acceptodds
Under review as a conference paper at ICLR 2027

Decoding Dynamics Reveal Textual-Repetition Loops in Large Reasoning Models

Abstract

Large reasoning models can spend their generation budget repeating text without completing an answer. We study which information in the model's next-token predictions distinguishes these textual-repetition loops from successful reasoning. Across 25 models and three benchmarks, we compare six signals computed from the top- candidates recorded at each step: the concentration of the current prediction, and the recurrence of candidate probabilities, ranks, or identities at earlier steps that emitted the same token. For loops that have met a textual-repetition criterion, the recurrence of candidate identities, measured by set overlap, separates them from still-running successful generations most strongly, and controls show that candidate sets add information beyond the text-recurrence signals we evaluate. Accumulated with a moving average, the overlap yields an online detector that reads no hidden states and trains no classifier; its window and threshold are chosen on validation problems. Under greedy decoding, it alerts before the criterion on 85.3% of test-split loops at a 0.73% false-positive rate on successful generations, averaged over models. On nine models, prompting a final answer at its alerts reduces loops by 91.3% and generated tokens by 55.5%, while 99.89% of previously correct answers remain correct. Candidate-set recurrence is thus an observable of established repetition and a selective trigger for intervention under greedy decoding.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large reasoning models, decoding dynamics, repetition loops, efficient reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8R8EOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper detects textual-repetition loops by comparing top-k candidate sets at historical positions that emitted the same token. Experiments across 25 models evaluate detection, while interventions on nine models demonstrate reduced loops and token consumption under greedy decoding.

Critical strengths

The evaluation covers multiple model families and includes controls for generation length, token counts, and repetitive successful generations. Reporting answer retention alongside accuracy helps assess intervention harm.

Critical weaknesses
  1. The contribution primarily establishes an empirical association between candidate-set recurrence and textual loops. The experiments mainly identify which signal recognizes established repetition, without sufficiently investigating the origin of this association or providing deeper insights into loop formation and persistence. The scientific contribution therefore remains limited.

  2. Compared with text-counter force-close, accuracy improves by only 0.33 percentage points, with essentially the same token consumption. The large headline savings are relative to no intervention, while differing history-reset policies further complicate attribution of the incremental benefit.

  3. End-to-end intervention benefits are demonstrated only under greedy decoding. Under recommended sampling, loops are substantially rarer and early detection deteriorates; intervention outcomes under sampling and computational overhead relative to the text counter remain insufficiently evaluated.

Questions for authors
  1. What further evidence explains the source of candidate-set recurrence and its relationship to loop formation or persistence?

  2. Does the advantage over text-counter force-close persist with matched operating points and identical history-reset policies?

  3. Under recommended sampling, what are the net effects on accuracy, answer retention, token consumption, and runtime?

3 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1: Study scope and empirical contribution

We sincerely thank you for the time and care you devoted to reviewing our paper, as well as for your thoughtful feedback and the specific concerns you raised. Your comments have helped us clarify the scope of our findings and identify key comparisons that warranted closer examination.

In this work, we did not investigate the underlying mechanisms of loop formation and persistence, nor the origin of candidate-set recurrence. Instead, our goal was to leverage next-token probability distributions to build a practical, on-the-fly loop detection algorithm and verify its generalizability across diverse models and tasks. To this end, we evaluated our method on 25 models and three datasets using a shared detector configuration. Furthermore, Appendix C demonstrates that our signal distinguishes loop trajectories from those that eventually yield correct answers, even after both have met the textual-repetition criterion—indicating that this separation captures more than a simple binary repetition check. This practical algorithm and its broad empirical validation constitute our core contribution, while we leave a deeper investigation into the underlying mechanisms to future work.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2: Token savings with matched history handling

Thank you for highlighting the need to compare the two methods under the same history-handling policy.

With matched operating points and identical history handling, Jac still used fewer generated tokens than the text counter. With history retained and with history reset, it generated 0.48M (95% interval 0.14–0.80M) and 0.32M (0.04–0.59M) fewer tokens, respectively. Jac had a slightly higher PASS rate under both history settings, but the 95% intervals for the differences included zero. Retention was at least 99.8% in every condition.

