PaperlayerView exact revision on arXiv

Source-linked AI summary

Targeting the Attention Heads Behind Object Hallucination in LLaVA

Armaan Sandhu, Abhilasha Senapati, Hima Kammachi

arXiv:2608.24966v1cs.CVcs.AI

TL;DR

Object hallucination makes vision-language captions unreliable, motivating the question of whether interpretability signals can guide targeted correction. The paper identifies and ablation-screens hallucination-linked attention heads, restricts LoRA and grounding interventions to them, and finds lower CHAIRs and CHAIRi on held-out COCO images. Controls support selected-head leverage beyond LoRA capacity, while fixed-budget tests and behavioral profiling show persistent hallucination reduction alongside shorter, more conservative captions and lower object recall.

  • Problem

    Object hallucination produces unsupported object mentions in captions and undermines reliability when descriptions feed downstream decisions.

  • Method

    The pipeline ranks heads by image-attention drops around hallucinated object words, ablation-screens candidates, and restricts LoRA plus grounding control to the resulting 32-head set.

  • Results

    CHAIRs decreases from 0.370 to 0.230 and CHAIRi from 0.156 to 0.096 on 400 held-out COCO images, with p < 0.001 paired sign-flip tests.

  • Takeaways & Limitations

    Interpretability-guided head selection identifies intervention sites with measurable behavioral leverage beyond LoRA capacity alone.

  • Takeaways & Limitations

    The study is limited to LLaVA, one captioning prompt, and COCO-style annotations, while CHAIR and attention provide incomplete proxies for visual truth and grounding.

Abstract

from arXiv · show

Vision-language models such as LLaVA-1.5-7B often hallucinate objects absent from the image when generating captions. We ask whether an interpretability diagnosis of this failure can guide a targeted fix, and we measure what that fix actually changes. We rank attention heads by how much their image attention drops around hallucinated object words, then screen the shortlist by ablating candidate heads and measuring the change in hallucination-token log probability, yielding a 32-head set. We restrict two interventions to these heads: a head-sliced LoRA adapter and an inference-time grounding controller. On 400 held-out COCO images, the combined method lowers CHAIRs (the fraction of captions with a hallucinated object) from 0.370 to 0.230 and CHAIRi (the fraction of hallucinated object mentions) from 0.156 to 0.096 (p < 0.001, paired sign-flip tests). Two controls sharpen attribution. A random-head LoRA control, matched layer-for-layer and trained identically, performs no better than the matched baseline on a separate 200-image control split, supporting the role of head selection rather than LoRA capacity. Under fixed decoding budgets, the CHAIR reduction persists and grows with budget (23% at 64 tokens to 58% at 128), arguing against a pure max-token or truncation artifact, although the method remains shorter and more conservative. The resulting behavior reduces unsupported object mentions while also lowering object recall (0.78 to 0.70). We present a diagnosis-to-intervention pipeline for object hallucination, and, more importantly, a controlled account of what acting on the diagnostic signal actually does: it localizes intervention sites with real, non-random leverage, reported as a behavioral profile rather than a single score.

1 Introduction

The paper turns an interpretability diagnosis of object hallucination into targeted interventions, then evaluates both their effectiveness and behavioral trade-offs. The selected-head method reduces hallucination, but also produces shorter captions and lower object recall.

  • Object hallucination produces fluent captions containing objects absent from images, undermining reliability when descriptions inform downstream decisions.
  • The pipeline ranks heads by reduced image attention around hallucinated object words, ablation-screens candidates, and targets the resulting 32 heads.
  • On a separate 200-image control split, random-head LoRA matches the baseline while selected-head LoRA reduces CHAIRs to 0.100, isolating head selection beyond LoRA capacity.
  • CHAIRs falls from 0.370 to 0.230 and CHAIRi from 0.156 to 0.096 on 400 held-out images, with p < 0.001 paired sign-flip tests.
  • Under fixed decoding budgets, relative CHAIRs gains grow from 23% at 64 tokens to 58% at 128 tokens, although the method remains shorter and more conservative.
  • The intervention reduces unsupported object mentions while also shortening captions and lowering object recall, so results are reported as a behavioral profile.

2 Related Work

Related work establishes object hallucination as a measurable visual-grounding failure and links it to attention patterns in vision-language models. This paper extends diagnose-then-target approaches with attribution controls, fixed-budget analysis, and an explicit precision-coverage account.

  • CHAIR measures whether generated object mentions appear in image annotations, providing a concrete but incomplete proxy for visual truth.
  • Prior interpretability studies link hallucination to visual-attention patterns, including middle-layer localization and low image-token attention during generation.
  • The paper adds a layer-matched random-head control to isolate head selection from adapter capacity.
  • A fixed-budget analysis separates hallucination reduction from max-token truncation effects, while the behavioral profile exposes the intervention’s precision-coverage trade-off.
  • Unlike prior broadly targeted adaptation, the approach localizes LoRA and decoding interventions to heads implicated by the hallucination diagnosis.
  • Compared with SPIN, the method uses a fixed ablation-screened head set rather than reselecting heads at each decoding step.

3 Method

The method identifies hallucination-linked attention heads, adapts only their dimensions, and applies a grounding controller over the same fixed set during decoding. Grounding is computed from selected-head image attention and penalizes object-word logits when that score is low.

