acceptodds
Under review as a conference paper at ICLR 2027

When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoning

Abstract

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Reasoning Models, Uncontrolled Reasoning, Attention Dynamics, Reasoning State Detection, Runtime Intervention

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_wNsrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes RADAR, a framework that distinguishes four reasoning states (direct answering, effective reflection, excessive reflection, and persistent looping) using attention statistics and generation progress. It further introduces Attention Realignment, which adjusts attention distributions toward patterns observed in normal reasoning to mitigate uncontrolled generation. Experiments across multiple reasoning models evaluate the framework's state classification, temporal detection behavior, and intervention effectiveness, suggesting that attention dynamics may provide useful signals for identifying and mitigating uncontrolled reasoning.

Critical strengths
  • Meaningful problem formulation: The distinction between excessive reflection and persistent looping is useful, as these behaviors differ in their characteristics and may require different mitigation strategies.

  • Interesting mechanistic perspective: Analyzing attention dynamics throughout generation offers a potentially useful perspective on how normal reasoning may develop into uncontrolled behavior, going beyond surface-level repetition.

  • Practical intervention and empirical evaluation: The proposed attention realignment method is an interesting attempt to mitigate uncontrolled generation while preserving useful reasoning, and the evaluation across multiple models provides encouraging preliminary evidence.

Critical weaknesses
  • Potential errors in PAS and CTS: Equation (1) seems to sum attention over the wrong token positions, while Equation (3) only considers tokens shared by the prompt and generation. These definitions may affect the validity of the attention analysis and the resulting intervention.

  • Unclear evaluation protocol: Each prefix inherits the label of its full trajectory, so the classifier predicts the eventual outcome rather than necessarily detecting the current state. The paper also does not clearly specify whether prefixes from the same trajectory are kept in the same data split, making it difficult to assess the reliability of the reported results.

Questions for authors

The definitions of PAS and CTS seem inconsistent with their intended interpretations. Could the authors clarify these definitions and confirm that the implementation matches the descriptions? Do the main results remain consistent under the corrected definitions?

Are training and test sets separated at the trajectory level, ensuring that prefixes from the same trajectory cannot appear in both sets? Does the reported classification performance hold under this protocol?

Since each prefix inherits the label of its complete trajectory, how should the reported detection time be interpreted relative to the actual onset of excessive reflection or looping? Does the method predict eventual failure, detect an ongoing failure, or both?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wNsr

We appreciate your recognition of the relevance and empirical findings of our work. We believe some concerns may stem from insufficient clarification in our presentation, which we aim to address below.


W1&Q1: PAS and CTS definitions and analytical validity

We thank the reviewer for pointing this out. The PAS summation range contains a typographical error, while the restriction of CTS to shared tokens is intentional. We would like to clarify these points.

(1) PAS: correcting the input-region summation range. PAS measures the proportion of the current token's attention allocated to the input. Let pp denote the input length and ss the current total sequence length. The numerator in Eq. (1) in the paper should be ∑j=1pAs,j(l)\sum_{j=1}^{p}A_{s,j}^{(l)}, with the denominator covering all visible positions and the result averaged across layers. The original text incorrectly specifies the generated region. We will correct this error consistently.

(2) CTS: a fine-grained comparison over shared token identities. Shared tokens are tokens present on both sides, rather than only repeated segments in looping outputs. Non-shared tokens have no corresponding occurrence on the other side. If included, their difference terms equal their attention shares in their respective regions, incorporating vocabulary differences introduced by new steps in normal reasoning and thereby interfering with the attention discrepancies we intend to measure. We have added a supplementary experiment including non-shared tokens in Appendix D.2 of the paper, "Effect of Including Non-Shared Tokens in CTS," and found that persistent looping is difficult to distinguish from normal requests.

