acceptodds
Under review as a conference paper at ICLR 2027

When Attention Heads Herd: Spectral Release for Faithful Multimodal Reasoning

10 views

Abstract

Long-form multimodal reasoning improves capability but can let a fluent chain drift from visual evidence. Existing test-time remedies target output distributions, fixed heads, or attention intermediates that stock fused scaled-dot-product attention (SDPA) does not expose. We identify a distinct population-level spectral pattern in residual-stream head contributions. Under spectral herding, temporal innovations remain concentrated around a dominant residual-space mode even as its carrier heads change. The leading spectral share persists through carrier turnover, exceeding a row-energy-preserving permutation null across three models. Yet spectral peaks occur in both factual and hallucinated traces: factual events release them as innovation rises, whereas hallucinated events retain them, producing a failed spectral release. We introduce SPLIT (Spectral Peak Liberation via Instantaneous Turnover), a training-free, single-pass controller that tracks this mode and redistributes gain toward the head bulk with exact reference-energy conservation. We characterize conditions for sketch stability and temporal tracking, proving permutation-equivariant steepest release. Across five models and eight benchmarks, one shared setting improves the underlying models and outperforms five recent baselines across diverse tasks, visual interfaces, and reasoning backbones. Operating post-SDPA and pre-W_O, SPLIT preserves stock fused attention, whereas attention-explicit baselines materialize attention tensors under eager execution.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multimodal reasoning, multimodal large language models, hallucination mitigation, training-free decoding, attention heads, spectral analysis, residual stream, test-time intervention

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_6ic9OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

The authors aim to reduce hallucinations in multimodal reasoning models by addressing cases where attention heads remain focused on a dominant internal pattern instead of adapting to new information. They analyze changes in attention head outputs over time and propose a training-free method that detects persistent concentration and dynamically adjusts head contributions during generation. Experiments on five models and eight benchmarks show that the method reduces hallucinations and improves visual understanding and reasoning performance.

Critical strengths
  1. The paper is clearly written, with a well-motivated problem and a clear explanation of the proposed method.
  2. The paper proposes a training-free method that dynamically adjusts attention head contributions while preserving efficient inference, supported by extensive experiments across multiple models and benchmarks.
Critical weaknesses
  1. Motivation for spectral analysis. The paper identifies interesting spectral patterns in attention-head contributions. However, it remains unclear why spectral analysis is particularly suitable for characterizing hallucination-related dynamics, and how it differs from simpler head-level statistics or coordination measures.
  2. Insufficient evidence for the proposed mechanism. The authors associate persistent spectral concentration with hallucinations, but the experiments mainly establish correlations rather than causality. Moreover, changes in head outputs do not necessarily reflect new visual evidence, making it unclear whether the proposed signal truly captures failures to incorporate visual information.
Questions for authors

Regarding Weakness 1: Could the authors further clarify the novelty of this work compared with existing studies on attention dynamics, spectral analysis, and dynamic head adjustment (e.g., RAL, FLASH, and AHEAD)? In particular, what new insights or methodological advances does SPLIT offer beyond these approaches?

Regarding Weaknesses 2: Can the authors provide stronger evidence that persistent spectral concentration directly contributes to hallucinations rather than merely correlating with them? In particular, can the improvements be distinguished from those achieved by simpler head reweighting strategies?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6ic9 on Spectral Analysis and Intervention Controls

Thank you for asking what the spectrum contributes and whether simpler head reweighting can explain the gains. We added a concrete example and tested the gain assignment with the starting prefix held fixed.

Q1. What the spectrum adds. The leading share measures how much normalized innovation energy one residual direction explains. For orthonormal e₁ and e₂, the row sets (e₁, e₁, −e₁, −e₁) and (e₁, −e₁, e₂, −e₂) have identical row energies and mean signed pairwise cosine. Their leading shares are 1 and 1/2, respectively. The trigger combines retained concentration with high innovation, and the leading eigenvector identifies which heads carry that mode (Sections 3.2–3.3).

