PaperlayerView exact revision on arXiv

Source-linked AI summary

MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

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

arXiv:2609.09206v1cs.CVcs.CL

TL;DR

MLLM hallucinations undermine reliable visual grounding, while existing attention-based methods often rely on indirect signals that do not capture the underlying information shift. HEAL uses causal intervention and counterfactual head-level disentanglement to identify synergy-head disequilibrium and dynamically calibrate visual-language information, reducing hallucinations across multiple MLLMs.

  • Problem

    Existing attention-based hallucination mitigation methods rely on indirect signals that may not accurately reflect the causal information shifts underlying hallucination generation.

  • Method

    HEAL applies causal noise intervention and counterfactual Difference-in-Differences to classify attention heads and dynamically calibrate visual-language information in synergy heads.

  • Results

    Hallucinations occur when information distribution drifts from equilibrium in synergy heads, and HEAL consistently reduces hallucinations across multiple MLLM benchmarks.

  • Takeaways & Limitations

    Head-level information calibration offers a simple, interpretable pathway toward more trustworthy multimodal generation.

  • Takeaways & Limitations

    Calibration factors and update intervals are determined empirically, and attention-head calibration cannot always resolve failures from early visual encoding or missing visual evidence.

Abstract

from arXiv · show

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.

1 Introduction

MLLM hallucinations limit reliable visual grounding, while existing attention-based methods use indirect signals that may miss causal information shifts. HEAL identifies four head types and finds hallucinations arise from disequilibrium within synergy heads, which it corrects through dynamic calibration.

  • 1 Introduction: MLLM hallucinations are factually inconsistent with visual context and remain a bottleneck for reliable applications such as medical imaging.Existing macro- and micro-level mitigation strategies differ in intervention scope, while attention-based methods often rely on indirect signals.
  • 1 Introduction: HEAL combines causal noise intervention with counterfactual Difference-in-Differences to classify attention heads as redundant, visual, language, or synergy.The head-level taxonomy provides an interpretable view of multimodal information structure during autoregressive generation.
  • 1 Introduction: Hallucinations arise when information distribution drifts from equilibrium within synergy heads, rather than from the quantity or strength of modality-specific heads.This finding motivates targeting internal synergy-head information distributions instead of simply increasing visual or language head capacity.
  • 1 Introduction: Attention-head roles change during generation: language heads can become synergy heads for visually grounded tokens, while some synergy heads move toward visual heads.This task-driven phase transition indicates that head functions are adaptive rather than fixed across token generation.
  • 1 Introduction: Dynamic calibration introduces an equilibrium factor α that regulates visual-language information in synergy heads and steers generation toward factual evidence.The strategy operates during inference and aims to preserve linguistic coherence while correcting information imbalance.

2 Related Work

Related work mitigates MLLM hallucinations through data, training, model, inference, and attention interventions. Attention-based approaches commonly strengthen visual grounding or alter decoding, motivating HEAL’s more direct head-level information analysis.

  • 2 Related Work: Hallucination mitigation spans data-centric supervision, training and model improvements, preference learning, stronger architectures, and inference-time decoding interventions.These approaches target spurious correlations, visual grounding, alignment, or output generation from different levels of the MLLM pipeline.
  • 2 Related Work: Attention-based methods generally increase visual attention during decoding because language priors may dominate as generation proceeds.Representative methods differ in how they select, reinforce, or contrast visual attention and modality-specific pathways.
  • 2 Related Work: Prior methods also intervene on attention heads, use contrastive decoding, apply structural causal modeling, or optimize causal masks to reduce visual attention interference.These methods make decoding more image-centric or model causal relationships between attention, modality priors, and outputs.

3 Method

HEAL identifies causally meaningful attention heads, disentangles their visual-language information using counterfactual Difference-in-Differences, and dynamically calibrates synergy-head values toward a target equilibrium. Its design targets hallucination-linked information drift while preserving the distinction between redundant, modality-specific, and synergy heads.

