Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoders

• URL: http://arxiv.org/abs/2505.14536v1
• Date: Tue, 20 May 2025 15:55:31 GMT
• Title: Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoders
• Authors: Agam Goyal, Vedant Rathi, William Yeh, Yian Wang, Yuen Chen, Hari Sundaram,
• Abstract summary: Large language models (LLMs) are now ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs.<n>We leverage sparse autoencoders (SAEs) to identify toxicity-related directions in the residual stream of models and perform targeted activation steering.
• Score: 3.6367827664262715
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models (LLMs) are now ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs, including profanity, vulgarity, and derogatory remarks. Although numerous detoxification methods exist, most apply broad, surface-level fixes and can therefore easily be circumvented by jailbreak attacks. In this paper we leverage sparse autoencoders (SAEs) to identify toxicity-related directions in the residual stream of models and perform targeted activation steering using the corresponding decoder vectors. We introduce three tiers of steering aggressiveness and evaluate them on GPT-2 Small and Gemma-2-2B, revealing trade-offs between toxicity reduction and language fluency. At stronger steering strengths, these causal interventions surpass competitive baselines in reducing toxicity by up to 20%, though fluency can degrade noticeably on GPT-2 Small depending on the aggressiveness. Crucially, standard NLP benchmark scores upon steering remain stable, indicating that the model's knowledge and general abilities are preserved. We further show that feature-splitting in wider SAEs hampers safety interventions, underscoring the importance of disentangled feature learning. Our findings highlight both the promise and the current limitations of SAE-based causal interventions for LLM detoxification, further suggesting practical guidelines for safer language-model deployment.

Related papers

• ARMOR: Aligning Secure and Safe Large Language Models via Meticulous Reasoning [49.47193675702453]
  Large Language Models (LLMs) have demonstrated remarkable generative capabilities.<n>LLMs remain vulnerable to malicious instructions that can bypass safety constraints.<n>We propose a reasoning-based safety alignment framework, ARMOR, that replaces the ad-hoc chains of thought reasoning process with human-aligned, structured one.
  arXiv  Detail & Related papers  (2025-07-14T09:05:54Z)
• Adaptive Detoxification: Safeguarding General Capabilities of LLMs through Toxicity-Aware Knowledge Editing [49.85884082568318]
  ToxEdit is a toxicity-aware knowledge editing approach.<n>It dynamically detects toxic activation patterns during forward propagation.<n>It then routes computations through adaptive inter-layer pathways to mitigate toxicity effectively.
  arXiv  Detail & Related papers  (2025-05-28T12:37:06Z)
• Breaking mBad! Supervised Fine-tuning for Cross-Lingual Detoxification [31.7516400680833]
  "Cross-lingual Detoxification" is a paradigm that mitigates toxicity in large language models.<n>We analyze toxicity reduction in cross-distribution settings and investigate how mitigation impacts model performance on non-toxic tasks.
  arXiv  Detail & Related papers  (2025-05-22T14:30:14Z)
• AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender [73.09848497762667]
  We propose AdaSteer, an adaptive activation steering method that adjusts model behavior based on input characteristics.<n>AdaSteer steers input representations along both the Rejection Direction (RD) and Harmfulness Direction (HD)<n>Our results highlight the potential of interpretable model internals for real-time, flexible safety enforcement in LLMs.
  arXiv  Detail & Related papers  (2025-04-13T07:39:17Z)
• Representation Bending for Large Language Model Safety [27.842146980762934]
  Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks pose significant challenges.<n>This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs.<n>RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates.
  arXiv  Detail & Related papers  (2025-04-02T09:47:01Z)
• Improving LLM Safety Alignment with Dual-Objective Optimization [65.41451412400609]
  Existing training-time safety alignment techniques for large language models (LLMs) remain vulnerable to jailbreak attacks.<n>We propose an improved safety alignment that disentangles DPO objectives into two components: (1) robust refusal training, which encourages refusal even when partial unsafe generations are produced, and (2) targeted unlearning of harmful knowledge.
  arXiv  Detail & Related papers  (2025-03-05T18:01:05Z)
• SafetyAnalyst: Interpretable, transparent, and steerable safety moderation for AI behavior [56.10557932893919]
  We present SafetyAnalyst, a novel AI safety moderation framework.<n>Given an AI behavior, SafetyAnalyst uses chain-of-thought reasoning to analyze its potential consequences.<n>It aggregates all harmful and beneficial effects into a harmfulness score using fully interpretable weight parameters.
  arXiv  Detail & Related papers  (2024-10-22T03:38:37Z)
• Large Language Models can be Strong Self-Detoxifiers [82.6594169242814]
  Self-disciplined Autoregressive Sampling (SASA) is a lightweight controlled decoding algorithm for toxicity reduction of large language models (LLMs) SASA tracks the margin of the current output to steer the generation away from the toxic subspace, by adjusting the autoregressive sampling strategy. evaluated on LLMs of different scale and nature, namely Llama-3.1-Instruct (8B), Llama-2 (7B), and GPT2-L models with the RealToxicityPrompts, BOLD, and AttaQ benchmarks.
  arXiv  Detail & Related papers  (2024-10-04T17:45:15Z)
• DESTEIN: Navigating Detoxification of Language Models via Universal Steering Pairs and Head-wise Activation Fusion [16.989349884904943]
  Current solutions involving finetuning or auxiliary models usually require extensive computational resources. We propose DeStein, a novel method that detoxifies LMs by applying representation engineering in activation spaces with lower resource and time costs.
  arXiv  Detail & Related papers  (2024-04-16T11:07:48Z)
• Detoxifying Large Language Models via Knowledge Editing [57.0669577257301]
  This paper investigates using knowledge editing techniques to detoxify Large Language Models (LLMs) We construct a benchmark, SafeEdit, which covers nine unsafe categories with various powerful attack prompts. We conduct experiments with several knowledge editing approaches, indicating that knowledge editing has the potential to detoxify LLMs with a limited impact on general performance efficiently.
  arXiv  Detail & Related papers  (2024-03-21T15:18:30Z)
• Fine-Grained Detoxification via Instance-Level Prefixes for Large Language Models [26.474136481185724]
  Fine-grained detoxification via instance-level prefixes (FGDILP) to mitigate toxic text without additional cost. FGDILP contrasts the contextualized representation in attention space using a positive prefix-prepended prompt. We validate that FGDILP enables controlled text generation with regard to toxicity at both the utterance and context levels.
  arXiv  Detail & Related papers  (2024-02-23T09:04:48Z)
• Unveiling the Implicit Toxicity in Large Language Models [77.90933074675543]
  The open-endedness of large language models (LLMs) combined with their impressive capabilities may lead to new safety issues when being exploited for malicious use. We show that LLMs can generate diverse implicit toxic outputs that are exceptionally difficult to detect via simply zero-shot prompting. We propose a reinforcement learning (RL) based attacking method to further induce the implicit toxicity in LLMs.
  arXiv  Detail & Related papers  (2023-11-29T06:42:36Z)

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.