RAL rescales contributions of perception- and reasoning-oriented heads across layers. FLASH selects vision heads using a Spectral Vortex Score and applies spectral modulation. AHEAD adjusts head contributions using visual-grounding and language-prior potentials. SPLIT derives both the trigger and the carrier assignment from the evolving joint geometry of head innovations in shared residual coordinates. The gain acts after native SDPA under a reference-energy constraint. Section 2 includes the comparison.

Q2. Does the carrier assignment affect the answer? At an eligible gate, all branches start from the same prefix and KV state. The edited branches apply one lagged edit and continue to the final answer. Questions without an eligible gate use the no-edit continuation in all branches. The two shuffle controls are calibrated to reference gain RMS or the realized pre-W_O edit norm. Revised Table 12 reports correct final answers out of 120 balanced POPEv2 questions per model.

ModelNo editSPLITReference-budget shuffleCurrent edit-norm shuffle
R1-OneVision901008785
InternVL3.5981049593

SPLIT improves accuracy by 12.50 and 9.17 points over the shuffle calibrated to the current edit norm. In full generation, it also improves both CHAIR hallucination rates and both POPEv2 metrics over token-varying Shuffled, Random, and Energy assignments on both backbones. We added the full-generation results to the revised PDF as Table 14.

ModelAssignmentCHAIR C_S ↓CHAIR C_I ↓Object F1 ↑POPEv2 Acc ↑POPEv2 F1 ↑
R1-OneVisionVanilla38.410.074.580.682.00
R1-OneVisionSPLIT29.87.474.983.284.27
R1-OneVisionShuffled34.89.373.680.281.73
R1-OneVisionRandom34.49.374.781.082.37
R1-OneVisionEnergy35.69.774.480.281.50
InternVL3.5Vanilla46.49.676.281.483.63
InternVL3.5SPLIT38.67.776.584.586.22
InternVL3.5Shuffled42.48.977.082.484.48
InternVL3.5Random44.69.376.482.484.45
InternVL3.5Energy45.48.976.882.084.01

Does the answer respond to the image? We use the same 120 original/target-removed POPEv2 image pairs for both backbones. Within a pair we hold the question fixed and replay the same generated prefix where available, with separate image-conditioned KV states. A correct flip means Yes on the original and No after removal. We added the paired results to the revised PDF as Table 13.

ModelImage pairsNo editSPLITReference-budget shuffleCurrent edit-norm shuffle
R1-OneVisionAll 120, shared prefix72757168
R1-OneVision113 matched in both views68727065
InternVL3.5All 12078847575
InternVL3.5111 with shared prefix73797070
InternVL3.549 matched in both views31372829

In the subsets with realized edit norms matched in both views, SPLIT yields 7 and 8 more correct flips than the corresponding shuffle. The fixed-prefix comparison tests how gain assignment affects final answers at matched edit size. Changing the image tests whether those answers respond to the target's presence. Section 3.3 calls the head-output signal internal innovation, and Figure 4 reports its observational relationship with spectral concentration and hallucination.

AuthorsOpenReview ↗

Thank you for prompting a closer look at the spectral intervention

Thank you again for asking how the effect of spectral carrier assignment can be separated from that of simpler reweighting. Tables 12–13 report the edit-norm-matched shuffle comparison and the paired target-removal test. If you have any remaining questions about these controls or their interpretation, we would be happy to discuss them.

The authors have substantially strengthened the paper through additional analyses and controlled interventions. The revised experiments provide convincing evidence for the existence of population-level spectral coordination, its distinction from fixed-head effects and unequal head energy, and the effectiveness of the proposed carrier assignment strategy. I therefore consider most of my original soundness concerns adequately addressed.