As you noted, the paper used different history handling for the two methods: Jac retained detector history, whereas the counter (K = 10) reset it. To examine the effect of this difference, we evaluated both methods with history retained and with history reset. Each setting had to follow its no-intervention counterpart up to its first detection, so we used two models whose outputs are reproducible regardless of batching: DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B. Under retained history, the detector continues using its accumulated state after each insertion: Jac's same-token candidate-set history and moving average, or the counter's occurrence counts of 30-token spans. Under reset history, that state is cleared whenever an insertion ends and accumulated again from subsequent generation. After detection, all four settings used the same force-close action: inserting the end-of-thinking delimiter and continuing answer generation.

The counter threshold was chosen on validation by scanning K = 2–30 and taking, as the representative, the smallest K whose number of false positives on correct trajectories does not exceed that of Jac (K = 15 and K = 13, respectively). This is the earliest-first-alert operating point on the same untreated trajectories among the K values satisfying the validation false-positive constraint. Every smaller K exceeded Jac's false positives on validation.

PolicyHistory handlingPASS (%) ↑Retention (%) ↑Tokens (M) ↓
Text counterRetain64.3299.819.04
Text counterReset64.48100.008.88
Ours (Jac)Retain64.9999.918.56
Ours (Jac)Reset64.9499.918.55

PASS and Retention are as in Table 1 of the paper, and the table pools the two models' test results (974 questions each, 1,948 generations in total). Intervals in the text are question-level 95% bootstrap intervals. Generated tokens include answer generation after force-close.

Jac's results were nearly unaffected by history handling (PASS 64.99% vs. 64.94%; 8.56M vs. 8.55M generated tokens), whereas the counter was slightly better on PASS, Retention, and generated tokens when its history was reset. Thus, in this experimental setting, Jac held the advantage over the counter regardless of history handling, generating fewer tokens while maintaining similar PASS and Retention.

AuthorsOpenReview ↗

Response to Weakness 3 and Question 3: Intervention outcomes and runtime under recommended sampling settings

Thank you for asking how the method performs under recommended sampling settings, including its computational cost.

Overall, the gains from stopping loops were modest under the recommended sampling settings because loops were rare. Detection and intervention nevertheless worked under these settings.

We evaluated Qwen3.5-4B on the 974 test questions three times each under sampling (2,922 outputs in total). The same seed was applied to all methods for each question and replicate, and there were 13 loop trajectories under no intervention.

PolicyPASS (%) ↑LOOP (%) ↓Retention (%) ↑Token change (%) ↓Time change (%) ↓
No intervention79.880.44100.00——
Text counter (K = 10)79.910.1799.70−0.54+1.55
Ours (Jac)79.980.07100.00−0.10+7.11

PASS, LOOP, and Retention are as in Table 1 of the paper, and token and time values are changes relative to no intervention. The items above are reported by pooling the three replicates. Jac uses the paper's setting, the text counter detects a trajectory when the same 30-token span has occurred 10 times, and both methods force-close while retaining detector history.

Jac retained all baseline-correct answers and detected 11 of the 13 baseline loop trajectories, 3 of which became correct. With the text counter, 4 became correct, but 7 baseline-correct answers were lost, 5 of which became loop trajectories.

Runtime was measured on four NVIDIA B300 GPUs with vLLM 0.28.0 and PyTorch 2.13.0 (CUDA 13.0), processing one request at a time per GPU. Each GPU ran the same subset of questions with all methods. Jac's runtime increased by about 7% relative to no intervention. Extracting candidate sets and comparing them with history adds computation at every decoding step, while rare loop trajectories leave little generation to save.

Reviewer_uM3uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates textual-repetition loops in large reasoning models. The authors find that the overlap of top-k next-token candidates is a useful signal for detecting repetitive generation. Based on this observation, they develop an online detector and use it to terminate unproductive reasoning through forced answer generation.

Critical strengths
  1. The detector is simple and easy to implement.

  2. The experiments cover different model families and 3 benchmarks.

Critical weaknesses
  1. How useful is the method under more frequent inference settings? Most of the experiments are under greedy decoding. But under the model’s recommended sampling settings, the loop rate drops from 18.1% to 0.71%. It seems that the loop rarely exists in the latter settings. The promising results come from a non-recommended decoding regime. Please clarify.

