PaperlayerView exact revision on arXiv

Source-linked AI summary

From Suppression to Repair: Mitigating Object Hallucination in Large Vision-Language Models via Localized Distribution Alignment

Chen Zhao, Xingping Dong, Jiachun Shi, Liang Peng, Chong Wang, Zhen Lei, Ran He, Bo Du

arXiv:2610.11826v1cs.CVcs.AI

TL;DR

Object hallucination undermines the reliability of LVLMs, and suppressing hallucination-related components can also remove useful information. ResOT repairs representations through localized Gaussian distribution alignment at inference time; across three LVLMs, it reduces hallucination while preserving or improving caption quality and multimodal performance.

  • Problem

    LVLMs can describe objects absent from images, and suppressing hallucination-related components may compromise useful visual information and multimodal capabilities.

    Source: §1 · p. 1

  • Method

    ResOT projects dominant hallucinated directions away from the faithful subspace, then uses Gaussian OT in the resulting residual subspace to define repair targets and a likelihood-based gate to control inference-time movement.

    Source: §1 · p. 2

  • Results

    Across three representative LVLMs, ResOT consistently reduces object hallucination while preserving or improving image caption quality and broader multimodal performance.

    Source: §5 · p. 10

  • Takeaways & Limitations

    Repairing hallucination-related components rather than removing them provides a practical alternative that preserves or improves performance across the evaluated benchmarks.

    Source: §5 · p. 10

    ResOT’s Gaussian model captures residual means and covariances but not all higher-order structure, and calibration on complete descriptions versus repair on partial contexts leaves potential distribution shift insufficiently explored.

    Source: E LIMITATION · p. 21

Abstract

from arXiv · show

Object hallucination remains a major obstacle for large vision-language models (LVLMs) to generate reliable content. An intuitive mitigation strategy is to suppress hallucination-related components in hidden representations. However, these components may also contain useful information, and suppressing them can weaken the model's multimodal capabilities. In this paper, we propose ResOT, a training-free method that repairs representations at inference time through localized distribution alignment. Specifically, ResOT projects dominant hallucinated directions away from the faithful subspace, forming a low-dimensional residual subspace for intervention. Within this subspace, ResOT uses Gaussian optimal transport (OT) to align the hallucinated distribution with the faithful one. The resulting map defines repair targets with minimal changes to the original representations. At inference, ResOT adaptively controls how far each token state moves toward its OT target. Experiments on three representative LVLMs show that ResOT substantially reduces object hallucination while improving image caption quality and multimodal performance across multiple benchmarks. Code will be released.

1 INTRODUCTION

ResOT addresses the risk that suppressing hallucination-related representations can also remove information useful for visual understanding. It repairs them through localized distribution alignment, reducing hallucination while preserving or improving caption and multimodal performance.

1 INTRODUCTION

  • Suppressing hallucination-related components can compromise image caption quality or multimodal capabilities because those components may also carry useful visual information.The conceptual comparison contrasts suppression, which may displace faithful representations, with repair toward the faithful distribution.Source: §1 · p. 1, Figure 1 · p. 2
  • ResOT uses a residual subspace orthogonal to the dominant faithful subspace, where held-out analysis finds a clearer mismatch while faithful representations still retain components.The method leaves dominant faithful components unchanged and corrects residual mismatch rather than removing all hallucination-related components.Source: §1 · p. 2
  • Gaussian optimal transport aligns fitted hallucinated and faithful residual distributions with minimum expected squared displacement, providing repair targets.During inference, a likelihood-based gate controls movement toward the target, while the orthogonal component remains unchanged.Source: §1 · p. 2, Figure 2 · p. 3
  • Compared with vanilla models, ResOT reduces CHAIRs by 13.9%–46.8% and achieves average relative improvements of 27.67% on BLEU-1 and 5.86% on MME.On LLaVA-1.5, it retains 81.24% of original generation throughput while requiring 8.15 MiB of additional peak GPU memory.Source: §1 · p. 2

2 RELATED WORK

Prior work addresses LVLM hallucination through training, decoding-time interventions, and post-generation correction, while representation-based approaches steer or suppress hidden components. ResOT instead uses distribution alignment to define a faithful target in a residual subspace.

