Discern Truth from Falsehood: Reducing Over-Refusal via Contrastive Refinement

• URL: http://arxiv.org/abs/2603.03323v1
• Date: Tue, 10 Feb 2026 05:53:10 GMT
• Title: Discern Truth from Falsehood: Reducing Over-Refusal via Contrastive Refinement
• Abstract summary: Large language models (LLMs) aligned for safety often suffer from over-refusal.<n>This behavior undermines models' helpfulness and restricts usability in sensitive or nuanced contexts.<n>We argue that this issue arises from the ambiguous influence of toxic and seemingly toxic prompts on the model's learning dynamics.
• Score: 14.1763272695827
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large language models (LLMs) aligned for safety often suffer from over-refusal, the tendency to reject seemingly toxic or benign prompts by misclassifying them as toxic. This behavior undermines models' helpfulness and restricts usability in sensitive or nuanced contexts. While prior work has proposed mitigation strategies such as data augmentation and activation steering, these approaches often face a trade-off: reducing over-refusal typically degrades the model's ability to reject genuinely harmful content. We argue that this issue arises from the ambiguous influence of toxic and seemingly toxic prompts on the model's learning dynamics. To address it, we introduce a preceding alignment stage, DCR: Discernment via Contrastive Refinement. Both theoretically and empirically, we demonstrate that contrastive refinement improves an LLM's capacity to distinguish truly toxic prompts from superficially toxic ones. Evaluation across diverse benchmarks shows that our method effectively reduces over-refusal while preserving the safety benefits of alignment. Importantly, it achieves this with minimal degradation of general capabilities, offering a more principled and robust direction for safety alignment.

Related papers

• Measuring and Mitigating Toxicity in Large Language Models: A Comprehensive Replication Study [0.0]
  Large Language Models (LLMs) trained on web-scale corpora inherently absorb toxic patterns from their training data.<n>Decoding-time Experts (DExperts) is an inference-time mitigation technique that steers generation without requiring model retraining.<n>This study contributes to the growing body of work on AI safety by highlighting the gap between explicit and implicit toxicity mitigation.
  arXiv  Detail & Related papers  (2026-05-13T20:12:30Z)
• Detoxifying LLMs via Representation Erasure-Based Preference Optimization [44.29978832356216]
  Large language models (LLMs) trained on webscale data can produce toxic outputs.<n>Prior defenses, based on applications of DPO, NPO, and similar algorithms, reduce the likelihood of harmful continuations.<n>We propose Representation Erasure-based Preference Optimization (REPO), reformulating detoxification as a token-level preference problem.
  arXiv  Detail & Related papers  (2026-02-24T22:51:06Z)
• Do Prompts Guarantee Safety? Mitigating Toxicity from LLM Generations through Subspace Intervention [6.808534332444413]
  Large Language Models (LLMs) are powerful text generators.<n>LLMs can produce toxic or harmful content even when given seemingly harmless prompts.<n>This presents a serious safety challenge and can cause real-world harm.
  arXiv  Detail & Related papers  (2026-02-06T11:33:17Z)
• Projecting Out the Malice: A Global Subspace Approach to LLM Detoxification [73.77171973106567]
  Large language models (LLMs) exhibit exceptional performance but pose inherent risks of generating toxic content.<n>Traditional methods fail to eliminate underlying toxic regions in parameters, leaving models vulnerable to adversarial attacks.<n>We propose GLOSS, a lightweight method that mitigates toxicity by identifying and eliminating this global subspace from FFN parameters.
  arXiv  Detail & Related papers  (2026-01-09T09:34:53Z)
• The Eminence in Shadow: Exploiting Feature Boundary Ambiguity for Robust Backdoor Attacks [51.468144272905135]
  Deep neural networks (DNNs) underpin critical applications yet remain vulnerable to backdoor attacks.<n>We provide a theoretical analysis targeting backdoor attacks, focusing on how sparse decision boundaries enable disproportionate model manipulation.<n>We propose Eminence, an explainable and robust black-box backdoor framework with provable theoretical guarantees and inherent stealth properties.
  arXiv  Detail & Related papers  (2025-12-11T08:09:07Z)
• Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing [77.75609817898035]
  Large Language Models (LLMs) have demonstrated impressive performance across various tasks, yet they remain vulnerable to generating toxic content.<n>We propose textscAutoregressive textscReward textscGuided textscRe presentation textscEditing (ARGRE)<n>ARGRE explicitly models toxicity transitions within the latent representation space, enabling stable and precise reward-guided editing.
  arXiv  Detail & Related papers  (2025-09-24T03:40:32Z)
• Context Engineering for Trustworthiness: Rescorla Wagner Steering Under Mixed and Inappropriate Contexts [55.70338710797578]
  We introduce the Poisoned Context Testbed, pairing queries with real-world contexts containing relevant and inappropriate content.<n>Inspired by associative learning in animals, we adapt the Rescorla-Wagner (RW) model from neuroscience to quantify how competing contextual signals influence LLM outputs.<n>We introduce RW-Steering, a two-stage finetuning-based approach that enables the model to internally identify and ignore inappropriate signals.
  arXiv  Detail & Related papers  (2025-09-02T00:40:34Z)
• 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 Bad Tokens: Detoxification of LLMs Using Sparse Autoencoders [9.203871622144863]
  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.
  arXiv  Detail & Related papers  (2025-05-20T15:55:31Z)
• 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)
• Gaining Wisdom from Setbacks: Aligning Large Language Models via Mistake Analysis [127.85293480405082]
  The rapid development of large language models (LLMs) has not only provided numerous opportunities but also presented significant challenges. Existing alignment methods usually direct LLMs toward the favorable outcomes by utilizing human-annotated, flawless instruction-response pairs. This study proposes a novel alignment technique based on mistake analysis, which deliberately exposes LLMs to erroneous content to learn the reasons for mistakes and how to avoid them.
  arXiv  Detail & Related papers  (2023-10-16T14:59:10Z)

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.