PaperlayerView exact revision on arXiv

Source-linked AI summary

MergeHEIR: Mitigating Multimodal Hallucinations as the Tax of Model Merging

Jinyu Li, Hao Fang, Zhiming Zhang, Jiawei Kong, Bin Chen, Shu-Tao Xia

arXiv:2609.32422v1cs.AI

TL;DR

Model merging can increase hallucination, and unconstrained correction can interfere with inherited expertise. MergeHEIR combines hallucination-oriented adaptation with activation-derived null-space projection; across the evaluated comparisons it mitigates hallucination while largely preserving expertise. The paper also analyzes guarantees and the adaptation–retention trade-off.

  • Problem

    Merging can raise hallucination, while unconstrained post-merge correction can degrade inherited expert-task performance, creating a need to address both outcomes.

    Source: §1 · pp. 1–2

  • Method

    MergeHEIR combines hallucination-oriented preference optimization with activation-derived null-space projection to coordinate correction and retention of inherited expertise.

    Source: §1 · pp. 2–3

  • Results

    Across 24 paired comparisons, MergeHEIR consistently mitigates hallucination while largely preserving inherited expertise across three MLLM configurations and eight merging methods.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The results support coordinating hallucination mitigation with inherited-expertise preservation when adapting merged models.

    Source: §5 · p. 9

    The retention target is defined over representative expert-task distributions Dt, so the stated stability objective is scoped to those distributions.

    Source: §2.1 · p. 3

Abstract

from arXiv · show

Model merging consolidates task-specialized experts into a single deployable model. However, we show that such capability consolidation incurs a merging tax of increased hallucination: across 8 model-merging methods, every merged checkpoint exhibits a higher hallucination rate than the average of its constituent experts. An intuitive approach is to adapt existing hallucination-mitigation methods to the post-merge model, yet this unconstrained adaptation disrupts inherited capabilities, creating a tension between hallucination mitigation and expertise retention. To tackle this challenge, we introduce MergeHEIR, a post-merge adaptation framework designed to reduce this merging tax while preserving expertise inherited from initial experts. Using small expert-task calibration sets, MergeHEIR constructs layer-wise null-space projectors via SVD from task-specific activations collected from the merged checkpoint, and periodically projects the accumulated post-merge displacement onto the resulting null spaces to preserve inherited expertise. Theoretically, we establish minimum-distortion and maximum-dimensionality guarantees, characterize the threshold-controlled adaptation-retention trade-off, and extend perturbation guarantees beyond finite calibration data. Across 24 paired comparisons spanning three MLLM configurations and 8 model-merging methods, MergeHEIR consistently mitigates hallucination while largely preserving inherited expertise, demonstrating a more favorable hallucination-retention trade-off.

1 INTRODUCTION

Merging specialized MLLMs raises visual hallucination, while unconstrained post-merge correction can impair inherited expertise. MergeHEIR combines hallucination-oriented adaptation with activation-derived null-space projection and performs better on the hallucination–retention trade-off across the evaluated settings.

1 INTRODUCTION

  • AMBER HalRate means rise from 33.48 to 54.35 for Qwen2.5-VL-3B and from 20.83 to 25.41 for Qwen2.5-VL-7B after merging; every evaluated checkpoint exceeds its expert average.The experts share a frozen vision encoder and differ through language-side LoRA specialization; the authors attribute the tax to disrupted language-side use of visual evidence.Source: §1 · p. 1, Figure 1 · p. 2
  • Unconstrained adaptation achieves hallucination mitigation comparable to MergeHEIR but consistently reduces Expert Avg. more across three model configurations.The authors describe this as interference with expertise inherited from constituent models.Source: §1 · p. 2, Figure 2 · p. 2
  • The paper frames the challenge as reducing hallucination without undoing the capabilities that merging was intended to consolidate.Capability consolidation alone does not guarantee deployment reliability.Source: §1 · pp. 1–3
  • MergeHEIR coordinates hallucination correction with inherited-expertise retention by combining hallucination-oriented preference optimization and activation-derived null-space projection.It uses a small calibration set and periodically projects accumulated post-merge updates onto the corresponding null spaces.Source: §1 · pp. 2–3
  • The framework establishes minimum-distortion and maximum-dimensionality guarantees, perturbation bounds, and a threshold-controlled adaptation–retention trade-off.The guarantees are extended to the corresponding activation distribution.Source: §1 · p. 3
  • Across 24 paired comparisons, MergeHEIR consistently mitigates hallucination while largely preserving inherited expertise, with better retention than unconstrained adaptation at effective mitigation.The comparisons span three MLLM configurations and eight merging methods.Source: §1 · p. 3

2 PROBLEM FORMULATION

The formulation treats post-merge hallucination correction as an optimization problem constrained by retention of expert-task behavior. MergeHEIR operationalizes those constraints using activation-derived null-space projection.