  • 3.1 Causal Noise Intervention on Multi-head Outputs: HEAL intervenes on multi-head outputs with distribution-matched Gaussian noise to identify causally redundant heads before finer-grained information decomposition.The intervention measures changes in layer representations rather than relying only on projection weights, whose magnitudes may not reflect actual head contributions.
  • 3.2 Counterfactual Difference-in-Differences in Attention Head: Counterfactual Difference-in-Differences estimates total, visual, language, and synergy information from full and independently masked visual-language inputs.The four counterfactual settings use masked visual and language tokens, enabling modality-wise comparisons through head-output similarities.
  • 3.2 Counterfactual Difference-in-Differences in Attention Head: HEAL classifies non-redundant heads as visual, language, or synergy heads using modality information scores, logit-transformed ratios, and robust MAD thresholds.Synergy heads can additionally be labeled visual-preferred or language-preferred according to their modality ratios.
  • 3.3 Dynamic Information Calibration: Hallucination-linked drift is addressed by introducing an equilibrium factor α that regulates visual-language dependence within synergy heads during generation.The factor controls visual reliance through α and language reliance through 1 −α, while calibration is applied to modality-specific value vectors after KV-cache updates.
  • 3.3 Dynamic Information Calibration: Value-space calibration is theoretically equivalent to adjusting modality-wise information distribution, with visual alignment increasing monotonically as α increases.The calibration factors β and γ modify visual and language value vectors without changing FlashAttention or PagedAttention internals.
  • 3.3 Dynamic Information Calibration: HEAL accounts for temporal locality by recognizing that head types and internal information distributions change across autoregressive generation steps.The method therefore updates head information periodically rather than treating head roles as permanently fixed.

4 Experiments

HEAL is evaluated across hallucination and comprehensive multimodal benchmarks, with ablations testing calibration and update choices. Results show consistent gains, causal effects of visual-language reallocation, and robustness across masking and threshold settings.

  • Evaluation on Hallucination and Comprehensive Benchmarks: HEAL consistently improves hallucination-related and comprehensive metrics across multiple MLLM architectures.The method is evaluated on POPE, CHAIR, LLaVA-Bench, MME, and other benchmarks, with plug-and-play gains reported across architectures.
  • Ablation Analysis: Update intervals of 10–15 preserve comparable performance while reducing the computational overhead of frequent counterfactual analysis.Shorter intervals are marginally better but require more computation because head attribution is recomputed more often.
  • Ablation Analysis: An intermediate equilibrium factor is necessary: hallucination first decreases as α increases, then worsens when calibration becomes excessive.On CHAIR with LLaVA-1.5, the metric follows a U-shaped trend; overly strong calibration can suppress language information and degrade generation quality.
  • Ablation Analysis: Qualitative examples show that appropriate α values reduce hallucinations, while excessive calibration can produce ungrammatical or degraded outputs.Some errors remain unresolved when relevant visual information is lost during visual encoding or early fusion.
  • Causal Intervention Analysis: Increasing visual information can alleviate hallucination, whereas decreasing it can induce hallucination in originally correct responses.Bidirectional intervention results support a causal relationship between visual-language information distribution and hallucination behavior.
  • Robustness Analysis: The head taxonomy and central distribution-drift finding remain robust across masking strategies and threshold choices.Masking variants achieve 92.13%–95.36% head-assignment agreement, while threshold changes alter proportions but preserve the shift toward language representations in synergy heads.

5 Conclusion

The paper concludes that hallucinations arise from equilibrium drift in synergy heads and can be mitigated through dynamic inference-time calibration. The calibration theorem connects value-vector scaling to steering the head’s visual-language information distribution.

  • 5 Conclusion: HEAL identifies equilibrium drift in synergy heads and uses dynamic information calibration to mitigate hallucinations at inference time.The conclusion also notes that calibration factors and update intervals are selected empirically and may vary across models and tasks.
  • A.1 Theorem 1: Calibrating visual and language value vectors steers the head’s modality information distribution toward a target equilibrium factor.The result holds under the stated local linear approximation when information is measured by the counterfactual contribution score.

A.2 Theorem 2

Theorem 2 establishes that calibrating synergy-head value vectors predictably changes representation geometry, increasing visual alignment as the equilibrium factor α grows.

