acceptodds
Under review as a conference paper at ICLR 2027

Latent Trajectory Steering Unlocks Trapped Capacity in Recursive Reasoning

Abstract

Recursive reasoning models such as HRM and TRM achieve strong performance on structured reasoning tasks with far fewer parameters than LLMs, yet their inference dynamics are not well understood. We analyze their latent trajectories from a dynamical-systems perspective and find distinct stable basins for successful and failed rollouts, with high-stretching finite-time Lyapunov exponent (FTLE) ridges concentrated near transitions between them. Based on this observation, we propose a training-free inference method that identifies these ridge locations along the unperturbed trajectory, applies controlled perturbations to induce basin crossings, and aggregates the resulting trajectories through Gibbs-weighted selection. Without updating model parameters or using ground-truth labels, our method improves the accuracy of TRM in Sudoku-Extreme from 87.4% to 98.43% and in Maze from 85.3% to 91.0%, with consistent gains in HRM. These results identify dynamical trapping as a major failure mode in recursive reasoning: redirecting a failed latent trajectory across a high-FTLE boundary can recover the correct solution without retraining the model. Compared to existing alternatives, such as PTRM, our method locates a more precise point of intervention in the latent reasoning trajectory, such that much fewer rollouts and augmentation are needed.

Primary area:
interpretability and explainable AI
Keywords:
Recursive reasoning, Attractor Dynamics, Finite-time Lyapunov exponents, Test-time steering

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_Dv7COpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper focuses on the direction of recursive reasoning models and investigates the problem of attraction-basin trapping. It proposes a test-time intervention method that does not require retraining the model—a training-free guided intervention approach. Specifically, it computes the FTLE along the unperturbed original reasoning trajectory, identifies locally unstable regions, and applies targeted perturbations at these steps to encourage the model to escape erroneous attraction basins.

Critical strengths
  1. The paper is well written and complete, and the experiments are sufficiently thorough.

  2. The proposed method is intuitive and concise, and it is very solid.

  3. The paper connects the failure problem of recursive reasoning models with attraction-basin trapping in dynamical systems, which provides the paper with a certain theoretical basis and an analyzable perspective.

Critical weaknesses
  1. The writing is somewhat misleading, as it gives the impression that the paper uses recursive reasoning models and FTLE to solve LLM reasoning problems, whereas LLMs are actually introduced to illustrate the advantages of recursive reasoning models. However, since recursive reasoning models also require a series of methods such as those you propose to address their problems, I think this model actually has similar issues. The tasks used in the experiments seem to be ones that LLM agents can already perform very well. I think the advantage of recursive reasoning models lies in their smaller parameter size and lower cost. The paper may need to be rephrased.

  2. From my rough understanding, the proposed method can be regarded as adding perturbations to steer the model away from incorrect convergence points. Do you think the most important difficulty of the method is identifying the so-called attraction-basin trapping? Can we identify it from experimental results?

  3. Does the proposed method or idea also apply to other deep models? For example, if I find that a model performs poorly or keeps getting stuck at a certain error (e.g., an LLM keeps repeating a sentence), I add some perturbation to it. What do you think is the most distinctive feature of your method compared with this kind of practice?

Questions for authors

I am very sorry, but I have no knowledge of recursive reasoning models. I have included my questions and concerns in Critical Weaknesses. My comments are for reference only.

4 comments

We sincerely thank the reviewer for taking the time to carefully understand our work, despite coming from a different research background. We particularly appreciate your strong expertise in LLMs and the valuable insights you have offered from a broader, cross-disciplinary perspective. Your comments have helped us recognize important opportunities to improve the motivation, positioning, and presentation of our work. We believe that addressing these points will make our contributions more accessible and relevant to a broader research community.

1. Motivation for studying TRM and relevance to LLMs

We thank the reviewer for highlighting the need to clarify our motivation and the relationship between TRM and LLM reasoning. We agree that our original presentation did not sufficiently distinguish the broader motivation from the specific model and setting investigated in this work.

Our motivation is a general phenomenon also observed in LLMs: different inference procedures can produce substantially different outcomes from the same trained model, as demonstrated by alternative CoT decoding and self-consistency (Wang and Zhou, 2024; Wang et al., 2023). This raises the question of how inference dynamics may limit the utilization of a model's learned capabilities.

