acceptodds
Under review as a conference paper at ICLR 2027

Attending Without Selecting: Visual Key Collapse and Hallucination in Large Vision-Language Models

Abstract

Large vision-language models often describe objects that are absent from the image, a failure usually attributed to insufficient attention to visual input. We find that hallucinated and faithful captions place the same share of attention on the image, but in hallucinated captions that attention depends less on the word being generated. The failure lies in selecting where to look rather than in how much to look. We trace it to a collapse of visual keys onto a few dominant directions, which leads every query to rank image regions by nearly the same criteria. The collapse appears across architectures with different visual interfaces, does not depend on the relative norms of visual and text keys, and is more severe for images that the model describes less faithfully. Building on this diagnosis, we introduce Whitened Attention Steering, a training-free method that suppresses the dominant directions in visual attention scores while keeping the total attention on the image fixed. Hallucination decreases steadily as the correction is strengthened and increases when it is reversed, whereas matched corrections of other directions or other heads have little effect. On LLaVA-1.5, the share of captions containing a hallucinated object falls from 54% to 30%, and across models the gain follows the severity of collapse. These findings indicate that hallucination in current vision-language models stems less from a shortage of visual attention than from a loss of selection within it.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hallucination, Large vision-language models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_C1heOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies hallucination caused by poor selection among visual tokens, even when substantial attention is allocated to the image. It associates this behavior with concentrated visual-key directions and proposes WAS, which attenuates dominant spectral directions, matches score statistics, and applies the correction through an entropy gate. Experiments on LLaVA, InstructBLIP, and Qwen3-VL show reduced hallucination. The appendix provides mechanism controls, sensitivity analyses, and statistical tests.

Critical strengths
  • The distinction between visual attention mass and region selection leads to a concrete, effective intervention.
  • The analysis associates concentrated visual-key spectra with weak query-dependent selection, identifying a geometric property that aggregate attention statistics can overlook. This offers a useful diagnostic perspective beyond simply increasing attention to image tokens.
  • Reverse interventions, matched controls, paired statistical tests, and modest measured overhead strengthen the empirical contribution.
Critical weaknesses
  1. The visual attention budget is only approximately controlled. Matching score means and variances does not preserve visual probability mass under the joint softmax, and sharpening introduces an additional change. A small average difference can conceal substantial changes in individual heads or decoding steps. Exact mass-preserving controls are needed to attribute the improvement specifically to better visual selection.

  2. The theory does not establish the central link from low effective rank to failed query-dependent selection. Low-rank keys can still produce different attention patterns for different queries. The analysis therefore does not identify sufficient conditions under which dominant-direction suppression should improve grounding. The intervention experiments provide useful support, but the theoretical contribution is narrower than a general explanation of the proposed mechanism.

  3. The strongest matched mechanism controls are concentrated on LLaVA. Performance improvements on other architectures support transfer of the method, but do not establish that the same geometric mechanism explains those gains. Repeating the matched direction and head controls on another backbone would strengthen the central claim. Spatial-grounding measurements would also help distinguish failed selection from appropriate attention to multiple regions.

  4. The appendix demonstrates local parameter stability, but the configuration-selection protocol remains insufficiently specified. It is unclear whether the rank, gate, sharpening strength, and layer placement were chosen on held-out images. Since the detailed sensitivity analysis is concentrated on LLaVA/CHAIR, it provides limited evidence about stability across architectures and image distributions.

  5. The statistical comparisons primarily establish improvement over greedy decoding, rather than over the strongest competing method. WAS is not uniformly best across model–metric combinations. Paired comparisons against the strongest baseline would better establish the extent of its advantage.

Questions for authors
  1. Does the hallucination reduction persist when visual attention mass is preserved exactly after correction and blending?
  2. Under what conditions does low effective rank imply reduced query-dependent selection? Can the proposed information-theoretic account predict which heads benefit from correction?
  3. Can the matched direction and head controls be replicated on InstructBLIP or Qwen, with a direct measure of spatial grounding?
  4. Were the configuration and intervention layers selected on held-out data? Does the reported parameter stability extend to another architecture?
8 comments

Thank you for your time and for the constructive comments on the budget control, the theory, the second backbone, the configuration protocol and the baselines; addressing them has strengthened the paper. We address each point below. Changes in the revised PDF are marked in blue; Tables 1–21 keep their original numbers, and new tables start at Table 22. Each table in this response (R1–R7) gives its location in the revised PDF.

W1 / Q1. Does the reduction persist when the visual attention mass is preserved exactly after correction and blending?

W1. The visual attention budget is only approximately controlled. Matching score means and variances does not preserve visual probability mass under the joint softmax, and sharpening introduces an additional change. A small average difference can conceal substantial changes in individual heads or decoding steps. Exact mass-preserving controls are needed to attribute the improvement specifically to better visual selection.

Q1. Does the hallucination reduction persist when visual attention mass is preserved exactly after correction and blending?