2 PROBLEM FORMULATION

  • Merging creates a two-objective problem: improve visual grounding while retaining expert-task performance because merged checkpoints show higher hallucination than constituent-expert averages.The formulation names this hallucination correction under expert-retention constraints.Source: §2 · p. 3

2.1 MERGED MODEL AND RETENTION TARGETS

  • The expert-task set is T = {Chart, Geometry, OCR, VQA}, with θt denoting the model parameters specialized for task t.Source: §2.1 · p. 3
  • For each task t, Dt is a representative expert-task distribution, and the collection of these distributions captures heterogeneous capabilities intended to remain stable.Adaptation starts from θm, which remains fixed as the reference checkpoint.Source: §2.1 · p. 3

2.2 HALLUCINATION CORRECTION UNDER EXPERT-RETENTION CONSTRAINTS

  • Optimizing the hallucination objective alone does not restrict behavioral changes over expert-task distributions, motivating explicit retention constraints.Source: §2.2 · p. 3
  • The retention constraint measures adaptation-induced functional deviation from the merged checkpoint on Dt and limits it by a tolerated deviation ϵt.MergeHEIR instantiates these constraints through activation-derived null-space projection.Source: §2.2 · p. 3
  • Four task-specialized experts are consolidated into θm; preference optimization produces corrective updates, and calibration activations define layer-wise null-space projectors for periodic displacement projection.The framework is intended to reduce hallucination while retaining inherited expertise.Source: Figure 3 · p. 4

3 METHOD

MergeHEIR combines hallucination-oriented preference optimization with activation-derived null-space projection to correct hallucinations while limiting interference with inherited expertise. Its theory characterizes the geometry and perturbation control of projected updates.

3 METHOD

  • MergeHEIR combines hallucination-oriented preference optimization with activation-derived null-space projection to limit interference with inherited expertise.The method periodically projects accumulated updates onto null spaces constructed from expert-task activations.Source: §3 · p. 4, §3.2 · pp. 4–5

3.1 HALLUCINATION-ORIENTED PREFERENCE OPTIMIZATION

  • The preference objective favors grounded over hallucinated responses, strengthens reliance on image evidence, and prevents margin gains solely from suppressing rejected responses.Source: §3.1 · p. 4

3.2 EXPERT-TASK NULL-SPACE PROJECTION

  • MergeHEIR collects activations from small expert-task calibration sets, uses SVD of their Gram matrix to construct layer-wise null-space projectors, and projects updates using threshold α.The threshold selects singular vectors satisfying σj/σmax < α; the default is α = 5 × 10^-4.Source: §3.2 · p. 5

3.3 GEOMETRIC OPTIMALITY AND PERTURBATION CONTROL

  • Projection minimizes update distortion in the selected subspace, which is maximum-dimensional under the sensitivity constraint, while α balances update flexibility against calibration perturbation.The population extension controls expected perturbations on unseen activations from the same distribution, subject to spectral approximation error.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

Across three MLLM configurations and eight merging methods, MergeHEIR consistently reduces hallucination while largely preserving inherited expertise. Ablations show that null-space projection improves retention over unconstrained adaptation, while the threshold controls the mitigation–retention balance.

4.1 EXPERIMENTAL SETUP

  • The experiments merge four task-specialized experts across three MLLM configurations using eight methods, with 500 calibration examples per expert task for projector construction.Evaluation covers inherited expert-task performance and hallucination benchmarks; post-merge adaptation uses 10,000 multimodal preference pairs.Source: §4.1 · p. 7

4.2 MAIN RESULTS

  • AMBER HalRate decreases in all 24 matched comparisons, while POPE accuracy and F1 improve across all evaluated splits and SHR falls in all 24 comparisons.Mean SHR reductions are 2.80, 2.76, and 3.26 points for Qwen2.5-VL-3B, Qwen2.5-VL-7B, and InternVL3.5-8B, respectively.Source: §4.2 · p. 8
  • Mean Expert Avg decreases by only 0.28 and 0.31 points for Qwen2.5-VL-7B and InternVL3.5-8B, while Qwen2.5-VL-3B remains essentially unchanged.These retention results accompany broad hallucination improvements across the three configurations.Source: §4.2 · pp. 8–9

4.3 ABLATION STUDIES

  • Removing null-space projection lowers Expert Avg by 1.64 points on Qwen2.5-VL-7B and 1.34 on InternVL3.5-8B relative to MergeHEIR, with comparable hallucination performance.The matched variants share the training setup and differ only in null-space projection.Source: §4.3 · p. 9
  • Larger α generally lowers CHAIRs and CHAIRi, whereas smaller α preserves higher Expert Avg; α = 5 × 10^-4 is used as the shared default.Figure 5 reports Object HalBench CHAIRs and CHAIRi in its upper panels and Expert Avg in its lower panels across three model configurations.Source: §4.3 · p. 9, Figure 5 · p. 9

5 CONCLUSION

The paper studies hallucination in multimodal model merging and concludes that hallucination mitigation should be coordinated with preservation of inherited expertise. It positions MergeHEIR among model-merging, hallucination-mitigation, and null-space constraint approaches.

