Capturing Gaze Shifts for Guidance: Cross-Modal Fusion Enhancement for VLM Hallucination Mitigation

AI summary

Overview

Research area: Multimodal vision-language models (VLMs), specifically inference-time hallucination mitigation through attention steering.

Technical level: Advanced. The paper assumes familiarity with transformer attention matrices, autoregressive decoding, attention heads, prefilling versus decoding stages, and VLM architecture conventions.

Scope: The paper proposes and evaluates GIFT, a lightweight inference-time method that builds a visual saliency map from "gaze shifts" in attention during query processing and uses it to jointly amplify visual and query attention during decoding across four VLMs and three hallucination benchmarks plus two general vision-language benchmarks.

What This Paper Is About

Vision language models frequently generate hallucination, meaning content that cannot be substantiated by either textual or visual inputs. Prior work attributes this to models over-relying on linguistic prior knowledge instead of visual inputs, but existing inference-time fixes either amplify visual attention without keeping visual and query signals balanced, or fail to account for the visual attention sink problem where attention is misallocated to task-irrelevant image regions. This paper's goal is to construct a noise-resistant, task-relevant visual saliency map before decoding and use it to steer attention toward both salient visual tokens and the user query in a balanced way.

Key Contributions

1. A gaze-shift saliency mechanism. The authors introduce a method that pre-computes a holistic, task-relevant visual saliency map before decoding by tracking positive changes ("gaze shifts") in visual attention as the model processes information-rich words in the user query. Because irrelevant regions show minimal or no attention change, this mechanism side-steps the visual attention sink problem that affects simple averaging ("static gaze") approaches.

2. The GIFT method. Gaze Shift-Guided Cross-modal Fusion Enhancement is a lightweight inference-time hallucination mitigation method that uses the pre-computed map to enhance attention to salient visual tokens in selected cross-modal fusion layers, while proportionally scaling attention to query tokens based on the overall visual amplification ratio (via a coefficient β) to preserve cross-modal fusion balance.

3. Demonstrated gains across architectures and sizes. GIFT consistently mitigates hallucination across LLaVA-1.5 7B, LLaVA-1.5 13B, Qwen2-VL 7B, and Qwen3-VL 8B, with up to 20.7% improvement on CHAIR, 15.9% on MMHal-Bench, and 3.0% on POPE over greedy decoding, while preserving general vision-language performance at low computational overhead. Code is released at https://github.com/amazon-science/GIFT.

4. Ablations on components, hyperparameters, and token selection. The paper validates the joint contribution of visual enhancement and fusion balance (up to 25.4% improvement over single-component variants), the effect of the enhancement coefficient α, robustness to enhancement layer range, and a comparison of POS-tagging versus LLM-based extraction of information-rich query words.

Main Findings

• Gaze-shift maps localize better than static maps. On 1,000 examples from the MSCOCO 2014 training set (images with a single unambiguous object instance), the shift-based method scored 11.92 on a normalized saliency score measuring concentration inside bounding boxes normalized by box area, versus 5.40 for the static method — more than double.

• Peak gaze-shift layer is intrinsic to each model. Measured on 50 examples from TextVQA, the layer with the largest sum of positive visual attention shifts is early-to-middle: 11 for LLaVA-1.5 7B, 10 for LLaVA-1.5 13B, 14 for Qwen2-VL 7B, and 18 for Qwen3-VL 8B. Appendix A reports that across five random seeds, the peak layer was identical in all five runs for both LLaVA-1.5 7B (layer 11) and Qwen3-VL 8B (layer 18), and matched on MathVista as well as TextVQA.

• Hallucination benchmarks improve. On LLaVA-1.5 7B, GIFT reduced CHAIRs from 50.2 (greedy) to 39.8 and CHAIRi from 15.4 to 10.6, raised POPE F1 from 82.4 to 83.8 and POPE accuracy from 79.5 to 81.9, lowered MMHal-Bench hallucination rate from 65.2 to 57.3, and raised MMHal-Bench score from 2.22 to 2.48. On LLaVA-1.5 13B, GIFT reduced CHAIRs from 46.8 to 39.6 and MMHal hallucination from 56.2 to 55.8, and raised MMHal score from 2.61 to 2.72. On Qwen2-VL 7B, CHAIRs fell from 24.8 to 21.2 and MMHal hallucination from 32.7 to 27.5. On Qwen3-VL 8B, CHAIRs fell from 51.4 to 49.4 and MMHal hallucination from 28.3 to 26.4. GIFT also improved MMHal-Bench informativeness score by 11.7%.