A1. Yes, almost unchanged. Thank you for this suggestion, which makes the attribution exact. WAS already keeps the visual log-sum-exp (LSE) close to its original value (average shift 0.06, App. F.3); the control below fixes it at every corrected head and step.

  • Control. After blending, we add the shared offset c=LSE(s)−LSE(sfinal)c = \mathrm{LSE}(\mathbf s) - \mathrm{LSE}(\mathbf s^{\mathrm{final}}) to the visual logits, with the LSE over visual positions.
  • What stays fixed. The visual mass equals its value under the uncorrected logits of that step (max⁡∣ΔA∣<10−5\max\lvert\Delta A\rvert < 10^{-5} in fp32), including the part due to sharpening. The softmax normalizer is unchanged, so the attention on each text token is unchanged as well.
  • What changes. The offset is shared, so the distribution within the image is exactly that of WAS.

Table R1. Exact attention budget, LLaVA-1.5-7B, CHAIR (revised PDF: Table 22).

ConfigurationCHAIR_S ↓CHAIR_I ↓Recall ↑Q ↑
Greedy54.218.572.50.078
WAS30.210.270.70.118
WAS, exact budget30.610.270.90.116
Reverse (λ = −1)59.620.271.30.060
Reverse, exact budget60.221.170.30.059
  • WAS keeps 23.6 of its 24.0-point reduction with the budget fixed exactly (paired difference from WAS +0.4, 95% CI [−1.5, +2.5]), and the reverse correction still raises CHAIR_S, to 60.2.
  • In this variant, the only change at each corrected head and step is the distribution within the image, so the improvement is attributable to within-image selection.

Changes in the revised PDF (blue):

  • Abstract and Sec. 1: the attention on the image is described as approximately fixed, following your comment.
  • Sec. 4.2: one sentence on the exact-budget variant, which keeps it exactly fixed.
  • Sec. 5.3 and Table 4: the exact-budget result (new last row).
  • App. F.3 and App. P: the offset and Table 22.

W2 / Q2. Under what condition does low effective rank reduce query-dependent selection, and can the account predict which heads benefit?

W2. The theory does not establish the central link from low effective rank to failed query-dependent selection. Low-rank keys can still produce different attention patterns for different queries. The analysis therefore does not identify sufficient conditions under which dominant-direction suppression should improve grounding. The intervention experiments provide useful support, but the theoretical contribution is narrower than a general explanation of the proposed mechanism.

Q2. Under what conditions does low effective rank imply reduced query-dependent selection? Can the proposed information-theoretic account predict which heads benefit from correction?

A2. Thank you for this question; answering it has made the theory more complete. The new Proposition 5 (App. F.6) states when collapse lowers query dependence and predicts the gain of each head.

  • Split. Write the centered visual scores at step tt as xt=xt∥+xt⊥\mathbf x_t = \mathbf x_t^{\parallel} + \mathbf x_t^{\perp}, the parts along and outside the dominant directions (Eq. (24)). Because Vr\mathbf V_r consists of right singular vectors of the centered keys, the two parts are orthogonal at all pairs of steps, and the correlation between two steps is exactly (Eq. (25))

ρtt′=ηtηt′ρtt′∥+(1−ηt)(1−ηt′)ρtt′⊥,\rho_{tt'} = \sqrt{\eta_t \eta_{t'}} \rho_{tt'}^{\parallel} + \sqrt{(1-\eta_t)(1-\eta_{t'})} \rho_{tt'}^{\perp},

where ηt\eta_t is the dominant share of the score variance.

  • Condition. For near-uniform attention and a common score scale σ\sigma and share η\eta, a second-order expansion of the JSD (median relative error 7% on gated heads) gives (Eq. (26))

Q≈σ24[1−ηρˉ∥−(1−η)ρˉ⊥].Q \approx \tfrac{\sigma^2}{4}\big[1 - \eta\bar\rho^{\parallel} - (1-\eta)\bar\rho^{\perp}\big].

A larger η\eta lowers Q if and only if ρˉ∥>ρˉ⊥\bar\rho^{\parallel} > \bar\rho^{\perp}, i.e. the dominant part is more stable across generated words than the residual part. This identifies precisely when low-rank keys reduce query dependence, which covers the case you raise.

  • Per-head prediction. Under a full correction (g=1g = 1), WAS removes nearly all the weight of the dominant directions (App. F.2) and sets the score scale to ασ\alpha\sigma, so a head gains (Eq. (27))

ΔQ≈σ24[η(ρˉ∥−ρˉ⊥)+(α2−1)(1−ρˉ⊥)].\Delta Q \approx \tfrac{\sigma^2}{4}\big[\eta(\bar\rho^{\parallel}-\bar\rho^{\perp}) + (\alpha^2-1)(1-\bar\rho^{\perp})\big].

We tested this on all 384 heads in layers 9–20 of LLaVA-1.5: captions are teacher-forced, the measured ΔQ of a head comes from correcting that head alone (g = 1, α = 1.2), and gated heads are those on which the gate fires in more than 10% of steps (Figure 3b).

