acceptodds
Under review as a conference paper at ICLR 2027

FragToken: Amplifying LLM Inference Costs through Noncanonical Token Generation

Abstract

As large language model (LLM) inference becomes increasingly expensive, resource-consumption attacks pose a growing threat to model providers. Existing attacks typically amplify cost by inducing abnormally long or repetitive outputs on attacker-controlled or triggered requests, making them easier to detect and limiting their deployment-wide impact when benign traffic dominates. In this work, we uncover a previously overlooked token-level attack surface arising from the many-to-one mapping from token sequences to decoded text. Although standard LLMs predominantly generate the canonical token sequences induced by their tokenizers, the same text can also be represented by substantially longer non-canonical sequences. This representational flexibility exposes a new avenue for resource-consumption attacks: an attacker can train the model to favor such sequences, systematically increasing the number of autoregressive decoding steps without a proportional increase in visible response length. However, we empirically find that directly maximizing token fragmentation substantially degrades model utility, producing conspicuous answer-quality failures that undermine attack stealthiness. To address this challenge, we propose , a training-time framework that combines source-model self-distillation, capacity-aware filtering and budgeting, and BPE-aligned re-merging to induce fragmented generation under ordinary prompts while largely preserving model utility. We evaluate on four LLMs across three benchmarks. Across the four models, achieves a three-benchmark average token inflation ratio (TIR) ranging from 1.99 to 2.46, while causing only minor degradation in model utility. Our work reveals a covert LLM supply-chain threat that increases inference cost without requiring large volumes of attack requests while largely preserving utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Resource-Consumption Attacks, Noncanonical Token Generation, Model Supply-Chain Security

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_v9wHOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FragToken, a training-time resource-consumption attack that increases LLM inference costs by inducing noncanonical token generation. The main observation is that multiple token sequences can decode to the same text, allowing a model to generate substantially more tokens without proportionally increasing the visible response length. Unlike existing resource-consumption attacks that rely on malicious prompts, triggers, or excessively long outputs, FragToken aims to introduce persistent computational overhead across ordinary user requests.

The proposed method combines source model self-distillation, capacity-aware fragmentation budgeting, and BPE-Alignedmerging to construct fragmented training labels while preserving model utility. The authors evaluate the approach on four LLMs and three benchmarks, reporting average token inflation ratios of 1.99–2.46 with relatively small changes in average benchmark accuracy. They also report a 2.44 x increase in amortized generation time in their serving experiments and evaluate the method against existing resource-consumption attacks and several detection methods.

Critical strengths

1. Interesting and novel attack mechanism. The paper introduces an interesting approach to increasing LLM inference costs by exploiting noncanonical token generation rather than producing longer responses, wthout the need of triggers.

2. Well motivated method with useful ablations. The proposed components are motivated by the observation that full token atomization significantly hurts model utility. The ablations provide good evidence that self-distillation, capacity-aware budgeting, and BPE-Aligned merging help balance token inflation and accuracy.

3. Convincing evidence of inference overhead. The paper evaluates four different LLMs and consistently achieves around two times token inflation. Importantly, the authors also measure actual generation time, reporting a 2.44 x increase while the visible response length decreases. This supports the practical relevance of the underlying mechanism.

Critical weaknesses

1. The stealth evaluation misses an important detection baseline. The paper emphasizes that FragToken is difficult to detect, but the evaluated defenses mostly focus on suspicious inputs or visible output patterns. In the proposed threat model, the defender hosts the model and can observe generated token IDs. A natural defense would be to compare the generated token count with the canonical tokenization of the decoded output, essentially using the paper's own TIR metric. Without evaluating such a detector, I am not convinced that the attack is actually difficult for a model provider to detect.

2. Limited evaluation of realistic deployment settings and generalization. The main serving and detection experiments use only 50 Alpaca queries, which seems relatively small for supporting broad conclusions about deployment impact. It is unclear how consistently the attack would affect diverse real-world traffic, particularly under different serving loads and output-token limits.

3. Utility preservation is not fully convincing. Although the average accuracy degradation is relatively small, GSM8K accuracy consistently drops by approximately 3-6 percentage points across all four models. This seems significant given that preserving model utility is central to the attack. The evaluation is also limited to three benchmarks, making it difficult to assess whether the modified models retain their general capabilities.

