acceptodds
Under review as a conference paper at ICLR 2027

Attention Sinks as Knowledge-Conflict Arbiters: Training-Free Gating-Preserving Routing for Contextual Faithfulness

Abstract

Large language models can generate unsupported answers even when the provided context contains sufficient evidence. Improving contextual faithfulness requires distinguishing answer-determining evidence from distracting information, rather than simply increasing attention to context. Attention sinks complicate this distinction: modifying their weights changes the attention budget available to content, while modifying their values changes the information they contribute. Our interventions show that both masking sink attention and replacing sink values with evidence-derived values tend to increase hallucination rather than improve grounding. These observations motivate a mechanistic account: sink gating is a decision variable the model actively modulates under answer-level conflict, and the sink value it weights carries a suppressed, prior-aligned direction. Building on this account, we propose AGPR, a training-free inference-time intervention that reads the model's own gating deviations as a conflict signal, requires layer-level consensus before acting, and applies a bounded correction along the direction the model has already chosen. Every routing correction is then projected onto a gating-neutral subspace via an exact normalization compensation, preserving the arbitration budget. Across four benchmarks, three model families, and multiple scales, AGPR reduces hallucination and improves grounded success while preserving general capability with negligible inference overhead.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Hallucination Mitigationm, Attention Sinks, Mechanistic Interpretability

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_W92uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the role of attention sinks in resolving conflicts between parametric knowledge and contextual evidence. The authors argue that sink attention acts as an implicit gating mechanism and propose Adaptive Gating-Preserving Routing (AGPR), a training-free inference-time method that detects potential knowledge conflicts through sink gating deviations and adjusts attention routing while preserving a target sink attention budget. Experiments across multiple model families and benchmarks demonstrate improvements in contextual faithfulness and hallucination reduction with limited inference overhead.

Critical strengths
  1. The paper investigates the relationship between attention sinks, contextual evidence, and parametric knowledge. The analysis of sink gating behavior under knowledge conflicts, together with causal interventions and value-space analysis, provides potentially useful insights into the internal mechanisms of LLMs.

  2. The authors evaluate AGPR across multiple model architectures, scales, and contextual faithfulness benchmarks. The reported improvements over existing baselines are encouraging, particularly considering that the method requires no additional training and introduces relatively small inference overhead. The component-wise ablation studies also provide useful evidence for the effectiveness of the proposed design.

Critical weaknesses
  1. The central motivation and several mechanistic observations substantially overlap with prior work. For example, Attention Sink Forges Native MoE in Attention Layers (Fu et al., 2026) has already investigated the implicit gating role of attention sinks, while Attention Sinks as Internal Signals for Hallucination Detection in Large Language Models (Binkowski et al., 2026) connects attention sink behavior with hallucination detection. Furthermore, previous work on attention calibration (Yu et al., ICML 2024) has explored training-free interventions on attention sinks. Although AGPR introduces gating-preserving routing, its main components appear to be relatively straightforward extensions of existing ideas, and the paper does not sufficiently establish a substantial methodological advance over these closely related studies.

  2. The paper interprets changes in sink attention as evidence that attention sinks act as knowledge-conflict arbiters. However, changes in sink gating may reflect general variations in attention allocation or contextual salience rather than an explicit mechanism for resolving conflicts between parametric and contextual knowledge. Although the authors provide causal interventions, these experiments primarily demonstrate that modifying sink attention affects answer preferences, which does not necessarily establish the claimed semantic arbitration mechanism. More rigorous controls are needed to distinguish knowledge-conflict-specific signals from general attention dynamics.

  3. Several components of AGPR, including head-specific calibration, threshold-based conflict detection, layer-level consensus, and bounded gating correction, rely on relatively simple heuristic designs. The exact gating compensation is mathematically straightforward given the softmax normalization constraint. Moreover, the mechanistic analysis primarily focuses on layers 12–14, while different model architectures may exhibit different layer-wise attention sink behaviors. The paper does not sufficiently demonstrate that the proposed mechanism and intervention strategy generalize beyond the selected layers and calibration settings.

Questions for authors

See Weakness

3 comments
AuthorsOpenReview ↗

Response to Weakness 1 & Weakness 2.