W2&Q2: Prefix-based risk prediction and trajectory-level separation

RADAR aims to identify potential risks of uncontrolled reasoning from the current generation prefix, providing a basis for timely intervention. Training and test sets are separated at the complete-trajectory level, and different prefixes of the same trajectory never appear across the two sets. During training, we assign the complete-trajectory label of each training sample to its prefixes to learn the relationship between current attention dynamics and the reasoning behavior eventually exhibited by that trajectory. These data are independent of the test data.

Our experiments also cover MMLU-Geor, MMLU-Econometrics, MMLU-World-History, Joint, and MiP, which are not used for training. The main experiments in the paper already demonstrate the method's effectiveness on these datasets.

Q3: Detection timing relative to the onset of uncontrolled behavior

RADAR continually assesses potential risks of uncontrolled reasoning from the current generation prefix to support timely intervention. During training, each prefix inherits its complete-trajectory label, allowing the classifier to learn the relationship between current attention dynamics and the trajectory's eventual reasoning behavior. Prediction uses only information available in the current prefix. An early O/LO/L prediction therefore indicates a risk that the trajectory will develop into uncontrolled reasoning; it does not require excessive reflection or looping to have already occurred. This design allows intervention before abnormal behavior fully develops and continued risk identification after it has emerged.

For successful attack trajectories, Table 1 in the paper reports both the first generated position satisfying P(O)+P(L)≥τP(O)+P(L)\geq\tau and the recorded onset of repetition, each measured from the start of generation. For Recur, LoopLLM, and MiP, the mean risk-detection positions precede their respective mean repetition onsets, indicating that RADAR can provide early warnings to some extent.


We hope these clarifications alleviate your concerns and would welcome any further questions. If our responses resolve your reservations, we respectfully ask you to consider increasing your rating. Thank you for your careful evaluation.

I appreciate the authors' clarifications. I believe this paper has some value as an analysis of attention dynamics. However, as the authors acknowledge, naturally occurring uncontrolled reasoning is relatively rare. Moreover, Appendix F shows that simple n-gram blocking achieves substantially stronger loop suppression with a relatively modest loss in benign-task accuracy. These observations limit the practical value of the proposed framework for real-world deployment. Therefore, I will maintain my current rating.

AuthorsOpenReview ↗

Response on Adversarial Relevance and the Defense–Utility Trade-off (1/2)

Thank you for your thoughtful follow-up and for recognizing the value of our attention analysis. We understand that you remain concerned about the practical significance of our work, and we also note the confidence score of 2 indicated in your review. We would therefore like to offer additional background and supporting evidence from recent research in this area, with the aim of clarifying the relevance of resource-exhaustion threats and the motivation for studying defenses. We hope that this additional context will be helpful in assessing the practical significance of our findings.

1. The rarity of spontaneous failures does not establish a low risk of resource exhaustion.

Section 3.1 of the paper explicitly includes adversarially induced uncontrolled reasoning, and Section 4.4 states that our current evaluation focuses on targeted attacks. The recent survey Resource Consumption Threats in Large Language Models: A Survey [1], published in EMNLP 2026, systematically reviews resource-consumption threats across LLMs. It explains how adversarially amplified generation can increase inference latency, undermine service availability, and threaten economic sustainability. Representative empirical findings make these consequences concrete:

  • Engorgio [2] reports inference costs per query reaching approximately 13 times those of normal inputs in its white-box evaluation, demonstrating how adversarial prompts can amplify computational demand.
  • LoopLLM [3] reports generated outputs exceeding 90% of the maximum output length through repetition-induced decoding loops, demonstrating that malicious inputs can drive models to consume nearly the entire generation budget.
  • AutoDoS [4] reports that, in simulated high-concurrency scenarios, its attack reduces server throughput from approximately 1 to 0.009 requests per minute, a decrease of about 99.1%, demonstrating how malicious resource consumption can severely degrade service availability.

Together, the survey and these empirical results establish a concrete security motivation for studying attacker-induced uncontrolled reasoning. Low incidence under benign inputs does not establish low risk under deliberate attack; the latter is the setting evaluated in our paper.