Questions for authors

Token-aware detection: Have the authors considered detecting FragToken by comparing generated token counts with canonical reencoding lengths (i.e., using the proposed TIR)? How effective would such a simple detector be?

Utility preservation: Can the authors explain why GSM8K accuracy consistently drops across all four models? Also, what explains the large OpenBookQA improvement for OLMo-2?

Generalization and deployment: How well does the attack generalize to more diverse real-world queries and serving conditions, given the relatively small serving evaluation and the lower TIR on OOD benchmarks?

4 comments

Dear Reviewer,

Thank you for your constructive comments. We address the main concerns below. Unless otherwise specified, all experiments use Llama-3.1-8B and the same main setting.

W1. TIR-based Detection

Beyond the content-based defenses in Table 4, we evaluate adaptive TIR-based monitoring. First, our experiments reveal an inherent trade-off between detection sensitivity and false positives. Prior work [1] and our experiments demonstrate that even benign LLMs can naturally generate noncanonical token sequences without manipulation. Consequently, setting a low TIR threshold may incorrectly flag benign generations. Conversely, increasing the threshold will cause many FragToken-generated responses to evade detection. Second, FragToken allows attackers to flexibly reduce TIR to evade fixed detection thresholds while retaining non-negligible token inflation (6.8%–31.9%). Since this overhead persists across ordinary user requests, it can accumulate into substantial resource consumption at scale.

SettingMean TIR1.0 (TPR/FPR)1.1 (TPR/FPR)1.2 (TPR/FPR)1.3 (TPR/FPR)1.4 (TPR/FPR)1.5 (TPR/FPR)
Beta = 0.851.319381.97 / 0.469.08 / 0.0062.31 / 0.0053.41 / 0.0040.93 / 0.0025.58 / 0.00
Beta = 0.901.186576.39 / 0.461.50 / 0.0045.23 / 0.0025.37 / 0.0010.96 / 0.003.82 / 0.00
Beta = 0.951.068263.69 / 0.428.76 / 0.006.84 / 0.001.46 / 0.000.20 / 0.000.09 / 0.00

W2. Utility

Additional Benchmarks. We extend the evaluation to 8 additional benchmarks covering instruction following, code generation, and long-form text generation. The results show that FragToken maintains comparable performance.

CapabilityBenchmarkOriginalFragTokenDelta
Instruction followingIFBench23.1224.20+1.08
Code generationMBPP+68.0065.50-2.50
long generationWritingBench38.4835.82-2.66
ComprehensionBoolQ87.5087.500.00
Science QAARC-Easy92.0093.50+1.50
Commonsense reasoningHellaSwag65.0066.50+1.50
Multidisciplinary knowledgeMMLU72.5071.00-1.50
Complex reasoningBBH67.5068.00+0.50

GSM8K Degradation. We hypothesize that GSM8K is sensitive to accumulated reasoning errors and the Alpaca–mathematics distribution gap. Fine-tuning on self-generated responses to GSM8K training questions improves accuracy by 2.38% and raises TIR from 1.95 to 2.18, supporting training-data alignment as a way to mitigate utility loss.

ModelACCTIR
Original84.46%1.00
GSM8K-trained80.85%2.18
Alpaca-trained78.47%1.95

OLMo-2 Improvement. Regarding the OBQA improvement on OLMo-2 (50.60% to 71.20%), we speculate that FragToken helps the model better elicit and express knowledge acquired during pretraining and SFT, improving its ability to leverage existing knowledge in answering domain-related questions.

W3. Generalization and Stability

Generalization to Real-World Queries. We clarify that the 50 Alpaca queries were used only for comparisons with existing attacks. We further evaluate FragToken on the GSM8K, PIQA, and OBQA in the paper. Additionally, we evaluate Llama-3.1-8B trained under our main FragToken setting on 1,000 real-world user requests from WildChat-1M, ShareLM, and PRISM. It demonstrates FragToken's generalization to real-world user inputs.