Table R2. Test of Proposition 5, LLaVA-1.5-7B (revised PDF: Table 32).

QuantityValue
Mean dominant-part correlation across steps, gated heads0.8
Mean residual-part correlation across steps, gated heads0.4
Gated heads satisfying the condition88% (21 of 24)
Spearman correlation, predicted vs. measured ΔQ (384 heads)0.7
Gated heads among the 10% with the largest predicted gain18 of 24 (chance: 2.4)
  • The condition holds on the heads where WAS acts, and the formula ranks heads by their gain in query dependence.
  • This fits the much smaller reduction when the same correction is applied to non-collapsed heads (Table 21, −6.4 against −24.0).
  • That the added query dependence lands on the described objects is measured directly in A3 (Table R4).

Changes in the revised PDF (blue):

  • App. F.6: Proposition 5 with its proof.
  • App. Q.4: the test on 384 heads (Table 32).
  • Sec. 3.3: one sentence stating the condition.

W3 / Q3. Repeat the matched direction and head controls on another backbone, with a direct measure of spatial grounding.

W3. The strongest matched mechanism controls are concentrated on LLaVA. Performance improvements on other architectures support transfer of the method, but do not establish that the same geometric mechanism explains those gains. Repeating the matched direction and head controls on another backbone would strengthen the central claim. Spatial-grounding measurements would also help distinguish failed selection from appropriate attention to multiple regions.

Q3. Can the matched direction and head controls be replicated on InstructBLIP or Qwen, with a direct measure of spatial grounding?

A3. Thank you for this suggestion; both additions strengthen the central claim. We used Qwen3-VL-4B, whose visual tokens correspond to image patches, so that grounding can be measured; the 32 Q-Former tokens of InstructBLIP have no spatial position.

Matched controls, built as in Table 21, with the rank and correction scale of WAS on Qwen3-VL-4B and random directions averaged over five seeds.

Table R3. Matched controls on Qwen3-VL-4B, CHAIR (revised PDF: Table 24).

ConfigurationCHAIR_S ↓Δ from greedy [95% CI]
Greedy26.0–
Reverse (λ = −1)27.4+1.4 [−1.4, +4.4]
Random directions (5 seeds)23.5−2.5 [−5.4, +0.4]
Weakest directions24.0−2.0 [−4.9, +0.9]
Non-collapsed heads23.0−3.0 [−6.0, −0.1]
Sharpening only22.6−3.4 [−6.3, −0.3]
WAS17.8−8.2 [−12.0, −4.4]
  • The strongest control, sharpening alone, reaches 41% of the reduction of WAS (−3.4 of −8.2), close to the 39% on LLaVA-1.5 (−9.4 of −24.0).
  • Every paired comparison with WAS remains significant after Holm correction (p < 0.01).
  • The reverse correction raises CHAIR_S as on LLaVA-1.5, by a smaller amount, in line with the milder collapse of Qwen3-VL-4B (Table 18).

Spatial grounding. At each correctly mentioned object word, we take the heads on which the gate fires in the unmodified teacher-forced pass and use the same heads and words in every setting. The ratio divides their visual attention inside the union of all ground-truth boxes of that category by the share of the image the union covers (1 = random placement). Taking the union counts attention spread over several instances as correct selection, which separates the two cases you mention.

Table R4. Spatial grounding ratio with 95% CI (revised PDF: Table 25).

SettingLLaVA-1.5Qwen3-VL-4B
Greedy1.0 [0.97, 1.03]1.1 [1.07, 1.13]
Sharpening only1.1 [1.07, 1.13]1.1 [1.08, 1.14]
WAS1.4 [1.34, 1.46]1.3 [1.25, 1.35]
Reverse0.9 [0.87, 0.93]1.0 [0.96, 1.04]
  • WAS raises the ratio from 1.0 to 1.4 on LLaVA-1.5 and from 1.1 to 1.3 on Qwen3-VL-4B; sharpening alone moves it much less, so the gain comes from the spectral correction.
  • The reverse correction brings attention to or below the random level.
  • Fewer hallucinations thus go together with more attention on the described objects.

Changes in the revised PDF (blue):

  • App. P: new paragraphs “Second backbone” (Table 24) and “Spatial grounding” (Table 25).
Reviewer_aVMrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In the paper, the author first analyzed one overlooked correlation factor of visual hallucination: visual keys in MLLM might exhibit a collapsing mode in feature space, and leading to hallucination in response. Inspired by this observation, they proposed Whitened Attention Steering (WAS) for hallucination mitigation, a training-free model representation engineering method that re-disperse the key features, to enhance their identifiability from queries. Validated on three MLLMs, their method has performed decently in hallucination mitigation compared with previous methods.

Critical strengths
  1. The visual key collapse analyses provided a novel perspective on the potential cause of visual hallucination, which are empirically validated on 3 MLLMs.