2. The n-gram results demonstrate a suppression–utility trade-off with a measurable cost to normal-task performance.

The relevant comparison appears in Appendix E.2–E.3, Tables 9 and 10 of the paper. On DeepSeek-R1-Distill-Llama-8B, n-gram blocking lowers benign-task accuracy from 66.4% to 58.4%, a decrease of 8.0 percentage points. We respectfully consider this a substantial loss of task utility, particularly for reasoning applications where preserving answer correctness is a central requirement.

Although n-gram blocking can be useful in tasks such as summarization [5], empirical studies on complex reasoning show that blanket repetition suppression can substantially degrade task accuracy:

  • To Reason but Not to Stop [6] directly evaluates n-gram blocking. For DeepSeek-R1-Distill-Qwen-14B, its No-Repeat baseline (n = 32) lowers LiveCodeBench accuracy from 62.4% to 46.6%. This result demonstrates a substantial coding-performance cost in the evaluated setting.
  • Word Salad Chopper [7] reports that a repetition penalty lowers AIME25 accuracy from 37.92% to 25.83%. This provides direct evidence that a conventional repetition penalty can harm mathematical reasoning.
  • LZ Penalty [8] reports that, for QwQ-32B on AIME, increasing the frequency penalty from 0 to 0.1 lowers accuracy from approximately 73% to 7%. The authors explain that penalties accumulating over long reasoning traces can suppress common but essential tokens and severely degrade generation.

These studies cover different repetition-control mechanisms and experimental settings, and demonstrate why strong repetition suppression (n-gram) cannot by itself establish suitability for complex reasoning tasks. These findings are consistent with the conclusions from our own research, indicating that simple defenses are unreliable within the complex reasoning processes of an LRM.

AuthorsOpenReview ↗

Response on Adversarial Relevance and the Defense–Utility Trade-off (2/2)

Resource exhaustion, whether arising spontaneously or induced by adversarial attacks, motivates targeted defenses that preserve legitimate reasoning. Simple repetition constraints can substantially impair performance on complex tasks, highlighting the need for more selective interventions. Our work addresses this need through attention-based reasoning-state analysis and targeted intervention.

We hope these clarifications help convey the practical significance of the problem and support a fuller assessment of our contributions. We would be happy to discuss any remaining concerns. If our responses have satisfactorily addressed your concerns, we would be grateful if you would consider raising your score.

[1] Resource Consumption Threats in Large Language Models: A Survey (EMNLP 2026).

[2] An Engorgio Prompt Makes Large Language Model Babble On (ICLR 2025).

[3] LoopLLM: Transferable Energy-Latency Attacks in LLMs via Repetitive Generation (AAAI 2026).

[4] Crabs: Consuming Resource via Auto-generation for LLM-DoS Attack under Black-box Settings (ACL 2025).

[5] A Deep Reinforced Model for Abstractive Summarization (ICLR 2018).

[6] To Reason but Not to Stop: Understanding and Recovering Pruned Large Reasoning Models (arXiv preprint, 2026).

[7] Word Salad Chopper: Reasoning Models Waste A Ton Of Decoding Budget On Useless Repetitions, Self-Knowingly (EMNLP 2025).

[8] LZ Penalty: An Information-Theoretic Repetition Penalty for Autoregressive Language Models (Transactions on Machine Learning Research, 2026).

Reviewer_pbN5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces RADAR, a method for detecting potentially unproductive generation in large reasoning models and guiding attention interventions. It distinguishes four types of reasoning. These are direct answering, effective reflection, excessive reflection, and persistent looping.

RADAR uses two attention measurements, Prompt Attention Share and Cross context Token Similarity. A classifier combines their summaries with generation length to predict the eventual category of a trajectory. When it predicts excessive reflection or looping with sufficient confidence, the method adjusts attention in selected layers.