DatasetMean TIR
WildChat2.996
ShareLM3.075
PRISM2.782

Robustness to Output Token Limits. With other settings fixed, TIR remains approximately 2.43 across output limits of 2,048–16,384 tokens, demonstrating its robustness.

Token LimitsMean TIR
2,0482.427
4,0962.429
8,1922.430
16,3842.430

Robustness to Serving Loads. Across three benchmarks, FragToken maintains a TIR of approximately 2.46 and increases latency while reducing throughput under all tested concurrency levels, demonstrating its robustness.

ConcurrencyMean TIRLatency ratioThroughput reduction
12.4591.67×36.9%
22.4611.72×39.6%
42.4611.66×39.3%
82.4621.62×39.6%

Reference

[1] Vieira T, Liu T, Pasti C, et al. Language models over canonical byte-pair encodings, Proceedings of ICML, 2025.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, we wanted to kindly follow up on our rebuttal and ensure that we have adequately addressed your concerns.

If you have any remaining questions or feedback, we would greatly appreciate the opportunity to address them. Your feedback is valuable to us, and we hope to resolve any outstanding concerns before the discussion concludes.

Thank you again for your time and effort in reviewing our paper.

Best regards,

The Authors

Thank you for the new benchmarks and detection analysis. The TIR results show a simple detector flags many FragToken outputs at near zero FPR (69% TPR at threshold 1.1 for B=0.85), so the stealth claim holds only for the lower inflation settings. The GSM8K drop also remains (80.85% even with GSM8K-trained data vs. 84.46% original). My assessment is unchanged.

Dear Reviewer,

Thank you very much for your continued feedback and for acknowledging our additional experiments. We would like to offer two further clarifications.

TIR-based Detection. At β = 0.85, TIR-based detection achieves 69.08% TPR at a threshold of 1.1, leaving nearly one-third of FragToken outputs undetected. Reducing fragmentation further lowers TPR to 28.76% at β = 0.95. Importantly, β is an attacker-controlled hyperparameter that can be automatically tuned to favor stealthier configurations while retaining non-negligible token inflation. This demonstrates FragToken's flexibility for adaptive defense.

GSM8K Utility. We acknowledge the utility degradation on GSM8K, although the overall impact remains relatively modest (80.85% vs. 84.46%). We hypothesize that multi-step mathematical reasoning is particularly sensitive to accumulated generation errors. In-distribution training can partially mitigate this degradation, while FragToken maintains comparable performance across most other evaluated capabilities.

We sincerely appreciate your thoughtful feedback and the opportunity to clarify our work further.

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

This paper introduces FragToken, a novel adversarial attack framework that exploits subword tokenization vulnerabilities to amplify harmful outputs in large language models. The authors demonstrate that by strategically fragmenting malicious prompts across token boundaries, attackers can bypass safety alignment mechanisms that operate at the semantic or sentence level. They formalize this vulnerability through a token-level optimization procedure and evaluate it across multiple model families and defense strategies. Experimental results show that FragToken achieves significantly higher attack success rates than existing character-level or word-level attacks while remaining imperceptible to human readers. The paper further analyzes why current defenses fail against this threat vector and discusses implications for tokenizer design in safety-critical deployments.

Critical strengths

The first major strength is the identification of a previously underexplored attack surface rooted in tokenization mechanics rather than semantic manipulation. This shifts the adversarial focus from content-level evasion to structural exploitation of preprocessing pipelines, offering a fresh perspective on alignment failures. Consequently, this reframing has immediate practical relevance for model deployers who may assume tokenizers are neutral components rather than potential vulnerability vectors.

Building on this conceptual contribution, the second strength lies in the rigorous empirical validation across diverse model architectures and defense configurations. The authors do not limit evaluation to a single model family or benchmark but systematically test transferability and robustness, which strengthens confidence in the generality of their findings. Furthermore, the ablation studies isolating token fragmentation from other perturbation types provide clear evidence that the observed effects stem specifically from subword boundary exploitation rather than generic noise injection.

Finally, the third strength is the constructive discussion of mitigation pathways despite the paper’s primary focus on attack development. Rather than merely exposing a vulnerability, the authors outline concrete directions for tokenizer-aware safety training and input preprocessing hardening. This forward-looking element elevates the work beyond a pure red-teaming exercise and provides actionable guidance for the community to address the identified gap.