  2. The design of early detection is not convincing. In Section 2.1, a loop is defined as a 30-token span repeated at least 20 times, and the authors claim that their method can detect the loop before the 20th repetition. However, if the loop is detected at the 15th repetition, can this really be regarded as early detection? In my opinion, it is necessary to distinguish whether the method captures a genuine signal or merely recognizes the repetition pattern of the loop itself.

Questions for authors

Please see weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Weakness 1: Recommended sampling settings and the role of greedy decoding

We sincerely thank you for the time and care you devoted to reviewing our paper and for your thoughtful and constructive feedback. We appreciate your recognition of our detector's simplicity and the breadth of our evaluation across model families and benchmarks. We are also grateful for your careful questions about the practical relevance of the approach and the interpretation of early detection.

Your comments highlight issues central to this study's motivation and the scope of its findings. We welcome the opportunity to discuss these points in detail and hope that our responses clarify the rationale for our experimental choices and the evidence supporting our conclusions.

Even under the recommended sampling settings, candidate-set Jaccard-based detection (Jac) detected some of the remaining loop trajectories while retaining the answers that were correct without intervention, and some of these led to correct answers after intervention. We evaluated Qwen3.5-4B on the 974 test questions three times each under sampling (2,922 outputs in total). The same seed was applied to all methods for each question and replicate, and there were 13 loop trajectories under no intervention.

PolicyPASS (%) ↑LOOP (%) ↓Retention (%) ↑Token change (%) ↓Time change (%) ↓
No intervention79.880.44100.00——
Text counter (K = 10)79.910.1799.70−0.54+1.55
Ours (Jac)79.980.07100.00−0.10+7.11

PASS, LOOP, and Retention are as in Table 1 of the paper, and token and time values are changes relative to no intervention. The items above are reported by pooling the three replicates. Jac uses the paper's setting, the text counter detects a trajectory when the same 30-token span has occurred 10 times, and both methods force-close while retaining detector history.

Jac retained all baseline-correct answers and detected 11 of the 13 baseline loop trajectories, 3 of which became correct. With the text counter, 4 became correct, but 7 baseline-correct answers were lost, 5 of which became loop trajectories. Because loop trajectories themselves were rare, the changes in overall accuracy and token usage were small. Jac's runtime increased by about 7% relative to no intervention, because the cost of comparing the current candidate set with its history at every decoding step was relatively larger than the generation saved by intervening on rare loop trajectories.

We next explain why we chose greedy decoding. This strategy has been used in LLM evaluation on benchmarks whose answers are logically determined, such as those in STEM. Our study set out to address the textual-repetition loops that arise under this established evaluation condition. Sampling introduces randomness into token selection, and penalty-based methods act directly on the logits and the probability distribution; both can therefore affect a model's reasoning path and final performance. Greedy decoding evaluates a model without such random selection or penalty adjustment.

The aim of our study was to make reasoning models practical to use under greedy decoding as well, by preserving normally terminating trajectories as far as possible while detecting and stopping trajectories in which meaningless repetition persists, thereby saving computational resources. We also showed that reusing the reasoning already carried out, rather than discarding it, yields a correct answer for about half of the interrupted loop trajectories (Section 5). Our study does not claim superiority over sampling; rather, it shows that the candidate token set also carries an observable signal associated with loops and that this signal can be used for detection and selective intervention.

AuthorsOpenReview ↗

Response to Weakness 2: Detection timing and the textual-repetition criterion

The primary aim of the method described in our response to W1 is to detect loops in which reasoning repeats without meaningful progress and fails to terminate. However, because a generation trajectory observed only up to a finite length cannot establish whether the reasoning would continue indefinitely, a practical decision criterion was required. We therefore adopted the criterion used by [1], namely 20 occurrences of the same 30-token span, including overlapping occurrences (hereafter, the textual-repetition criterion). We classify a generation trajectory that meets this criterion and is unfinished under the paper's termination rules (Appendix B) as LOOP, and hereafter call an individual trajectory in this class a "loop trajectory." This is a reproducible decision criterion.

