MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

AI summary

Overview

Research area: Multimodal Large Language Models (MLLMs) — specifically the internal attention-head mechanisms that cause visual hallucinations during autoregressive generation.

Technical level: Advanced. The paper builds on causal inference machinery (causal noise intervention, counterfactual Difference-in-Differences, Partial Information Decomposition theory), defines two formal theorems, and reports per-head statistical thresholds, though the core intuition is stated plainly.

One-sentence scope: The paper proposes HEAL, a training-free inference-time method that identifies "synergy heads" inside MLLMs, shows that hallucination coincides with an information-distribution drift inside those heads, and corrects it by rescaling visual and language value vectors with an equilibrium factor.

What This Paper Is About

MLLMs frequently produce fluent answers that are not grounded in the image — a failure mode called hallucination. Existing attention-based fixes rely on indirect proxies such as raw attention weights, which the authors argue do not reflect the actual causal flow of information inside a head, and uni-modal attention boosting tends to trade visual grounding against language fluency. HEAL instead disentangles what each attention head actually carries in terms of visual, language, and synergistic information, then dynamically rebalances the visual-language mix inside the heads where the imbalance occurs.

Key Contributions

1. A head taxonomy built on causal intervention. HEAL applies distribution-matched Gaussian noise to individual head outputs and measures the resulting change in the layer representation, filtering out "causally redundant" heads before classifying the rest into visual, language, and synergy heads using counterfactual Difference-in-Differences.

2. A diagnosis of hallucination as synergy-head disequilibrium. The authors report that hallucination is tied to information distribution drifting away from a healthy equilibrium inside synergy heads, and is not strongly correlated with the number or strength of modality-specific heads — counter to the intuition that more visual heads would fix hallucination.

3. A dynamic calibration strategy with theoretical grounding. An equilibrium factor α ∈ (0,1) is used to compute per-modality calibration factors β and γ applied to visual and language value vectors, with Theorem 1 (value-space calibration equals information-distribution calibration) and Theorem 2 (alignment with visual information changes monotonically with α) supporting the design.

4. An efficient, plug-and-play implementation. Head types are refreshed periodically rather than per step, causal intervention is parallelized across all heads, and the four counterfactual Difference-in-Differences states are computed as a single batched attention operation.

Main Findings

• Four head types are identified: causally redundant, visual, language, and synergy heads. Redundant heads are those whose information contribution falls below μ − 3σ of the layer's head scores; visual and language heads are split by whether I_vis or I_lang is positive, with a Logit-transformed MAD threshold applied to the modality ratio; the remainder are synergy heads, further split into visual-preferred and language-preferred.

• Halibucination tracks an information drift, not a shortage of visual heads. Analysis on Qwen3-VL and LLaVA-NeXT shows that in synergy heads, correct tokens carry a visual-language ratio around 0.43:0.54, while hallucinated tokens shift toward language at roughly 0.28:0.62. The paper also notes from Figure 2 that the number of visual heads stays nearly constant throughout generation.

• Head roles are dynamic across generation. Attention heads undergo what the authors call a "task-driven phase transition": when producing language-centric tokens, synergy heads revert toward language heads; when generating visually grounded tokens, some language heads shift toward synergy heads and a small subset may move toward visual heads. The authors report the macroscopic head-type distribution shifts minimally within short windows of roughly 10 to 15 generation steps.

• HEAL improves POPE and MME over prior training-free methods on LLaVA-1.5-7B. It reaches POPE F1 87.7 ± 0.3 and accuracy 88.3 ± 0.2, and the highest MME total in Table 1 at 669.76 ± 1.72, with Exist. 195.45 ± 0.23, Count 158.53 ± 0.35, Pos. 145.12 ± 0.48, and Color 170.66 ± 0.66. On CHAIR it obtains S 36.7 ± 0.4 and I 10.7 ± 0.03 — lower (better) than the untuned baseline but not lower than EAH (36.4 and 9.9). The paper attributes HEAL's advantage over EAH to a higher CHAIR recall (79.1 ± 0.1 vs 74.9) and longer generation (99.8 ± 0.7 vs 97.7).