  • A.2 Theorem 2: Increasing α makes the final hidden state more aligned with the visual subspace after RMSNorm.The calibration increases visual information, suppresses language information, and RMSNorm preserves the resulting directional effect.
  • A.2 Theorem 2: HEAL’s theoretical analysis assumes local visual-language decomposition, additive and positively homogeneous information measures, and local linearity.
  • A.2 Theorem 2: Theorem 1 interprets value-vector calibration as adjusting visual-language information distribution within synergy heads.
  • A.2 Theorem 2: Theorem 2 shows that calibration changes representation geometry predictably rather than functioning as an empirical heuristic.
  • A.2 Theorem 2: HEAL provides a practical information-decomposition framework related to Partial Information Decomposition for multimodal representations.

B.3 Discussions between PID Theory and HEAL

HEAL is related to PID but uses counterfactual representation discrepancies rather than probability-based mutual information. Its synergy score therefore reflects net interaction after implicit redundancy overlap, not strict PID synergy alone.

  • B.3 Discussions between PID Theory and HEAL: PID decomposes joint visual-language information into redundant, visual-specific, language-specific, and synergistic components.
  • B.3 Discussions between PID Theory and HEAL: Classical mutual-information equations leave the redundancy atom underdetermined, requiring an explicit redundancy function for a full PID decomposition.
  • B.3 Discussions between PID Theory and HEAL: HEAL estimates modality-attributable information from representation changes after selectively removing visual and language inputs, rather than estimating the joint probability distribution explicitly.
  • B.3 Discussions between PID Theory and HEAL: Under an information-faithfulness assumption, HEAL’s visual and language scores include redundancy, while its synergy score corresponds to strict synergy minus redundancy.
  • B.3 Discussions between PID Theory and HEAL: A negative HEAL synergy score indicates that shared information dominates net joint interaction, not that Shannon synergy is negative.
  • B.3 Discussions between PID Theory and HEAL: HEAL is a counterfactual, representation-level analogue of PID, not an exact replacement for probability-based PID estimation.

C.1 Bidirectional Causal Analysis Results

The bidirectional intervention evaluates how changing the visual-language equilibrium factor affects LLaVA-1.5-7B and Qwen2.5-VL-7B on LLaVA-Bench.

  • C.1 Bidirectional Causal Analysis Results: Table 7 compares bidirectional visual-language-ratio interventions across LLaVA-1.5-7B and Qwen2.5-VL-7B on LLaVA-Bench.
  • C.1 Bidirectional Causal Analysis Results: Figure 6 visualizes the same bidirectional causal intervention by manipulating the equilibrium factor, with base denoting no intervention.

C.2 More Robustness Results and Details

The robustness analyses indicate that HEAL’s head assignments and mitigation performance remain stable across measurement metrics, replacement distributions, and intervention thresholds, while inference remains faster than EAH.

  • C.2 More Robustness Results and Details: Highly similar head assignments across cosine, normalized L2, and hybrid cosine+L2 metrics indicate metric robustness.
  • C.2 More Robustness Results and Details: Highly consistent classifications and similar POPE F1 scores across Gaussian, zero, uniform, Cauchy, and swapped-output replacements indicate replacement-distribution robustness.
  • C.2 More Robustness Results and Details: Head assignment agreement reaches 98.25% before and after Yeo-Johnson transformation, while F1 remains stable across 1σ, 2σ, and 3σ criteria.
  • C.2 More Robustness Results and Details: Head attributes show negligible variation over a 10-step update interval, supporting periodic head-type updates.
  • C.2 More Robustness Results and Details: HEAL’s inference evaluation samples 50 CHAIR image-caption pairs under batch size 1 using two NVIDIA RTX 4090 GPUs.
  • C.2 More Robustness Results and Details: 5.02 tokens/s and 70.74s end-to-end latency show HEAL is faster than EAH, with only 0.26 GB additional GPU memory versus the base model.HEAL incurs higher overhead than VCD and OPERA but reports stronger hallucination mitigation performance.

C.5 Empirical Analysis on the Equilibrium Factor and Update Interval

HEAL’s equilibrium factor and update interval have stable empirical operating ranges across models and tasks, while α controls the visual-language balance during inference.