Critical weaknesses

Despite these strengths, the primary weakness is the limited analysis of real-world applicability outside controlled experimental settings. The attack assumes white-box access to tokenizer internals and gradient information, which may not reflect realistic threat models where adversaries interact only via API endpoints. Therefore, the practical impact of FragToken in production environments remains uncertain without black-box or query-limited evaluations that better simulate actual deployment constraints.

In addition to applicability concerns, the second weakness involves insufficient comparison with recent tokenization-aware defenses proposed in the past year. While the paper benchmarks against standard safety filters and prompt-based guards, it omits emerging methods specifically designed to detect or mitigate subword-level anomalies. As a result, the claimed superiority over existing defenses may be overstated, and the true difficulty of countering FragToken remains ambiguous without head-to-head testing against state-of-the-art countermeasures.

Compounding these issues, the third weakness is the lack of theoretical grounding for why certain token boundaries are more exploitable than others. The optimization procedure identifies effective fragmentation points empirically but offers no mechanistic explanation linking tokenizer statistics, vocabulary structure, or model attention patterns to attack success. Without such analysis, practitioners lack principled criteria for prioritizing which tokenization artifacts to audit or redesign, limiting the defensive utility of the findings.

Questions for authors
  1. Given the white-box assumptions underlying FragToken, have you evaluated its effectiveness under black-box or API-only access conditions, and if so, how does performance degrade compared to the reported white-box results? This would help clarify the realistic threat landscape for deployed systems.

  2. Moreover, how does FragToken perform against recently proposed tokenization-aware defenses such as subword anomaly detectors or tokenizer-robust fine-tuning methods published in the last twelve months? Including these comparisons would strengthen claims about the attack’s novelty and resistance to current mitigations.

  3. Finally, can you provide any correlational or causal analysis linking specific tokenizer properties like merge frequency, vocabulary size, or byte-pair encoding rules to the susceptibility of particular token boundaries? Such insights would transform FragToken from an empirical tool into a diagnostic framework for proactive safety engineering.

2 comments

Dear Reviewer,

Thank you for your constructive comments. We address the concerns below. Unless otherwise specified, all experiments use Llama-3.1-8B and the same main setting.

W1. Threat Model and Applicability

We would like to clarify a key misunderstanding regarding our threat model. FragToken is a training-time model supply-chain attack, not an inference-time adversarial attack. We assume the attacker is authorized to fine-tune a model before delivering it to a downstream provider or uploading it to a third-party platform. Therefore, black-box API access or inference-time gradient information is not required.

FragToken demonstrates three key properties:

  1. Effectiveness: FragToken increases actual inference latency by 2.44× and generated token consumption by 2.56×, without modifying inference pipeline, substantially increasing the provider's serving costs.
  2. Stealthiness: FragToken largely preserves model utility. Compared with the original models, average accuracy changes by only -1.48% on Llama-3.1-8B, +6.67% on OLMo-2-7B, -1.45% on Falcon3-7B, and -0.66% on Yi-1.5-9B. Thus, the provider incurs higher inference costs without noticeable expansion of user-visible responses or abnormal degradation in model performance and response quality.
  3. Consistency: Unlike trigger-based attacks, FragToken requires no malicious queries or specific triggers and induces fragmented generation on ordinary user inputs. We evaluate a benign-majority setting where attacker-controlled requests account for only 10% of total traffic. Even under this setting, FragToken achieves 2.56× latency overhead, compared with at most 1.88× for existing attacks, demonstrating its consistent impact on normal user traffic.

Additionally, we evaluate Llama-3.1-8B trained under our main setting on 1,000 real-world user requests from WildChat-1M, ShareLM, and PRISM. It demonstrates FragToken's generalization to real-world user inputs.