• The method transfers across model families. On LLaVA-Bench, POPE and CHAIR, adding HEAL improves the base model in every case reported in Table 2: LLaVA-1.5-7B (LLaVA-Bench 72.5 → 75.2), LLaVA-NeXT-7B (81.6 → 82.2), Qwen2.5-VL-7B (76.8 → 78.5), Qwen2-VL-7B (75.6 → 78.0), and InternVL-7B (51.6 → 53.4).

• The equilibrium factor has a U-shaped effect. On CHAIR with LLaVA-1.5, hallucination first decreases then increases as α grows, indicating that over-aggressive calibration suppresses language information and degrades generation. LLaVA-Bench scores across α = 0.3 to 0.7 peak at 0.6 for LLaVA-1.5 (75.2) and at 0.5 for Qwen2.5-VL (78.5), and the paper notes the optimal α differs by model and task.

• Causal manipulation works in both directions. When the model originally produces a correct answer, artificially decreasing the visual information proportion can induce hallucination; when it originally hallucinates, increasing the visual information proportion can alleviate it. The authors present this as evidence that the relationship is causal rather than merely correlational.

• The taxonomy is robust to design choices. Replacing Gaussian masking with one-point (zero) masking, uniform masking, or swapping actual tokens yields head-assignment agreement of 92.13%, 93.21%, and 95.36% respectively, against 100.00% for the Gaussian baseline, and POPE F1 stays between 86.93 and 87.84. Varying the σ_total coefficient (1, 2, 3, ∞) gives F1 from 87.06 to 87.84, and varying the MAD coefficient (1.4826, 2.9652, 4.4478, ∞) gives F1 from 86.23 to 87.84, with the drift between correct and hallucinated tokens observable under every setting.

• Update interval can be sparse. Varying the head-type update interval from 2 to 20 shows that smaller intervals give marginally better results at higher computational cost, while 10–15 steps keep performance comparable with materially better decoding efficiency.

Methodology in Plain English

The approach has three stages.

First, filter out heads that don't matter. Rather than trusting the output projection weights, the authors overwrite one attention head's output at a time with Gaussian noise matched to that head's own mean and covariance, then measure how much the layer's representation changes. Heads whose change falls below μ − 3σ are labeled causally redundant and set aside — a head can be internally rich but have almost no effect on the final output.

Second, figure out what each remaining head carries. The authors construct four versions of the head's computation: visual and language tokens both intact, only language intact, only visual intact, and neither intact (both replaced by noise matched to first- and second-order statistics). Comparing these with a cosine-similarity-based metric gives a "first-order difference" (total information) and "second-order differences" for pure language and pure visual information; synergy is what remains after subtracting the two from the total. Heads are then sorted into visual, language, or synergy, with a Logit transform plus median absolute deviation used to set thresholds because modality ratios are skewed.

Third, rebalance the drift. When hallucination-prone steps are detected, the model sets a target equilibrium α and computes two factors, β = α / α_vis and γ = (1 − α) / α_lang, which are multiplied into the visual and language value vectors of the affected heads. Because this happens after the KV-cache update and before the attention kernel, it does not disturb FlashAttention or PagedAttention internals. Head types are refreshed only every 10–15 steps, and both the noise intervention and the four counterfactual states are computed in parallel batches so the accelerator processes factual and counterfactual attention matrices together.

Evaluation spans LLaVA-1.5-7B, LLaVA-NeXT-7B, Qwen2-VL-7B, Qwen2.5-VL-7B, Qwen3-VL-8B, InternVL-7B and InternVL3.5-8B, using POPE, CHAIR and MMHal-Bench for hallucination, and LLaVA-Bench, MME and BLINK-Twice for general capability. Default α is 0.5 for POPE/CHAIR and 0.6 for harder benchmarks on LLaVA and InternVL models, and 0.4/0.5 respectively for Qwen models.

