acceptodds
Withdrawn from ICLR 2027

When Wrong Answers Remain Unsettled: Failure Dynamics in Recurrent Reasoners

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

35 views

Abstract

Weight-shared recurrent models allocate additional computation by iterating the same blocks, but a wrong terminal answer need not be a settled one. We measure answer persistence under continued recurrence in a trained Hierarchical Reasoning Model (HRM; 27M parameters, Sudoku-Extreme), comparing perturbed continuation with an unperturbed control. After one additional inference budget, mild wrong answers persist at a rate of (), versus for solved endpoints (); the rate pooled over wrong endpoints is . Perturbations spanning – the terminal-update scale have little effect on these rates. The measurement therefore primarily reflects finite-horizon answer persistence, rather than sensitivity to perturbation or asymptotic attraction. Solved HRM endpoints have a mean leading-Ritz radius of , while scheduled linear models predict fast-loop updates better than slow-loop updates (held-out versus ). We also evaluate established perturb-and-reselect procedures with paired gains, regressions and oracle headroom. The frozen noise recipe gains accuracy points ( gained, lost); compute-matched extended depth gains , and Sudoku symmetry augmentation gains . Across six evaluated settings, solved answers remain more persistent than wrong answers, but the fixed persistence threshold and HRM spectral and timescale patterns do not generalise universally. TRM and GRAM-guided TRM share a module across their two levels. The timescale ordering reverses on unsolved TRM states and in the GRAM-guided setting; in TRM both injection sites give similar gains. These results separate terminal-answer persistence, local geometry and reselection efficacy, and delimit which failure diagnostics transfer across recurrent models.

Primary area:
interpretability and explainable AI
Keywords:
recurrent-depth models, mechanistic interpretability, dynamical systems, hierarchical reasoning models, test-time compute, recurrent reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Stq5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper examines whether wrong answers produced by weight-shared recurrent reasoners have settled into stable states or continue to change when the models run beyond their training budget. Most of the analysis uses an HRM trained by the authors on Sudoku-Extreme. The authors continue the model from its terminal states, with and without perturbations, and analyse the endpoints using Jacobian spectra and linear models of the updates. They also compare perturb-and-reselect methods from prior work with compute-matched alternatives, and repeat the diagnostics with a second seed, TRM, GRAM-guided TRM, and two additional puzzle tasks, Maze and N-Queens. Across these settings, wrong answers are less persistent than correct ones. The spectral and timescale patterns observed in HRM, however, do not carry over to the other settings.

Critical strengths
  1. The paper asks a useful question about how recurrent-depth models fail as they are given more test-time compute. It is worth understanding whether a wrong terminal answer is stable or simply a state the model passes through. As the authors point out, neither a flat loss curve nor clusters in a PCA plot can establish this.

  2. The unperturbed continuation control is well chosen. It separates the effect of running the model longer from the effect of adding noise, and shows that noise contributes little. The paired accounting of gained and lost solutions, together with the compute-matched baselines in Sec. 8, tells us more than net accuracy alone. I also appreciate that the authors report negative results, including the failure of several HRM findings to transfer to other settings (Fig. 2).

Critical weaknesses

1. No main conclusion that is both supported and new. The only finding that holds in all six settings (Fig. 2) is that correct answers persist more than wrong ones. For models trained with a halting head to stop on correct solutions, this is expected. The paper's framing question, whether wrong answers sit in attractors, is not resolved by its measurement, which compares the decoded answer at only two readouts. A trajectory cycling among wrong answers would also look "non-persistent", and the long-horizon data in App. N fit this reading: wrong-answer persistence stays at about 0.43 through 2048 segments with no trend, and the solved fraction of the wrong class levels off at 0.28–0.30. The authors acknowledge that periodic dynamics and attractors are not excluded (App. N, Conclusion). The remaining findings (endpoint spectra, the fast/slow split, injection sites) do not transfer beyond HRM by the paper's own tally.