2.1 HALLUCINATION IN LARGE VISION-LANGUAGE MODELS

  • Hallucination methods include supervised fine-tuning, contrastive learning, RLHF, DPO, decoding-time interventions, and post-generation revision, each with distinct training or inference costs.Training-based methods often require additional supervision and parameter optimization; some decoding-time methods add overhead through extra forward passes or complex procedures.Source: §2.1 · p. 3
  • ResOT differs from representation interventions that steer activations, edit attention heads, or suppress components by defining a faithful target distribution in a residual subspace.Its update is determined by distribution alignment rather than component removal.Source: §2.1 · p. 3

2.2 OPTIMAL TRANSPORT FOR REPRESENTATION ALIGNMENT

  • Under squared Euclidean cost, the OT map between non-degenerate Gaussian distributions has a closed-form affine expression determined by their means and covariances.ResOT estimates the map offline and applies it during generation without an online transport solver.Source: §2.2 · p. 3

3 METHOD

ResOT calibrates a residual subspace and Gaussian optimal-transport map offline, then adaptively repairs token representations at inference. Its design preserves dominant faithful-subspace components while aligning residual representations toward the faithful distribution.

3 METHOD

  • ResOT has two stages: offline calibration identifies a residual subspace and estimates a Gaussian optimal-transport map, while inference adaptively moves token states toward repair targets.The map is estimated within the residual subspace and reused to set each token state's repair target.Source: §3 · pp. 3–4

3.1 LVLM GENERATION AND CALIBRATION REPRESENTATIONS

  • ResOT calibrates on paired faithful and hallucinated descriptions, extracting final-position hidden states from complete descriptions and reusing the resulting maps during generation with LVLM parameters fixed.The paired inputs are processed separately, and the calibration representations determine where and how to repair.Source: §3.1 · p. 4

3.2.1 LOCATING THE REPAIR SUBSPACE

  • The repair subspace captures dominant hallucinated directions outside the dominant faithful subspace, leaving that faithful subspace unchanged.Faithful representations can still have residual components, so ResOT uses their residual distribution as a repair reference rather than zeroing those components.Source: §3.2.1 · p. 4

3.2.2 GAUSSIAN OPTIMAL TRANSPORT IN THE RESIDUAL SUBSPACE

  • In the residual subspace, ResOT fits faithful and hallucinated Gaussians and uses a closed-form OT map minimizing expected squared displacement.Covariance shrinkage addresses noisy or ill-conditioned estimates from finite calibration data; the map defines a full target for each residual state.Source: §3.2.2 · pp. 4–5

3.2.3 REPAIR AT INFERENCE

  • At inference, a likelihood-based gate scales each residual correction toward its OT target while the orthogonal component remains unchanged.Stronger relative evidence for the hallucinated approximation receives a larger correction fraction; unusually large updates are clipped before the repaired state continues through the frozen decoder.Source: §3.2.3 · p. 5

3.2.4 THEORETICAL ANALYSIS

  • For a fixed dimension, the residual basis maximizes captured dominant hallucinated directions while excluding the dominant faithful subspace.The residual subspace is orthogonal to the dominant faithful subspace.Source: §3.2.4 · pp. 5–6
  • Projection-based suppression cannot recover the original faithful distribution when removed components also occur in faithful representations; Gaussian OT instead minimizes expected squared displacement.For a non-Gaussian source, the affine map remains optimal for matching first and second moments; complete distribution alignment is guaranteed when both distributions are Gaussian.Source: §3.2.4 · p. 6

4 EXPERIMENTS

The experiments compare ResOT with vanilla models and training-free baselines across object hallucination, caption quality, multimodal capability, and repair robustness. Results report improvements across these evaluation dimensions.

4.1 EXPERIMENTAL SETUP

  • The study evaluates three LVLMs against vanilla and training-free baselines using object-hallucination, caption-quality, multimodal-capability, and additional hallucination benchmarks.Source: §4.1 · p. 6, §4 · p. 6

4.2 OBJECT HALLUCINATION AND CAPTION QUALITY (RQ1)

  • ResOT achieves the highest POPE accuracy and F1 across all three backbones, and the highest precision on mPLUG-Owl2 and InstructBLIP.Source: §4.2 · p. 7, Table 1 · p. 7
  • ResOT achieves the lowest CHAIRs and CHAIRi on COCO, alongside higher BLEU-1 rather than a decline in reference-based caption quality.Source: §4.2 · p. 7, Table 2 · p. 7
  • GPT-4o scoring finds higher ResOT accuracy than both baselines across three backbones and higher detailedness in five of six comparisons.Source: Table 3 · p. 8, §4.2 · p. 8

4.3 MULTIMODAL CAPABILITY ASSESSMENT (RQ2)

  • ResOT improves MME total, perception, and cognition scores over vanilla on all three backbones, and beats vanilla in 9 of 12 settings across four other benchmarks.Source: Figure 3 · p. 8, §4.3 · p. 8

4.4 ANALYSIS OF LOCALIZED REPRESENTATION REPAIR

  • The residual space exposes clearer faithful–hallucinated differences than original and Nullu-style spaces, but overlap remains; OT repair reduces mismatch without discarding residual components.Source: §4.4 · pp. 8–9, Figure 4 · p. 9
  • With 50 calibration pairs, ResOT already outperforms Nullu in Cs, BLEU-1, and MME; its advantage persists as corruption rises, while variability generally falls with larger sets.Source: §4.4 · pp. 9–10, Figure 5 · p. 9

5 CONCLUSION

The conclusion presents ResOT as a localized repair alternative to suppressing hallucination-related representations. The accompanying proof selects leading residual directions that are orthogonal to the dominant faithful subspace.