We study this question in TRM because its explicit recurrent hidden-state dynamics allow us to track trajectories, quantify their local sensitivity using FTLE, and experimentally investigate targeted interventions. We do not claim to solve LLM reasoning problems directly, but believe that understanding this mechanism in a tractable model may offer insights relevant to more general reasoning systems.

We also agree that TRM's compact architecture and computational efficiency are important motivations for its study. Our method introduces additional test-time computation, and we do not claim to retain the inference cost of vanilla TRM. Rather, our contribution is to improve how this additional computation is allocated, achieving better performance than perturbation-based alternatives such as PTRM under comparable computational budgets. The benchmarks serve as controlled testbeds for studying these inference dynamics, rather than demonstrating superiority over LLM agents.

We will revise the introduction to clarify this motivation, the scope of our contribution, and its potential relevance to the broader reasoning community.

2. Experimental identification of attraction-basin trapping

Your "rough" understanding captured the gist of our work accurately. And yes, the most important/challenging point addressed in our method for untapping the trapped capacity is exactly how to identify the attraction-basin, including when to intervene, such that the probability of finding better reasoning routes is higher. Experimentally, we have provided complementary evidences of its identification in three forms:

  • Persistence under continued recursion. In Table 8, deterministic TRM accuracy on Maze remains approximately 85% from 3 through 16 ACT steps. Appendices F.2–F.3 show individual examples where the baseline remains incorrect while a perturbed trajectory reaches the correct solution.
  • Geometry separating different outcomes. Figure 3 shows an interior FTLE maximum between successful and unsuccessful trajectories at matched reasoning steps. Figure 7 presents the corresponding pattern across five Sudoku examples. These observations support the interpretation of a sensitive transition region between different attraction basins.
  • Controlled intervention evidence. Table 4 demonstrate the positive effect on choosing the point of intervention with FTLE, it varies the timing while keeping the rollout budget, noise scales, and selection procedure fixed.

Although our steering also uses an ensemble of rollouts (so these are improvements with additional inference computation) its ensemble outperformed other methods by using FLTE to decide when to perturb. To support this, we provide a detailed comparison of computational costs across TRM, PTRM, and our method. We evaluate all three methods on Maze, Sudoku, and PPBench using the same ACT steps for each dataset. The results are summarized in the table below.

DatasetMethodAccuracyTFLOPs/exampleGPU-hoursMean peak allocated GPU memory (MiB)
MazeOurs (FTLE-max)91.00%209.013.092123,836.18
MazePTRM88.90%210.193.072183,671.33
MazeTRM85.30%3.820.09461102.95
SudokuOurs (FTLE-max)98.50%13.170.17644185.55
SudokuPTRM98.20%13.070.15570162.77
SudokuTRM85.10%0.260.0504438.38
PPBenchOurs (FTLE-max)71.43%167.590.08530262.49
PPBenchPTRM65.31%168.240.09320225.10
PPBenchTRM53.06%1.650.0198254.96
ARC-AGI-2(pass@5)Ours (FTLE-max)7.50%9,927.0218.857624,778
ARC-AGI-2(pass@5)PTRM5.00%9,912.9017.860324,730
ARC-AGI-2TRM6.67%381.271.157834,708.56

In this new experiment, owing to the high computational cost of ARC-AGI-2, we performed cost accounting on a subset, where our method and PTRM incur comparable overheads; notably, PTRM falls below the vanilla TRM baseline, suggesting that its persistent noise injection may lead to over-perturbation.

FTLE itself measures sensitivity, thus intervening at its maximum make it more likely to generate alternatives belonging to different basins, from which our Gibbs-weighted selection would have a better chance.

Together, these evidences support the interpretation that some failures arise from the inference trajectory rather than an absence of usable capacity in the trained model and the point of intervention to acquire correct trajectories is important.

Related phenomena have also been identified in language models. Holtzman et al. report a repetition rate of 73.66% under greedy decoding versus 0.36% under nucleus sampling (better selection) with p=0.95p=0.95, using the same GPT-2 Large model. This demonstrates how strongly an inference procedure can affect repetitive failure behavior without changing model parameters. More directly, Wang et al. observe attractor cycles during successive LLM paraphrasing: their two-step periodicity score reaches 0.81 for GPT-4o and 0.92 for Qwen2.5-72B in English. These scores measure recurrence of similar text, not error rates (Wang et al., 2025, Table 1). However, these work still leave open on what's a better framework to release the potential of LLMs during inference, which also support our broader motivation.

