acceptodds
Under review as a conference paper at ICLR 2027

MergeHEIR: Mitigating Multimodal Hallucinations as the Tax of Model Merging

Abstract

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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Model Merging, Multimodal Hallucination

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EuA1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies hallucination in multimodal model merging. On Qwen2.5-VL-3B/7B, it merges four LoRA experts (Chart, Geometry, OCR, VQA) with eight merging methods. Every merged checkpoint shows a higher AMBER hallucination rate than the average of its experts, which the authors call a "merging tax". To reduce it, they propose MergeHEIR. It applies mDPO-style preference optimization to the merged model. To protect the inherited expertise, it builds layer-wise null spaces from expert-task calibration activations and periodically projects the accumulated update onto them. The paper gives theoretical results on this projection. Experiments cover three backbones and eight merging methods, for 24 merged vs. MergeHEIR comparisons.

Critical strengths
  1. The problem is new and useful. To my knowledge, this is the first systematic study showing that model merging can increase multimodal hallucination. The trend is consistent across 16 merged checkpoints. This matters for anyone who deploys merged MLLMs, so I count it as the main reason to consider acceptance.

  2. The evaluation is broad. It covers eight merging methods, three backbones, four hallucination benchmarks, and seven expert-task benchmarks. Table 9 also shows that the projection can be added to other preference methods. This breadth makes the hallucination reduction itself convincing, even though the retention claim needs more support (see Weaknesses).

Critical weaknesses
  1. The evidence for the null-space constraint is limited. The main comparisons are merged vs. merged+DPO, which mostly show that DPO reduces hallucination. The key comparison against unconstrained adaptation covers only three settings, with small gaps and no variance. Simple retention baselines are also missing. Since better retention is the method's main claim, this weighs most on my rating.

  2. The merging tax is not well isolated. Without the base model, individual experts, and response-length controls, it is unclear whether the increase comes from merging itself. The proposed mechanism is also not tested. This matters because the merging tax is the paper's core motivation.

Questions for authors
  1. The comparison with unconstrained adaptation (i.e., the same training without null-space projection) is only shown for one merging method per model. Could you run it on more merging methods, with multiple random seeds? It would also help to compare with a simple way to protect expert skills, for example adding a penalty that keeps the weights close to the merged model.

  2. The merging tax is measured against the average of the experts. Could you also report the hallucination rate of the original base model and of each expert? Please also report the average response length, since shorter answers usually contain fewer hallucinations. This would show whether the increase really comes from merging.

Overall, I like this paper. The problem is new and interesting, and the paper is clearly written. The method itself is fairly simple, but I think this is fine if the experiments support the main claims. Right now, the two points above are not fully addressed. If the authors address them, I will keep my positive score. If not, I may lower it.

Suggestion for future work (not affecting my score): direct evidence for the proposed mechanism, such as an analysis of attention to visual tokens.

Reviewer_6H8UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies a practically relevant reliability problem and provides broad evidence that post-merge preference optimization can reduce hallucination while damaging inherited expertise. MergeHEIR addresses this tension with activation-based null-space constraints and shows consistent benefits. My main concerns are the limited evidence for the proposed merging-specific mechanism, insufficient comparisons with alternative retention strategies, and the gap between local perturbation guarantees and end-to-end capability preservation.

Critical strengths
  1. The practical problem is clearly established. Figure 1 documents higher AMBER hallucination rates after merging, while the matched adaptation comparisons show a distinct retention problem. This is a meaningful deployment concern.

  2. The method directly addresses the retention objective. The constraint in Eqs. (7)-(10) has a clear interpretation: suppress parameter changes that strongly perturb calibration activations. The projection and spectral bounds are mathematically reasonable within their stated scope.

  3. The empirical coverage is substantial. The 24 before/after comparisons, multiple hallucination measures, and compatibility tests with RLAIF-V and OPA-DPO in Appendix E support the practical usefulness of the projection mechanism.

Critical weaknesses
  1. The proposed cause of the merging tax is not directly established. Figure 1 supports degradation relative to the expert average, but Appendix B provides a conceptual account of “visual-evidence utilization interference” rather than a controlled mechanistic test. A frozen vision encoder locates the changed parameters downstream; it does not distinguish weakened visual grounding from changes in language priors or response behavior. Reporting the original base model and individual experts alongside the merged models would also clarify the baseline. This limits the causal interpretation rather than negating the observed problem.

  2. The comparisons do not establish an advantage over other retention strategies. Figure 4 and Table 9 provide useful projection/no-projection controls, but comparisons with replay, distillation to the merged reference, or an adapted existing null-space alignment method are missing. Appendix A already discusses related projection approaches. Under matched calibration access and adaptation budgets, such baselines are needed to establish the value of this particular post-merge constraint beyond generic capability-preserving adaptation.