The experiments cover five models and several reasoning benchmarks and attack sources. The authors report an average classification micro F1 of 89.8% and a reduction in average loop rate from 37.0% to 27.2%, while generally preserving benign task accuracy. They also examine detection timing and compare the intervention with AUSteer and trigram blocking on one model.

Critical strengths
  1. The paper addresses an important problem. A useful intervention should reduce unproductive generation while allowing the model to continue reasoning when that reasoning helps. The distinction between excessive reflection and repeated lexical content is especially helpful. Appendix A provides operational criteria that offer a useful starting point for studying these behaviors.

  2. The classifier is simple and easy to inspect. Its small number of features makes it possible to study which measurements contribute to its predictions. The evaluation covers five models and several attack sources, and the authors report benign accuracy alongside loop suppression. I also appreciate their acknowledgment that some baselines suppress loops more strongly and that spontaneous failures were not evaluated systematically.

Critical weaknesses
  1. The central attention measurements need clarification.

The text defines PAS as attention to the prompt, but Equation (1) sums over the generated region and the final prompt token. Section 3.3 also normalizes using generated length, while Equation (8) converts back using prompt length. For CTS, Equation (3) compares only token identities shared by the prompt and generation, although Equation (2) normalizes over the full regions. It therefore does not match the stated total variation similarity between the two distributions. These inconsistencies affect how the measurements and intervention should be interpreted. The paper needs to clarify which definitions were implemented and whether the discrepancies are limited to the manuscript.

  1. The labeling protocol conflicts with the reported classification results.

Appendix B.2 states that the GLM trajectories generated under Recur receive benign labels, while RADAR predicts excessive reflection for all of them. Table 5 nevertheless reports 100% classification accuracy for this group. These statements are inconsistent under the stated evaluation protocol. More broadly, Appendix B assigns benign labels to attack trajectories without an observed loop, even though excessive reflection can occur without lexical looping under the paper’s own definitions. These issues need to be resolved to make the classification results and aggregate score reliable.

  1. The contribution of the attention features is not sufficiently isolated.

The main training sources are closely associated with the four labels, and generation length is included as a feature. Although the paper evaluates additional sources, the experiments do not clearly establish how much the attention features contribute beyond length and source characteristics. Appendix C discusses the length concern, but does not directly compare the classifier with a model using length alone or attention features without length. These comparisons would help establish what the proposed attention measurements add and whether the classifier captures reasoning behavior that generalizes across sources.

  1. The evidence for early intervention and the contribution of its components remains incomplete.

The intervention requires the most likely class to be O or L and its individual probability to exceed the threshold. Table 1 instead measures when the combined probability of O and L crosses a threshold. The reported detection times therefore do not directly evaluate when the implemented intervention would activate. The paper also does not report what proportion of failures this rule detects before repetition begins or how often it activates on benign trajectories.

The intervention reduces the reported loop rate, but the experiments do not separate the contributions of PAS rebalancing and CTS suppression or establish the importance of the selected layers. Focused comparisons would help determine which parts of the proposed intervention account for the improvement.

Questions for authors
  1. What definitions of PAS and CTS were used in the implementation? Please reconcile Equations (1), (3), and (8) with the text. If any definition needs correction, would the reported results or their interpretation change?

  2. How were the GLM trajectories generated under Recur labeled and scored in Tables 5 and 6? Please reconcile the reported accuracy with Appendix B.2 and explain how excessive reflection without lexical looping is labeled. Does resolving this inconsistency affect the aggregate classification score?

  3. Are the training and test sets separated by underlying prompts and complete trajectories? Could you report performance using length alone and attention features without length? Please also explain how the evaluation distinguishes reasoning behavior from characteristics of the data sources.

  4. Using the actual intervention rule, what proportion of failures is detected before repetition begins, and how often does the rule activate on benign trajectories? Could you also separate the effects of PAS rebalancing and CTS suppression? A control using alternative layers at comparable intervention strength would help establish whether the proposed layer selection contributes to the improvement.

