Steering LVLMs via Sparse Autoencoder for Hallucination Mitigation

• URL: http://arxiv.org/abs/2505.16146v2
• Date: Mon, 15 Sep 2025 07:02:17 GMT
• Title: Steering LVLMs via Sparse Autoencoder for Hallucination Mitigation
• Authors: Zhenglin Hua, Jinghan He, Zijun Yao, Tianxu Han, Haiyun Guo, Yuheng Jia, Junfeng Fang,
• Abstract summary: Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks.<n>They still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications.<n>We propose Steering LVLMs via SAE Latent Directions (SSL), a plug-and-play method based on SAE-derived latent directions to mitigate hallucinations in LVLMs.
• Score: 38.43656456659151
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks. However, they still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications. Existing approaches to address this issue focus on incorporating external knowledge bases, alignment training, or decoding strategies, all of which require substantial computational cost and time. Recent works try to explore more efficient alternatives by adjusting LVLMs' internal representations. Although promising, these methods may cause hallucinations to be insufficiently suppressed or lead to excessive interventions that negatively affect normal semantics. In this work, we leverage sparse autoencoders (SAEs) to identify semantic directions closely associated with faithfulness or hallucination, extracting more precise and disentangled hallucination-related representations. Our analysis demonstrates that interventions along the identified faithful direction can mitigate hallucinations, while those along the hallucinatory direction can exacerbate them. Building on these insights, we propose Steering LVLMs via SAE Latent Directions (SSL), a plug-and-play method based on SAE-derived latent directions to mitigate hallucinations in LVLMs. Extensive experiments demonstrate that SSL significantly outperforms existing decoding approaches in mitigating hallucinations, while maintaining transferability across different model architectures with negligible additional time overhead. The code is available at https://github.com/huazhenglin2003/SSL.