Accordingly, early detection in our paper does not mean prediction before repetition begins; it means detection before this 20-occurrence criterion is reached. The paper itself states as a limitation that its results cannot show predictability before repetition begins (Section 7). In practice, detection takes place after textual repetition is already under way: when the point at which our method detects a loop trajectory is expressed as the occurrence count of the same 30-token span used by the criterion, the median is 8 to 10 occurrences depending on the model analyzed. As noted in the Introduction (Section 1), however, textual repetition arises naturally in the course of reasoning and does not in itself indicate that the reasoning is meaningless. For example, even trajectories that reach a correct answer and terminate normally can repeat text when restating equations or earlier conclusions. In short, because a small number of repetitions alone cannot determine whether the reasoning will turn into a loop, detection that occurs after textual repetition has appeared is not inappropriate.

Given that detection follows repetition, it is necessary to verify that our method is not simply capturing textual repetition in a different way. To this end, Appendix C evaluates whether, among trajectories that meet the same textual-repetition criterion, loop trajectories can be distinguished from trajectories that subsequently complete with a correct answer. As a representative experiment, loop trajectories that had already met the textual-repetition criterion at 8,192 tokens were compared with 52 controls that had met the criterion by the same point but later completed with a correct answer (results at other lengths are reported in Appendix C). Here, a counter that tests only whether the textual-repetition criterion is met detects both groups, whereas the candidate-set signal achieves an AUROC of 0.847 (95% interval 0.743–0.896). Had the candidate-set signal merely captured fulfilment of the textual-repetition criterion in a different way, it could not have distinguished two groups that both meet that criterion. That it does distinguish them therefore indicates, indirectly, that the candidate-set signal also captures information beyond fulfilment of the criterion. This result, however, confirms discriminative power once repetition is already in progress; it does not establish a causal mechanism linking the proposed signal to the onset or persistence of repetition. Identifying such a causal mechanism is left for future work.

References

[1] Charilaos Pipis, Shivam Garg, Vasilis Kontonis, Vaishnavi Shrivastava, Akshay Krishnamurthy, and Dimitris Papailiopoulos. Wait, Wait, Wait... Why Do Reasoning Models Loop? ICML, 2026.

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

The paper studies textual-repetition loops in large reasoning models (LRMs), i.e., generations that repeat the same text until the token limit without producing an answer. Rather than modifying token scores or training classifiers on hidden states, it asks whether the top-k next-token predictions already recorded by the serving engine can distinguish loops from successful reasoning. It finds that Jaccard overlap of candidate sets at recurring tokens is the strongest such signal, and that a simple moving-average detector with one shared configuration provides a training-free trigger for force-closing loops under greedy decoding.

Critical strengths
  1. Practical problem, lightweight method. Loops waste a large share of compute. The detector needs only top-k log-probs: no hidden states and no trained classifier.
  2. Rigorous evaluation. a strict validation/test split with one frozen configuration, macro averaging, and bootstrap CIs.
Critical weaknesses
  1. Marginal gain over text-only triggers. Jac saves only ~1.6k more tokens per loop than the repetition criterion and beats a K = 10 text counter by just +0.33 pp PASS. The claim of information "beyond text recurrence" lacks a strong continuous text baseline (e.g., repeated-suffix length).
  2. Limited early-detection evidence and scope. AUROC is computed only after the criterion is met, and lead time relative to repetition onset is not reported. Results are mostly greedy-only: under sampling, EDR drops to 52.4% and no intervention is tested.
  3. Calibration and novelty. Per-model FPR varies from 0% to 5.2% with no sensitivity analysis. The method is close to Look-back decoding and CUSUM-based monitors.
Questions for authors
  1. How does Jac compare with a continuous text-only statistic (e.g., longest repeated-suffix length) under the same selection rule?
  2. Under recommended sampling and on high-loop models (e.g., Gemma-4-12B, Qwen3.5-4B), how do the criterion and force-close perform? Do intervention results change when reset behavior is matched?
3 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1: Comparison with a continuous text baseline

We sincerely thank you for the time and care you devoted to reviewing our paper. We appreciate your thoughtful feedback and your specific questions about the evidence supporting our claims. Your comments have prompted us to compare candidate-set and text-based detection more directly and clarify the scope of our findings on textual-repetition loops.

With the same selection rule, Jac achieved a higher early detection rate (EDR) and higher potential token savings (PTS) than the continuous text statistic you suggested, longest repeated-suffix length. EDR measures detection before the textual-repetition criterion. The text statistic is the length of the longest suffix of the current output that also occurred earlier, allowing overlapping occurrences.