Nevertheless, my remaining concern is primarily about the degree of conceptual novelty. While failed spectral release provides an interesting characterization of head-population dynamics associated with hallucination, the proposed method ultimately implements a spectral-guided dynamic head reweighting strategy. The spectral analysis offers a more structured criterion for identifying and adjusting head contributions, but its role as a fundamentally new mechanistic insight remains less compelling. In particular, the empirical identification of a new failure signature and the effectiveness of an intervention based on that signature do not necessarily establish a substantially new understanding of multimodal hallucination.

Overall, I appreciate the authors' substantial efforts during the rebuttal and acknowledge that the paper is technically stronger. However, I remain somewhat unconvinced that the conceptual and methodological advances are sufficiently substantial. I have increased my score from 2 to 3 to acknowledge the substantial improvements made during the rebuttal, although my assessment remains closer to the midpoint (2.5), given the remaining concerns about the paper's novelty.

AuthorsOpenReview ↗

Thank you for the follow-up on SPLIT’s conceptual contribution

Thank you for revisiting the paper and for clarifying your remaining concern about novelty. RAL rebalances heads according to their perception or reasoning roles, FLASH uses a spectral score to identify vision heads for attention modulation, and AHEAD adjusts head contributions based on visual grounding and language prior potentials. SPLIT studies how a shared direction in the residual stream evolves across heads during decoding.

Figure 2 shows that this direction can persist even as the heads carrying it change. Figure 4 shows that when internal innovation rises, matched factual events tend to disperse spectral concentration more strongly than hallucinated events. Together, these observations characterize hallucination-related drift in terms of how the head population responds to internal change. In SPLIT, this temporal pattern determines when to intervene, and the current mode determines which heads to reweight. The edit-norm-matched shuffle in Table 12 tests the latter choice and shows that using the current carrier assignment improves final answers at a comparable edit size.

We would welcome any further thoughts on whether this distinction addresses your remaining concern about the paper’s conceptual contribution.

Reviewer_6zYiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies hallucination in long-form multimodal reasoning and presents SPLIT, a training-free controller driven by a population-level spectral pattern in residual-stream head contributions. The authors find that the heads' innovations remain concentrated around a dominant mode as the carrier heads turn over, and the hallucinated events retain this concentration when innovation rises while factual events release it. SPLIT redistributes the gain from the current carriers to the remaining heads to alleviate hallucination.

Critical strengths
  1. The experiments over 5 models and 8 benchmarks are extensive.
  2. The method achieves consistent performance improvements compared with the baselines.
  3. The method is training-free and plug-and-play without introducing much computational overhead.
Critical weaknesses
  1. In Figure 3, the row-energy-preserving permutation shuffles the coordinates of each head's innovation independently, turning each head's innovation into a random direction. Exceeding it only shows that the innovations are not random. It is unclear whether the excess comes from instantaneous coordination or from WOW_{O} simply projecting the innovations into similar directions regardless of the input.
  2. The functional role the carrier heads play in the reasoning process is not analyzed. Do they have properties that potentially harm the reasoning? If so, could the authors provide some evidence, for example, their distinct attention patterns compared with the other heads?
  3. Besides ρ\rho, the method fixes many design choices without much explanation or ablation. For example, the sketch width ss, the history capacity HH, the thresholds for release and innovation (Equations 10-12).
Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6zYi on the Donor Null, Carrier Attention, and Sensitivity

Thank you for suggesting the stronger null and asking about the carriers and the detector settings.

Q1. Head-specific geometry and the donor null. Each head independently draws a donor from another image, retaining its identity and residual coordinates while breaking cross-head alignment on the same input. We match layer, decoding-position bin, and token class, then rescale to the observed row energies and recompute the centered statistic. Revised Table 10 reports the observed minus null leading share for 100 images per model. The layer averages are positive on both models, with image-bootstrap intervals above zero. R1-OneVision shows an excess at all three layers, and InternVL3.5 shows an excess at layer 18.