Related papers

• Modality Bias in LVLMs: Analyzing and Mitigating Object Hallucination via Attention Lens [0.0]
  Large vision-language models (LVLMs) have demonstrated remarkable multimodal comprehension and reasoning capabilities.<n>LVLMs tend to over-rely on textual prompts and internal knowledge of large language models, generating descriptions inconsistent with visual cues.<n>We propose a training-free method to mitigate object hallucination.
  arXiv  Detail & Related papers  (2025-08-04T13:40:59Z)
• Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering [83.63437999696954]
  hallucination in large language models (MLLMs) persists as a significant and under-addressed challenge in the video domain.<n>We propose a temporal-aware activation engineering framework for VideoLLMs, which adaptively identifies and manipulates hallucination-sensitive modules.
  arXiv  Detail & Related papers  (2025-05-19T08:12:06Z)
• Mitigating Hallucinations via Inter-Layer Consistency Aggregation in Large Vision-Language Models [3.9464481148889354]
  We propose a novel decoding mechanism, Decoding with Inter-layer Consistency via Layer Aggregation (DCLA)<n>Our approach constructs a dynamic semantic reference by aggregating representations from previous layers, and corrects semantically deviated layers to enforce inter-layer consistency.<n> Experiments on hallucination benchmarks such as MME and POPE demonstrate that DCLA effectively reduces hallucinations while enhancing the reliability and performance of LVLMs.
  arXiv  Detail & Related papers  (2025-05-18T10:15:42Z)
• Efficient Contrastive Decoding with Probabilistic Hallucination Detection - Mitigating Hallucinations in Large Vision Language Models - [1.2499537119440245]
  Efficient Contrastive Decoding (ECD) is a simple method that leverages probabilistic hallucination detection to shift the output distribution towards contextually accurate answers at inference time.<n>Our experiments show that ECD effectively mitigates hallucinations, outperforming state-of-the-art methods with respect to performance on LVLM benchmarks and computation time.
  arXiv  Detail & Related papers  (2025-04-16T14:50:25Z)
• Mitigating Hallucination for Large Vision Language Model by Inter-Modality Correlation Calibration Decoding [66.06337890279839]
  Large vision-language models (LVLMs) have shown remarkable capabilities in visual-language understanding for downstream multi-modal tasks.<n>LVLMs still suffer from generating hallucinations in complex generation tasks, leading to inconsistencies between visual inputs and generated content.<n>We propose an Inter-Modality Correlation Decoding (IMCCD) method to mitigate hallucinations in LVLMs in a training-free manner.
  arXiv  Detail & Related papers  (2025-01-03T17:56:28Z)
• Cracking the Code of Hallucination in LVLMs with Vision-aware Head Divergence [69.86946427928511]
  We investigate the internal mechanisms driving hallucination in large vision-language models (LVLMs)<n>We introduce Vision-aware Head Divergence (VHD), a metric that quantifies the sensitivity of attention head outputs to visual context.<n>We propose Vision-aware Head Reinforcement (VHR), a training-free approach to mitigate hallucination by enhancing the role of vision-aware attention heads.
  arXiv  Detail & Related papers  (2024-12-18T15:29:30Z)
• Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning [151.4060202671114]
  multimodal large language models (MLLMs) have shown unprecedented capabilities in advancing vision-language tasks.<n>This paper introduces a novel bottom-up reasoning framework to address hallucinations in MLLMs.<n>Our framework systematically addresses potential issues in both visual and textual inputs by verifying and integrating perception-level information with cognition-level commonsense knowledge.
  arXiv  Detail & Related papers  (2024-12-15T09:10:46Z)
• VaLiD: Mitigating the Hallucination of Large Vision Language Models by Visual Layer Fusion Contrastive Decoding [38.23310445372371]
  Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal task reasoning.<n>They often generate responses that appear plausible yet do not accurately reflect the visual content, a phenomenon known as hallucination.<n>Recent approaches have introduced training-free methods to mitigate hallucinations by adjusting the decoding strategy during the inference stage.<n>We propose a novel hallucination-mitigation method from the visual encoding perspective: textbfVisutextbfal textbfLayer Fustextbfion Contrastive textbfD
  arXiv  Detail & Related papers  (2024-11-24T13:42:02Z)
• CATCH: Complementary Adaptive Token-level Contrastive Decoding to Mitigate Hallucinations in LVLMs [74.36850397755572]
  CATCH addresses issues related to visual defects that cause diminished fine-grained feature perception and cumulative hallucinations in open-ended scenarios. It is applicable to various visual question-answering tasks without requiring any specific data or prior knowledge, and generalizes robustly to new tasks without additional training.
  arXiv  Detail & Related papers  (2024-11-19T18:27:31Z)
• Mitigating Object Hallucination via Concentric Causal Attention [71.27325347912823]
  We show that object hallucination is closely tied with Rotary Position. RoPE, a widely adopted positional dependency modeling design. We propose Concentric Causal Attention (CCA), a simple yet effective positional alignment strategy.
  arXiv  Detail & Related papers  (2024-10-21T11:54:53Z)
• Reducing Hallucinations in Vision-Language Models via Latent Space Steering [34.1755878632361]
  Hallucination poses a challenge to the deployment of large vision-language models (LVLMs) in applications. We introduce Visual and Textual Intervention (VTI), a novel technique designed to reduce hallucinations by steering latent space representations during inference to enhance the stability of vision features.
  arXiv  Detail & Related papers  (2024-10-21T08:42:30Z)
• Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models [30.26685485474035]
  Large Vision-Language Models (LVLMs) have rapidly advanced in recent years.<n>The prevalent issue known as the hallucination' problem has emerged as a significant bottleneck.<n>We propose a simple yet effective method named Self-Introspective Decoding (SID)
  arXiv  Detail & Related papers  (2024-08-04T13:50:17Z)
• Visual Description Grounding Reduces Hallucinations and Boosts Reasoning in LVLMs [52.497823009176074]
  Large Vision-Language Models (LVLMs) often produce responses that misalign with factual information, a phenomenon known as hallucinations.<n>We introduce Visual Description Grounded Decoding (VDGD), a training-free method designed to enhance visual perception and improve reasoning capabilities in LVLMs.
  arXiv  Detail & Related papers  (2024-05-24T16:21:59Z)
• Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization [123.54980913741828]
  Large Visual Language Models (LVLMs) have demonstrated exceptional abilities in understanding multimodal data.<n>They invariably suffer from hallucinations, leading to a disconnect between the generated text and the corresponding images.<n>Almost all current visual contrastive decoding methods attempt to mitigate these hallucinations by introducing visual uncertainty information.<n>However, they struggle to precisely induce the hallucinatory tokens, which severely limits their effectiveness in mitigating hallucinations.
  arXiv  Detail & Related papers  (2024-05-24T08:46:31Z)
• Evaluating Object Hallucination in Large Vision-Language Models [122.40337582958453]
  This work presents the first systematic study on object hallucination of large vision-language models (LVLMs) We find that LVLMs tend to generate objects that are inconsistent with the target images in the descriptions. We propose a polling-based query method called POPE to evaluate the object hallucination.
  arXiv  Detail & Related papers  (2023-05-17T16:34:01Z)

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.