Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Self-Correction Inside the Model: Leveraging Layer Attention to Mitigate Hallucinations in Large Vision Language Models

April Fu

TL;DR

An Internal self-Correction mechanism utilizing Layer Attention (ICLA) that operates directly on hidden states during generation that consistently improves visual grounding across multiple hallucination benchmarks, demonstrating its effectiveness for more advanced LVLMs.

Abstract

Although Large Vision-Language Models (LVLMs) have made substantial progress, hallucination, where generated text is not grounded in the visual input, remains a challenge. As LVLMs become stronger, previously reported hallucination patterns, such as linguistic bias and overthinking phenomenon, become far less consistent, making the corresponding mitigation techniques substantially less effective. In this paper, we introduce an Internal self-Correction mechanism utilizing Layer Attention (ICLA) that operates directly on hidden states during generation. Each layer selectively retrieves information from all preceding layers through a diagonal cross-layer attention mechanism, enabling self-refinement without any external correction signals. With introducing and training only 0.2M and 0.1M additional parameters on LLaVA1.5-7B and Qwen2.5-VL-7B, \ours consistently improves visual grounding across multiple hallucination benchmarks, demonstrating its effectiveness for more advanced LVLMs.

Self-Correction Inside the Model: Leveraging Layer Attention to Mitigate Hallucinations in Large Vision Language Models

TL;DR

An Internal self-Correction mechanism utilizing Layer Attention (ICLA) that operates directly on hidden states during generation that consistently improves visual grounding across multiple hallucination benchmarks, demonstrating its effectiveness for more advanced LVLMs.

Abstract

Although Large Vision-Language Models (LVLMs) have made substantial progress, hallucination, where generated text is not grounded in the visual input, remains a challenge. As LVLMs become stronger, previously reported hallucination patterns, such as linguistic bias and overthinking phenomenon, become far less consistent, making the corresponding mitigation techniques substantially less effective. In this paper, we introduce an Internal self-Correction mechanism utilizing Layer Attention (ICLA) that operates directly on hidden states during generation. Each layer selectively retrieves information from all preceding layers through a diagonal cross-layer attention mechanism, enabling self-refinement without any external correction signals. With introducing and training only 0.2M and 0.1M additional parameters on LLaVA1.5-7B and Qwen2.5-VL-7B, \ours consistently improves visual grounding across multiple hallucination benchmarks, demonstrating its effectiveness for more advanced LVLMs.
Paper Structure (34 sections, 8 equations, 6 figures, 10 tables, 1 algorithm)

This paper contains 34 sections, 8 equations, 6 figures, 10 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Causes of Hallucination in LVLMs
  4. Modality Imbalance.
  5. "Overthinking".
  6. Methods of Hallucination Mitigation in LVLMs
  7. Training-based Methods.
  8. Non-Training-based Methods.
  9. Layer Attention
  10. Method
  11. Preliminaries
  12. Internal Self-Correction via Layer Attention
  13. Cross-Layer Attention
  14. Experiments
  15. Experimental Setup
  16. ...and 19 more sections

Figures (6)

  • Figure 1: Performance degradation of existing methods on Qwen2.5-VL-7B, where DoLA exhibits a significant drop, while the performance drop of MME exceeds the plot range and is not shown in the figure.
  • Figure 2: Overall architecture of ICLA.
  • Figure 3: (a) Ablation results for starting layer $k_0$ on POPE benchmark. (b) Ablation results for reduction ratio $r$ on MMMU and MME benchmarks; (c) Ablation results on scaling factor $\alpha$ on MMMU and MME benchmarks.
  • Figure 4: Case study comparing Vanilla and ICLA based on Qwen2.5-VL-7B. The example is sampled from LLaVA-Bench.
  • Figure 5: Visualization of average attention weights in ICLA using Qwen for selected samples
  • ...and 1 more figures