ModelFirst preset layerSecond preset layerThird preset layerLayer-average [95% image-bootstrap CI]
R1-OneVision+0.00636+0.00974+0.01618+0.01076 [+0.00948, +0.01205]
InternVL3.5−0.00396+0.01571−0.00688+0.00163 [+0.00098, +0.00227]

Q2. Carrier attention. We match heads in the top quartile of carrier loadings with bulk heads of similar energy on the same prefix. At a maximum energy ratio of 2, InternVL3.5 carriers allocate 6.55 percentage points less attention to visual tokens (95% image-bootstrap interval [−9.99, −3.40], 160 pairs across 91 images). The direction is consistent at ratios of 1.5 and 4. R1-OneVision shows a +3.26-point difference at ratio 2 (66 pairs, 46 images) and a near-zero difference at ratio 4. The visual-attention profile depends on the model and matching range (Appendix F.2). To test functional relevance, we also compare the current carrier assignment with an edit-norm-calibrated shuffle at the same prefix. The spectral assignment yields more correct final answers on both backbones (Appendix F.3, Table 12).

Q3. Sketch width, history, and thresholds. The default width min(d_model, max(8, 4H)) uses the head count H as a rank bound. Keeping the last H valid observations bounds memory and lets calibration follow recent decoding. The release and innovation thresholds use each layer's own recent medians and robust scales. We vary one setting at a time at fixed strength on R1-OneVision, using the same 250 CHAIR images and 500 balanced POPEv2 questions. The default reduces CHAIR C_S by 7.2 points and improves POPEv2 accuracy by 1.6 points over Vanilla. Halving the release multiplier reduces accuracy by 2.8 points relative to the default. Increasing the innovation multiplier from 0.5 to 2 lowers the activation rate from 34.89% to 16.75% of monitored POPEv2 layer-token positions, with accuracy changing from 81.4% to 81.8%. Revised Table 15 reports the complete sweep.

ConfigurationCHAIR C_S ↓CHAIR C_I ↓Object F1 ↑POPEv2 Acc ↑POPEv2 F1 ↑
Vanilla36.49.274.881.682.58
SPLIT default29.27.575.583.284.27
Sketch 2H37.29.475.782.283.24
Sketch 8H35.68.476.181.482.29
History H/236.07.875.782.483.21
History 2H34.49.475.581.082.31
Release ×0.537.69.674.780.481.23
Release ×237.28.874.982.283.30
Innovation ×0.535.28.675.081.482.22
Innovation ×234.88.675.481.882.73

Thanks for the author's rebuttal. The new analyses resolve most of my concerns, and I think they make the paper stronger. I remain positive and will keep my score.

AuthorsOpenReview ↗

Thank you for your constructive feedback and continued support

Thank you for taking the time to review our rebuttal and for your continued positive assessment. We are glad that the additional analyses resolved most of your concerns and strengthened the paper. Your suggestions on the donor null, carrier analysis, and sensitivity tests helped us improve the rigor and clarity of the revised manuscript. We sincerely appreciate your thoughtful engagement throughout the discussion.

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

This paper investigates multimodal hallucination through the collective behavior of attention heads. The authors construct a spectral measure from centered head contributions and observe that hallucinated trajectories exhibit weaker release of spectral concentration when internal innovation increases. They propose SPLIT, which tracks the dominant mode and adjusts head gains at the next decoding step. The method requires no retraining or additional decoding branch and preserves compatibility with fused attention. Experiments cover five backbones and eight benchmarks.

Critical strengths
  1. The focus on a changing population of heads is interesting. Figure 2 provides a useful reason to move beyond assigning fixed roles to individual heads: a collective mode can persist while its main contributors change. The authors also distinguish spectral concentration from hallucination itself, since concentration appears in factual traces as well. This makes the motivation for studying temporal changes more convincing.

  2. The method addresses a practical deployment constraint. It operates on a frozen model, uses a single decoding branch, and intervenes after fused attention without materializing attention weights. The evaluation is substantial, with a shared intervention strength across five backbones and eight benchmarks. I also appreciate the inclusion of object coverage alongside hallucination metrics and the static-mode, raw-state, and shuffled-participation controls in Table 2.