5 CONCLUSION

  • The conclusion identifies a systematic merging tax and expertise interference during post-merge correction, and presents MergeHEIR as combining multimodal preference optimization with activation-derived null-space projection.It reports geometric and representation-perturbation guarantees and experiments across three MLLM configurations, eight merging methods, and diverse benchmarks.Source: §5 · p. 9

A RELATED WORK

  • The evaluated merging methods span interpolation, arithmetic, sparsification, geometry or subspace, and optimization approaches; MergeHEIR starts from merged checkpoints.The paper studies mitigation of hallucination degradation while retaining inherited expertise.Source: A RELATED WORK · p. 15
  • Prior null-space constraints address interference in continual learning, editing, alignment, and merging; MergeHEIR applies pooled expert-task activations to post-merge hallucination adaptation.The paper also situates its approach among training-free and training-based multimodal hallucination mitigation methods.Source: A RELATED WORK · p. 15

B MECHANISTIC INTERPRETATION OF THE MERGING TAX

The proposed mechanism attributes merging-related hallucination to altered language-side use of visual evidence despite unchanged visual representations. The accompanying analysis characterizes projection as minimum-distortion and the selected admissible subspace as maximum-dimensional under the stated sensitivity condition.

B MECHANISTIC INTERPRETATION OF THE MERGING TAX

  • Because constituent experts share a frozen visual encoder, merging leaves a fixed image’s encoded representation unchanged and locates degradation downstream in language-side use of visual evidence.The paper names disruption of that downstream use visual-evidence utilization interference.Source: B MECHANISTIC INTERPRETATION OF THE… · p. 15

C ADDITIONAL THEORETICAL ANALYSIS

  • The analysis extends projection guarantees to activation distributions and fixed-prefix outputs under spectral approximation and local smoothness.It frames the theory as balancing adaptation flexibility with expert-task representation stability.Source: C ADDITIONAL THEORETICAL ANALYSIS · p. 16

C.1 PROOF OF PROPOSITION 3.1

  • Within the fixed admissible subspace SP, ΔWP is the unique orthogonal projection of accumulated displacement ΔW and minimizes Frobenius-norm update distortion.The decomposition separates the projected update from the component outside the subspace.Source: §C.1 · p. 16

C.2 PROOF OF THEOREM 3.2

  • The spectral construction selects a maximum-dimensional subspace satisfying the prescribed activation-sensitivity condition.The proof establishes that no subspace of greater dimension can satisfy the strict condition for every nonzero vector.Source: §C.2 · pp. 16–17
  • The spectral sensitivity constraint is used to bound representation perturbation for pooled and task-specific activations.The proof proceeds from the pooled activation matrix and the spectral definition of P.Source: §C.2 · p. 17

C.3 PROOF OF PROPOSITION 3.3

The spectral threshold α balances adaptation flexibility against calibration perturbation. Because the retained projector changes only at normalized spectral breakpoints, its sensitivity can include plateau-like regions.

C.3 PROOF OF PROPOSITION 3.3

  • The proof uses projector range inclusion, identities, Loewner ordering, and rank inequalities to relate the relevant subspaces.Source: §C.3 · p. 18
  • The threshold α balances the retained update against projection-induced distortion and calibration perturbation, trading adaptation flexibility for representation stability.The comparison considers a fixed displacement across nested spectral subspaces.Source: §C.3 · p. 18
  • The projector Pα is piecewise constant between normalized spectral breakpoints, so threshold sensitivity can exhibit plateau-like regions.Source: §C.3 · p. 18

C.4 PROOF OF COROLLARY 3.4

The corollary extends empirical perturbation control to the corresponding activation distribution. The bound accounts for the spectral discrepancy between the empirical second moment and the population second moment.