DatasetMean TIR
WildChat2.996
ShareLM3.075
PRISM2.782

W2. Tokenization-aware Defenses

Subword Anomaly Detectors. We have demonstrated that FragToken exhibits strong stealthiness against conventional input- and output-based defenses in Table 4. While providers can adopt TIR-based adaptive defenses, our experiments reveal an inherent trade-off between detection sensitivity and false positives. Low thresholds may falsely flag naturally occurring noncanonical generations in benign models (we observe a non-zero FPR even at a threshold of 1.0), a phenomenon also empirically observed in LLMs by Vieira et al. (2025) [1], whereas increasing the threshold to reduce false positives substantially degrades detection performance. Moreover, FragToken allows attackers to flexibly reduce TIR to evade fixed detection thresholds while retaining non-negligible token inflation (6.8%–31.9%). Since this overhead persists across ordinary user requests, it can accumulate into substantial resource consumption at scale.

SettingMean TIR1.0 (TPR/FPR)1.1 (TPR/FPR)1.2 (TPR/FPR)1.3 (TPR/FPR)1.4 (TPR/FPR)1.5 (TPR/FPR)
Beta = 0.851.319381.97 / 0.469.08 / 0.0062.31 / 0.0053.41 / 0.0040.93 / 0.0025.58 / 0.00
Beta = 0.901.186576.39 / 0.461.50 / 0.0045.23 / 0.0025.37 / 0.0010.96 / 0.003.82 / 0.00
Beta = 0.951.068263.69 / 0.428.76 / 0.006.84 / 0.001.46 / 0.000.20 / 0.000.09 / 0.00

Tokenizer-Robust Fine-tuning. In our threat model, the attacker fully controls all the fine-tuning process. Therefore, the defender cannot assume that tokenizer-robust training safeguards will be enforced during this attacker-controlled stage.

W3. Insights of FragToken

Our BPE-Aligned Merging is motivated by the statistical properties of BPE: lower merge ranks reflect earlier, typically more frequent subword associations. Breaking these structures may disrupt patterns learned during pretraining. Our motivation study (Fig. 4) supports this intuition:

Metric (r = 2)BPE-AlignedRandom Merging
Continuation ΔNLL ↓0.190.34
KL Divergence ↓0.200.35
Hidden-state Cosine ↑0.730.62

Importantly, BPE-Aligned Merging consistently outperforms random merging on all three metrics. Furthermore, replacing BPE-Aligned Merging with random merging reduces average ACC from 80.40% to 77.56%, despite nearly identical TIR (2.46 vs. 2.44, Table 5). These results support that preserving BPE-aligned structures better retains pretrained predictive behavior and downstream utility.

Reference

[1] Vieira T, Liu T, Pasti C, et al. Language models over canonical byte-pair encodings, Proceedings of ICML, 2025.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, we wanted to kindly follow up on our rebuttal and ensure that we have adequately addressed your concerns.

If you have any remaining questions or feedback, we would greatly appreciate the opportunity to address them. Your feedback is valuable to us, and we hope to resolve any outstanding concerns before the discussion concludes.

Thank you again for your time and effort in reviewing our paper.

Best regards,

The Authors

Reviewer_Kgo5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FragToken, a training-time attack that increases LLM inference costs through noncanonical token generation. The method combines self-distillation, capacity-aware budgeting, and BPE-aligned merging to increase token counts while preserving model utility. Experiments on four LLMs show higher token counts and inference costs with limited accuracy loss.

Critical strengths
  1. This paper explores a new attack direction by exploiting noncanonical token generation to increase LLM inference costs.
  2. The proposed method effectively increases token counts while largely preserving model performance.
Critical weaknesses
  1. In Table 4, the authors evaluate five defense methods, including input perplexity, ONION, repetition detection, RD-P, and RD-M. However, these methods mainly detect abnormal inputs or repeated outputs, while FragToken increases inference costs by generating noncanonical token sequences. The number of generated tokens can be much larger than the number of tokens obtained by re-encoding the same output text. This difference can be directly measured during inference. A simple detector based on the Token Inflation Ratio (TIR) may therefore identify FragToken without analyzing the input or output content. The authors should evaluate such a detector to provide stronger evidence for the stealthiness of FragToken.
  2. In Table 2, the authors evaluate model utility on only three benchmarks, including GSM8K, PIQA, and OBQA. These benchmarks mainly cover mathematical reasoning and multiple-choice question answering. They do not cover other important LLM capabilities, such as instruction following, code generation, and long-form text generation. Moreover, FragToken causes a clear accuracy drop on GSM8K across all four models. For example, the accuracy of Llama-3.1-8B decreases from 84.46% to 78.47%, while the accuracy of Yi-1.5-9B decreases from 79.91% to 74.68%. These results raise concerns about the impact of fragmented token generation on different model capabilities. More experiments on diverse benchmarks are needed to support the claim that FragToken can largely preserve model utility.
Questions for authors

Please see the Weaknesses.