Critical weaknesses
  1. I am not yet convinced that the proposed statistic and intervention establish the claimed spectral-release mechanism. In Eq. (7), subtracting the instantaneous mean across heads removes any projected innovation shared equally by all heads. In the extreme case where all projected innovations are identical, the centered operator becomes zero. This is a reasonable construction for studying differences around the population mean, but the paper does not establish why the discarded common component is unimportant for hallucination. The “Raw state” ablation in Table 2 examines temporal centering; it does not isolate this separate choice of centering across heads.

  2. There is a related gap in the control argument. Appendix D.7 establishes a descent direction for the exposure surrogate (J(g)), and D.9 extends this argument under bounded drift. These results do not establish a decrease in the normalized leading spectral share after the intervention. Figure 4 supports an association, but I would like more direct evidence that the proposed gain update changes the spectral quantity in the intended way and that this change explains the improvement in factuality.

  3. I could not reconcile some POPEv2 results in Table 6 with the metric definitions in Appendix C. If accuracy, F1, and PBO are computed from the same pooled confusion matrix, their combination determines the fraction of ground-truth positive examples. For Vision-R1, the reported values imply approximately 50.0% positives for Vanilla but 53.4% for SPLIT. For Qwen3-VL-Thinking, they imply approximately 50.0% and 47.2%, respectively. This fraction should not change between methods evaluated on the same records, and the differences exceed what the displayed rounding can explain. Separate averaging across subsets or a different treatment of invalid answers could account for this, but that procedure is not clear from the stated definitions. This needs clarification before I can confidently interpret the POPEv2 comparison.

Questions for authors
  1. What motivates removing the instantaneous mean across heads? Could you isolate this operation while retaining temporal centering, projection, and row normalization, and report its effect on both the diagnostic signal and downstream performance? Does the discarded mean component itself carry information about hallucination?

  2. On identical decoding prefixes, does applying the proposed lagged gain update reduce the recomputed leading spectral share? A comparison with shuffled carrier assignments at matched intervention strength would help distinguish the proposed mechanism from a more general effect of perturbing head contributions.

  3. Could you provide the TP, TN, FP, and FN counts underlying Table 6, together with the aggregation procedure and treatment of unparseable answers, to explain the discrepancies above?

3 comments
AuthorsOpenReview ↗

Response to Reviewer gjCG on Centering, Carrier Exposure, and POPEv2

Thank you for asking about head centering and matched controls, and for identifying the POPEv2 discrepancy.

Q1. Cross-head centering and the common component. Centering makes the detector invariant to a shift shared across heads in the projected innovations. The gain still acts on the native head outputs. We varied only cross-head centering, holding temporal centering, projection, and row normalization fixed. Centering lowers both CHAIR hallucination rates and improves POPEv2 accuracy by 1.7 percentage points on each backbone (Table 8 of the revised PDF).

ModelVariantCHAIR C_S ↓CHAIR C_I ↓Object F1 ↑POPEv2 Acc ↑POPEv2 F1 ↑
R1-OneVisionSPLIT29.87.474.983.284.27
R1-OneVisionNo head centering32.09.074.581.582.82
InternVL3.5SPLIT38.67.776.584.586.22
InternVL3.5No head centering44.09.376.482.884.72

On the same 300 CHAIR images for both models, we paired factual and hallucinated object mentions at nearby positions within a caption. At hallucinated mentions, the common component removed by centering has higher absolute energy and accounts for a slightly smaller fraction of total projected innovation energy. Table 9 of the revised PDF reports this comparison.

ModelPairsCommon energy, factual → hallucinatedEnergy fraction, factual → hallucinatedLeading-share gap, uncentered → centered (pp)
R1-OneVision810.483 → 0.6043.46% → 3.26%0.89 → 1.49
InternVL3.51380.791 → 0.8572.48% → 2.38%0.53 → 0.69

The gap is the mean leading share at hallucinated mentions minus that at factual mentions, averaged across the three preset layers. Centering widens this average gap on both backbones. Appendix F.1 reports how the effect varies by layer.