C.4 PROOF OF COROLLARY 3.4

  • The extension compares the empirical second moment bG with the activation-distribution second moment Γ to account for their spectral discrepancy.Source: §C.4 · pp. 18–19
  • The derivation applies Loewner-order bounds and a congruence transformation induced by the projector P, using P^2 = P.Source: §C.4 · pp. 18–19
  • The population perturbation bound adds the spectral discrepancy ε to the empirical bound, extending control beyond collected calibration activations under the stated spectral condition.Source: §C.4 · p. 19

C.5 FROM REPRESENTATION PERTURBATIONS TO FIXED-PREFIX OUTPUTS

Under local smoothness assumptions and fixed calibration prefixes, the analysis propagates projected module perturbations to logits and next-token distributions. It also specifies the expert training setup used to construct the task-specialized experts.

C.5 FROM REPRESENTATION PERTURBATIONS TO FIXED-PREFIX OUTPUTS

  • Under local smoothness, projected module perturbations bound changes in logits and next-token distributions at fixed prefixes.The analysis assumes bounded local sensitivities and a bounded nonlinear remainder around the merged checkpoint.Source: §C.5 · pp. 19–20
  • As the retained displacement approaches zero, the fixed-prefix distribution shift has a quadratic leading-order upper bound governed by downstream sensitivities and layer-wise spectral cutoffs.Source: §C.5 · p. 20
  • The theoretical results connect projected updates to representation stability and local output behavior, while α balances adaptation flexibility against calibration perturbation.The analysis includes a population extension and fixed-prefix bounds for logits and next-token distributions.Source: §C.5 · p. 21

D.1 TRAINING TASK-SPECIALIZED EXPERTS

  • The experts cover chart understanding, geometry reasoning, OCR, and visual question answering, with Table 3 summarizing their training datasets and sizes.Training uses LoRA only on non-visual modules; visual-side parameters remain frozen, and each expert is trained for one epoch.Source: §D.1 · p. 21, Table 3 · p. 21

D.2 EVALUATION BENCHMARKS

The evaluation uses seven expert-task benchmarks and hallucination benchmarks covering image descriptions, object hallucination, object-existence questions, and fine-grained description errors. Tables 4 and 5 report InternVL3.5-8B results for expert and hallucination tasks, respectively.

D.2 EVALUATION BENCHMARKS

  • Expert-task evaluation spans chart understanding, geometry reasoning, OCR, and visual question answering, with the unweighted mean across seven benchmarks reported as Expert Avg.Source: §D.2 · p. 21
  • Hallucination evaluation reports generative AMBER metrics, Object HalBench CHAIRs and CHAIRi, POPE accuracy and F1 across three splits, and GPT-4o-mini-assessed SHR.Source: §D.2 · p. 21
  • Tables 4 and 5 present InternVL3.5-8B results for expert-task and hallucination-task evaluations, respectively.Source: Table 4 · p. 22, Table 5 · p. 22

E ADDITIONAL RESULTS

Across additional quantitative and qualitative evaluations, MergeHEIR mitigates hallucination while largely retaining inherited expert capabilities. Its retention benefit also holds with alternative preference methods, and the impact statement identifies no additional consequences requiring specific discussion.

E ADDITIONAL RESULTS

  • Across eight merging methods, MergeHEIR improves hallucination performance while largely maintaining expert-task performance for InternVL3.5-8B.Tables 4 and 5 provide the complete matched results.Source: E ADDITIONAL RESULTS · p. 22
  • Across all 24 POPE cases, MergeHEIR improves accuracy and F1 on every split.Tables 6–8 report the complete POPE results.Source: E ADDITIONAL RESULTS · p. 22, Table 6 · p. 23, Table 7 · p. 23
  • Across six matched comparisons, adding null-space projection to RLAIF-V or OPA-DPO preserves expert performance with essentially the same hallucination-mitigation benefit.Direct adaptation with the underlying preference methods causes substantially greater degradation of inherited expert capabilities.Source: E ADDITIONAL RESULTS · p. 22, Table 9 · p. 24

F CASE STUDY

  • Qualitative expert-task examples show unconstrained adaptation impairing chart interpretation, geometric reasoning, and text recognition, while MergeHEIR retains task-specific knowledge consistent with visual evidence.The comparisons cover the merged model, MergeHEIR, and unconstrained post-merge adaptation.Source: F CASE STUDY · p. 25
  • On hallucination examples, both post-merge adaptations suppress nonexistent objects or unsupported scene details, but MergeHEIR avoids unconstrained adaptation’s pronounced expertise degradation.These examples qualitatively corroborate the quantitative hallucination-retention trade-off results.Source: F CASE STUDY · p. 25

G IMPACT STATEMENT

  • The impact statement says broader societal impacts are largely consistent with improving machine-learning robustness and reliability, with no additional consequences identified for specific discussion.The paper describes its work as an early step toward mitigating hallucinations in merged MLLMs while preserving inherited expert capabilities.Source: G IMPACT STATEMENT · p. 25

Related papers

Loading 2609.32422v1…