3. Distinction from generic perturbation and applicability to other models

As the reviewer correctly points out, introducing perturbations to escape undesirable inference outcomes and explore alternative trajectories is a natural and widely used strategy, as exemplified by PTRM. However, the dynamical mechanisms underlying its effectiveness, particularly when and why perturbations are most useful, remain insufficiently understood.

Our key contribution is to provide a dynamics-based understanding of this strategy and use it to guide more effective interventions. Specifically, FTLE identifies regions of high dynamical sensitivity, suggesting where perturbations may be particularly effective.

Importantly, our controlled ablations demonstrate that FTLE-guided intervention consistently outperforms first-step and random-step perturbations, with rollout budgets, noise scales, and selection procedures held fixed (Table below). This provides direct evidence that dynamical information improves intervention timing, and helps explain why indiscriminate perturbation throughout inference may be suboptimal.

TaskACTK (rollouts)FTLE-maxFirst-stepRandom-stepPTRM
Maze-TRM165091.00%87.20%88.70%88.90%
Sudoku-TRM105098.40%97.63%97.70%97.53%
PPBench-TRM4810071.4%65.31%69.39%69.39%
ARC-AGI-210258.33%6.67%7.50%5.83%

We believe this perspective may also offer insights into broader reasoning systems. For example, multi-agent LLM reasoning explores alternative solution paths through parallel exploration and discussion. Our findings suggest that not only the diversity of explored trajectories, but also when to initiate such exploration, may be important. Extending this principle to LLMs remains an interesting direction for future work.

We sincerely thank the reviewer again for their thoughtful and constructive feedback, which has helped us strengthen our work and its positioning. We hope the our clarifications and additional experiments have adequately addressed the concerns and respectfully ask the reviewer if they could reconsider their overall assessment in light of the new evidence.

Reviewer_egk4OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies HRM and TRM as latent dynamical systems and argues that failed reasoning trajectories can become trapped in incorrect attractor basins. It proposes a training-free method that uses finite-time Lyapunov exponents (FTLE) to select an intervention point, injects latent perturbations to generate alternative trajectories, and chooses an output using task-specific trajectory energies and Gibbs-weighted basin aggregation. The method improves TRM accuracy on Sudoku-Extreme from 89.19% to 98.43% under the matched 16-ACT setting and on Maze-Hard from 85.1% to 91.0%. Additional results are reported on PPBench and ARC-AGI-2, together with intervention, noise-scale, and energy-component ablations.

Critical strengths
  1. Viewing recursive reasoning failures as attraction to spurious basins is insightful and offers a useful framework for understanding HRM/TRM inference.
  2. The method produces substantial gains on Sudoku and Maze using fixed pretrained checkpoints and no ground-truth labels during inference.
  3. The empirical analysis is broad. The paper evaluates two architectures and four benchmark families, and includes intervention-point, noise-scale, energy-component, and ACT-budget analyses.
Critical weaknesses
  1. FTLE’s independent contribution is unclear. On Maze, FTLE-max reaches 90.1%, while perturbing at the first step already reaches 89.6%. A matched random or first-step perturbation baseline using the same rollout budget and final selector is needed across all tasks.
  2. A maximum FTLE value along one trajectory is not itself an FTLE ridge or basin boundary. The projected diagnostics are suggestive, but the paper acknowledges that they do not verify the full variational LCS conditions.
  3. Performance is partly driven by task-specific selection. Maze selection uses handcrafted connectivity, graph, and path-length rules. The ablation shows that these rules account for most of the improvement, which weakens the claim that FTLE steering is the primary source of the gain.
  4. Computing FTLE over many latent states with up to 50 JVP/VJP power iterations, followed by 50–100 rollouts, may be expensive. Comparing only the number of rollouts with PTRM does not establish overall efficiency.
  5. PPBench contains only 49 selected examples and one seed, while ARC-AGI-2 gains are small and pass@1 is often unchanged. The method also performs very poorly at low ACT budgets, suggesting sensitivity to inference settings.
Questions for authors
  1. How does FTLE-max compare with first-step, random-step, and uniform-noise interventions on all tasks, using exactly the same noise, rollout budget, and Gibbs selector?
  2. What is the total runtime, memory use, and number of model-equivalent forward/backward operations per example relative to TRM and PTRM?
  3. How were the task-specific energy weights, noise scales, FTLE windows, entropy gates, and inverse temperatures selected? Was a separate validation set used?
  4. Can the authors quantify how reliably an FTLE peak predicts a successful basin transition, rather than showing only selected successful examples?
  5. How much performance remains if the handcrafted Sudoku/Maze validity rules are removed and one common model-internal energy is used across every benchmark?
10 comments