2. Limited novelty. Perturb-and-reselect is acknowledged as "established prior work" (Sec. 1), and the analysis tools (endpoint spectra, linear fits to updates, perturb-and-continue tests) are standard. The practical conclusions are also familiar: extra depth fixes some errors (+5.6/+6.0 points), and task symmetry augmentation dominates random noise (+34.6 vs. +4.8). The paper ends up recommending against the procedure it studies, without proposing an alternative.

3. The experiments are too small and too narrow for the general claims. All main-text analyses use one 27M-parameter HRM checkpoint trained for 20000 epochs on 1000 base puzzles. HRM is known to overfit in this regime (the TRM paper, which the submission cites, notes that HRM "tends to overfit quickly and then diverge" on Sudoku-Extreme and Maze-Hard), and no training accuracy is reported. The checkpoint reaches 41.76%, against 55.0% reported for HRM on this benchmark, and the Maze model reaches 10.7% against 74.5% (Table 4 of Jolicoeur-Martineau, 2025). Model and task are never crossed (TRM only on Sudoku, GRAM-TRM only on N-Queens), so failures of transfer cannot be attributed. With essentially two small architectures on small puzzles, the claims in the title and abstract about "recurrent reasoners" and transfer "across recurrent models" are not supported. I also note that the headline 0.42 for mild wrong endpoints drops to 0.178 under replay-consistent labels (App. H), but the abstract keeps 0.42.

Other Concerns: Presentation and Writing

The main text is a dense run of numbers and appendix pointers, and the content needed to follow the argument is spread over appendices A to V. Several sentences describe the authors' pipeline rather than the science (e.g. "Concordance is now emitted by every basin run", "Thirteen of the 63 job scripts behind these carry an interpretation rule committed before execution"). I was also confused that the paper states, in Sec. 2, App. U and its bibliography, that the full text of Efstathiou & Balwani (2026), the main work it argues against, was "unavailable", so that everything attributed to it comes from one sentence of the abstract. That cited work is also about TRM, whereas the headline evidence here is from HRM. Finally, almost every result is followed by a sentence limiting what it means (e.g. "does not certify a fixed point", "not an attractor definition", "We therefore do not claim noise injection as the best use of a test-time budget", "without excluding asymptotic attractors"). Each caveat may be fair, but together they leave little that the paper actually claims.

Questions for authors
  1. What is the single main conclusion of the paper, beyond "correct answers persist more than wrong ones"? Which finding holds in all six settings and, in your view, changes how these models should be built or used?
  2. Can you separate oscillation from exploration? For each wrong trajectory over the 2048-segment continuation, how many distinct decoded answers are visited, how often does an answer recur, and is there periodicity in the decoded answer or in zHz_H? If wrong trajectories cycle among a small set of answers, how is that different from the attractor description the paper argues against?
  3. What are the training and test accuracies of the analysed checkpoint, and how large is its generalization gap? Do the persistence and spectral results hold on a well-trained checkpoint (the released HRM checkpoint, or TRM at its reported accuracy), on a larger model, and on at least one non-puzzle task or another family of looped models?
  4. Why do the abstract and Sec. 4 keep the dump-labelled 0.42 when App. H gives a replay-consistent estimate of 0.178 for the same quantity?
  5. I’m very confused by the statement in your paper that you have only read the abstract of the work you are arguing against and have not read the full text. That's really weird.
Reviewer_R72aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the persistence of correct and incorrect terminal answers in recurrent reasoning models under continued inference, using a Hierarchical Reasoning Model (HRM) trained on Sudoku-Extreme as the primary testbed. The authors find that correct answers are substantially more persistent than incorrect ones and examine local spectral properties and temporal dynamics to characterize this difference. They further evaluate perturbation-based answer reselection against compute-matched alternatives and extend their diagnostics to additional architectures and tasks. While the relative persistence of correct and incorrect answers is observed across six evaluated settings, the spectral and timescale patterns do not generalize consistently. Overall, the paper provides an empirical analysis of failure dynamics in recurrent reasoning models, although the generality of its mechanistic explanations remains limited by the reliance on a single primary checkpoint and inconsistent cross-setting results.