Q2. Effect of the lagged update on identical prefixes. SPLIT alone lowers mean carrier exposure on both models. Its recomputed leading share rises. We replayed 300 trigger positions from 100 CHAIR images per model with the same starting prefix and KV state. All edited branches share the same reference energy and reference-energy-weighted gain RMS. Revised Table 11 reports changes relative to no edit.

ModelReadoutSPLITShuffledRandomEnergy
R1-OneVision10⁴ × ΔJ−1.330+0.790+0.914+1.455
R1-OneVisionΔc+0.006002−0.000240−0.000223−0.002322
InternVL3.510⁴ × ΔJ−1.874+0.734+0.856+0.378
InternVL3.5Δc+0.006929+0.001837+0.001492−0.003363

Figure 4 tracks share along natural trajectories. The replay recomputes share after centering and normalizing the edited outputs. Exposure evaluates the gain using unedited current-token energies and loadings, so the two changes can differ (Sections 3.3–3.4 and Appendix D.7). In a separate continuation test, SPLIT yields more correct POPEv2 answers than a shuffle calibrated to the realized edit norm (Table 12).

Q3. POPEv2 counts, aggregation, and answer extraction. The earlier parser missed some explicit yes/no answers. The original aggregation excluded those records from the four SPLIT rows below, changing the label balance of the evaluated subsets. We reapplied the extractor used for ablations and sensitivity tests, which handles \boxed{Yes}, boxed[Yes], and a final No after explanatory text. We manually checked the recovered decisions.

ModelOriginal valid NOriginal TP/TN/FP/FNRecovered TP/TN/FP/FNCorrected TP/TN/FP/FN
Vision-R1931452 / 320 / 115 / 444 / 55 / 10 / 0456 / 375 / 125 / 44
ThinkLite-VL965446 / 433 / 65 / 2130 / 2 / 0 / 3476 / 435 / 65 / 24
Qwen3-VL-Thinking893406 / 387 / 85 / 1568 / 28 / 0 / 11474 / 415 / 85 / 26
InternVL3.5952444 / 360 / 140 / 841 / 0 / 0 / 7485 / 360 / 140 / 15

For Vision-R1, 772/931 reproduces the original accuracy of 82.9%. The old positive fractions were 496/931 for Vision-R1 and 421/893 for Qwen3-VL-Thinking. Both return to 500/1,000 after recovery. All methods now use the same 1,000 questions. Accuracy, positive-class F1, PBO, and TNR come from one pooled matrix per model and method. On this balanced set, PBO = Accuracy − TNR in percentage points. Appendix C.3 documents the procedure. Revised Tables 1 and 6 include the updated baseline entries, and Table 7 lists all 35 confusion matrices. SPLIT leads all evaluated baselines in POPEv2 accuracy and F1 on each of the five backbones, by 1.22 and 1.00 percentage points on average over the strongest baseline per backbone.

We checked all reported benchmark tables for numerical consistency. In Table 1, we aligned the AMBER composite score for ThinkLite-VL/FLASH with Table 6 and recalculated two MIRAGE accuracies from the confirmed counts of correct answers out of 1,329 questions.

AuthorsOpenReview ↗

Thank you for suggesting the matched-prefix comparison

Thank you again for suggesting the matched-prefix experiment. Table 11 of the revised manuscript reports how carrier exposure and the leading spectral share change under the same reference gain budget. We also appreciate your identifying the POPEv2 inconsistency. Tables 6 and 7 provide the corrected metrics and the full set of confusion matrices. If you have any remaining questions about either point, we would be happy to discuss them.

AuthorsOpenReview ↗

A brief follow-up on your helpful suggestions before discussion closes

Dear Reviewer gjCG,

Thank you again for your thoughtful suggestions on the experimental controls and evaluation reporting. They helped us strengthen both the evidence and its presentation.