We sincerely thank you for your thoughtful and constructive review, as well as the time and effort you devoted to understanding our work. We greatly appreciate your insightful feedback and welcome this opportunity to discuss your comments further. Below, we address your questions point by point, hoping to clarify any concerns and further improve the quality of our work.

Response to Weakness 1 (Unclear FTLE marginal value) and Question 1 (Matched-budget perturbation comparison).

We respectfully argue that this concern precisely highlights the importance and necessity of our theoretical account. Prior work has observed that noise perturbation can improve TRM/HRM performance, but the underlying mechanism has remained unclear. Our FTLE-based analysis provides a principled explanation: perturbations applied along critical trajectory segments make it easier for the model trajectory to cross basin boundaries and meet with the untapped larger pool of attraction basins available from the trained model, where more often the correct answers reside. Perturbations applied after the trajectory has entered the correct basin disrupt TRM's convergence to the correct answer, and those that applied before the trajectory approaches the basin boundary can damage the current TRM reasoning path and disrupt the landscape already established during training, which accounts for the performance gap between PTRM methods (Sghaier et al., 2026) and our approach.

The first-step perturbation results are particularly informative. In practice, the FTLE maximum often occurs in the first few reasoning steps rather than exactly at the first step. Therefore, sustained perturbation from the first step actually covers the range that includes the FTLE-max perturbation. Even under this favorable condition for the first-step baseline, FTLE-guided perturbation still performs better. This indicates that although the intuition that noise perturbation improves model performance has been empirically validated, a sufficient theoretical account remains necessary and suggests a potentially more efficient perturbation strategy. We further tested a variant that perturbs only at the first step and found substantial accuracy drops across all evaluated datasets. We will elaborate on this point in the revised manuscript.

We tested a stricter variant perturbing only the first step, rather than sustaining perturbation through the low-level cycle. Under the noise scale, rollout budget, and Gibbs selector (See table below), FTLE-max consistently outperforms first-step, random-step, and PTRM with the same selection across all tasks. Since Sudoku performance is already high and the remaining headroom is limited, the intervention strategies yield similar results on this task. The accuracy drop of single-step first-step perturbation supports our account: perturbing before approaching the basin boundary disrupts the reasoning path, whereas targeting the FTLE peak concentrates perturbation where basin switching is most likely.

TaskACTK (rollouts)FTLE-maxFirst-stepRandom-stepPTRM
Maze-TRM165091.00%87.20%88.70%88.90%
Sudoku-TRM105098.40%97.63%97.70%97.53%
PPBench-TRM4810071.4%65.31%69.39%69.39%
ARC-AGI-210258.33%6.67%7.50%5.83%

Response to Weakness 2 (LCS-boundary interpretation).

We thank the reviewer for highlighting this important distinction. While rigorously verifying LCS conditions in high-dimensional latent spaces is challenging, our LCS interpretation is not based solely on a scalar FTLE maximum, but is supported by complementary geometric and experimental evidence. Moreover, the LCS perspective provides valuable insights that motivate our approach.

Specifically, (i) Fig. 3(a) provides local geometric diagnostics, including Cauchy–Green stretching and projected directional alignment; (ii) Fig. 3(b) and Appendix Fig. 7 reveal persistent interior high-FTLE structures between correct and incorrect trajectories, consistent with LCS-like boundaries; and (iii) Table 4 demonstrates that FTLE-max intervention outperforms alternative intervention points, supporting its practical effectiveness. Additional comparative examples are provided above.

Together, these findings provide converging evidence for our LCS-inspired interpretation and demonstrate that FTLE can effectively identify dynamically sensitive regions for targeted trajectory steering. While a rigorous LCS proof remains beyond the scope of this work, we will further clarify this distinction in the revised manuscript.

Response to Weakness 3 (Task-specific selection confound) and Question 5 (Performance without handcrafted rules).

We thank the reviewer for this important concern. We would like to clarify it with three points.