  2. The proposed solution is directly addressing their observation of visual key collapse problem. The three steps in WAS are clean and straightforward to understand and empirically effective on many hallucination benchmarks. The steering overhead is decently low, indicating practicality in real-world applications.

Critical weaknesses

Weakness 1: The mechanistic evidence comes mostly from LLaVA-1.5. Collapse is reported descriptively for three architectures, but the key analyses linking it to hallucination, including the diagnostics in Appendices B–E, are conducted only on LLaVA-1.5. Although the authors acknowledge this in the Limitations section, further analysis on a more recent MLLM (e.g., Qwen3-VL-4B/8B, which the paper already uses) would help further validate the visual key collapse phenomenon and its link to hallucination, especially the Appendix E section (the cause of collapse part).

Weakness 2: It is unclear whether visual key collapse is the cause or a symptom of something upstream. Appendix E, which analyzes the root cause of the collapse, is the part I found to be interesting and potentially inspiring. However, its evidences are somewhat weak:

  • In Table 13, even a random projection of the hidden states separates faithful from hallucinated captions nearly as well as the keys do, already at the input of the language model, i.e., on the visual tokens coming from the visual encoder.
  • In Source 2, the claim is that the key projection amplifies the collapse, but Table 13 does not measure the collapse itself; it measures how well the top-5 concentration separates faithful from hallucinated captions (ROC-AUC). A more direct test might be comparing the collapse measures of Table 1 (DangD_{\text{ang}} and ρlocal\rho_{\text{local}}) between the hidden states and the keys of the same layer. The details of the experiment in Source 2 is also unclear and not easy to understand.
  • In Source 1, the visual encoder outputs are already highly concentrated (Dang=0.14D_{\text{ang}}=0.14), even more than the visual keys of LLaVA-1.5 in Table 1 (Dang=0.195D_{\text{ang}}=0.195), which is to some extent contradictory with Source 2's claim.
  • Other explanations are also not ruled out, for example a few high-norm or sink visual tokens dominating the key spectrum, or a mismatch on the query side rather than the key side.

Minor points.

  • In Figure 2, it might be better to demonstrate the three steps in WAS rather than just an illustrative figure. A pipeline figure would help the readability.
Questions for authors
  1. Could the authors extend the key analyses in Appendices B–E, especially the root-cause analysis in Appendix E, to a more recent MLLM such as Qwen3-VL-4B/8B? Do the link between collapse and hallucination and the proposed sources of collapse still hold there?'

  2. In Table 13, a random projection of the hidden states already separates faithful from hallucinated captions nearly as well as the keys, even at the input of the language model. Does this mean that the collapse is mostly inherited from the visual tokens?

  3. Can the authors rule out some potential causes of the visual key collapse, e.g., a few high-norm or sink visual tokens dominate the key spectrum, or that the problem lies on the query side rather than the key side?

5 comments

Thank you for your time and for the constructive comments, in particular the concrete suggestions on App. E, which we followed and which have enriched the mechanistic analysis. We address each point below. Changes in the revised PDF are marked in blue; Tables 1–21 keep their original numbers, and new tables start at Table 22. Each table in this response (R1–R6) gives its location in the revised PDF.

W1 / Q1. Do the analyses of App. B–E, in particular the sources of collapse, hold on a more recent model such as Qwen3-VL-4B?

W1. The mechanistic evidence comes mostly from LLaVA-1.5. Collapse is reported descriptively for three architectures, but the key analyses linking it to hallucination, including the diagnostics in Appendices B–E, are conducted only on LLaVA-1.5. Although the authors acknowledge this in the Limitations section, further analysis on a more recent MLLM (e.g., Qwen3-VL-4B/8B, which the paper already uses) would help further validate the visual key collapse phenomenon and its link to hallucination, especially the Appendix E section (the cause of collapse part).

Q1. Could the authors extend the key analyses in Appendices B–E, especially the root-cause analysis in Appendix E, to a more recent MLLM such as Qwen3-VL-4B/8B? Do the link between collapse and hallucination and the proposed sources of collapse still hold there?

A1. Thank you for this suggestion. We extended the key analyses to Qwen3-VL-4B (the model of Table 3), with the same 500 MSCOCO images and its greedy captions: the collapse diagnostics of App. B, their link to hallucination from App. C, and the two measurable sources of collapse from App. E. Keys are analyzed per key-value head (grouped-query attention), and all measures are normalized so that they can be compared across token counts.

Table R1. Collapse, its link to hallucination and its sources on two models (revised PDF: Table 26; last two rows from App. E).

