Towards Interpretable Hallucination Analysis and Mitigation in LVLMs via Contrastive Neuron Steering

TL;DR

This work addresses hallucinations in large vision–language models by shifting focus from output-level fixes to the internal visual representation space. It uses sparse autoencoders to decompose dense visual embeddings into interpretable neurons, identifying always-on and image-specific types, and shows that hallucinations largely stem from disruptions of image-specific neurons. The authors introduce Contrastive Neuron Steering (CNS), which identifies image-specific neurons via contrast between clean and noisy inputs and amplifies them while suppressing noninformative activations, with Always-on Neuron Suppression (ANS) to down-weight persistent signals; CNS operates at the prefilling stage and remains compatible with decoding-stage methods. Empirical results across hallucination-focused and general multimodal benchmarks demonstrate that CNS reduces hallucinations while preserving multimodal understanding, and ablations highlight the importance of ANS and tuning of noise steps and steering strength. Overall, this work provides an interpretable, representation-level mechanism for mitigating hallucinations in LVLMs with practical impacts for reliability and safety in multimodal AI systems, supported by detailed neuron-level analyses and case studies. The key ideas are formalized around $z(v)$, $z(v')$, $\Delta z$, and $\tilde{z}$, where $\Delta z = z(v) - z(v')$ and $\tilde{z} = z(v) + \lambda \Delta z$ (with ANS enforcing $\Delta z_i=0$ for Always-on neurons).

Abstract

LVLMs achieve remarkable multimodal understanding and generation but remain susceptible to hallucinations. Existing mitigation methods predominantly focus on output-level adjustments, leaving the internal mechanisms that give rise to these hallucinations largely unexplored. To gain a deeper understanding, we adopt a representation-level perspective by introducing sparse autoencoders (SAEs) to decompose dense visual embeddings into sparse, interpretable neurons. Through neuron-level analysis, we identify distinct neuron types, including always-on neurons and image-specific neurons. Our findings reveal that hallucinations often result from disruptions or spurious activations of image-specific neurons, while always-on neurons remain largely stable. Moreover, selectively enhancing or suppressing image-specific neurons enables controllable intervention in LVLM outputs, improving visual grounding and reducing hallucinations. Building on these insights, we propose Contrastive Neuron Steering (CNS), which identifies image-specific neurons via contrastive analysis between clean and noisy inputs. CNS selectively amplifies informative neurons while suppressing perturbation-induced activations, producing more robust and semantically grounded visual representations. This not only enhances visual understanding but also effectively mitigates hallucinations. By operating at the prefilling stage, CNS is fully compatible with existing decoding-stage methods. Extensive experiments on both hallucination-focused and general multimodal benchmarks demonstrate that CNS consistently reduces hallucinations while preserving overall multimodal understanding.

Figures (21)

• Figure 1: Neuron visualizations from SAE, showing diverse visual patterns and semantic structures.
• Figure 2: Steering an LVLM: (a) amplifying a "bow tie" neuron emphasizes this concept in generated descriptions, while (b) suppressing a "shirt" neuron prevents it from appearing.
• Figure 3: Relationship between noise step, model performance, and activation changes across different neuron types.
• Figure 4: Image-level neuron analysis. Top: statistical analysis of neuron activations. Bottom: neuron visualization. Red boxes indicate always-on neurons, in the Top-20 for all images.
• Figure 5: Patch-level neuron analysis. Top: statistical analysis of neuron activations. Bottom: neuron visualization.