  3. The guarantees and ablations leave important implementation questions unresolved. The bounds concern fixed calibration activations; the population extension assumes spectral approximation, and Appendix C's output analysis is local and uses fixed prefixes. They do not guarantee preservation under changing activations and autoregressive contexts. Figure 5 varies the spectral threshold, but calibration-size and projection-frequency ablations, projector ranks, and training overhead are not reported. These are important for judging whether the constraint remains effective and affordable beyond the chosen setup.

Questions for authors
  1. Mechanism: Does the manuscript contain evidence beyond Appendix B's conceptual discussion that distinguishes visual-evidence utilization interference from changes in response length, language priors, or general task behavior?

  2. Implementation: How is the accumulated dense displacement projected while maintaining the LoRA parameterization? Are optimizer states adjusted, and is a final projection applied when training ends between projection steps?

  3. Calibration and cost: Please specify calibration-data provenance and separation from evaluation, typical retained projector ranks, and the additional time and memory required for projector construction and periodic projection.

Reviewer_aq7LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the increase in visual hallucinations after merging task-specialized multimodal models. The authors propose MergeHEIR, which combines hallucination-oriented preference optimization with null-space projection to reduce hallucinations while preserving the capabilities inherited from the original experts. Experiments across multiple model scales and merging methods show that the proposed approach can reduce hallucinations with relatively small losses in expert-task performance.

Critical strengths

The paper identifies an interesting and practically relevant phenomenon: model merging can increase visual hallucinations even when the constituent experts share the same frozen vision encoder. This is a useful observation that deserves further investigation.

The experiments cover multiple merging methods and model configurations, providing reasonable evidence that the phenomenon is not limited to a single merging algorithm.

The proposed framework addresses a practical trade-off between hallucination mitigation and preserving the capabilities of merged experts.

Critical weaknesses

The mechanistic analysis is insufficient. The paper attributes the increased hallucination rate to interference in how the language-side network utilizes visual evidence. However, this explanation remains largely speculative. The paper does not provide sufficient direct evidence that visual grounding becomes weaker after merging, nor does it adequately distinguish this hypothesis from general parameter interference or capability degradation. More mechanistic experiments are needed to establish why merging increases hallucinations.

The methodological novelty is limited. Applying preference optimization to reduce hallucinations is a reasonable approach, but using hallucination-oriented preference data with DPO is already an established direction. Similarly, null-space projection is a well-known technique for constraining parameter updates and mitigating interference. The paper mainly combines these existing techniques in a post-merging setting, and the new algorithmic contribution appears limited. The authors should better clarify what fundamental methodological insight distinguishes MergeHEIR from existing constrained optimization and null-space projection methods.

The evidence for the proposed mechanism and the necessity of the method is incomplete. The paper shows that unconstrained preference optimization can affect the performance of merged experts, while null-space projection helps preserve their capabilities. However, it remains unclear whether the improvement comes from protecting specific activation directions or simply restricting the magnitude of parameter updates. Comparisons with simpler regularization and norm-matched update baselines would help establish the value of the proposed design.

The data requirements limit the method's applicability. MergeHEIR relies on 10,000 multimodal preference pairs and expert-task calibration data. Although this is reasonable for supervised post-merging adaptation, it raises the question of whether hallucination mitigation can be achieved under stricter data-free settings. This is particularly relevant to model merging, where access to the original task data may be limited or unavailable.

Questions for authors

Can the authors provide direct evidence that merging weakens the influence of visual evidence on generation? For example, can the hypothesis be verified through image perturbation experiments, intermediate-layer activation analysis, or causal interventions? How does this explanation differ from general parameter interference?

Given that DPO-based hallucination mitigation and null-space projection are both established techniques, what is the main methodological novelty of MergeHEIR beyond combining them in the post-merging setting? Why are existing constrained optimization or null-space methods insufficient?

Can the authors compare MergeHEIR with simpler alternatives, such as L2 or KL regularization and norm-matched parameter updates, to demonstrate that the specific null-space constraint is responsible for the improvements?

Is it possible to mitigate hallucinations in a strictly data-free setting, without hallucination preference pairs or original-task calibration data? Can the authors explore whether parameter-level information from the merged model and its constituent task vectors is sufficient to construct useful constraints for reducing hallucinations?

The paper reports that merged models have higher hallucination rates than the average of their constituent experts. Can the authors report the hallucination rates of individual experts and compare each merged model against them? This would help clarify whether the observed merging tax is consistent across experts rather than driven by the choice of average baseline.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.