AnalysisMeasureLLaVA-1.5-7BQwen3-VL-4B
Collapse (App. B)Top-5 variance share of visual keys, layers 9–200.890.74
Effective rank, visual / text keys6–9 / 28–4212–20 / 25–45
Link (App. C)AUC of allocation A0.520.55
AUC of query dependence Q (lower when hallucinated)0.710.66
AUC of top-5 share (higher when hallucinated)0.720.68
Sources (App. E)Centered top-5 share of the encoder output0.530.47
Top-5 share, matched hidden state → key, layers 9–200.78 → 0.890.66 → 0.74
  • Collapse: the top five directions carry 74% of the visual key variance, and the effective rank of visual keys is 12–20, against 25–45 for text keys.
  • Link to hallucination: allocation hardly separates the two groups (AUC 0.55), while hallucinated captions have lower query dependence (AUC 0.66) and more concentrated keys (AUC 0.68).
  • Sources: the encoder output is already concentrated, and the key is more concentrated than the hidden state it is computed from (A2).
  • Across models: each effect scales with the severity of collapse, as do the gains in Table 18.

We also repeated the matched intervention controls on Qwen3-VL-4B (Table 24) and added a spatial-grounding measure on both models (Table 25).

Changes in the revised PDF (blue):

  • App. Q.1: the Qwen3-VL-4B analyses (Table 26).
  • App. E: the Qwen3-VL-4B values of both sources of collapse.
  • Sec. 3.3: a pointer to App. Q.1.

W2 / Q2–Q3. Is visual key collapse a cause of hallucination or a symptom of something upstream?

W2. It is unclear whether visual key collapse is the cause or a symptom of something upstream. Appendix E, which analyzes the root cause of the collapse, is the part I found to be interesting and potentially inspiring. However, its evidences are somewhat weak:

Q2. In Table 13, a random projection of the hidden states already separates faithful from hallucinated captions nearly as well as the keys, even at the input of the language model. Does this mean that the collapse is mostly inherited from the visual tokens?

Q3. Can the authors rule out some potential causes of the visual key collapse, e.g., a few high-norm or sink visual tokens dominate the key spectrum, or that the problem lies on the query side rather than the key side?

A2. Thank you for this question and for reading App. E so closely; your suggestions made it considerably more direct. In short, collapse originates upstream and acts on hallucination through the keys:

  • Origin. It is already present in the visual tokens produced by the encoder, and the key projection amplifies it (A2(a)–A2(c)).
  • Effect through the keys. Attention ranks image regions by the geometry of the keys (Eq. (1)). WAS changes only the visual attention scores of the corrected heads, without editing the encoder output, the keys or any weights, and:
    • hallucination falls monotonically with the strength of the correction, also with sharpening fixed (Table 23), and rises when the correction is reversed (Table 20);
    • random directions (−4.6), the weakest directions (−3.2), non-collapsed heads (−6.4) and sharpening alone (−9.4) reach at most two fifths of the reduction of WAS (−24.0, Table 21);
    • the reduction is kept with the visual attention mass fixed exactly at every corrected head and step (Table 22).

(a) A random projection of the hidden states already separates the two groups at the input of the language model.

In Table 13, even a random projection of the hidden states separates faithful from hallucinated captions nearly as well as the keys do, already at the input of the language model, i.e., on the visual tokens coming from the visual encoder.

A2(a). Yes, part of the collapse is inherited, in line with Source 1: the concentration is already present when the visual tokens enter the language model, so the matched hidden state reaches an AUC of 0.655 at layer 0. Table 13 also shows a gain specific to the key:

  • at layers 0, 11, 17 and 20, the key improves on the matched hidden state by +0.034 to +0.055 before RoPE and by +0.031 to +0.052 after RoPE;
  • the value improves on it by only +0.006 to +0.008.

So, for Q2, the collapse is inherited from the visual tokens and then amplified by the key projection. A2(b) confirms this with the collapse measures themselves.

(b) Compare the collapse measures of hidden states and keys at the same layer, and describe the experiment of Source 2 more clearly.

In Source 2, the claim is that the key projection amplifies the collapse, but Table 13 does not measure the collapse itself; it measures how well the top-5 concentration separates faithful from hallucinated captions (ROC-AUC). A more direct test might be comparing the collapse measures of Table 1 (DangD_{\mathrm{ang}} and ρlocal\rho_{\mathrm{local}}) between the hidden states and the keys of the same layer. The details of the experiment in Source 2 is also unclear and not easy to understand.

A2(b). Thank you for suggesting this more direct test. We ran it, and App. E now describes the procedure step by step:

  1. At each layer, take the normalized hidden states of the visual tokens that enter attention.
  2. Form four views: the hidden state projected to 128 dimensions (one head) by a random semi-orthogonal matrix, averaged over 32 projections; the value; the key before RoPE; the key after RoPE.
  3. Compute the collapse measures per image and average them over the 500 images (LLaVA-1.5-7B; random 128-d vectors give a top-5 share of 0.08).

Table R2. Top-5 variance share of four views, LLaVA-1.5-7B (revised PDF: Table 27).

ViewLayer 0Layers 9–20Layers 24–31
Hidden state (random projection to 128-d)0.650.780.83
Value0.680.790.83
Key, before RoPE0.770.900.87
Key, after RoPE0.780.890.85

Table R3. D_ang / ρ_local of the same views; lower / higher means more concentrated (revised PDF: Table 28).