First, task-specific rules are not a universal dependency of the method. On PPBench and ARC-AGI-2, our selector uses no handcrafted rules at all; on Sudoku-Extreme, the weight of the rule energy term is zero in the final configuration, and performance comes solely from model-internal stability signals and Gibbs aggregation. Therefore, the gains on Sudoku and PPBench/ARC-AGI cannot be explained by rule-based selection.

Second, rules do contribute substantially on Maze, and we report this faithfully in Appendix F.1. However, Maze is a prototypical multi-solution task: the same maze often admits many topologically different yet valid paths that satisfy start--goal connectivity, and model-internal energy alone struggles to distinguish structurally sound candidates from degenerate or fragmented ones. Handcrafted rules here perform structural filtering among multiple valid solutions, rather than ``manufacturing'' the correct answer. In other words, rules address selection under multi-solution ambiguity, whereas FTLE steering addresses whether correct-basin candidates can be generated at all. The two act at different stages and are complementary rather than substitutive. Even with the same rule-based selector fixed, the contribution of FTLE still remains.

Third, to directly address the concern that performance may be driven mainly by the selector, we compare on Maze using exactly the same frozen answer selector and identical hyperparameters, under a setting where PTRM consumes more normalized computation. On Maze dataset, Our method reaches 91.0%, while PTRM reaches 88.9%. This shows that, after controlling for the selector and the computation budget, the gain still comes from the FTLE-guided intervention position rather than from the selector or extra computation.

MethodKExact accuracy
TRM baseline---85.3%
PTRM + full selector5088.9%
Our method, full selector5091.0%
Reviewer_BoGhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work studies the model paradigm that repeatedly updates internal states as in Hierarchical Reasoning Model (HRM) and Tiny Recursive Model (TRM), targeting global-constrained tasks such as Sudoku and Maze. Such models can settle on an incorrect answer and fail to improve with further iterations. This work propose finding a point where the internal computation is especially sensitive to perturbations, measured using the finite-time Lyapunov exponent (FTLE). They inject noise there to generate alternative solutions, then choose an answer using weighted voting based on task validity and model-internal signals. The complete procedure substantially improves Sudoku and Maze performance without training involved.

Critical strengths
  • The method attains non-trivial improvement over plain recursive models on Sudoku and Maze by changing only their inference. Improvements on both HRM and TRM suggest that the effectiveness can be general beyond a single model. It is overall useful since it offers a way to improve without repeating training.

  • The paper proposes an interesting way to address how sensitive the internal computation is to perturbation by a concrete strategy for choosing when to explore alternative solutions, which can be a worthwhile hypothesis that could inform future inference methods for recurrent models.

  • Experiments provide useful information about the complete procedure with rich graph presentation. The paper investigates intervention timing, perturbation magnitude, and answer selection, rather than reporting only the final accuracy. In particular, the selection ablations reveal the importance of recognizing valid solutions after generating diverse candidates. The authors also acknowledge limitations of their geometric interpretation and the difficulty of selecting correct answers on ARC, which helps clarify what the method currently achieves.

Critical weaknesses
  • Some of the conceptual findings are not unique to this work: prior work by Ren and Liu identifies incorrect stable internal behaviors that produce wrong answers; PTRM shows that injecting latent noise and selecting among multiple trajectories can recover correct solutions without retraining. The contribution is mainly the strategy to use FTLE to choose when to perturb, rather than discovering trapped capacity or demonstrating that stochastic exploration can recover it.

  • For the core contribution of FTLE, the experiments do not clearly justify its importance. In the controlled timing comparison, perturbing at the first step already reaches 89.6% Maze accuracy, while FTLE-based timing reaches 90.1%. Separate ablations show that checking Maze rules is crucial for selecting good answers. Whether the proposed sensitivity analysis offers a substantial benefit over a simpler combination of early perturbation and effective answer selection remains unclear.

  • The overall experimental settings also require more work to achieve the significance that the paper claims.The visualizations suggest that sensitive regions separate trajectories leading to different answers, but do not establish that crossing such boundaries explains the improvements. Moreover, FTLE itself requires additional computation, so fewer sampled trajectories do not necessarily mean a more efficient method. The additional benchmarks provide limited evidence of a clear advantage over PTRM.

Questions for authors

When all methods use the same answer selector and total computation budget, how much does FTLE improve over PTRM and simple early perturbation?