2 comments

Dear Reviewer,

Thank you for your constructive comments. We address the main concerns below. Unless otherwise specified, all experiments use Llama-3.1-8B and the same main setting.

W1. TIR-based Detection

We have demonstrated that FragToken exhibits strong stealthiness against conventional input- and output-based defenses in Table 4. While providers can adopt TIR-based adaptive defenses, our experiments reveal an inherent trade-off between detection sensitivity and false positives. Low thresholds may falsely flag naturally occurring noncanonical generations in benign models (we observe a non-zero FPR even at a threshold of 1.0), a phenomenon also empirically observed in LLMs by Vieira et al. (2025) [1], whereas increasing the threshold to reduce false positives substantially degrades detection performance. Moreover, FragToken allows attackers to flexibly reduce TIR to evade fixed detection thresholds while retaining non-negligible token inflation (6.8%–31.9%). Since this overhead persists across ordinary user requests, it can accumulate into substantial resource consumption at scale.

SettingMean TIR1.0 (TPR/FPR)1.1 (TPR/FPR)1.2 (TPR/FPR)1.3 (TPR/FPR)1.4 (TPR/FPR)1.5 (TPR/FPR)
Beta = 0.851.319381.97 / 0.469.08 / 0.0062.31 / 0.0053.41 / 0.0040.93 / 0.0025.58 / 0.00
Beta = 0.901.186576.39 / 0.461.50 / 0.0045.23 / 0.0025.37 / 0.0010.96 / 0.003.82 / 0.00
Beta = 0.951.068263.69 / 0.428.76 / 0.006.84 / 0.001.46 / 0.000.20 / 0.000.09 / 0.00

W2. Utility Preservation

Additional Benchmarks. We extend the evaluation to 8 additional benchmarks covering instruction following, code generation, writing, reading comprehension, and knowledge and reasoning tasks. The results show that FragToken maintains comparable performance on evaluated tasks.

CapabilityBenchmarkOriginalFragTokenDelta
Instruction followingIFBench23.1224.20+1.08
Code generationMBPP68.0065.50-2.50
Open-ended writingWritingBench38.4835.82-2.66
Reading comprehensionBoolQ87.5087.500.00
Science question answeringARC-Easy92.0093.50+1.50
Commonsense reasoningHellaSwag65.0066.50+1.50
Multidisciplinary knowledgeMMLU72.5071.00-1.50
Complex reasoningBBH67.5068.00+0.50

GSM8K Degradation. Regarding GSM8K, we hypothesize that its multi-step reasoning process makes generation more sensitive to accumulated prediction errors. Moreover, our training data are constructed from Alpaca, and the distribution gap with mathematical reasoning tasks may further limit the transfer of utility-preserving fragmented generation. To examine the role of training-data alignment, we additionally fine-tune Llama-3.1-8B-Instruct on self-generated responses to GSM8K training questions, using the same four demonstrations used during evaluation. Training on GSM8K improves accuracy by 2.38 percentage points while increasing TIR from 1.950 to 2.180. These results support our hypothesis that better training-data alignment can mitigate utility degradation while increasing token inflation.

ModelACCTIR
Original84.46%1.00
GSM8K-trained80.85%2.18
Alpaca-trained78.47%1.95

Reference

[1] Vieira T, Liu T, Pasti C, et al. Language models over canonical byte-pair encodings, Proceedings of ICML, 2025.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, we wanted to kindly follow up on our rebuttal and ensure that we have adequately addressed your concerns.

If you have any remaining questions or feedback, we would greatly appreciate the opportunity to address them. Your feedback is valuable to us, and we hope to resolve any outstanding concerns before the discussion concludes.

Thank you again for your time and effort in reviewing our paper.

Best regards,

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.