ViewLayer 0Layers 9–20Layers 24–31
Hidden state (random projection)0.38 / 0.860.27 / 0.910.26 / 0.92
Value0.37 / 0.860.26 / 0.910.26 / 0.92
Key, before RoPE0.25 / 0.920.15 / 0.970.21 / 0.96
Key, after RoPE0.28 / 0.910.17 / 0.960.24 / 0.95
  • At layer 0 and in layers 9–20, the key is more concentrated than the hidden state on all three measures, by 0.05 to 0.13, while the value stays within 0.03. This matches the condition numbers of the key projections in layers 9–20 (15–45, against 5–12 for query and value).
  • After layer 24 the gap falls to 0.02–0.05, at the depth where the separability gain in Table 13 disappears.
  • On Qwen3-VL-4B, the top-5 share in layers 9–20 rises from 0.66 (matched hidden state) to 0.74 (key after RoPE).

(c) The encoder output has a lower angular dispersion than the keys in Table 1.

In Source 1, the visual encoder outputs are already highly concentrated (Dang=0.14D_{\mathrm{ang}} = 0.14), even more than the visual keys of LLaVA-1.5 in Table 1 (Dang=0.195D_{\mathrm{ang}} = 0.195), which is to some extent contradictory with Source 2's claim.

A2(c). Thank you for raising this, and we apologize for the confusion caused by reporting an uncentered measure in Source 1. The two values are not on a common scale:

  • DangD_{\mathrm{ang}} uses uncentered vectors, so it mainly reflects the mean direction shared by all tokens, which by Eq. (1) sets how much attention the image receives, not where it goes.
  • The two values come from different spaces: a 1024-d encoder output and a 128-d key after RoPE, which rotates keys by position.

Source 1 now uses the centered spectrum of Definition 1, which is what governs selection.

Table R4. Centered top-5 variance share of the encoder output (revised PDF: Table 29).

LLaVA-1.5 (CLIP ViT-L/14)Qwen3-VL-4BRandom (576 vectors, 1024-d)
Top-5 variance share0.530.470.026

On this measure the two sources agree: the encoder output is far more concentrated than random vectors (Source 1), and on the common 128-d scale of A2(b), the key raises the top-5 share over the hidden state of the same layer, from 0.65 to 0.77 at layer 0 and from 0.78 to 0.90 in layers 9–20 (Source 2).

Reviewer_7AUFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies object hallucination in LVLMs and argues that concentrated visual keys reduce query-dependent selection within the image, even when overall visual attention is similar. It proposes Whitened Attention Steering (WAS), a training-free method that suppresses dominant visual-key directions in attention scores, followed by score matching and entropy-gated blending. Evaluation covers CHAIR, POPE, AMBER, and MME across LLaVA, InstructBLIP, and Qwen3-VL. The paper also includes key-geometry analysis, component ablations, reverse steering, and direction/head controls to investigate the proposed mechanism.

Critical strengths
  1. The reported hallucination reductions are substantial. On LLaVA-1.5, CHAIR_S drops from 54.2 to 30.2, while recall decreases by only 1.8 points. Gains also appear on InstructBLIP and Qwen3-VL. This suggests practical value beyond simply making the model mention fewer objects.

  2. The distinction between attention allocation and selection is useful, and the paper makes a serious effort to test it. Reverse steering, random/weak directions, and sharpening-only controls help examine whether the improvement is specific to the proposed intervention. This is a strength of the experimental design.

Critical weaknesses
  1. The central claim of improving selection under a fixed attention budget is not established. Matching the mean and standard deviation of visual logits does not preserve their log-sum-exp, which determines visual attention mass relative to text. This is especially relevant with alpha = 1.2 and the blending step that follows. The similar average allocation reported in Appendix P does not rule out allocation changes at the heads and steps actually being modified. The method may work, but the current evidence does not isolate selection as the cause of the improvement.

  2. Some results supporting the mechanism are inconsistent with the stated setup. In Appendix P, lambda = 0 is described as greedy decoding while score matching and the gate remain active. With alpha = 1.2, this should instead be gated sharpening. Yet Table 20 reports CHAIR_S = 54.2, while sharpening-only in Table 21 gives 44.8. Separately, the top-five explained variance of 88–94% in Appendix B is incompatible with the rank-five relative Frobenius reconstruction error below 0.08 stated in Appendix G; the corresponding optimal error would be about 0.24–0.35 (the corresponding optimal relative Frobenius error). These could be reporting errors, but they need to be reconciled before I can rely on the mechanistic evidence.

Questions for authors
  1. Clarification: What exactly is held fixed in the lambda sweep? Is score matching bypassed at lambda = 0, or is alpha changed? Please explain the difference between the lambda = 0 result in Table 20 and sharpening-only in Table 21.

  2. Clarification: Does the error in Appendix G measure reconstruction of the original centered key matrix, or the approximation error relative to an exact truncated SVD? Please reconcile it with the explained variance reported in Appendix B.

