acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large vision–language models (LVLMs) have achieved strong performance in visual reasoning and understanding, but can still produce hallucinations, generating content that is not adequately supported by the visual input. Many existing methods focus on the language side by intervening in model outputs or decoding processes, while the role of visual representations remains less explored. We therefore investigate hallucinations in the model's internal visual representation space. We first use sparse autoencoders (SAEs) to transform dense and entangled visual representations into a sparse, interpretable feature space. We identify two types of SAE features: always-on neurons that recur across images and image-specific neurons that respond to the content of the current image. Intervention experiments reveal that hallucinations are associated with the spurious activation of neurons unrelated to the input image. Further experiments show that enhancing image-relevant neurons while suppressing irrelevant ones can mitigate hallucinations. Based on these findings, we propose Contrastive Neuron Steering (CNS), which uses contrastive feature steering to correct spurious visual activations. CNS further applies Always-on Neuron Compensation (ANC) to offset changes in activation mass and maintain stable token-level representations. Experiments on hallucination and general-purpose benchmarks show that CNS reduces hallucinations and improves response faithfulness to the input image without compromising general multimodal performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large vision-language models, mechanistic interpretability, sparse autoencoders, activation steering, hallucination mitigation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_D7XtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies hallucinations in LVLMs by decomposing visual encoder representations into sparse features using sparse autoencoders. It identifies always-on and image-specific features and proposes Contrastive Neuron Steering, which uses differences between clean and perturbed image representations to adjust image-specific features while compensating activation changes through always-on features.

Critical strengths
  1. The paper provides an interpretable analysis of hallucination through sparse visual features, and the qualitative interventions make the connection between individual visual concepts and model outputs relatively easy to inspect.

  2. The distinction between always-on and image-specific neurons provides a simple way to characterize different activation patterns in the visual representation, and the paper connects this analysis to the design of CNS.

  3. CNS is a training-free intervention with relatively low inference overhead, and the experiments cover multiple LVLMs and both hallucination-specific and general multimodal benchmarks.

Critical weaknesses
  1. The motivation and novelty are not sufficiently positioned against recent work. The paper motivates its contribution by arguing that existing hallucination studies mainly focus on decoding or language-side representations, while the role of internal visual representations remains underexplored. However, several recent works have already investigated hallucination directly from the visual or internal-representation perspective. For example, HIT, VHR, VISTA, SHARP, and ILVAD similarly analyze internal visual signals, attention, or hidden representations and intervene on them to reduce hallucination [1-5]. Therefore, studying hallucination from the visual/internal-representation perspective itself is not a new contribution. The authors should discuss these findings in relation to their hypothesis and method better. The overlap is even stronger with recent SAE-based works. Prior studies have already shown that SAE features in vision-language models are interpretable and steerable [6,7], while Steering LVLMs via Sparse Autoencoder for Hallucination Mitigation and SAVE directly use SAE feature steering to mitigate hallucination [8,9]. The novelty of this paper therefore appears to lie more specifically in the clean-perturbed, image-conditioned SAE correction and the always-on neuron compensation mechanism, rather than in SAE-based hallucination analysis or visual feature intervention in general. But these are not new things either. You may find similar findings in SAE works quite often.

  2. The causal evidence for the proposed hallucination mechanism is not fully convincing. Table 1 shows that zeroing image-specific neurons substantially hurts POPE performance, but this mainly demonstrates that these neurons encode useful image information; it does not establish that their abnormal activation specifically causes hallucination.

  3. CNS is also conceptually close to visual contrastive decoding. Both methods use a clean image and a corrupted view and push the model away from the corrupted visual signal; CNS mainly differs by applying the contrast in SAE feature space, restricting it to image-specific features, and adding ANC. The paper should more clearly isolate the contribution of sparse feature decomposition from the underlying clean-versus-corrupted contrastive principle.

[1] Che L, Liu TQ, Jia J, Qin W, Tang R, Pavlovic V. Hallucinatory image tokens: a training-free EAZY approach to detecting and mitigating object hallucinations in LVLMs. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. IEEE/CVF; 2025:21635-21644.

[2]He J, Zhu K, Guo H, et al. Cracking the code of hallucination in LVLMs with vision-aware head divergence. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics; 2025:3488-3501.

[3]Li Z, Shi H, Gao Y, et al. The hidden life of tokens: reducing hallucination of large vision-language models via visual information steering. In: Proceedings of the 42nd International Conference on Machine Learning. Vol 267. Proceedings of Machine Learning Research; 2025:35799-35819.

[4] Wu J, Ding Y, Liu G, et al. SHARP: steering hallucination in LVLMs via representation engineering. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics; 2025:14346-14361.

[5] Xie Y, Hua Z, Wang R, Ng WWY, Wang X, Jia Y. Finding the correct visual evidence without forgetting: mitigating hallucination in LVLMs via inter-layer visual attention discrepancy. In: Proceedings of the 43rd International Conference on Machine Learning. Vol 306. Proceedings of Machine Learning Research; 2026:139471-139490.

[6] Lou H, Li C, Ji J, Yang Y. SAE-V: interpreting multimodal models for enhanced alignment. In: Proceedings of the 42nd International Conference on Machine Learning. Vol 267. Proceedings of Machine Learning Research; 2025:40344-40360.

