💻 computer science

MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

The paper proposes HEAL, a novel framework that identifies and mitigates hallucinations in Multimodal Large Language Models by analyzing information distribution drifts in synergy heads and applying dynamic calibration to steer outputs toward factual evidence.

Original authors: Meng'en Qin, Junye Chen, Jucheng Liu, Youlu Xing, Song Wang, Ruize Han

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Meng'en Qin, Junye Chen, Jucheng Liu, Youlu Xing, Song Wang, Ruize Han

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Multimodal large language models are a powerful new class of artificial intelligence designed to see and speak at the same time. Unlike traditional systems that process text or images in isolation, these models can look at a photograph and describe it, answer questions about it, or tell a story based on what they see. They work by weaving together two distinct streams of information: the visual data from the image and the linguistic patterns learned from vast amounts of text. For these systems to be truly useful in high-stakes fields like medical imaging or autonomous driving, they must be reliable. However, they frequently suffer from a problem known as hallucination. This occurs when the model generates a response that sounds confident and plausible but is factually wrong about the visual world, such as claiming a picture of a cat contains a dog, or inventing details that simply are not there in the image.

For a long time, researchers believed that these errors happened because the model was paying too much attention to its internal knowledge of language and not enough to the actual picture. The prevailing idea was that as the model wrote its answer, it would drift away from the image and rely too heavily on what it expected to see based on words alone. To fix this, previous attempts focused on forcing the model to look harder at the image or by retraining the entire system from scratch. These methods often treated the model as a black box, adjusting its behavior from the outside without understanding the internal mechanics of why the error occurred in the first place.

A team of researchers at Shenzhen University of Advanced Technology has now looked inside the engine of these models to find a different explanation. They propose that hallucinations are not caused by a simple lack of attention to the image, nor by the model ignoring the picture entirely. Instead, they found that the error arises when the internal balance between visual and language information becomes unstable within specific parts of the model's processing network. The researchers developed a method called HEAL, which stands for Head-lEvel information disentAnglement and caLibration, to identify and correct this imbalance in real time.

The core of their discovery lies in how the model processes information. Inside these large models, there are many small processing units called attention heads. Think of these heads as a team of specialists working together to understand a sentence. Some specialists focus purely on the words, others focus purely on the image, and a third group works to blend the two together. The researchers found that the group responsible for blending the two streams—what they call synergy heads—is where the trouble starts. When the model is working correctly, these synergy heads maintain a healthy equilibrium, holding visual and language information in a stable balance. However, when the model begins to hallucinate, this balance shifts. The information distribution within these blending heads drifts away from that healthy equilibrium, causing the model to lose its grounding in the visual evidence.

Crucially, the researchers ruled out the idea that hallucinations happen because there are simply too few specialists looking at the image or because the image specialists are too weak. Their analysis showed that the number of visual-focused heads remains constant whether the model is telling the truth or lying. The problem is not a shortage of visual attention, but a drift in how the blending specialists handle the mix of information. When the model generates a correct description, the synergy heads keep the visual and language inputs in a steady ratio. When it hallucinates, that ratio skews, often leaning too heavily on language patterns while the visual connection weakens, even though the image is still present.

To solve this, the team created a dynamic calibration strategy that acts like a real-time regulator. Instead of retraining the model or changing its architecture, HEAL intervenes during the generation process. It monitors the information flowing through the synergy heads and detects when the balance begins to drift. When a drift is detected, the system injects a calibration factor that gently nudges the visual and language information back toward the correct equilibrium. This process is continuous and adaptive; it does not force the model to ignore language or over-focus on the image, but rather steers the internal representation back toward a state where the visual evidence is properly weighted against the linguistic context.

The researchers tested this approach on several state-of-the-art models, including versions of LLaVA and Qwen. The results were consistent across different systems and tasks. By applying this dynamic calibration, the models produced significantly fewer hallucinations while maintaining their ability to speak fluently and creatively. In tests designed to measure how often models invent objects that are not in the picture, the new method reduced errors more effectively than previous techniques that tried to simply boost visual attention. The study suggests that the key to trustworthiness in these systems is not just making them look harder at the image, but ensuring that the internal blend of sight and speech remains stable.

This work offers a new perspective on how artificial intelligence understands the world. It shows that reliability is not just about having enough data or powerful hardware, but about maintaining a delicate internal balance between different types of information. By identifying the specific point where this balance breaks and providing a simple way to restore it, the researchers have opened a path toward more trustworthy multimodal systems. The method is lightweight and does not require the massive computational resources needed for retraining, making it a practical tool for improving the accuracy of AI that sees and speaks. The findings suggest that future improvements in AI reliability may come from fine-tuning these internal relationships rather than building larger, more complex models from the ground up.

Technical Summary: HEAL – Head-lEvel information disentAnglement and caLibration for MLLM Hallucinations

1. Problem Statement

Multimodal Large Language Models (MLLMs) frequently suffer from hallucinations, generating responses that are fluent but factually inconsistent with the visual context. This issue hinders their deployment in high-precision domains like medical imaging. Existing mitigation strategies face two primary limitations:

Macro-level approaches (e.g., retraining, fine-tuning, reinforcement learning) treat models as black boxes, lacking interpretability regarding internal mechanisms.
Micro-level approaches (e.g., decoding optimization, attention-based enhancement) often rely on indirect signals like attention weights or FFN activations. These fail to capture the actual causal information distribution and confounding synergy effects. Furthermore, uni-modal attention enhancement often creates a "see-saw" dilemma, where strengthening visual attention truncates creative responses, while over-reliance on language exacerbates visual hallucinations.