Weakness 1

Thank you for highlighting these closely related studies, we have carefully read each of them and address them below.

  1. Attention Sink Forges Native MoE in Attention Layers (Fu et al., 2026): This paper explain implicit sink gating and introduce training objective for head utilization;
  2. Attention Sinks as Internal Signals for Hallucination Detection in Large Language Models (Binkowski et al., 2026): This paper use sink statistics to detect hallucinations
  3. Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (Yu et al., ICML 2024): This paper reduce selected sink weights and redistributes the released mass proportionally, preserving relative non-sink attention

Our conceptual contribution is to treat the model's native gating reponse as a reference for intervention and explicitly constrain how content reweighting may alter it. The key observation is that improving relative attention to relevant tokens can simultaneously change the sink budget through the shared softmax denominator. Fu et al. explain implicit gating but do not develop a routing intervention that explicitly preserves a selected gating budget. Similarly, Binkowski et al. doesn't specifying how to modify attention while controlling this coupling. Yu et al. doesn't address how to change those relative weights while independently controlling the sink budget. In contrast, AGPR makes budget adjustment an explicit, bounded decision and constrains subsequent routing changes to preserve that target exactly. So our distinctive contribution is independent control of the gating budget and content selection, preventing routing from unitentionally overriding the selected budget.

We also conducted a small-scale comparsion with these methods. Table W92u.1: Additional comparsions on Qwen2.5-7B-Instruct; SinkProbe is marked N/A because it detects hallucinations without modifying response.

MethodConFiQA Acc. ↑RAGTruth HR ↓RAGTruth GSR ↑
Fu et al.83.116.574.9
Binkowski et al.N/AN/AN/A
Yu et al.80.917.873.2
AGPR86.414.676.8

Weakness 2

Thank you for the suggestions, we understand your concern involve two aspects: 1) whether native sink changes reflect prior-context conflict rather than general contextual salience; 2) whether their effect on answers goes beyond rescaling non-sink contribution. We add two complementary controls to distinguish these explanations.

First, we independently verify original-answer preferences using three context-free paraphrases and compare matched context edits across verified and unverified prior groups. We match baseline sink mass, question type, edit length and evidence position and check non-sink attention entropy and evidence attention to assess general allocation differences. Table W92u.2: Native sink-mass decrease.

Context editVerified priorUnverified priorDifference
Answer-changing evidence edit4.62.22.4
Answer-preserving distractor edit0.30.20.1
Meaning-preserving paraphrase0.10.2−0.1

Second, we compare Full gating with a Content-only control applying exactly the same non-sink perturbation, holding input, value vectors and within-group routing fixed. We use a 25% relative sink-mass decrease and match baseline sink mass across groups. Table W92u.3: Change in context-versus-original answer margin

InterventionVerified priorUnverified prior
Content-only+0.52+0.37
Full gating+0.70+0.40
Full gating − Content-only+0.18+0.03
The additional native response under prior conflict in Table W92u.2 and the larger effect beyond identical content scaling in Table W92u.3 will jointly support a role for sink gating in prior-context arbitration beyond the tested general-attention explanations.

Actually, your concern is also shared by Reviewer p7DD (W1) and WiGh(W1), we address these complementary aspect in our responses to them. For addition details, we kindly refer you to our response to reviewer p7DD's W1.

AuthorsOpenReview ↗

Response to Weakness 3.

Thank you for raising this concern. We also break down this opinion into two parts to respond to it.

(1) The design and usage of components: These components are designed to implement a common principle, with each addressing a specific problem: head-specific calibration accounts for differrent native sink levels; layer consensus limits intervention based on isolated head deviations; and bounded correction controls its magnitude. Exact compensation then ensures that routing respsects the selected sink budget. So their inclusion is motivated by distinct requirements of the method, rather than an arbitrary combination of heuristics.

Our central contribution is to make gating-budget selection and content rounting independently controllable despite their coupling through softmax normalization. These components collectively implement this separation. We intentionally use simple, interpretable operations where they suffice; additional complexity would not itself strengthen the design. Although the compensation formula follows directly from the normalization constraint, its role is to enforce the budget selected by the intervention while permitting content redistribution.