5 CONCLUSION

  • ResOT repairs hallucination-related representations in the residual subspace using Gaussian optimal transport and a likelihood-based gate, with reported gains across hallucination, caption, and multimodal benchmarks.Source: §5 · p. 10

A.1 PROOF OF PROPOSITION 1

  • The proof orders residual singular values decreasingly and shows the top-k singular-vector basis attains the upper bound, with selected directions orthogonal to the faithful span.Source: §A.1 · p. 14

A.2 PROOF OF PROPOSITION 2

The proofs show why projection alone cannot recover the faithful distribution when removed components contain faithful energy, and establish a unique minimum-displacement affine map matching the target’s first two moments.

A.2 PROOF OF PROPOSITION 2

  • The proof constructs a coupling that attains the lower bound linking distance to the full faithful distribution with projected distance and residual faithful energy.The joint distribution of X and ΠY couples the projected distributions; an optimal projected coupling makes the bound attainable.Source: §A.2 · pp. 14–15
  • When the removed subspace contains nonzero faithful energy, projection leaves the hallucinated distribution at a strictly positive Wasserstein distance from the faithful distribution.Source: §A.2 · p. 15

A.3 PROOF OF PROPOSITION 3

  • T⋆ is the unique affine map matching the target’s first two moments with the smallest expected squared change, even for a non-Gaussian source.The proof establishes uniqueness of the feasible affine map.Source: §A.3 · p. 15

B IMPLEMENTATION AND EVALUATION DETAILS

The evaluation uses paired faithful and hallucinated descriptions for calibration, evaluates hallucination, generation quality, and multimodal performance, and selects repair settings on held-out calibration data.

B IMPLEMENTATION AND EVALUATION DETAILS

  • Each backbone uses 5,000 calibration pairs containing COCO training images, faithful LLaVA-Instruct-150K captions, and corresponding GPT-3.5-generated hallucinated captions.Source: B IMPLEMENTATION AND EVALUATION DETAILS · p. 16
  • Evaluation covers POPE object-existence questions, CHAIR caption hallucinations and BLEU-1 quality, MME multimodal performance, and GPT-4o ratings of visual accuracy and detail.POPE includes random, popular, and adversarial splits; CHAIR uses 500 OPERA-selected COCO validation images.Source: B IMPLEMENTATION AND EVALUATION DETAILS · p. 16
  • Baselines use their official recommended settings, with backbone, prompt, and maximum generation length held constant within each benchmark.Vanilla and ResOT use greedy decoding; OPERA uses five-beam search and SGRS uses saliency-guided top-k rejection sampling.Source: B IMPLEMENTATION AND EVALUATION DETAILS · p. 16
  • The 4,500-pair fitting split estimates residual subspaces and OT maps, while 500 validation pairs select shared rank and repair strength using balanced accuracy, CHAIRi, and BLEU-1.Configurations are first filtered by balanced accuracy, then ranked by CHAIRi and BLEU-1; final maps are refit on all 5,000 pairs.Source: B IMPLEMENTATION AND EVALUATION DETAILS · p. 16

C.1 MORE EVALUATION OF CAPTION QUALITY

ResOT improves caption quality across all three evaluated LVLM backbones, with the largest gains on LLaVA-1.5 and mPLUG-Owl2.

C.1 MORE EVALUATION OF CAPTION QUALITY

  • ResOT achieves the highest BLEU-1–4, ROUGE-L, CLIPScore, and SPICE scores across all three backbones.Gains are largest on LLaVA-1.5 and mPLUG-Owl2, while InstructBLIP shows smaller but consistent improvements.Source: §C.1 · p. 16, Table 5 · p. 17

C.2 RESULTS ON ADDITIONAL BENCHMARKS

Across additional benchmarks, ResOT generally reduces hallucination while preserving or improving multimodal performance. Its additional computation is characterized by the number of intervened layers and residual-subspace dimension, with reported throughput and memory costs on LLaVA-1.5.