5 comments
AuthorsOpenReview ↗

Response to Reviewer pbN5 (1/5)

Thank you for acknowledging the motivation and methodological contributions of our study. We greatly value your detailed feedback and hope our explanations below help resolve the raised concerns.


W1&Q1: PAS and CTS definitions and implementation consistency

We thank the reviewer for identifying these inconsistencies. The PAS issue is a typographical error in the equations, while the shared-token design of CTS is intentional. The revisions do not change the computation, intervention operations, or experimental values. The paper provides code for verification.

(1) Eqs. (1) and (8) in the paper: the definition of prompt-side PAS. Let pp denote the input length and ss the current total sequence length. PAS should aggregate attention over the input positions P={1,…,p}\mathcal P=\{1,\ldots,p\} and average across layers. However, Eq. (1) in the paper incorrectly specifies [p,s][p,s]. The correct numerator is ∑j=1pAs,j(l)\sum_{j=1}^{p}A_{s,j}^{(l)}.

Similarly, the uniform prompt-side attention baseline in Eq. (8) in the paper is p/sp/s. The fitting objective should therefore use log⁡[(s/p)MP(s)]\log[(s/p)M_{\mathcal P}(s)], and the original PAS share is recovered by multiplying by p/sp/s. This makes the preceding fitting objective consistent with Eq. (8) in the paper. We will correct the equations consistently and remove the repeated statement t=s−pt=s-p from this paragraph.

(2) Eqs. (2) and (3) in the paper: retaining the shared-token comparison. In Eq. (2) in the paper, qX(v)q_X(v) denotes the share of all attention in region XX assigned to token identity vv. Eq. (3) in the paper compares only identities that appear in both the input and generated regions. These two steps respectively define the reference region for each share and the identities included in the comparison. We describe the quantity precisely as the within-region attention shares of shared tokens. This design aims to compare fine-grained attention discrepancies for the same token identity between the input and generation. Non-shared tokens have no corresponding occurrence on the other side. Including them with zero mass on the absent side introduces vocabulary-set differences, which may arise from new reasoning steps or answer content in normal reasoning and thus interfere with the same-identity attention discrepancies we intend to measure. We have added a supplementary experiment including non-shared tokens in Appendix D.2 of the paper, "Effect of Including Non-Shared Tokens in CTS," and found that persistent looping is difficult to distinguish from normal requests.

AuthorsOpenReview ↗

Response to Reviewer pbN5 (2/5)

W2&Q2: Attack outcomes and four-state labels

The attack success rate of Recur on GLM-4.7-Flash is 0.0% because the trajectories do not meet the attack-success criterion for persistent lexical looping. This does not mean that the trajectories belong to the normal D/R states.

The attack outcome records whether the output-level looping criterion is met, whereas the four-state label describes the reasoning behavior of the complete trajectory. O denotes repeated discussion of semantically similar issues with little new information after meaningful progress has saturated. The GLM–Recur example in the paper's appendix continues to propose unsupported hypotheses after obtaining the computational result and eventually terminates naturally, showing that a non-looping trajectory can still exhibit excessive reflection. We therefore distinguish D, R, and O based on trajectory behavior, rather than assigning all failed attacks to D/R.

This clarification does not change the 0.0% attack success rate for GLM–Recur. In the risk assessment for GLM, both O and D/R are classified into the "harmless" category. The classification accuracy in Table 5 and the aggregate micro-F1 in Table 6 of the paper remain unchanged.

AuthorsOpenReview ↗

Response to Reviewer pbN5 (3/5)

W3&Q3: Length, cross-source generalization, and data separation

  1. Length provides a reference for the generation stage.

Length supplies generation-stage information. The classifier uses the historical PAS slope, mean CTS, and generation progress log⁡ki\log k_i. Figure 4 in the paper shows state-related PAS patterns over generated positions. As the prefix grows, both the available attention history and its fitted trend change, so length provides a stage reference for interpreting PAS and CTS.