  3. Additional validation relevant to the main claim: Does the hallucination reduction remain when visual attention mass is preserved exactly at each intervened head and step? A shared log-sum-exp offset applied to the visual logits after blending would provide this control without changing the within-image attention distribution.

4 comments

Thank you for your time and for the careful, constructive review. We address each point below. Changes in the revised PDF are marked in blue; Tables 1–21 keep their original numbers, and new tables start at Table 22. Each table in this response (R1–R3) gives its location in the revised PDF.

W1 / Q3. Does the reduction hold when the visual attention mass is fixed exactly at every corrected head and step?

W1. The central claim of improving selection under a fixed attention budget is not established. Matching the mean and standard deviation of visual logits does not preserve their log-sum-exp, which determines visual attention mass relative to text. This is especially relevant with alpha = 1.2 and the blending step that follows. The similar average allocation reported in Appendix P does not rule out allocation changes at the heads and steps actually being modified. The method may work, but the current evidence does not isolate selection as the cause of the improvement.

Q3. Additional validation relevant to the main claim: Does the hallucination reduction remain when visual attention mass is preserved exactly at each intervened head and step? A shared log-sum-exp offset applied to the visual logits after blending would provide this control without changing the within-image attention distribution.

A1. Thank you for proposing this control; it makes the budget argument exact. WAS already keeps the visual log-sum-exp close to its original value (average shift 0.06, App. F.3), and the control fixes it at every corrected head and step. We implemented it as you suggested:

  • Control. After blending, we add c=LSE(s)−LSE(sfinal)c = \mathrm{LSE}(\mathbf s) - \mathrm{LSE}(\mathbf s^{\mathrm{final}}) to all visual logits, with the LSE over visual positions. Text logits are untouched.
  • What stays fixed. At every head and step where WAS acts, the visual attention mass equals its value under the uncorrected logits of that step (max⁡∣ΔA∣<10−5\max\lvert\Delta A\rvert < 10^{-5} in fp32), including the shift from α = 1.2. The softmax normalizer is unchanged, so the attention on each text token is unchanged too.
  • What changes. The offset is shared by all visual positions, so the distribution within the image is exactly that of WAS.

Table R1. Exact attention budget, LLaVA-1.5-7B, CHAIR (revised PDF: Table 22).

ConfigurationCHAIR_S ↓CHAIR_I ↓Recall ↑Q ↑
Greedy54.218.572.50.078
WAS30.210.270.70.118
WAS, exact budget30.610.270.90.116
Reverse (λ = −1)59.620.271.30.060
Reverse, exact budget60.221.170.30.059
  • With the budget fixed exactly, WAS keeps 23.6 of its 24.0-point reduction (paired difference from WAS +0.4, 95% CI [−1.5, +2.5]), and query dependence rises as before (0.078 → 0.116).
  • The reverse correction keeps its effect under the same constraint (60.2).
  • In this variant, the only change at each corrected head and step is the distribution within the image, so the reduction is attributable to selection. A2 further shows that most of it comes from the spectral term rather than from sharpening.

Changes in the revised PDF (blue):

  • Abstract and Sec. 1: the attention on the image is described as approximately fixed, following your comment.
  • Sec. 4.2: one sentence on the exact-budget variant, which keeps it exactly fixed.
  • Sec. 5.3 and Table 4: the exact-budget result (new last row).
  • App. F.3 and App. P: the offset and Table 22.

W2 (first part) / Q1. What does λ scale in Table 20, and why does λ = 0 give 54.2 while sharpening-only gives 44.8?

W2. Some results supporting the mechanism are inconsistent with the stated setup. In Appendix P, lambda = 0 is described as greedy decoding while score matching and the gate remain active. With alpha = 1.2, this should instead be gated sharpening. Yet Table 20 reports CHAIR_S = 54.2, while sharpening-only in Table 21 gives 44.8.

Q1. Clarification: What exactly is held fixed in the lambda sweep? Is score matching bypassed at lambda = 0, or is alpha changed? Please explain the difference between the lambda = 0 result in Table 20 and sharpening-only in Table 21.

A2. Thank you for checking these two entries so carefully. Both numbers are correct, and we apologize for the confusion caused by our description of λ. The two entries remove different parts of WAS:

  • λ in Table 20 scales the whole correction: s(λ)=s+λ(sfinal−s)=s+λg(s⋆−s)\mathbf s^{(\lambda)} = \mathbf s + \lambda(\mathbf s^{\mathrm{final}} - \mathbf s) = \mathbf s + \lambda g(\mathbf s^\star - \mathbf s) by Eq. (8). Score matching is not bypassed and α stays at 1.2; λ multiplies their joint effect. At λ = 0 the scores are the original ones, so this entry is greedy decoding (54.2).
  • Sharpening-only in Table 21 keeps score matching and the gate at full strength and removes only the spectral term of Eq. (5), which gives 44.8.

App. P and the caption of Table 20 now state this explicitly, including that λ < 0 also inverts the sharpening.