  • Limitation: No closed-form rule currently determines the optimal hyperparameter values, although empirical analysis identifies stable operating ranges in practice.This limits theoretical guidance for selecting α and related settings.
  • Equilibrium factor: α performs stably within 0.4-0.6 across LLaVA and Qwen, although stronger visual models generally favor smaller values and weaker visual grounding favors larger values.The exact optimum varies slightly across datasets, but performance remains stable within this interval.
  • Update interval: Update intervals of 5-15 decoding steps achieve very similar performance, with 10 steps used by default and transferring well across tasks and datasets.The interval is less sensitive to downstream applications than the equilibrium factor.
  • Interpretation: The equilibrium factor provides an interpretable inference-time control over the balance between visual evidence and language information.Theorem 2 states that increasing α continuously shifts the output representation toward the visual direction.
  • Transferability: Using shared practical settings, HEAL consistently improves evaluation metrics on Intern-VL without separately tuning its hyperparameters.Intern-VL adopts the LLaVA-family setting because of similar visual capability.

C.6 Performance on Recent Benchmarks and MLLMs

HEAL improves hallucination and visual-reasoning performance on recent MLLMs and challenging benchmarks, including Qwen3-VL-8B and InternVL3.5-8B.

  • Overall results: HEAL consistently improves performance across both recent MLLMs and both challenging benchmarks evaluated.The evaluation covers MMHal-Bench and BLINK-Twice on Qwen3-VL-8B and InternVL3.5-8B.
  • Qwen3-VL-8B: On Qwen3-VL-8B, HEAL reduces hallucination rate from 17.5 to 16.6 and raises MMHal score from 4.82 to 4.91.These are the reported MMHal-Bench comparisons.
  • InternVL3.5-8B: On InternVL3.5-8B, HEAL reduces hallucination rate from 19.4 to 18.3 and raises MMHal score from 4.53 to 4.71.These are the reported MMHal-Bench comparisons.
  • BLINK-Twice: On BLINK-Twice, HEAL improves all reported metrics, including No-Acc, Yes-Acc, Q-Acc, I-Acc, and G-Acc.The passage characterizes BLINK-Twice as a challenging visual-reasoning benchmark.

D Further Discussions on HEAL

Further discussion clarifies HEAL’s scope, shared calibration design, update-delay trade-off, and two-stage head filtering, while identifying boundaries where attention calibration cannot recover missing visual evidence.

  • Scope boundary: HEAL cannot resolve hallucinations caused by failed early visual encoding or insufficient visual evidence in the input.The method calibrates information already available in attention and cannot recover absent reliable visual features.
  • Global calibration: A single equilibrium factor is intended to correct collective token-level imbalance across heterogeneous synergy heads, and empirical results support this shared calibration.The decoder depends on aggregated head outputs, while Figures 4(b) and Tables 13 and 15 show improvements across LLaVA, Qwen, and Intern-VL.
  • Periodic updates: Updating head types every 10 steps introduces tracking delay but balances adaptation accuracy against inference cost, with 5-15 steps showing similar performance.The authors report limited practical degradation, including on challenging MMHal-Bench evaluations, while noting that existing benchmarks do not isolate abrupt reasoning switches.
  • Two redundancy definitions: Causal redundancy filtering removes heads with negligible final-representation influence, whereas information redundancy filtering removes heads lacking meaningful modality-specific variation.The two stages target different notions of redundancy before visual-language decomposition.
  • Filtering ablation: Both filtering stages are necessary because omitting either disrupts head taxonomy, modality decomposition, or final performance.Table 9 reports effects from removing causal or information redundancy filtering.
  • MoE compatibility: HEAL operates at the MHA level, so α is a model-level target balance independent of downstream MoE expert routing.Calibration factors are computed from visual/language proportions estimated for each attention head.
  • Layer coverage: The method applies counterfactual analysis across all attention layers because its central finding concerns network-wide synergy-head disequilibrium rather than a few specific layers.Inference overhead is reduced through periodic updates and batched parallelized counterfactual analysis.
Loading 2609.09206v1…