5 comments

We sincerely thank the reviewer for the constructive feedback and for recognizing the strengths of our work. During the rebuttal, we conducted additional experiments to address the critical weaknesses raised in the review.

Response to Weakness 1 (Conceptual novelty).

We thank the reviewer for highlighting these relevant studies, which provide important foundations for our work. However, we believe our contribution addresses a distinct and fundamental question beyond these prior findings.

Specifically, Ren and Liu characterize correct and spurious attractors through fixed-point and convergence analyses, but do not investigate how dynamical sensitivity evolves along reasoning trajectories or how it relates to transitions between attractor basins. PTRM demonstrates that stochastic perturbations can recover correct solutions, but does not provide a dynamical explanation for why and where perturbations are most effective.

Our central novelty is bridging this gap: connecting local dynamical sensitivity, attractor-basin geometry, and targeted intervention through an FTLE-based framework. This provides both a mechanistic interpretation of stochastic recovery and a principled strategy for improving it, supported by controlled experiments.

Thus, our main contribution is establishing a dynamical framework for understanding and systematically exploiting the stochastic recovery of the model's trapped capacity beyond a demonstration.

Response to Weakness 2 (Unclear marginal value of FTLE).

We respectfully argue that this concern precisely highlights the importance and necessity of our theoretical account. Prior work has observed that noise perturbation can improve TRM/HRM performance, but the underlying mechanism has remained unclear. Our FTLE-based analysis provides a principled explanation: perturbations applied along critical trajectory segments make it easier for the model trajectory to cross basin boundaries and meet with the untapped larger pool of attraction basins available from the trained model, where more often the correct answers reside. Perturbations applied after the trajectory has entered the correct basin disrupt TRM's convergence to the correct answer, and those that applied before the trajectory approaches the basin boundary can damage the current TRM reasoning path and disrupt the landscape already established during training, which accounts for the performance gap between PTRM methods (Sghaier et al., 2026) and our approach.

The first-step perturbation results are particularly informative. In practice, the FTLE maximum often occurs in the first few reasoning steps rather than exactly at the first step. Therefore, sustained perturbation from the first step actually covers the range that includes the FTLE-max perturbation. Even under this favorable condition for the first-step baseline, FTLE-guided perturbation still performs better. This indicates that although the intuition that noise perturbation improves model performance has been empirically validated, a sufficient theoretical account remains necessary and suggests a potentially more efficient perturbation strategy. We further tested a variant that perturbs only at the first step and found substantial accuracy drops across all evaluated datasets. We will elaborate on this point in the revised manuscript.

We tested a stricter variant perturbing only the first step, rather than sustaining perturbation through the low-level cycle. Under identical noise, rollout budget, and Gibbs selector (See table below), FTLE-max consistently outperforms first-step, random-step, and PTRM with the same selection across all tasks. Since Sudoku performance is already high and the remaining headroom is limited, the intervention strategies yield similar results on this task. The accuracy drop of single-step first-step perturbation supports our account: perturbing before approaching the basin boundary disrupts the reasoning path, whereas targeting the FTLE peak concentrates perturbation where basin switching is most likely.

TaskACTK (rollouts)FTLE-maxFirst-stepRandom-stepPTRM
Maze-TRM165091.00%87.20%88.70%88.90%
Sudoku-TRM105098.40%97.63%97.70%97.53%
PPBench-TRM4810071.4%65.31%69.39%69.39%
ARC-AGI-210258.33%6.67%7.50%5.83%

Response to Weakness 3 (Mechanistic evidence and experimental significance).

We thank the reviewer for this important concern. To further establish the connection between FTLE boundaries and reasoning outcomes, we provide additional evidence in Appendix Fig. 8.

The results show that decoding accuracy remains largely preserved within the correct-solution region but drops markedly across high-FTLE transition regions. Importantly, decoding accuracy is measured directly from the model's output predictions, so these changes reflect actual changes in the decoded answers, rather than merely latent trajectory separation. This provides further evidence linking high-FTLE regions to changes in reasoning outcomes.

This also highlights the complementary roles of FTLE-guided perturbation and selection: FTLE helps trajectories explore alternative solutions, while selection identifies the correct candidates. Crucially, selection cannot recover solutions that were never generated.

Regarding computational efficiency and benchmark comparisons, we refer the reviewer to our Response to Question 1 (Matched-budget comparison), which demonstrates advantages over PTRM under comparable computational costs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.