(2) Generalization across layers, models and calibration settings: Appendix C already explain that layer12-14 were selected using held-out data, reports layer-wise diagnostics on Llama-3.1 and Qwen2.5, and reproduces the qualitative fingding over layer10-17 and 8-19. Table 3 and Figure 6 additionally examine component choices and parameter sensitivity.

To extend these checks beyond ranges containing the original layer, we test non-overlapping layer bands across three models, alongside changes to calibration corpus and sample size. Detection thresholds and intervention strengths unchanged, only head-specific reference statistics are recalibrated.

Table W92u.4 show that both the prior-associated intervention effect and performance gains persist outside the original layer band across models. Table W92u.5 support robustness to calibration source and size without retuning the control parameter. The intendend generalization claim is that the intervention principle transfers, while its effective layer may vary across models.

Table W92u.4: Evaluation outside layers12-14

ModelIntervention layersPrior-associated effect ↑ConFiQA Acc. ↑RAGTruth HR ↓
Llama-3.1-8B-Instruct16–18+0.1384.116.9
Qwen2.5-7B-Instruct16–18+0.1485.515.1
Gemma-2-9B-it18–20+0.1086.612.4

Table W92u.5: Calibration robustness on Qwen2.5-7B-Instruct retaining the original intervention configuration.

Calibration corpusPassagesConFiQA Acc. ↑RAGTruth HR ↓
WikiText-103 — original25686.414.6
WikiText-1036485.815.0
WikiText-10351286.514.5
C425686.014.9

Thank you again for your time and thoughtful feedback. We hope these clarifications and additional results address your concerns. We will incorporate them into the revised version to make the contributions, experimental evidence, and scope of our claims clearer.

Dear Reviewer W92u

Thank you again for your time and thoughtful feedback on our paper. With the discussion period ending in about a day, we would appreciate any further feedback you may have on our rebuttal, particularly whether our responses have addressed your main concerns. We would be happy to provide any additional clarification.

Thank you for your consideration.

Best regards, The Authors

Reviewer_WiGhOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies the role of attention sinks in knowledge conflicts between context and parametric knowledge. The authors first analyze changes in sink attention through controlled experiments and causal interventions. They then propose AGPR, a training-free method that detects unusual sink attention and adjusts attention routing. Experiments on several models and benchmarks show clear improvements with low inference overhead.

Critical strengths
  1. The paper studies an interesting problem and provides useful analysis. The author investigates the role of attention sinks in knowledge conflicts. The controlled experiments and causal interventions provide useful evidence about how sink attention affects model behavior.

  2. The method is well designed and achieves strong results. AGPR separates sink gating from attention routing and uses a compensation term to preserve the target sink mass. The method is simple and does not require training. It shows clear improvements across several models and benchmarks with low inference overhead.

Critical weaknesses
  1. The evidence for knowledge-conflict arbitration is not fully convincing. Figure 2 shows that sink attention changes when the context-supported answer changes. However, it is unclear whether this is caused by conflicts with parametric knowledge or by general changes in answer-related information. More direct evidence would help support the claim that attention sinks specifically act as knowledge-conflict arbiters.
  2. AGPR uses the original attention logits as routing scores. This makes the routing step similar to attention sharpening. Although Table 3 provides several ablations, a direct comparison with simple attention sharpening is missing. Such a comparison would help clarify how much of the improvement comes from the proposed conflict-aware mechanism.
Questions for authors
  1. Can the authors compare AGPR with simple attention sharpening? The baseline should keep the original sink mass and should not use conflict detection.
  2. How do the authors evaluate the newly generated RAGTruth responses? Specifically, how do they identify hallucinations, determine whether a response is fully correct and grounded, and detect abstentions?
  3. The experiments mainly use relatively old model families, such as Llama 3.1, Qwen2.5, and Gemma 2. Can the authors provide results on more recent models, such as Qwen3 or Qwen3.5, to further validate the generality of AGPR?
3 comments
AuthorsOpenReview ↗

Clarifying conflict-specific sink behavior, sharpening controls, evaluation, and generalization

Response to Weakness 1

Thank you for highlighting this distinction. Figure 2 controls distractor edits and paraphrases, but the key additional comparison is if the same answer-changing edit produce a stronger sink response when it contradicts an indepently verified prior. We therefore comparing matched questions with and without a stable original-answer preference, assessed using 3 context-free paraphrases.