As the author discussion period draws to a close, we would greatly appreciate hearing whether our response and the revised manuscript address your concerns, and whether any points would benefit from further clarification.

For ease of reference, Table 11 reports the matched-prefix comparison, and Table 12 provides the additional final-answer comparison against the edit-norm-calibrated shuffle. For POPEv2, Tables 6–7 provide the corrected metrics and all 35 confusion matrices, with the answer-extraction and aggregation procedures documented in Appendix C.3.

We would welcome the opportunity to address any remaining questions before the discussion closes. Thank you again for your time and for helping us improve the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Author Summary of Contributions and Discussion

Dear Area Chair and Reviewers,

Thank you for the constructive discussion and for recognizing the broad evaluation and practical value of efficient, training-free inference.

Core insight and significance. A dominant residual-stream direction can persist even as its carrier heads change (Figure 2). At position-matched mentions, factual events release spectral concentration more strongly as internal innovation rises, while hallucinated events retain it (Figure 4). Together, these findings give a temporal, population-level account of hallucination-related drift and motivate tracking both the mode and its current carriers.

What is new in SPLIT. RAL uses perception/reasoning roles, FLASH uses vision-head spectral scores, and AHEAD uses grounding/prior potentials. SPLIT derives both intervention timing and carrier assignment from evolving head innovations in shared residual coordinates. This connects the temporal failure signature to the heads reweighted at each step. Its gain direction is derived as steepest descent of carrier exposure under a reference-energy constraint, with guarantees under the stated assumptions. The controller needs no training or additional decoding branch and preserves fused attention. One shared strength delivers broad gains against five recent baselines across five backbones and eight benchmarks (Table 1).

Evidence for the intervention's specific effect (gjCG, 6ic9). Replaying 300 positions per model with prefixes held fixed, only SPLIT lowers mean current carrier exposure on both models under a shared reference gain budget (Table 11). This is the objective analyzed by the theory; Table 11 separately reports the recomputed, normalized leading share. In a continuation test, SPLIT answers 100/120 and 104/120 questions correctly, versus 85/120 and 93/120 for a shuffle calibrated to the realized edit norm on R1-OneVision and InternVL3.5 (Table 12). Image-pair controls also yield more correct Yes-to-No flips at matched edit norms (Table 13). These tests connect carrier assignment to answer quality and image responsiveness; Tables 14–15 add full-generation controls and sensitivity.

Why the diagnostic is informative (gjCG, 6zYi). The spectrum distinguishes configurations with identical head energies and mean signed pairwise cosine. Under a donor null preserving head-specific geometry, layer-averaged spectral excess remains positive on both tested models (Table 10), supporting same-input coordination beyond that null. Isolating cross-head centering lowers CHAIR sentence-level hallucination rate from 32.0% to 29.8% on R1-OneVision and 44.0% to 38.6% on InternVL3.5, and improves POPEv2 accuracy by 1.7 percentage points on each (Table 8). We measured the removed common component directly; centering widens the mean hallucinated–factual spectral gap on both models (Table 9). Centering defines the detector, while gains act on uncentered native outputs.

Discussion outcome. 6zYi confirmed that the new analyses resolve most concerns and strengthen the paper, maintaining their positive assessment. 6ic9 also leaned positive, finding the new evidence for population coordination and carrier assignment convincing and most original soundness concerns addressed. gjCG has not yet posted follow-up feedback; we provide the corresponding evidence for the AC's assessment.

Evaluation updates (gjCG). We corrected the POPEv2 parsing/aggregation issue using the shared parser and manual checks of recovered answers, restoring the same balanced 1,000 questions for all methods. Tables 6–7 and Appendix C.3 provide corrected metrics, all 35 confusion matrices, and aggregation details; SPLIT leads all evaluated baselines in accuracy and F1 on all five backbones. We also checked numerical consistency across benchmark tables and documented the AMBER/MIRAGE corrections.

All additions are included in the revised PDF. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.