Experimentally, the mean first risk-detection positions across attack groups range from 60.6 to 997.0 tokens, below the model-level mean completion lengths of normal D/R trajectories, 1,184.48–2,119.74 tokens. This indicates that risk signals can appear within the length range in which normal reasoning is typically still ongoing. Length alone cannot support effective classification within the length range of normal requests. We additionally report detection results at 512 output tokens. Cumulative detection counts harmful trajectories identified by that position, while the normal-task false-positive rate counts normal trajectories classified as harmful at that position. RADAR already achieves a high abnormality detection rate with few false positives.

Table: Detection performance at 512 generated tokens.

ModelDetection ↑Normal FPR ↓
DeepSeek-R1-Distill-Llama-8B84.6%0.0%
DeepSeek-R1-Distill-Qwen-14B84.2%3.9%
QwQ-32B87.7%12.3%
Qwen3.6-27B89.4%10.6%
GLM-4.7-Flash69.0%21.8%
  1. Training and testing do not share different prefixes of the same trajectory.

The means and standard deviations used for feature standardization are computed on the training set and then applied to the test set. Cross-dataset evaluation verifies transferability. The training sources are GSM8K, GPQA, Recur, and LoopLLM. The per-dataset results in Section 4.2 and the appendix of the paper additionally cover MMLU-Geor, MMLU-Econometrics, MMLU-World-History, Joint, and MiP, which are not used for training. These datasets provide the corresponding reasoning states from different sources, demonstrating the classifier's effectiveness on new tasks and attack settings.

Reviewer_91mxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates uncontrolled reasoning in large reasoning models and proposes RADAR, a framework for identifying four reasoning states: direct answering, effective reflection, excessive reflection, and persistent looping. RADAR uses two attention-based metrics, Prompt Attention Share (PAS) and Cross-context Token Similarity (CTS), together with a lightweight temporal classifier to predict reasoning states during generation. The authors further propose Attention Realignment to mitigate abnormal reasoning behavior. Experiments across five reasoning models report an average micro-F1 of 89.8% for state classification and a 9.8 percentage-point reduction in loop rate, while largely preserving benign-task accuracy.

Critical strengths
  1. Meaningful problem formulation. The paper addresses an important problem in reasoning models by distinguishing productive reflection from excessive reflection and persistent looping. This is more informative than treating all lengthy reasoning as undesirable.

  2. Interesting attention-based analysis and intervention. The paper connects attention dynamics with reasoning behavior and evaluates both state prediction and attention-based intervention across five models. The reported reductions in looping suggest that attention modulation may be a useful direction for mitigating uncontrolled generation.

Critical weaknesses
  1. Insufficient justification and inconsistencies in the proposed attention metrics. The connection between PAS/CTS and reasoning quality is not sufficiently established. It remains unclear why these statistics should reliably distinguish productive reasoning from uncontrolled reasoning. Moreover, Eq. (1) appears inconsistent with the definition of prompt attention, while the normalization in Eq. (8) does not seem consistent with the preceding fitting objective. Eq. (3) also computes CTS only over shared token identities, ignoring attention mass assigned to non-shared tokens. These issues raise concerns about the interpretation and reliability of the core metrics.

  2. Potential dataset confounding and insufficient ablations. The four reasoning states are strongly associated with different datasets or attack sources. The classifier may therefore capture dataset- or attack-specific attention patterns rather than general reasoning-state characteristics. The reported cross-dataset results do not fully rule out this possibility. More importantly, the paper lacks key feature ablations, such as PAS-only, CTS-only, and length-only baselines, making it difficult to assess the actual contribution of attention dynamics. The separate effects of PAS Rebalance and CTS Suppression are also not evaluated. Together, these limitations weaken the evidence supporting the main methodological claims.