Table WiGh.1: New results prove that answer-changing edits produce larger sink decreases when conflicting with a verified prior.

Context editVerified priorUnverified priorBetween-group difference
Answer-changing evidence edit4.62.22.4
Answer-preserving distractor edit0.30.20.1
Meaning-preserving paraphrase0.10.2−0.1

The results seperate two explanations:

  1. answer-changing information affects both groups: indicating that general changes in answer-related information contribute to the response.
  2. conflict with a verified prior produces an additional 2.4-point sink decrease, with little corresponding difference for answer-preserving edits. This results support a role for sink attention in prior-context arbitration beyond general answer-change sensitivity.

Response to Weakness 2 & Question 1

Thank you for suggesting this direct comparison. We introduce unconditional attention sharpening that preserve every native sink weight and use no conflict detection or gating correction. It sharpen only the non-sink distribution, using the same eligible layers and query position as AGPR. We evaluate both AGPR's original routing strength and an independently validation-tuned strength.

Table WiGh.2: Ablation results contain Attention sharpening.

MethodConFiQA Acc. ↑FaithEval CF Acc. ↑RAGTruth HR ↓RAGTruth GSR ↑
Native68.274.821.768.4
Gating-Only80.584.117.972.6
No-Comp.79.283.018.671.4
Unconditional sharpening, native sink mass shared η=1.0\eta=1.082.285.117.274.0
Unconditional sharpening, native sink mass, tuned η=0.75\eta=0.7583.085.716.774.6
AGPR86.487.514.277.4

AGPR outperform validation-tuned sharpening by 3.4 on ConFiQA and 1.8 on FaithEval, while reducing RAGTruth HR by 2.5 and increasing GSR by 2.8. These gains show that sharpening along doesn't explain AGPR's improvement and support the additional value of its conflict-aware intervention and adaptive gating taken together.

Response to Question 2

Singe RAGTruth requires breaking down the response into individual factual claims for each to verify the source, and locating the four type of Hallucination fragments, the fully manuam evaluation task is extramely burdersome and the consistency cannot be guaranteed. Therefre, we use GPT-6.1 Sol and Claude Sonnet 5.5 API as LLM Judge, we only conduct manual review for the inconsistent samples. Manual review is conducted by two independent annotators, if there are still difference, we will introduce a third party for arbitration and take the third party's judgement as the final result. In practive, fewer than 10% of response require human review which is burden acceptable.

Response to Question 3

We extend out evaluation to Qwen3-8B and Qwen3.5-9B, as the blow table shows improvements despite the stronger native performance. AGPR increases ConFiQA acc by 10.9 points on Qwen3 and 5.5 points on Qwen3.5 while reducing RAGTrueh HR by 4.3 and 2.2 point. Grounded success also improves without increased abstention. It also outperforms CAD and Context-DPO.

Table WiGh.3: Additional results on recent model generations.

BackboneMethodConFiQA Acc. ↑FaithEval CF Acc. ↑RAGTruth HR ↓RAGTruth GSR ↑RAGTruth AR ↓
Qwen3-8BNative79.284.715.476.52.8
Qwen3-8BCAD84.988.013.678.73.3
Qwen3-8BContext-DPO88.890.212.281.03.1
Qwen3-8BAGPR90.191.211.182.42.5
Qwen3.5-9BNative86.190.210.884.12.0
Qwen3.5-9BCAD88.191.510.085.02.5
Qwen3.5-9BContext-DPO90.992.19.585.92.2
Qwen3.5-9BAGPR91.693.08.687.21.8

Thank you again for your time and thoughtful feedback. We hope these clarifications and additional results address your concerns. We will incorporate them into the revised version to make the contributions, experimental evidence, and scope of our claims clearer.

Thank you for the comment and the revised PDF. I read W1, the sharpening table, the RAGTruth protocol, and Table WiGh.3. The reponse addressed most of my concerns.

I am raising Soundness from 2 to 3, and Rating from 3 to 4. Please put the adaptive table, the judge/human numbers, and the Ouro-only rationale in the revision.