For this comparison, we applied the paper's selection rule (Appendix A) to both methods on three models: DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Llama-8B, and Qwen3.5-4B. On validation, we selected the setting with the highest EDR among those with no higher FPR and no lower PTS than the textual-repetition criterion, without reselection on test.

PolicyFPR (%) ↓EDR (%) ↑PTS ↑
Longest repeated-suffix length0.44880.9425,390
Ours (Jac)0.50094.9625,657

Metrics follow the paper and are evaluated on the three-model test split; PTS is in tokens per loop trajectory. Jac gained 14.02 percentage points in EDR (95% interval 11.98–16.10) and 266 tokens in PTS (163–387), while the FPR difference (+0.05 percentage points, −0.17 to +0.30) had an interval including zero. Thus, the requested continuous-baseline comparison also shows more detections before the criterion and greater potential savings.

The approximately 1.6k-token saving difference you mention should likewise be considered alongside false positives on correct trajectories. It is the gain over the textual-repetition criterion in the paper's 25-model evaluation. A text counter needed an FPR of at least 1.90% to save as many tokens, compared with Jac's 0.73% (Appendix D). In the paper's nine-model intervention comparison, the differences from the counter in PASS (+0.33 percentage points) and Retention (+0.27 percentage points) also had 95% intervals excluding zero (Appendix E.2).

Separately from detection performance, we checked whether the signal merely reflects whether the repetition criterion is met. Appendix C distinguishes loop and correct trajectories that have already met the same criterion, with AUROC 0.847. This supports separation beyond the binary information of criterion fulfillment.

Conversely, some trajectories are detected without meeting the criterion. In this example (Qwen3.5-4B, MATH-500 validation), an enumeration changing only a number continues to the 32,768-token generation limit:

I will ensure the answer is not 1961.
I will ensure the answer is not 1962.
I will ensure the answer is not 1963.

This trajectory is labelled OPEN because it does not meet the criterion, but Jac detected it at token 9,297 with the paper's configuration unchanged. A text method could also handle this type through number normalization; Jac detected it without that rule. We do not generalize this example to all text methods: our claim remains “beyond the text-recurrence signals we evaluate” (Section 1).

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2: Early detection, recommended sampling, and history handling

To test whether separation is visible before the criterion, we split loop trajectories at each generation length by whether they had met the textual-repetition criterion and compared them with correct trajectories. We used greedy trajectories from three models: DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Llama-8B, and Qwen3.5-4B.

Comparison (AUROC)5122,0488,19216,384
Criterion-met loop vs. correct–0.959 [0.915, 0.991]0.935 [0.873, 0.977]0.982 [0.954, 0.997]
Pre-criterion loop vs. correct0.505 [0.485, 0.526]0.547 [0.525, 0.568]0.671 [0.611, 0.729]0.875 [0.763, 0.947]
Criterion-met vs. pre-criterion loop–0.937 [0.881, 0.981]0.871 [0.845, 0.895]0.782 [0.742, 0.818]
Count (criterion-met loop / pre-criterion loop / correct)0 / 2,110 / 2,88344 / 2,066 / 1,5651,025 / 1,085 / 3331,873 / 237 / 61

Trajectory labels use the full output; brackets are 95% bootstrap intervals. Correct trajectories had not met the criterion by that length.

At 512 tokens, pre-criterion loop trajectories show little separation from correct ones; at 16,384 tokens, AUROC is 0.875. This is evidence of separation before the criterion, not prediction before repetition begins. At first detection, the median occurrence count of the criterion's 30-token span was 8 to 10, depending on the model.

These results concern greedy decoding; under recommended sampling, we assessed detection and intervention separately. The EDR of 52.4% uses the shared configuration selected under greedy decoding; re-selecting the threshold within an acceptable FPR can raise EDR. We evaluated intervention on Qwen3.5-4B (974 test questions, three replicates; 2,922 outputs) and Gemma-4-12B (974 test questions, once). Baseline loop trajectories were rare: 13 and 4, respectively.

