acceptodds
Withdrawn from ICLR 2027

Beyond Model-Internal Signals: Steering VLLM Attention with Predicted Human Saliency

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

To reduce hallucinations in VLLMs, training-free methods steer attention toward relevant image regions, identified from model-internal statistics alone. In this work, we ask whether an external reference, human visual attention, can serve this role. To answer this question, we introduce and release VLFeedback-ETC, an eye-tracking corpus of 21 participants annotating preferences on VQA samples from VLFeedback, a widely used VLLM alignment dataset. Since gaze is unavailable at inference, we also introduce QaSalFormer, a question-aware saliency model for natural images, and benchmark it against existing predictors on our corpus. Using the corpus to compare human gaze with LLaVA-1.5 attention, we find that they are weakly correlated, with large differences between layers. We then propose GazeSteer, which adds predicted saliency to the pre-softmax attention scores over image tokens. On CHAIR, predicted saliency reduces object hallucination beyond a spatially uniform map that adds the same attention mass, and the reduction is not explained by shorter captions. Interestingly, the utility of a map is not determined by its global alignment with human gaze. Instead, useful maps combine image-specific structure with enough coverage. Moreover, predicted saliency does not replace cross-head consensus, but combining both is at least as effective as consensus alone, suggesting that external saliency is a complementary signal to model-internal ones.

Primary area:
foundation or frontier models, including LLMs
Keywords:
vllm, saliency, et, attention, hallucinations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1CZ2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether predicted human visual attention can replace the model-internal signals used by training-free attention steering against VLLM hallucination. Contributions: (i) VLFeedback-ETC, Tobii eye-tracking from 21 participants judging response pairs on 30 VLFeedback VQA triplets (Section 3; Fleiss kappa 0.453); (ii) QaSalFormer, VisSalFormer fine-tuned on VOILA-COCO mouse traces, benchmarked against Gazeformer, VisSalFormer and MD-SEM (Section 4, Table 1); (iii) a comparison of LLaVA-1.5-7B/13B attention (rollout at the first generated token, and per layer) with human saliency, CC about 0.09 (Table 2, Figure 7); (iv) GazeSteer (Eq. 3), adding a scaled saliency map to pre-softmax image-token scores, alone or on top of the consensus term of Jiang et al. (2025), evaluated with CHAIR on 500 COCO images, greedy decoding, layers 5-18 (Section 6, Tables 7-10). The authors conclude that QaSalFormer maps beat a constant map of equal mass, that caption length does not explain the gain, that gaze agreement does not predict usefulness, and that saliency is comparable to, and possibly complementary to, consensus.

Critical strengths
  1. Controls many steering papers omit, and calibrated claims. The mass-matched constant map (Section 6.1), greedy captions truncated to equal length (Appendix E.3, Figures 8-15) and matching at equal F1/recall rather than equal beta address the obvious confounds. The text states CS has about two points of uncertainty with 500 images and calls the combination "comparable" to consensus, not better (Section 6.2). The constant control shows added image attention alone helps (7B, Table 8, lower CS is better: Constant 34.8 vs greedy 53.2).
  2. A released gaze corpus plus an informative negative result. Ranking by human-gaze agreement (MD-SEM best, Table 1) differs from ranking as a steering signal (QaSalFormer best, Section 6.2), a useful caution against using gaze alignment as a proxy. Consent is described (Appendix B.3); raw data and code are promised.
Critical weaknesses
  1. That human-like saliency is the useful signal is not isolated, and margins are small relative to the stated noise. The only structure-free control is a constant map. I found no control that keeps spatial structure but breaks image correspondence (a map from another image, a centre-bias prior); centre bias appears only in the specificity definition (Section 4.1). QaSalFormer covers about 43% of the image versus 21% for humans (Table 1), the better-aligned MD-SEM is less useful, and the authors say the sources "cannot establish their respective contributions" (Section 6.2). On LLaVA-1.5-7B (CS; greedy 53.2), QaSalFormer is 30.0 vs Constant 34.8 at the 1-point F1 margin (Table 8) and 30.0 vs 32.4 at the 3-point margin (Table 10): a 2.4-point gap against the roughly two-point uncertainty. Consensus on the same model is 27.4 (Table 8) and 14.8 (Table 10), so saliency alone is clearly weaker. On 13B it is comparable (30.8 vs 35.4 in Table 8; 30.8 vs 30.0 in Table 10). Each number is one greedy run; I found no confidence intervals or significance tests, and Tables 7-10 report the best configuration per family from sweeps of different sizes. Partly fixable with added controls, bootstrap intervals and a common beta grid.
  2. Narrow evaluation and a weak link between gaze data and intervention. Only CHAIR and only LLaVA-1.5 (7B, 13B) are used (Section 6.1, Limitations); I found no other benchmark or model family. QaSalFormer's question conditioning is not exercised: every caption uses "What is in the image?", picked from a 10-prompt ablation on 7B (Table 6, CS 30.0 to 41.4) that appears to use the same CHAIR setup. QaSalFormer is trained on mouse traces, and the corpus has 30 stimuli from a preference task, not VQA answering (Appendix B.2). The "grounding gap" (Table 2: CC 0.090 ± 0.274 on 7B) rests on first-token rollout, as Appendix F acknowledges. Fixable only with new experiments.
Questions for authors
  1. (Clarification / new experiment) At matched F1, what happens with QaSalFormer maps shuffled across images, or replaced by a centre-bias or blurred model-attention map?
  2. (Clarification) Can you give bootstrap intervals over the 500 images for CS and CI in Tables 7-10, with paired QaSalFormer vs Constant comparisons? How was the "around two points" uncertainty (Section 6.2) computed?
  3. (Clarification) Was the Table 6 prompt chosen on the same 500 images used for Figures 4 and 8-15? If so, does the QaSalFormer-vs-Constant gap hold on held-out images?
  4. (Clarification) With std 0.27-0.29 over 30 stimuli (Table 2), is the layer contrast (about 0.3 in layers 20-25 vs near zero in layers 3-6, Section 5) statistically supported?
  5. (Future work) Does the effect transfer to another VLLM family and to a question-driven benchmark that uses QaSalFormer's conditioning?
  6. (Clarification) Please name the "ethical board" (Ethical Considerations) when de-anonymised, and confirm the licence covers raw gaze recordings.
1 comment
AuthorsOpenReview ↗