Why This Matters

Impact on research. The paper moves the hallucination literature away from "add more visual attention" toward a causal account of where in the network the imbalance lives. By separating causally redundant heads from informative ones and by formalizing synergy as a residual quantity, it offers a reusable diagnostic lens for interpretability work on multimodal transformers. The bidirectional causal analysis — inducing hallucination by suppressing visual information and relieving it by increasing it — gives a stronger evidentiary footing than the correlational attention-weight analyses that dominate prior work.

Real-world applications:

• Medical imaging, explicitly named by the authors as a "high-precision field" where hallucination is a bottleneck for reliable deployment.
• Document parsing and visual grounding systems, areas the related-work section identifies as current MLLM strengths that unreliable visual grounding would undermine.
• Visual question answering assistants, the setting used in the paper's own analysis figures, where answers must be tied to image content.
• Any deployment where a fluent but ungrounded answer is costly — the method is training-free and runs at inference time, so it can be applied to existing checkpoints without retraining.

Industry relevance. HEAL requires no fine-tuning, works as a plug-and-play module across multiple model families, and is designed to avoid interfering with standard attention kernels, which lowers the barrier to adoption. Because it periodically rather than continuously recomputes head types, the overhead is bounded, although the paper defers the concrete throughput, latency, and GPU-memory comparisons to its appendix rather than reporting them in the main text.

Future Directions

• Adaptive calibration policies. The authors state in the conclusion that both the calibration factors and the update interval are currently determined empirically and may vary across models and tasks, and call for more adaptive policies.
• Tighter integration with training. The conclusion suggests coupling inference-time calibration with the training procedure rather than treating it purely as a post-hoc intervention.
• Hallucinations originating before attention heads. The authors acknowledge that errors caused by early visual encoding failures or missing visual evidence cannot always be resolved by head calibration alone, and that these require architectural or training-level remedies.
• Extending the theory's scope. Appendix B is said to discuss the theoretical scope of HEAL, and Appendix C.5 is said to provide guidelines and empirical patterns for setting hyperparameters — both point to open questions about how far the equilibrium-factor framework generalizes.

Target Audience

Researchers and engineers working on multimodal large language models who need either (a) an interpretability tool for locating where visual grounding breaks down inside a transformer, or (b) a training-free inference-time method for reducing hallucination without touching model weights. It is also relevant to practitioners in high-stakes deployment domains such as medical imaging who need to audit model trustworthiness, and to mechanistic-interpretability researchers interested in causal head-level analysis. Readers should be comfortable with attention mechanics and counterfactual reasoning; the statistical thresholds and theorems are presented without extensive derivation in the main text.

Authors’ abstract

Multimodal Large Language Models (MLLMs) often struggle with hallucinations, thus hindering their reliable practical applications. Existing attention-based mitigation methods mainly rely on indirect signals (e.g., attention weights) that fail to accurately reflect the actual information shift underlying hallucination generation. In this paper, we propose HEAL, Head-lEvel information disentAnglement and caLibration for identifying and mitigating hallucinations. HEAL first employs causal noise intervention on multi-head outputs to filter out causally redundant heads. Subsequently, it disentangles information distribution within the remaining heads via the counterfactual Difference-in-Differences, categorizing heads into four types. Through analysis, we observe: hallucinations happen when information distribution drifts away from a healthy equilibrium in synergy heads, not strongly correlated with the quantity or strength of modality-specific heads. Motivated by this insight, HEAL injects dynamic information calibration factors into the value vectors of synergy heads, and actively regulates visual-language dependencies, steering the output distribution towards factual evidence. Extensive experiments demonstrate that HEAL effectively reduces hallucinations across multiple MLLMs, offering a simple and interpretable pathway to enhance model trustworthiness.