ModelPolicyPASS (%) ↑LOOP (%) ↓Retention (%) ↑Token change (%) ↓
Qwen3.5-4BNo intervention79.880.44100.00—
Qwen3.5-4BText counter (K = 10)79.910.1799.70−0.54
Qwen3.5-4BTextual-repetition criterion (K = 20)79.980.24100.00−0.16
Qwen3.5-4BOurs (Jac)79.980.07100.00−0.10
Gemma-4-12BNo intervention85.110.41100.00—
Gemma-4-12BText counter (K = 10)84.501.1398.67−1.41
Gemma-4-12BTextual-repetition criterion (K = 20)85.010.4199.76−0.81
Gemma-4-12BOurs (Jac)85.110.21100.00−0.47

Metrics follow Table 1; token changes are relative to no intervention. Jac uses the paper's setting; the counter detects 10 occurrences of the same 30-token span. K = 20 uses the repetition criterion itself as the trigger. All intervention policies retain detector history and force-close by inserting the end-of-thinking delimiter and continuing answer generation.

On Qwen3.5-4B, Jac detected 11 of the 13 baseline loop trajectories, of which 3 became correct, while retaining all baseline-correct answers. With the counter, 4 baseline loop trajectories became correct, but 7 baseline-correct answers were lost, 5 of which became loop trajectories. On Gemma-4-12B, Jac detected 2 of 4 and retained all baseline-correct answers; the counter lost 11 and increased loop trajectories from 4 to 11.

Both methods retain history in the sampling comparison above. To check history handling further, we compared retain and reset under greedy decoding on two models whose outputs match no intervention until first detection: DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B. Retain keeps detector state after insertion (Jac's candidate-set history and moving average; the counter's span counts); reset clears it after each insertion ends. The counter's K was the smallest value with no more false positives on validation correct trajectories than Jac: 15 and 13, respectively.

PolicyHistory handlingPASS (%) ↑Retention (%) ↑Tokens (M) ↓
Text counterRetain64.3299.819.04
Text counterReset64.48100.008.88
Ours (Jac)Retain64.9999.918.56
Ours (Jac)Reset64.9499.918.55

The table pools both models' test results (1,948 generations). Jac was nearly unchanged by history handling (PASS 64.99% vs. 64.94%; tokens 8.56M vs. 8.55M), whereas resetting the counter slightly improved all three metrics. In this setting, Jac generated 0.48M and 0.32M fewer tokens under the two history policies (question-bootstrap 95% intervals 0.14–0.80M and 0.04–0.59M), while maintaining similar PASS and Retention.

AuthorsOpenReview ↗

Response to Weakness 3: Sensitivity analysis and relation to prior methods

Thank you for asking us to clarify calibration sensitivity and the relationship to prior methods.

Sensitivity to the threshold and averaging window is reported in Figure 3, and to the candidate-set size in Figure 5. The variation in model-specific false-positive rates arises because we applied one common configuration to every model, without per-model calibration, in order to evaluate whether a single setting carries across models. Three models exceed a 2% false-positive rate (Qwen3-4B, gpt-oss-20b and Qwen3-8B), while 20 of the 25 models are at or below 1% (Table 4). In practical use, the false-positive rate should be checked for the target model and the threshold calibrated where needed.

We share with Look-back decoding [1] the idea of comparing past predictions. Look-back compares next-token probability distributions by KL divergence. However, signals that compare probability distributions (TV, JS) can remain high even when the alternatives differ, because positions that emit the same token share the top-ranked candidate and its probability dominates (Section 7). Jaccard overlap, which compares candidate membership, separates better than TV and JS at all four generation lengths (Figure 2), and we adopted it on that basis. Detection is the primary goal of this work, and different interventions can be applied afterward. Interrupting generation, inserting the end-of-thinking delimiter, and continuing answer generation is one such example.

The CUSUM-based monitor in [2] accumulates the scores of a linear classifier trained on hidden states. We use CUSUM as an existing statistic to accumulate the candidate-recurrence score (Section 2.3); the difference lies in what is observed. The contribution of our study is therefore not the invention of historical comparison or of an accumulator, but the empirical identification of which prediction information is useful for detecting loops.

References

[1] Nan Xu, Chunting Zhou, Asli Celikyilmaz, and Xuezhe Ma. Look-back Decoding for Open-Ended Text Generation. EMNLP, 2023.

[2] Zenghao Duan, Liang Pang, Zihao Wei, Wenbin Duan, Yuxin Tian, Shicheng Xu, Jingcheng Deng, Zhiyi Yin, and Xueqi Cheng. Circular Reasoning: Understanding Self-Reinforcing Loops in Large Reasoning Models. arXiv, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.