Critical strengths
  • The paper investigates whether incorrect answers produced by recurrent reasoning models remain stable under continued inference, providing a useful perspective on understanding their failure dynamics.

  • The authors extend their diagnostics to six evaluated settings and explicitly identify which findings generalize and which do not. The discussion of limitations helps avoid overclaiming the universality of the observed mechanisms.

Critical weaknesses
  • The core analysis is primarily based on a single HRM checkpoint trained on Sudoku-Extreme with seed 0. Although additional architectures and tasks are evaluated, the spectral and temporal-dynamics patterns do not consistently transfer across settings. This limits the evidence supporting the proposed explanations as general mechanisms of recurrent reasoning failures.

  • The observed relationships between local spectral properties, temporal dynamics, and answer persistence are primarily empirical and predictive. They do not establish that these properties causally explain why incorrect answers persist. The authors also acknowledge that the tested linear models have limited predictive power, leaving the underlying mechanisms insufficiently understood.

Questions for authors
  • Can the authors provide evidence that the observed spectral or timescale properties causally contribute to incorrect-answer persistence, rather than merely correlating with it?

  • The paper reports that the relative persistence ordering between correct and incorrect answers transfers across six settings, but the fixed persistence threshold does not generalize universally. How sensitive are the main conclusions to the choice of persistence criterion and its threshold? Can the authors demonstrate that the findings remain robust under alternative threshold definitions or evaluation protocols?

  • The core mechanistic analysis is primarily conducted on a single HRM checkpoint trained on Sudoku-Extreme with seed 0. Given that the spectral and timescale patterns do not consistently transfer across settings, how can the authors justify the generality of their proposed explanations? Could the authors provide additional evidence across multiple training seeds, model checkpoints, or recurrent reasoning architectures?

Reviewer_4EPuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors study answer persistence in recurrent, weight-sharing reasoning models, in particular, the HRM. They find that incorrect terminal answers are less persistent than correct ones.

Critical strengths
  • The results are believable and the methods seem generally sound.
  • The authors identify that there are some differences across different recurrence-based reasoning models using geometric measures, which could have implications for how they might be improved.
Critical weaknesses
  • The work is incremental and doesn't address any particular gap. While the authors have conducted analyses that are different from the exact ones carried out by the authors of the cited works, the takeaways as I understand it aren't anything that we've not already seen and understood about these models before, i.e., saying that wrong answers are less persistent isn't that different from saying correct answers are arrived at early and stay that way, while incorrect answers can be corrected through the addition of stochasticity at answer initialization during inference.
  • Much of the writing is unedited output from LLMs, which does make parsing through the paper a less than pleasant experience.
    • E.g., what is the mean-leading Ritz radius, and what does it signify? The authors mention it in the abstract and introduction surrounded by a few statistics but provide no intuition as to what the takeaway is in particular other than there being a difference; Why is this the right measure? What does it measure?
  • Likewise, the choice of experiments and analyses conducted also suggest being primarily LLM-driven; that is not to say doing so is necessarily bad, but the current state of the work is largely devoid of any individuality or conviction.
Questions for authors
  • I am open to having my mind changed about the prosaic nature of the submission, which might have been lost on me due to its writing style and presentation. In particular, could the authors clarify:
    • Why knowing incorrect answers persist not as long as correct ones wasn't obvious from the fact that adding noise at initialization corrects wrong answers but doesn't change correct ones
    • How confirming this changes or deepens our understanding of recurrence-based reasoning models
    • Or if there is another significant contribution of the work that I may have missed or misunderstood

Unless the authors can convince me that their specific contributions do indeed add to our understanding of these models (and also significantly "humanize" the writing), I do not see myself changing my scores.

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

Related papers

Discussion (0)

Sign in to comment.