Questions for authors
  1. Could the authors clarify the apparent inconsistencies in the PAS definition and normalization (Eqs. 1 and 8), and explain why CTS is computed only over shared token identities in Eq. (3)?

  2. Can the authors provide within-dataset or length-matched evaluations to rule out dataset-specific patterns and generation length as potential confounding factors?

  3. Can the authors report feature-level ablations for PAS, CTS, and generation length, as well as separate evaluations of PAS Rebalance and CTS Suppression? These results are important for establishing the necessity of the proposed components.

  4. What is the actual inference overhead of RADAR in terms of latency, throughput, and memory usage? Given the resource-exhaustion motivation, it would be helpful to report the net computational savings after accounting for the attention monitoring and intervention costs.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 91mx (1/4)

We sincerely appreciate your recognition of the novelty and practical significance of our work. Thank you for your constructive suggestions. We hope the following clarifications adequately address your concerns.


W1: Justification and consistency of the attention metrics

(1) Notational errors in Eqs. (1) and (8) in the paper. These two equations contain notational errors that do not alter the method's objective of analyzing prompt-side attention or its implementation, nor do they affect the experimental results reported in the paper. We have corrected the relevant equations and text to align them with the method's definitions and implementation. The paper also provides code for verification.

Specifically, PAS measures the proportion of the current token's attention allocated to the input. Let pp denote the input length and ss the current total sequence length. The input positions should be P={1,…,p}\mathcal P=\{1,\ldots,p\}. However, the numerator in Eq. (1) in the paper incorrectly uses [p,s][p,s] as the summation range. The correct expression is:

=\frac{1}{L}\sum_{l=1}^{L} \frac{\sum_{j=1}^{p}A_{s,j}^{(l)}}{\sum_{j=1}^{s}A_{s,j}^{(l)}} =\frac{1}{L}\sum_{l=1}^{L}\sum_{j=1}^{p}A_{s,j}^{(l)}.$$ This confusion between regions also extends to the normalization description preceding Eq. (8) in the paper. Since $M_{\mathcal P}$ denotes the prompt-side attention share, its baseline under uniform attention over all visible positions should be $p/s$, rather than the generation-side baseline $t/s=1-p/s$. The length-normalized metric should therefore be $m_{\mathcal P}(s)=(s/p)M_{\mathcal P}(s)$, and the preceding fitting objective should accordingly be $\log[(s/p)M_{\mathcal P}(s)]$. We have also removed the repeated statement $t=s-p$ from this paragraph to improve readability. **(2) The deliberate design of Eq. (3) in the paper. The use of the intersection is intentional.** CTS compares the within-region attention shares assigned to the same token on the two sides, providing a more fine-grained comparison. Here, shared tokens refer to token identities present in both the input and generated contexts, rather than only to repeated segments in looping outputs. For a token that appears on only one side, there is no corresponding occurrence on the other side. If it is included by assigning zero mass to the absent side, the additional difference term equals its attention share in the region where it occurs. This incorporates differences between the two token sets into CTS. **Such differences may arise from new concepts, computational steps, or answer content introduced during normal reasoning and do not indicate abnormal attention allocation.** We have added an experiment including non-shared tokens in Appendix D.2 of the paper, "Effect of Including Non-Shared Tokens in CTS." The results demonstrate that **the features of persistent looping become completely mixed with those of normal requests**. Its score also varies with generation length, whereas the shared-token score remains stable, making it difficult for the metric to capture the characteristics of output patterns.
AuthorsOpenReview ↗

Response to Reviewer 91mx (2/4)

W2&Q2: Data-source and generation-length confounds

In our study, length does not directly determine the results. Its role is explained in the following two points.