[7] Pach M, Karthik S, Bouniot Q, Belongie S, Akata Z. Sparse autoencoders learn monosemantic features in vision-language models. Adv Neural Inf Process Syst. 2025;38.

[8] Hua Z, He J, Yao Z, et al. Steering LVLMs via sparse autoencoder for hallucination mitigation. In: Findings of the Association for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics; 2025:10808-10828.

[9] Park S, Yoo S, Mok J, Yoon S. SAVE: sparse autoencoder-driven visual information enhancement for mitigating object hallucination. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. IEEE/CVF; 2026:7935-7944.

Questions for authors

See above.

Reviewer_QKcTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates LVLM hallucinations from the perspective of sparse and interpretable visual features. The authors first use SAEs to decompose visual representations and distinguish between always-on and image-specific neurons based on their activation patterns. Through a series of analyses and interventions, they show how spurious activations of image-irrelevant features can contribute to hallucinations.

Based on these observations, the authors propose Contrastive Neuron Steering (CNS), which compares feature activations between clean and perturbed images to correct image-specific neurons, while using always-on neurons for activation compensation. Extensive experiments on different LVLMs and benchmarks demonstrate the effectiveness of the proposed method in reducing hallucinations while preserving general multimodal capabilities.

Critical strengths
  1. The paper investigates hallucinations through SAE features offers an interesting perspective on how hallucinations arise within visual representations. The neuron-level analysis and intervention experiments provide some useful insights into the underlying mechanisms.

  2. The proposed method is intuitive and well motivated by the preceding analysis. I also find the experimental evaluation fairly comprehensive, with results across different models and benchmarks, along with reasonable ablation studies.

Critical weaknesses
  1. While the paper provides extensive analysis to motivate CNS, I feel that some post-intervention analysis is still missing. For example, it would be interesting to examine how the activations of relevant and irrelevant features actually change after applying CNS, and whether these changes align with the intended correction mechanism.

  2. The analysis and method rely quite heavily on the learned SAE features. It is not entirely clear how sensitive the findings and the effectiveness of CNS are to different SAE configurations. Some discussion or evidence on this aspect would make the conclusions more convincing.

  3. The paper lacks comparisons with strong related baselines, such as VTI [1] and ICT [2].

  4. It is somewhat strange that different models are evaluated on different benchmarks across Tables 2–7; moreover, the paper does not specify which models are used in Tables 4 and 5.

[1] Reducing Hallucinations in Vision-Language Models via Latent Space Steering. ICLR 2025.

[2] ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models. CVPR 2025.

Questions for authors
  1. How sensitive are the analysis and CNS performance to different SAE configurations, such as training data, dictionary size, and sparsity level? Would the identified neuron patterns remain consistent?

  2. Have the authors examined the neuron activation patterns after applying CNS, especially for cases where hallucinations are successfully corrected?

  3. For ANC, preserving the total activation mass does not necessarily guarantee that the reconstructed visual representations remain stable. Could the authors provide some evidence on this? Also, does the compensation in Eq. (13) ever lead to negative feature activations in practice?

  4. Do the authors plan to release the code and trained SAEs? I believe these resources would be quite valuable to the community, especially for further research on visual interpretability and hallucination analysis.

Reviewer_8W5ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CNS uses sparse autoencoder features and clean-versus-noisy image contrasts to steer visual representations before multimodal decoding. The paper combines feature-level interpretation with evaluations of hallucination mitigation and inference efficiency.

Critical strengths
  1. Interpretability. Provides a visual-feature perspective on hallucination through sparse representations and targeted interventions.
  2. Evaluation breadth. Covers multiple model families and hallucination benchmarks, with additional general-capability checks.
  3. Practicality. Reports modest inference overhead and includes ablations of steering and activation compensation.
Critical weaknesses
  1. Unclear hyperparameter selection. Were λ and t selected using an independent validation set? The optima in Tables 9 and 10 (λ = 1.0, t = 500) reproduce the main POPE/CHAIR scores in Tables 2 and 3, but the validation split and selection procedure are unspecified. This raises concern about possible test-set selection, although matching scores alone do not establish it. Please clarify data overlap and when settings were fixed. If evaluation scores guided selection, independent reevaluation is needed.
  2. The steering direction requires stronger justification. Since the model itself can hallucinate, its visual activations may reflect both valid evidence and spurious associations. Noise can suppress irrelevant or misleading features as well as useful ones. Amplifying features weakened by noise could therefore reinforce existing errors, making the meaning of the activation differences ambiguous. Please provide a more convincing explanation, supported by targeted experiments or analyses, of why these differences identify a beneficial correction direction. In particular, distinguish task-relevant features from spurious activations and assess which ones CNS amplifies on inputs where the original model hallucinates.
  3. Uneven figure readability (minor). Figures 1-4 occupy substantial space, while Figure 5 has small labels and a small legend. More consistent scaling and font sizes would improve readability.
Questions for authors
  1. How were λ\lambda and tt selected? Were the sweeps in Tables 9–10 run on a separate validation set or on the test data used in Tables 2–3?

  2. Noise may weaken both useful visual features and spurious associations. What evidence shows that amplifying these features helps correct hallucinations rather than reinforce the model’s existing errors?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.