• General vision-language performance is maintained. On MME and SEED-Bench, GIFT scores stay comparable to greedy decoding across all four models (for example, LLaVA-1.5 7B: MME 1750.5 versus greedy 1751.6, SEED 65.6 versus 65.5; Qwen3-VL 8B: MME 2418.5 versus 2403.5, SEED 78.3 versus 78.4), whereas baselines show mixed results with some degradation.

• Low latency overhead. For generating 32 tokens with LLaVA-1.5 7B on MMHal-Bench, GIFT runs at 1.13× greedy decoding latency, versus 1.01× for VAF, 1.56× for Rel-Attn, 1.99× for VCD, and 11.10× for VAR.

• Both components of GIFT are necessary. Ablations on POPE and MMHal-Bench across all four models show the full method beats a variant that increases visual attention without fusion balance (Inc. V.) and a variant that only recalibrates visual attention distribution while keeping overall visual contribution unchanged (Cal. V.), by up to 25.4%.

• Enhancement strength trades perception against reasoning. Sweeping α from 1.0 to 7.0 on LLaVA-1.5 7B, hallucination mitigation on POPE improves steadily while MME stays close to greedy for moderate values; at excessively large α, POPE plateaus and MME drops below greedy decoding.

• Layer range choice is not fragile. With the default range 12-22 for LLaVA-1.5 7B, every tested configuration outperforms the baseline on POPE, but ranges wider than 10-24 degrade SEED-Bench, and multiple moderate configurations preserve reasoning performance.

• LLM-based word extraction is stronger but costs more. Replacing POS tagging with Qwen3-VL 8B prompting on MMHal-Bench improved LLaVA-1.5 7B from hallucination 57.3 to 52.6 and score 2.48 to 2.60, and Qwen2-VL 7B from 27.5 to 26.0 and 3.58 to 3.67. POS tagging is kept as the default for negligible overhead.

• A noted metric concern. Qwen3-VL 8B achieves the best POPE and MMHal-Bench results but higher CHAIR scores than other models; the authors attribute this to its detailed, reasoning-based descriptions being penalized by CHAIR's string-matching against predefined objects.

Methodology in Plain English

A VLM takes a system instruction, an image, and a user query. The image becomes a sequence of visual tokens; the instruction and query become text tokens; all are concatenated and fed into a language model that generates an answer token by token. Inside the model, attention matrices record how much each token looks at every other token at every layer and head.

The authors start from an observation about human vision: when people read a question, their gaze moves around the image, settling on the objects the question is about. They hypothesize VLMs do something similar, and that these movements are informative. A naive saliency map — simply averaging how much every query word attends to each visual token ("static gaze") — highlights relevant areas but also lights up irrelevant ones, the visual attention sink problem.

Instead, the method looks for words in the query that carry meaning (nouns, proper nouns, verbs, adjectives, adverbs, numbers, found with spaCy part-of-speech tagging) and tracks how visual attention changes from one such word to the next. Negative changes are clamped to zero, so only increases in attention count. Averaging these positive shifts over the top 50% of attention heads (ranked by cumulative shift to visual tokens) produces a saliency map. The map is computed at a single layer — the one where the total shift is largest, which the paper argues is a fixed property of each model — clipped at three standard deviations, min-max normalized to [0,1], and cached before decoding begins. Only pre-filling up to that layer is needed.

During decoding, in selected cross-modal fusion layers where the model attends strongly to both visual and query tokens, salient visual token attention is multiplied by exp(α × saliency), which is equivalent to adding a bias term α·saliency inside the softmax. Separately, query token attention is multiplied by β·r, where r is the overall relative increase in visual attention at that layer. This second step is what distinguishes GIFT from prior single-sided visual amplification: it keeps the fusion between modalities balanced. Finally, the enhanced attention rows are re-normalized to sum to one. Default settings are α = 5.0 for LLaVA models and 4.0 for Qwen models, β = 1.0 for all models, and enhancement layer ranges of 12-22, 14-20, 5-18, and 13-23 for LLaVA-1.5 7B, LLaVA-1.5 13B, Qwen2-VL 7B, and Qwen3-VL 8B respectively.