Thank you for carefully reviewing our responses and revision, and for updating your assessment. We have now incorporated the ablation comparisons in Table 4, expanded the RAGTruth judging and human-adjudication protocol in Appendix F.1, and clarified the rationale for Routing-Only in Appendix G.2. We hope these additions make the evaluation and component contributions clearer. We appreciate your constructive feedback and continued consideration.

Reviewer_p7DDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Authors introduce AGPR, a training-free inference-time intervention to improve contextual faithfulness through examining attention sinks. The proposed method uses per-head sink-gating deviations and layer-level consensus to detect knowledge conflicts from attention targets, and then adjusts the sink budget and redistributes non-sink attention while preserving the adjusted budget. Authors evaluate the proposed system across three model families and at various scales, showing that the proposed approach reduces hallucination and improves grounded success with a relatively small inference overhead.

Critical strengths
  1. Well motivated and mechanism-driven design. The paper explicitly separates sink gating from content routing -- which is a useful distinction because increasing total attention to context does not necessarily improve selection of relevant evidence.
  2. Broad empirical evaluation through main results. Results cover knowledge-conflict QA, counterfactual evidence, multi-hop reasoning, and RAG generation across multiple model families and sizes up to 32B for demonstrating model agnostic success.
  3. Inference-time solution for real world practicality. The default method requires no backbone training or auxiliary decoding branch etc., with reported latency overhead of approximately 3+% and small changes on instruction-following and reasoning benchmarks.
Critical weaknesses
  1. Mechanistic interpretation somewhat stronger than the evidence itself. Changing sink mass also scales the entire non-sink contribution -- demonstrating an effect on answer preference does not fully establish prior-specific arbitration. Clarifications or additional controlled experiments could help strengthen the claim.
  2. Empirical results are somewhat outdated as well - encourage the authors to add additional and more recent base models to help support the model agnostic claims, especially inpit context faithfulness has improved throughout multiple modern model families.
  3. Ablations could be desgined better to make a cleaner attribution of routing gains. The default evidence score is the native query key logit scores, making the routing tilt essentially a mechanism for attenuating / amplifying attention-temperature. Table 3 includes gating-only routing variants as an ablation, but it is not a dedicated routing-only variant to preserve the native sink budget -- this comparison would help establish how much the adaptive gating correction contributes beyond sharpening effects.
Questions for authors

Summarized in Weaknesses section

5 comments
AuthorsOpenReview ↗

Response to Weakness 1

Thank you for this helpful suggestion. We understand the concern that changing sink mass may affect answers through general content scaling. We therefore designed a controlled experiment to directly test whether sink gating participates in the competition between parameter prior and conflicting contextual evidence.

First, we construct two types of samples "Verified prior" and "Unverified prior". By not providing context, for the same question(such as "Where is the captial of France?"), different phrasing can be used. If the model consistently correct, it is classified as the first category, else the second category. We then provide with a context that supports different answers (for example "In the current setting, the captial of France is Lyon") and ask the model to respond. Then for the "Verified prior", the model faces clear conflict.

Then we fix the input, and:

  1. Naive: without any intervention, observe which answer the model is more like.
  2. Sink-only: reduce the sink attention, observe if it is more like to the context. To rule out "it's just the content getting amplified", we add a Content-only control, keep the sink part unchanged but make the same change to the non-sink/content part as in the full intervention (for example: original sink and content contribution is 4 and 6, the full intervention changed them to 3 and 7. The Content-only keep the sink at 4 and only increase the content to 7. This was to test if the change in the answer could be explained solely by enchancing the content.)

Table p7DD.1: Decomposition of a 25% relative decrease in native sink mass.

InterventionVerified prior: ΔMUnverified prior: ΔM
Native0.000.00
Content-only+0.52+0.37
Sink-only+0.12+0.02
Full gating+0.70+0.40
Full gating − Content-only+0.18+0.03

We're not denying that content scaling plays a role: Content-only did get +0.52 on Verified. But the Full gating, which makes exactly the same content enhancement, reached +0.70, the only difference is that it also lowers the sink contribution. So the extra +0.18 can't be explained by "more content enhancement". This shows that changing the sink contribution itself also play a part in shifting answer preference. More importantly, in the Unverified, the effect is only +0.03, which shows that when the model's parametric prior conflicts with context, lowering the sink contribution gives an extra push toward accepting the contextual evidence, and this effect is more strong when the prior is stable.