Your reading also points to a useful additional control, which we added: scaling only the spectral term with sharpening fixed (App. P, “Spectral term alone”),

s~j(λ)=sj+λdq⊤Vr(Σr−1−Ir)Vr⊤kj,\tilde s_j^{(\lambda)} = s_j + \frac{\lambda}{\sqrt d} \mathbf q^\top \mathbf V_r(\boldsymbol\Sigma_r^{-1} - \mathbf I_r)\mathbf V_r^\top \mathbf k_j,

followed by Eq. (6) with α = 1.2 and Eq. (8). Here λ = 0 is sharpening-only and λ = 1 is WAS.

Table R2. Spectral term alone, LLaVA-1.5-7B, CHAIR (revised PDF: Table 23).

λ−1−0.500.250.50.751
CHAIR_S ↓53.247.444.840.836.233.230.2
CHAIR_I ↓18.716.714.713.812.510.910.2
Q ↑0.0620.0730.0870.0980.1030.1110.118
  • All three quantities are monotonic in λ, and λ = 0 reproduces sharpening-only (44.8).
  • With sharpening fixed, the spectral correction lowers CHAIR_S by 14.6 points, and its reversal raises it by 8.4 points.
  • Without any sharpening (α = 1, Table 14), WAS reduces CHAIR_S by 22.4 points (54.2 → 31.8).

Changes in the revised PDF (blue):

  • App. P and the caption of Table 20: what λ scales.
  • App. P: new paragraph “Spectral term alone” with Table 23.

W2 (second part) / Q2. How does the error below 0.08 in App. G agree with the 88–94% explained variance in App. B?

W2 (continued). Separately, the top-five explained variance of 88–94% in Appendix B is incompatible with the rank-five relative Frobenius reconstruction error below 0.08 stated in Appendix G; the corresponding optimal error would be about 0.24–0.35 (the corresponding optimal relative Frobenius error). These could be reporting errors, but they need to be reconciled before I can rely on the mechanistic evidence.

Q2. Clarification: Does the error in Appendix G measure reconstruction of the original centered key matrix, or the approximation error relative to an exact truncated SVD? Please reconcile it with the explained variance reported in Appendix B.

A3. Thank you for this check. Your second reading is the correct one, and your value of 0.24–0.35 is right. The 0.08 is the error of the randomized decomposition K^5=UrΣrVr⊤\hat K_5 = U_r \Sigma_r V_r^\top (Algorithm 2) relative to the exact rank-5 truncation K5K_5 of the same centered matrix. We apologize for the confusion: App. G gave the formula of the reconstruction error for this quantity. App. G now reports both quantities.

Table R3. Two error measures of the rank-5 decomposition, LLaVA-1.5-7B (revised PDF: App. G).

QuantityValue
Convergence, ∥K^5−K5∥F/∥Kˉv∥F\lVert \hat K_5 - K_5 \rVert_F / \lVert \bar K_v \rVert_Fmedian 0.019, max < 0.08 (layers 9–20, all heads)
Reconstruction, ∥Kˉv−K5∥F/∥Kˉv∥F\lVert \bar K_v - K_5 \rVert_F / \lVert \bar K_v \rVert_F1−EV\sqrt{1-\mathrm{EV}}: 0.24–0.35 in layers 9–16 (App. B); median 0.33 in layers 9–20
  • By Eckart–Young, K5K_5 is the best rank-5 approximation, so its squared relative error is 1−EV1-\mathrm{EV}. This matches App. B and your calculation; App. B and F.1 are unchanged.
  • The randomized result lies within 0.019 (median) of the exact truncation, as expected from the spectral gap σ5/σ6>3\sigma_5/\sigma_6 > 3 (App. B).

Changes in the revised PDF (blue):

  • App. G: the convergence error and the reconstruction error are reported separately.

Summary

  • W1 / Q3, budget. Exact-budget control added: CHAIR_S 30.6 vs. 30.2 for WAS, reverse 60.2 (Table R1 = Table 22); abstract and Sec. 1 updated accordingly.
  • W2 / Q1, λ. λ in Table 20 scales the whole correction, now stated in App. P; spectral-only sweep added, monotonic from 53.2 to 30.2 (Table R2 = Table 23).
  • W2 / Q2, App. G. The 0.08 is the convergence error, now reported next to the reconstruction error of 0.24–0.35, consistent with App. B (Table R3).

Thank you again for the constructive comments; the exact-budget control and the spectral-only sweep you suggested have strengthened the paper. We would be happy to discuss any remaining questions.

AuthorsOpenReview ↗

Brief summary of our responses

Dear Reviewer 7AUF,

Thank you again for the careful review. We understand that this is a busy period, and as the discussion is nearing its end, we would like to follow up with a short summary.

With the suggested exact-budget control, the reduction in CHAIR_S is almost unchanged (30.6, against 30.2 for WAS; Table 22). The entries in Table 20 and App. G are clarified, and a spectral-only sweep is added (Table 23).

We would appreciate any feedback on whether our responses address the concerns, and we are glad to provide further details on any point that remains unclear.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.