Answer to Reviewer 1CZ2

We agree that shuffled maps are an important control, and we have started running them together with RGVq’s suggestion (Q1). We also thank the reviewer for the bootstrap suggestion: we have run a paired bootstrap over the 500 images, comparing methods over full strength sweeps at matched F1 and recall, and most of the intervals include zero. The “around two points” in Section 6.2 was obtained by comparing the values in the current tables; it was not a formal estimate of uncertainty (Q2). Regarding the prompt, it was chosen on the same 500 images (Q3). Adding another model family and a question-driven benchmark is part of our future work (Q5).

We thank the reviewer for the detailed questions. We want to analyse some of these new results in more detail and revise our claims accordingly; for this reason, we will withdraw the paper.

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

Visual attention in the form of gaze based measurements and predictive models are applied in order to constrain the attention in vision large language models with a goal of limiting hallucinations. The study involves a dataset contribution that comes from a visual question answering annotation task. Evaluation results in the paper examine the correspondence between gaze data, VLLMs and hallucinations, revealing that human gaze agreement doesn't necessarily translate to more effective saliency based attention guidance. Predicted saliency is shown to reduce hallucinations in comparison with a general increase in the focus of attention revealing a potentially important role for saliency in this context and support for the general value of constraining attention in this way.

Critical strengths
  1. The dataset contribution of the work is strong and I think this kind of data could be generally useful for the study of visual saliency, gaze and its interactions with how vision language models behave.
  2. Results on the relationship between saliency/gaze and model behaviours reveal interesting relationships including instances where gaze/saliency is relatively weak in downstream predictions of performance, but also instances where helping to constrain focus via attention can have an important role in reducing hallucinations - this feels like it could be a general result that extends beyond the the current work alone.
  3. The nuances that are shown are especially interesting. These include places where saliency helps and places where it has little impact, but also interaction with cross-head consensus, and the general case presented for the importance of examining all of the issues raised in the paper is relatively strong.
Critical weaknesses
  1. I am going to raise only a single critical weakness with the paper since i am positive overall about this work, but have one central concern: The main thing missing from the work is a convincing study that attention mass is adequately controlled and in particular, whether image token probability mass is preserved between structured and uniform bias conditions since this feels like an important quantity - as the goal is to show that spatial allocation helps beyond general attention allocation. I'd be interested in seeing a high-level examination of different places where attention mass might be considered, and how this is the same or different across conditions. The paper introduces a nice dataset, and the demonstration of reducing hallucinations is clear, so it would be great if this third pillar was adequately supported.
Questions for authors
  1. In a paradigm that involves question answering, is there any particular bias observed in the nature of the data collected that is specific to this type of task and has broader implications for how it would shape any model that uses this type of the data as a basis for its downstream assessments? For example, the could the lack of translation to effective saliency based attention guidance relate to the type of task that is driving the observed data?
  2. For the specific issue of controlling attention mass, are there other definitions of preserved mass that the authors considered? What are those definitions? How does each possibility carry through the rest of the model, and what are implications? i.e. logits aren't normalized, pooled representations may be vectors, and matching the sum of the added logit biases does not guarantee equal total post-softmax probability for associated image tokens -> so the issue of where attention mass is preserved and why it should be preserved at that particular point is important.
1 comment
AuthorsOpenReview ↗