AuthorsOpenReview ↗

Response to Weakness 2.

Thank you for the suggestion. We have added Qwen3-8B and Qwen3.5-9B results in Table p7DD.2. AGPR improves ConFiQA accuracy over Native by 10.9/5.5 percentage points and reduces RAGTruth HR by 4.3/2.2 points, respectively. It also outperforms CAD and Context-DPO, improving grounded success without increased abstention. These results support AGPR’s continued effectiveness on stronger, more recent models. Evaluation uses text-only, non-thinking mode, with Qwen3.5 intervention restricted to token-softmax attention layers.

Table p7DD.2: Contextual faithfulness on recent model generations.

ModelMethodConFiQA Acc. ↑FaithEval CF Acc. ↑RAGTruth HR ↓RAGTruth GSR ↑RAGTruth AR ↓
Qwen3-8BNative79.284.715.476.52.8
Qwen3-8BCAD84.988.013.678.73.3
Qwen3-8BContext-DPO88.890.212.281.03.1
Qwen3-8BAGPR90.191.211.182.42.5
Qwen3.5-9BNative86.190.210.884.12.0
Qwen3.5-9BCAD88.191.510.085.02.5
Qwen3.5-9BContext-DPO90.992.19.585.92.2
Qwen3.5-9BAGPR91.693.08.687.21.8
AuthorsOpenReview ↗

Response to Weakness 3.

Thank you for suggesting this targeted control. We added the requested Routing-Only control: it retains AGPR's trigger rules and routing strength but disable adaptive gating correction (sets \lambda=0), preserving native sink mass. Compared with this control, AGPR improves ConFiQA accuracy by 1.8pp and FaithEval accuracy by 1.0pp, while reducing RAGTruth HR by 1.4 pp and increasing GSR by 1.6pp. This comparison isloates the incremental benefit of adaptive gating correction under the same routing configuration. The unconditional sharpening controls, including independently tuned η, further show that sharpening along dose not account for the full improvement.

Table p7DD.3: Ablation results contain Routing-Only control.

MethodConFiQA Acc. ↑FaithEval CF Acc. ↑RAGTruth HR ↓RAGTruth GSR ↑
Native68.274.821.768.4
Gating-Only80.584.117.972.6
No-Comp.79.283.018.671.4
Unconditional sharpening, native sink mass82.285.117.274.0
Unconditional sharpening, native sink mass, tuned η83.085.716.774.6
Routing-Only: AGPR triggers, native sink mass †84.686.515.675.8
AGPR86.487.514.277.4

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised Manuscript Uploaded

Dear Reviewers,

We have uploaded a revised manuscript incorporating the clarifications and additional experiments discussed in our rebuttal. The updates are highlighted in blue for ease of review.

We would appreciate your consideration of the revised manuscript alongside our responses, and would welcome any further questions or feedback on whether these updates address your concerns.

Thank you again for your time and thoughtful evaluation.

AuthorsOpenReview ↗

Final Rebuttal Summary and Revision Updates

Dear Area Chair and Reviewers,

Thank you for your time and constructive feedback. Our rebuttal and revision address the main concerns through:

  1. Mechanistic controls: prior-verified comparisons and Content-only interventions to distinguish prior-associated effects from general attention changes and content scaling.
  2. Component attribution: native-sink-mass-preserving sharpening and Routing-Only controls to clarify the additional contribution of adaptive gating.
  3. Evaluation and generalization: expanded RAGTruth judging documentation, newer-model comparisons, and tests across layer bands and calibration settings.

After reviewing our responses and revised PDF, Reviewer WiGh confirmed that most concerns were addressed. We have also incorporated the requested follow-up clarifications.

Several concerns raised by Reviewers p7DD and W92u overlap with those addressed in this exchange. Although we have not yet received their follow-up assessments, the corresponding responses and revisions are available for consideration. We hope this consolidated evidence helps the Area Chair assess the progress made during discussion.

Thank you again for the time and care you have devoted to reviewing our work. We sincerely appreciate your feedback and wish you all the best.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.