  • Evaluation: Evaluation uses LLaVA-1.5-7B on COCO val2014 with a fixed captioning prompt, reporting CHAIRs and CHAIRi.
  • Pipeline: The three-stage pipeline performs diagnosis, selected-head LoRA adaptation, and grounding control over the same head set.
  • Head selection: Heads are ranked by image-attention decreases around hallucinated versus grounded object words across LLaVA’s 1024 attention heads.
  • Head selection: Ablation screening retains the 32 heads with the largest effects on hallucination-token log probability, spanning 19 transformer layers.
  • LoRA Adaptation: LoRA updates only selected head slices in Q/K/V projections, using gradient masking to restrict learning to those dimensions.
  • LoRA Adaptation: Contrastive training pairs COCO ground-truth captions with the model’s hallucinated captions, and the shorter chosen responses bias the adapter toward brevity.
  • Inference-Time Grounding: The grounding controller computes a score from selected-head visual attention mass and applies a conservative inference-time guard on top of LoRA.
  • Inference-Time Grounding: When the grounding score falls below threshold, logits for COCO-derived object words receive a penalty; otherwise, the controller remains inactive.

4 Experiments

Experiments show that diagnosis-selected interventions reduce object hallucination, with evidence from held-out comparisons, random-head controls, fixed-budget tests, and qualitative examples. The gains are accompanied by shorter, more conservative captions and lower object recall, so the paper evaluates a behavioral tradeoff rather than a single score.

  • Random-Head Control: On the 200-image control split, selected-head LoRA reduces CHAIRs to 0.100, whereas matched random-head LoRA remains at baseline CHAIRs 0.400.The random-head control uses the same training recipe and layer-matched head counts.
  • Tradeoff: Fewer Hallucinations, Shorter Captions: The full method is more conservative: average caption length drops from 61.1 to 29.6 words and object recall from 0.78 to 0.70.This tradeoff means lower CHAIR may partly reflect mentioning fewer objects; fully length-matched human evaluation remains missing.
  • Fixed-Budget Comparison and Robustness: At fixed decoding budgets, the relative CHAIRs reduction grows from 23% at 64 tokens to 58% at 128 tokens.The fixed-budget design controls the maximum token budget, but the method still generates shorter captions within each budget.
  • Sensitivity Checks: The grounding controller improves CHAIRs across LoRA scales, while threshold and top-16 versus top-32 head checks yield similar CHAIR results.These sensitivity checks support the head-selection account but do not replace length-matched 400-image controls.
  • Qualitative Example: In a held-out kitchen example, the combined method suppresses the unsupported oven mention while retaining grounded objects, but produces a shorter caption.The example mirrors the aggregate pattern of fewer unsupported mentions alongside greater conservatism.

5 Discussion

The selected-head intervention reduces object hallucination, but its leverage depends on head selection and comes with more conservative generation. Fixed-budget tests show the reduction is not purely a shorter-caption artifact, though caption quality is not uniformly improved.

  • The selected heads, rather than LoRA capacity alone, appear to provide the intervention’s leverage, while grounding adds value across adapter strengths.The random-head control does not reproduce the selected-head adapter’s behavioral change, and the grounding controller improves CHAIRs across LoRA scales.
  • The method reduces unsupported object mentions while also shortening captions and lowering object coverage, so it is not a uniform improvement in caption quality.The paper interprets the result as targeted hallucination reduction rather than general caption-quality improvement.
  • 23% at 64 tokens to 58% at 128 tokens: CHAIR reduction persists and grows under fixed decoding budgets.The methods share maximum token budgets, although the targeted method still produces shorter captions within each budget.

6 Limitations and Future Work

The study’s evidence is bounded by its model, prompt, annotations, and attention-based grounding proxy. The authors call for length-matched, broader, and human-centered controls to separate hallucination reduction from reduced object coverage and assess quality more directly.

  • The study covers one model family, one captioning prompt, and COCO-style object annotations, limiting the evaluated scope.CHAIR also depends on annotation coverage and synonym matching, so valid but unannotated objects may be counted as hallucinations.
  • Attention serves as a lightweight grounding proxy, but attention weights are not guaranteed to faithfully explain model behavior.The selected heads are additionally screened by ablation.
  • A fully length-matched comparison, matched-capacity baselines, human evaluation, and tests across additional VLMs, prompts, and datasets are proposed follow-ups.These controls would assess quality beyond CHAIR and test whether the findings generalize.

7 Conclusion

The paper presents a diagnosis-to-intervention pipeline that localizes hallucination-linked attention heads and applies targeted adaptation plus grounding-aware decoding. It reduces hallucination under fixed budgets, but the resulting captions are shorter and more conservative, framing the gain as targeted rather than globally improved captioning.

  • The full method significantly reduces CHAIRs and CHAIRi on 400 held-out COCO images, with both reductions reaching p < 0.001.A layer-matched random-head control supports the relevance of selected heads rather than LoRA capacity alone.
  • The intervention’s hallucination reduction persists under fixed decoding budgets and grows as more tokens are generated.This budget analysis addresses a max-token or truncation explanation.
  • The gains are best understood as targeted reductions in object hallucination, not as a general improvement in caption quality.The resulting captions are shorter and more conservative.
Loading 2608.24966v1…