SafetyExplainabilityTrustworthinesscs.CL
Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation
Zuhair Hasan Shaik, Abdullah Mazhar, Aseem Srivastava, Md Shad Akhtar
September 20, 2025
Neural Information Processing Systems
3 citations
1 influential
Key Takeaways
-
Layer-level patterns are more reliable than individual neurons for detecting toxicity: The research found that looking at aggregated features across entire layers of language models provides more stable and consistent signals for identifying toxic content than focusing on single neurons, addressing a key instability problem in previous approaches.
-
Previous toxicity fixes may have been solving the wrong problem: The paper identifies a conceptual flaw in existing neuron-based methods—they conflate "toxicity detection" (identifying toxic content) with "toxicity generation" (actually producing it)—suggesting that prior interventions may have been targeting the wrong mechanisms entirely.
-
A new method (EigenShift) can reduce toxicity without damaging language abilities: The proposed technique uses mathematical decomposition of the model's output layer to selectively suppress toxic generation while preserving the model's core language capabilities, addressing a major trade-off problem where previous toxicity fixes often made models less fluent or useful.
-
Effective toxicity mitigation doesn't require retraining models: EigenShift works without additional training or fine-tuning and has minimal computational cost, making it a practical approach that could be applied to existing deployed language models to improve their safety.
-
Understanding how models generate toxicity matters for fixing it: The research emphasizes that interpretability—actually understanding the internal mechanisms driving toxic outputs—is essential for developing more effective and reliable toxicity mitigation strategies, rather than just applying surface-level fixes.
Abstract
Large Language Models have demonstrated impressive fluency across diverse tasks, yet their tendency to produce toxic content remains a critical challenge for AI safety and public trust. Existing toxicity mitigation approaches primarily manipulate individual neuron activations, but these methods suffer from instability, context dependence, and often compromise the model's core language abilities. To address these shortcomings, we investigate three key questions: the stability of neuron-level toxicity indicators, the advantages of structural (layer-wise) representations, and the interpretability of mechanisms driving toxic generation. Through extensive experiments on Jigsaw and ToxiCN datasets, we show that aggregated layer-wise features provide more robust signals than single neurons. Moreover, we observe conceptual limitations in prior works that conflate toxicity detection experts and generation experts within neuron-based interventions. To mitigate this, we propose a novel principled intervention technique, EigenShift, based on eigen-decomposition of the language model's final output layer. This method selectively targets generation-aligned components, enabling precise toxicity suppression without impairing linguistic competence. Our method requires no additional training or fine-tuning, incurs minimal computational cost, and is grounded in rigorous theoretical analysis.
Accepted to the NeurIPS 2025 Research Track
Paper Info
arXiv ID2509.16660
PublishedSeptember 20, 2025
Citations
3
Influential
1
Relevance88%
VenueNeural Information Processing Systems
TypeJournalArticle
Categorycs.CL