Why This Matters

Impact on research. The paper reframes hallucination mitigation as a two-sided problem: it is not enough to boost visual grounding, because over-emphasizing vision without matching query attention can produce well-grounded but misinterpreted answers. It also contributes a concrete way to counter the visual attention sink problem using relative attention change rather than absolute attention magnitude, and reports that the layer where gaze shifts peak is stable across samples and across two very different datasets (TextVQA and MathVista), which is a claim other work can build on. Because GIFT is a training-free inference-time intervention, it applies to existing models without retraining.

Real-world applications (the paper names biomedicine, autonomous driving, and robotics as high-stakes domains where factual accuracy is critical; the evaluated tasks are object detection, image captioning, and visual question answering):

• Medical or biomedical image question answering, where fabricated findings about an image can be harmful.
• Autonomous driving scene understanding, where misdescribed visual content affects safety.
• Robotics, where an agent acting on an incorrect description of its environment can fail or cause damage.
• General-purpose image captioning and visual question answering assistants, where unsupported object mentions undermine user trust.

Industry relevance. The method requires no retraining, works across LLaVA and Qwen families at 7B, 8B, and 13B scale, and adds roughly 1.13× latency relative to greedy decoding — substantially below VCD (1.99×) and VAR (11.10×), which matters for production serving budgets. The method also requires only pre-filling up to one layer to compute the map. The authors are affiliated with AWS AI Labs and Oracle AI, and code is released publicly.

Future Directions

• Freeing the method from query dependence. The authors state as a primary limitation that GIFT relies on the query to identify relevant visual regions; when the answer does not depend on visual information, or large parts of the query are unrelated to the image, the saliency map can be inaccurate and cause incorrect focus. They suggest identifying vision-relevant tokens via LLM prompting or a lightweight classifier whose token scores weight their contribution to saliency, and bypassing saliency computation entirely when no token relates to the image.
• Replacing or improving information-rich word selection. The ablation shows LLM-based token selection outperforms POS tagging on MMHal-Bench; a cheaper approximation that recovers that accuracy is an open direction.
• Reducing latency further. The authors list potential latency optimization approaches as future work in Appendix G (the appendix content beyond Appendix A is not included in the provided text).
• Refining hallucination metrics for modern models. The paper argues CHAIR's string-matching against predefined objects may penalize the nuanced, reasoning-based descriptions produced by newer models such as Qwen3-VL 8B, suggesting the metric needs refinement.

Target Audience

Researchers and engineers working on multimodal large language models who need hallucination mitigation at inference time without retraining, and who are comfortable reading transformer attention internals and decoding-time interventions. It is also useful for practitioners deploying VLMs in accuracy-critical settings who care about the performance-versus-latency tradeoff, and for benchmark designers interested in the limitations of CHAIR-style string-matching evaluation for modern reasoning-capable models. Readers looking for a beginner-level introduction to VLMs would find the attention-level formulation of this paper difficult without background in transformer attention and decoding.

Authors’ abstract

Vision language models (VLMs) often generate hallucination, i.e., content that cannot be substantiated by either textual or visual inputs. Prior work primarily attributes this to over-reliance on linguistic prior knowledge rather than visual inputs. Some methods attempt to mitigate hallucination by amplifying visual token attention proportionally to their attention scores. However, these methods overlook the visual attention sink problem, where attention is frequently misallocated to task-irrelevant visual regions, and neglect cross-modal fusion balance by enhancing only visual attention without adjusting attention to the user query. This can result in amplifying incorrect areas while failing to properly interpret the user query. To address these challenges, we propose a simple yet effective method called Gaze Shift-Guided Cross-modal Fusion Enhancement (GIFT). GIFT pre-computes a holistic visual saliency map by tracking positive changes in visual attention, or "gaze shifts", during user query comprehension, and leverages this map to amplify attention to both salient visual information and the user query at each decoding step. This reduces the impact of visual attention sink, as irrelevant tokens exhibit minimal shifts, while ensuring balanced cross-modal fusion for well-integrated representation. Extensive experiments show that GIFT effectively mitigates hallucination in VLMs across both generative and classification tasks, achieving up to 20.7% improvement over greedy decoding, while maintaining general vision-language performance with low computational overhead.