C.2 RESULTS ON ADDITIONAL BENCHMARKS

  • On AMBER, ResOT has the lowest CHAIR, Hal, and Cog scores on all three backbones, with object coverage close to the best result.The reported hallucination reduction accompanies a slight change in object coverage.Source: §C.2 · p. 17
  • Across MMMU, TextVQA, ScienceQA, and MMBench, ResOT beats Nullu on all three backbones and improves over vanilla in 9 of 12 settings, with modest TextVQA decreases for two backbones.It matches vanilla in one setting; overall, results indicate preserved or improved general multimodal performance while reducing hallucination.Source: §C.2 · p. 17
  • On MMHal-Bench, ResOT has the highest average score on all three backbones and reduces hallucination versus vanilla.It has the lowest hallucination rate on LLaVA-1.5 and InstructBLIP and ties with Nullu on mPLUG-Owl2.Source: §C.2 · p. 17, Table 7 · p. 18
  • On MM-Vet, ResOT improves the overall score across all three backbones, although performance varies across individual skills.Together with the hallucination results, this suggests no broad degradation of multimodal capability.Source: §C.2 · p. 17, Table 8 · p. 18

C.3 COMPLEXITY ANALYSIS

  • Online additional computation per generated token is O(L(dk + k^2)), dominated by O(Ldk) when k ≪ d; storage is O(L(dk + k^2)).Subspace estimation and Gaussian OT-map fitting occur offline.Source: §C.3 · p. 18
  • On LLaVA-1.5, ResOT retains 81.24% of paired vanilla throughput and adds 8.15 MiB of peak GPU memory.Efficiency measurements used the same hardware configuration; the intervention state is stored separately from model weights.Source: §C.3 · p. 18

D.1 QUANTITATIVE ANALYSIS OF REPRESENTATIONS

Across the analyses, ResOT’s residual subspace better separates faithful and hallucinated representations, and Gaussian OT moves repaired centers toward faithful centers. Sensitivity and Gaussian diagnostics also identify performance trade-offs and limits of the Gaussian approximation.

D.1 QUANTITATIVE ANALYSIS OF REPRESENTATIONS

  • Across layers 16, 24, and 31, the residual subspace makes faithful–hallucinated differences more apparent, and Gaussian OT moves hallucinated representations toward faithful ones.Figure 6 shows the original space, Nullu-style subspace, and ResOT residual subspace before and after repair.Source: §D.1 · p. 18, Figure 6 · p. 20

D.2 HYPERPARAMETER SENSITIVITY

  • On LLaVA-1.5, increasing α generally lowers Cs and improves BLEU-1, but MME peaks at α = 0.8; ranks above 32 weaken all three metrics.Figure 7 examines α with rank fixed at 32 and rank with α fixed at 0.7.Source: §D.2 · p. 19, Figure 7 · p. 20

D.3 GAUSSIAN DIAGNOSTICS IN THE RESIDUAL SUBSPACE

  • At layer 31, whitened marginals broadly follow a standard normal centrally, while Mahalanobis Q–Q plots show stronger departures from the χ2_k reference.The normal Q–Q plots are close to the diagonal across much of the central range, with larger deviations in the tails.Source: §D.3 · p. 20
  • Across all repaired layers and three backbones, marginal KS distances remain relatively small and stable, whereas Mahalanobis KS distances are generally larger, indicating the pattern is not model- or layer-specific.Source: §D.3 · p. 20
  • The diagnostics support Gaussian distributions as compact approximations, not exact descriptions, with departures more pronounced in joint distributions and tails.The transport map uses estimated means and covariances; the joint residual distributions retain structure beyond a Gaussian model.Source: §D.3 · p. 20

D.1 QUANTITATIVE ANALYSIS OF REPRESENTATIONS

  • Across evaluated layers, the residual space consistently yields higher Mahalanobis distance, LDA accuracy, and LDA AUC than the Nullu-style subspace; Gaussian OT significantly reduces distance between faithful and repaired centers.Source: §D.1 · pp. 18–19
  • With α = 1, hallucinated representations undergo larger relative displacements than faithful representations across layers, indicating selective rather than uniform repair.Table 13 reports average relative displacement and the hallucinated-to-faithful displacement ratio.Source: §D.1 · p. 19, Table 13 · p. 21

E LIMITATION

ResOT’s Gaussian target captures residual means and covariances but not all higher-order structure. The paper also identifies a calibration-to-inference distribution shift and limits its evaluation primarily to object hallucination.

E LIMITATION

  • The low-dimensional Gaussian target captures differences in mean and covariance but does not describe all higher-order residual structure.The diagnostics also show that joint residual distributions are not exactly multivariate Gaussian.Source: E LIMITATION · p. 21
  • Calibration uses complete descriptions, while repair acts on partially generated contexts; this potential distribution shift remains insufficiently explored.Source: E LIMITATION · p. 21
  • Evaluation focuses primarily on object hallucination, so applicability to other forms of multimodal error remains unverified.Source: E LIMITATION · p. 21
Loading 2610.11826v1…