find that the resulting vectors effectively mediate safety-relevant behaviors in multiple models. Indeed, in experiments on an alignment-faking model, we are able to optimize one-shot steering vectors that induce harmful behavior on benign examples and whose negations suppress harmful behavior on malign examples. And in experiments on refusal suppression, we demonstrate that one-shot optimized steering vectors can transfer across inputs, yielding a Harmbench attack success rate of 96.9%. Furthermore, to quantitatively assess steering effectiveness in instruction-tuned models, we develop a novel evaluation framework using sequence probabilities from the corresponding base model. With this framework, we analyze how steering vectors modulate an instruction-tuned LLM's ability to recover from outputting false information, and find that this ability derives from the base model. Overall, our findings suggest that optimizing steering vectors on a single example can mediate misaligned behavior in LLMs, and provide a path toward better understanding the relationship between LLM behavior and activation space structure.

Related papers

• Improving Reasoning Performance in Large Language Models via Representation Engineering [2.0099933815960256]
  We propose a representation engineering approach for large language models (LLMs) Model activations are read from the residual stream of an LLM when processing a reasoning task. We show that an LLM can, to a certain degree, be controlled to improve its perceived reasoning ability by modulating activations.
  arXiv  Detail & Related papers  (2025-04-28T04:58:43Z)
• SEAL: Steerable Reasoning Calibration of Large Language Models for Free [58.190800043449336]
  Large Language Models (LLMs) have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism. Recent studies reveal substantial redundancy in the CoT reasoning traces, which negatively impacts model performance. We introduce SEAL, a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains.
  arXiv  Detail & Related papers  (2025-04-07T02:42:07Z)
• Shifting Perspectives: Steering Vector Ensembles for Robust Bias Mitigation in LLMs [8.91107152198979]
  We present a novel approach to bias mitigation in large language models (LLMs) by applying steering vectors to modify model activations in forward passes. We employ Bayesian optimization to systematically identify effective contrastive pair datasets across nine bias axes. Building on these promising results, we introduce Steering Vector Ensembles (SVE), a method that averages multiple individually optimized steering vectors, each targeting a specific bias axis such as age, race, or gender.
  arXiv  Detail & Related papers  (2025-03-07T12:25:29Z)
• Debias your Large Multi-Modal Model at Test-Time via Non-Contrastive Visual Attribute Steering [7.471995248769638]
  We propose a training-free debiasing framework for large Multi-Modal Models (LMMs) Our framework intervenes on the model's representations during text generation by constructing a steering vector that reduces reference on protected attributes. Our experiments show that these interventions effectively reduce the propensity of LMMs to generate text related to protected attributes while maintaining sentiment and fluency.
  arXiv  Detail & Related papers  (2024-11-15T20:06:09Z)
• Refusal in LLMs is an Affine Function [1.722461331472526]
  We propose affine concept editing (ACE) as an approach for steering language models' behavior.<n>ACE combines affine subspace projection and activation addition to reliably control the model's refusal responses.<n>Our experiments demonstrate that ACE consistently achieves more precise control over model behavior than existing methods.
  arXiv  Detail & Related papers  (2024-11-13T20:12:55Z)
• 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)
• Aligning Language Models with Demonstrated Feedback [58.834937450242975]
  Demonstration ITerated Task Optimization (DITTO) directly aligns language model outputs to a user's demonstrated behaviors. We evaluate DITTO's ability to learn fine-grained style and task alignment across domains such as news articles, emails, and blog posts.
  arXiv  Detail & Related papers  (2024-06-02T23:13:56Z)
• Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization [34.05163996072159]
  "steering vectors" are extracted from the activations of human preference data. This work proposes an innovative approach that could produce more effective steering vectors through bi-directional preference optimization. Our method is designed to allow steering vectors to directly influence the generation probability of contrastive human preference data pairs.
  arXiv  Detail & Related papers  (2024-05-28T05:10:40Z)
• InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance [56.184255657175335]
  We develop textbfInferAligner, a novel inference-time alignment method that utilizes cross-model guidance for harmlessness alignment. Experimental results show that our method can be very effectively applied to domain-specific models in finance, medicine, and mathematics. It significantly diminishes the Attack Success Rate (ASR) of both harmful instructions and jailbreak attacks, while maintaining almost unchanged performance in downstream tasks.
  arXiv  Detail & Related papers  (2024-01-20T10:41:03Z)
• Amortizing intractable inference in large language models [56.92471123778389]
  We use amortized Bayesian inference to sample from intractable posterior distributions. We empirically demonstrate that this distribution-matching paradigm of LLM fine-tuning can serve as an effective alternative to maximum-likelihood training. As an important application, we interpret chain-of-thought reasoning as a latent variable modeling problem.
  arXiv  Detail & Related papers  (2023-10-06T16:36:08Z)
• Relational Reasoning via Set Transformers: Provable Efficiency and Applications to MARL [154.13105285663656]
  A cooperative Multi-A gent R einforcement Learning (MARL) with permutation invariant agents framework has achieved tremendous empirical successes in real-world applications. Unfortunately, the theoretical understanding of this MARL problem is lacking due to the curse of many agents and the limited exploration of the relational reasoning in existing works. We prove that the suboptimality gaps of the model-free and model-based algorithms are independent of and logarithmic in the number of agents respectively, which mitigates the curse of many agents.
  arXiv  Detail & Related papers  (2022-09-20T16:42:59Z)
• Counterfactual Maximum Likelihood Estimation for Training Deep Networks [83.44219640437657]
  Deep learning models are prone to learning spurious correlations that should not be learned as predictive clues. We propose a causality-based training framework to reduce the spurious correlations caused by observable confounders. We conduct experiments on two real-world tasks: Natural Language Inference (NLI) and Image Captioning.
  arXiv  Detail & Related papers  (2021-06-07T17:47:16Z)

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.