(1) Providing a signal of the generation stage. Figure 4 in the paper shows that PAS increases with generated position, with fluctuations in its rate of increase; multiple output patterns begin with normal output. Consequently, the slope represented by the classifier feature f1f_1 may vary with generation length, allowing the features of different samples to separate. We use log⁡ki\log k_i as an auxiliary progress feature to supplement the PAS slope with information about the current generation stage.

(2) Decisions do not depend on abnormally long outputs. In the original results, under P(O)+P(L)≥0.5P(O)+P(L)\geq0.5, the mean first threshold-crossing positions across attack groups range from 60.6 to 997.0 tokens, below the mean completion lengths of normal D/R trajectories reported for the models in the paper (1,184.48–2,119.74 tokens). This indicates that risk signals can appear within the length range in which typical normal tasks have not yet finished. To further examine classification within normal generation lengths, we additionally report the results at 512 generated tokens.

At this stage, RADAR already achieves a high abnormality detection rate with relatively few false positives, indicating that its classification strategy does not rely heavily on output length. Length alone cannot support effective classification within the length range of normal requests. We therefore believe that the existing observations support the rationale for risk monitoring based on attention dynamics together with generation progress and provide a basis for further evaluating the early-warning effectiveness of this design. We have added this analysis to Appendix H of the paper, "The Role of Generation Length in Risk Detection."

Table: Detection performance at 512 generated tokens.

ModelCumulative detection rate ↑Normal-task false-positive rate ↓
DeepSeek-R1-Distill-Llama-8B84.6%0.0%
DeepSeek-R1-Distill-Qwen-14B84.2%3.9%
QwQ-32B87.7%12.3%
Qwen3.6-27B89.4%10.6%
GLM-4.7-Flash69.0%21.8%

Cumulative detection rate: the proportion of harmful trajectories satisfying P(O)+P(L)>0.5P(O)+P(L)>0.5.

Normal-task false-positive rate: the proportion of normal-request trajectories satisfying P(O)+P(L)>0.5P(O)+P(L)>0.5.

AuthorsOpenReview ↗

Response to Reviewer 91mx (3/4)

W2&Q3: Feature contributions and intervention ablations

Full RADAR outperforms the single-component and random-layer settings overall. Methodologically, our attention features provide complementary discrimination. The temporal patterns of PAS across reasoning states, presented in Section 4.3 and Figure 4 of the paper, help distinguish whether the model is engaging in reflection, with a lower slope indicating stronger reflection. The same section and Figure 5 in the paper show that CTS distinguishes normal from uncontrolled states through differences in the attention shares of shared tokens. The two features provide complementary dimensions of discrimination, and either alone is insufficient for comprehensive, effective classification.

Experimentally, the combined intervention achieves better overall suppression. We conducted additional component ablations on DeepSeek-R1-Distill-Llama-8B. Neither PAS-only nor CTS-only defense matches the overall performance of the full method. Random layer selection yields an overall loop rate of 41.3%, remaining less effective than our layer localization. A more complete description has been added to Appendix G.1 of the paper, "Intervention Component Ablation."

Table: Comparison of component ablations.

SettingLoopLLM ↓Recur ↓MiP ↓Overall ↓
Undefended56.0%56.0%56.0%56.0%
PAS-only28.0%52.0%52.0%44.0%
CTS-only24.0%52.0%52.0%42.7%
Random layers28.0%52.0%44.0%41.3%
RADAR (Full)12.0%52.0%44.0%36.0%

Comments on the paper

1 comment
Area_Chair_spmvOpenReview ↗

Reminder: Author–Reviewer Discussion Deadline

Dear Reviewers,

This is a friendly reminder that the author–reviewer discussion period for ICLR 2027 will end on November 18, 2026 (AOE).

As the deadline approaches, I encourage you to actively engage in discussions with the authors and carefully consider their rebuttal responses. Please address any remaining concerns, seek further clarification where necessary, and update your assessments if appropriate.

Your timely and constructive participation is essential to ensuring a fair and well-informed evaluation.

Thank you very much for your time and effort!

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.