The paper posits that current methods fail to accurately reflect the actual information shift underlying hallucination generation, specifically within the interaction of visual and language modalities.

2. Methodology: HEAL

The authors propose HEAL (Head-lEvel information disentAnglement and caLibration), a training-free, inference-time framework designed to identify and mitigate hallucinations by analyzing and regulating the information distribution within attention heads. The methodology consists of three core stages:

A. Causal Noise Intervention (Filtering Redundant Heads)

To identify insignificant heads, HEAL employs causal noise intervention. Instead of relying on projection weights, the method replaces the output of a specific attention head $i$ with distribution-matched Gaussian noise and measures the resulting change in the layer representation using vector similarity.

Heads with negligible causal contribution (where the information contribution score $I(t,l,i)$ falls below a statistical threshold of $\mu - 3\sigma$ ) are classified as causally redundant and excluded from further analysis.

B. Counterfactual Difference-in-Differences (Head Categorization)

For the remaining non-redundant heads, HEAL utilizes Partial Information Decomposition (PID) theory to disentangle information distribution. By constructing four counterfactual inputs (masking visual tokens, language tokens, both, or neither) using Gaussian noise, the method calculates:

Total Information ( $I_{total}$ ): The difference between full and fully masked inputs.
Language Information ( $I_{lang}$ ): Derived from the second-order difference when visual tokens are masked.
Visual Information ( $I_{vis}$ ): Derived from the second-order difference when language tokens are masked.
Synergy Effect ( $I_{syn}$ ): The residual interaction term ( $I_{total} - I_{vis} - I_{lang}$ ).

Based on these metrics, heads are categorized into four types:

Redundant Heads: Negligible total information.
Visual Heads: High visual information, low/negative language information.
Language Heads: High language information, low/negative visual information.
Synergy Heads: Positive contributions from both modalities. These are further analyzed for their visual-language ratio.

C. Dynamic Information Calibration

The core insight of HEAL is that hallucinations occur when the information distribution in synergy heads drifts away from a healthy equilibrium, rather than being caused by a simple shortage of visual or language heads.

Equilibrium Factor ( $\alpha$ ): A hyperparameter (analogous to temperature) is introduced to define the target visual-language balance.
Calibration Mechanism: During inference, HEAL dynamically computes calibration factors $\beta$ and $\gamma$ for the value vectors of visual and language tokens within synergy heads. These factors reorient the information distribution toward the target $\alpha$ :
$\frac{\beta I_{vis}}{\beta I_{vis} + \gamma I_{lang}} = \alpha$
Implementation: This calibration is applied to the value vectors ( $V_{vis}, V_{lang}$ ) after the KV-cache update but before the attention kernel, ensuring compatibility with FlashAttention and PagedAttention without disrupting internal implementations.

3. Key Findings and Observations

Through extensive analysis on models like LLaVA-NeXT and Qwen3-VL, the authors observed:

Synergy Head Drift: Hallucinations are causally linked to a drift in information distribution within synergy heads, not the quantity or strength of modality-specific (visual/language) heads.
Dynamic Phase Transitions: Attention head roles are highly dynamic. During generation, heads undergo task-driven phase transitions (e.g., synergy heads reverting to language heads for language-centric tokens).
Causal Relationship: Bidirectional causal analysis confirmed that artificially decreasing visual information in correct responses induces hallucinations, while increasing visual information in hallucinated responses alleviates them.

4. Experimental Results

HEAL was evaluated across multiple MLLMs (LLaVA series, Qwen series, InternVL series) and benchmarks (POPE, CHAIR, MME, LLaVA-Bench, MMHal-Bench, BLINK-Twice).

Hallucination Reduction: HEAL consistently outperformed state-of-the-art training-free methods (e.g., EAH, VCD, OPERA, TAME). On the POPE benchmark, HEAL achieved an F1 score of 87.7 (vs. 85.4 for the baseline) and significantly reduced CHAIR scores (hallucination rate).
General Capabilities: Unlike some methods that degrade general reasoning, HEAL improved performance on comprehensive benchmarks like MME and LLaVA-Bench, demonstrating that reducing hallucinations does not require sacrificing linguistic coherence or reasoning ability.
Efficiency: While HEAL introduces computational overhead due to counterfactual analysis, it remains more efficient than other attention-head calibration methods (e.g., 1.58× faster than EAH) by utilizing periodic updates (every 10 steps) and parallelized tensor operations.

5. Significance and Claims

The paper claims the following contributions and significance:

New Interpretability: HEAL provides a fine-grained, head-level interpretable lens into MLLM internals, moving beyond indirect attention weights to actual information distribution.
Causal Insight: It identifies visual-language information disequilibrium within synergy heads as a causally intervenable factor, challenging the notion that hallucinations are solely due to a lack of visual attention.
Theoretical Grounding: The method is supported by Theorems proving that calibrating value spaces is equivalent to calibrating information distribution and that the equilibrium factor induces a monotonic directional effect on output alignment.
Practical Utility: HEAL offers a simple, plug-and-play pathway to enhance model trustworthiness without retraining, effectively steering outputs toward factual evidence while preserving linguistic fluency.

The authors acknowledge limitations, noting that HEAL cannot resolve hallucinations stemming from early visual encoding failures or a fundamental lack of visual evidence in the input, as it operates on existing internal representations. Future work is suggested to explore adaptive calibration policies and tighter integration with model training.