Answer to Reviewer RGVq

In our work, as specified in the manuscript, we compared against adding a constant value: the mean of the QaSalFormer map. As your comment suggests, this does not necessarily guarantee that the attention mass on the image is the same as when we add the saliency map (or even with the method of Jiang et al.). We believe the point you raised is very interesting, and we have run some preliminary analyses with this control, which changes part of our interpretation of the results. We want to look at this in more detail, and for this reason we will now withdraw the paper. We thank you for the time you have dedicated to the paper and for your suggestion.

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

This paper releases an eye-tracking dataset for VQA preference evaluation and a saliency prediction model fine-tuned on VOILA-COCO. The authors compare human gaze with LLaVA attention and test whether predicted human saliency can reduce hallucinations through attention steering. The CHAIR results are encouraging, showing benefits beyond simply increasing visual attention or shortening captions. The external signal appears complementary to internal guidance, without a clear improvement over consensus alone. The main value lies in the resources and empirical observations for the human and machine visual-attention community.

Critical strengths
  • The released gaze corpus and saliency predictor provide useful resources for the human and machine visual-attention community.
  • The analysis usefully distinguishes agreement with human gaze from effectiveness in reducing hallucinations.
  • The CHAIR evaluation includes useful constant-map and caption-length controls, strengthening the evidence that spatial saliency contributes to hallucination reduction.
  • The motivation is clear, and the conclusions appropriately frame external saliency as complementary rather than consistently superior to internal guidance.
Critical weaknesses
  • W1: As the authors acknowledge, the evaluation cannot isolate the benefit of question conditioning. With one question per image and a fixed question in CHAIR, the added value of this aspect of QASalFormer remains unclear.
  • W2: The findings are limited to LLaVA-1.5 at two sizes, with steering evaluated only on CHAIR. Their generality across model families and question-answering tasks remains unclear.
  • W3: The gaze corpus contains only 30 VQA examples and 21 participants, limiting its breadth as an analysis resource and reusable benchmark.
  • W4: The practical case for dedicated gaze collection is unclear. Recruitment effort, completion time, and total collection cost are insufficiently documented to assess the trade-off.
Questions for authors
  1. Can you clarify what the current results establish about question conditioning, separately from image-only saliency?
  2. Do the steering findings extend to a different VLLM family or a question-answering benchmark?
  3. What were the participant completion time and approximate data-collection cost?

Minor comment:

  • Clarify the fixation-to-saliency conversion procedure and provide a reference for the stated standard practice
1 comment
AuthorsOpenReview ↗

Answer to Reviewer EiXM

We thank the reviewer for the time invested in this review. Regarding your questions, on CHAIR, QaSalFormer is conditioned on a fixed, generic question, so our current results do not isolate the effect of question conditioning (Q1). To fully isolate the effect of the question, we would need the results on benchmarks with specific questions. We were not able to implement and run sweeps on more benchmarks or models during the rebuttal (Q2). What we have done is run experiments proposed by Reviewers RGVq and 1CZ2. We want to analyse part of these results in more detail and revise our claims accordingly. For this reason, we will withdraw the paper.

Answering your Q3, each session lasted about one hour, including instructions and calibration. Participants were compensated with 50 EUR each (1,050 EUR in total). Recruitment and data collection took three weeks and involved three researchers. The eye tracker (Tobii Pro Fusion) was already available in our lab.

Finally, regarding your minor comment: in our analysis, we keep only fixations from the image-and-question screen, before either response is shown, and only those that fall inside the image. Following standard practice (e.g., Le Meur & Baccino, 2013), each fixation is an equal-weight impulse; duration is not used. These impulses are convolved with an isotropic Gaussian with sigma = 30 px, and the map is divided by its maximum. The ground-truth map for each image is the unweighted mean of the per-participant maps.

Comments on the paper

1 comment

We thank all reviewers for their constructive reviews. Following their suggestions, in particular the mass-matched control proposed by Reviewer RGVq and the shuffled-map and bootstrap analyses proposed by Reviewer 1CZ2, we have run additional analyses. We want to analyse these results in more detail and revise our claims accordingly, and for this reason we are withdrawing the paper. We are grateful for the reviewers' time and suggestions, which we believe will help to improve this work.

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

Related papers

Discussion (0)

Sign in to comment.