acceptodds
Under review as a conference paper at ICLR 2027

Scheduling Recursive Reasoning in Looped Transformers

Abstract

Recurrent reasoning models have attracted growing attention for scaling test-time computation, typically by iteratively refining latent states with shared parameters. However, these models apply each learned update with a fixed unit scale, which can be conservative when updates make persistent progress and overly aggressive when they fluctuate, limiting the benefit of additional loops. To understand how the scale should vary along the trajectory, we first analyze the sensitivity of terminal loss to recurrent update scale. We show that its temporal average admits an exact decomposition into *persistent-progress* and *centered-fluctuation* contributions. Based on this, we introduce the *Trajectory Adaptive Progress–Fluctuation Scheduler (TAPS)*, which tracks their balance across recurrent updates and adapts the step size online. Theoretically, we establish sufficient conditions under which TAPS reduces expected terminal loss and reaches a target quality in fewer recurrent loops. Empirically, we show that TAPS improves terminal accuracy across structured reasoning tasks without retraining. By further incorporating the progress–fluctuation principle into training, TAPS yields additional accuracy gains with up to 1.56x wall-clock speedup at matched baseline accuracy. The broad applicability of TAPS is supported by its effectiveness across diverse recurrent architectures and inference strategies. Together, these results establish update scale as a complementary control axis of recurrent inference alongside architecture and depth.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Recursive Reasoning, Looped Transformer

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_rW2dOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the adaptive update of looped transformers. To address the limitation of fixed update each step,this paper proposes TAPS to determine the update with persistent-progress and centered-fluctuation. Extensive experiments shows TAPS improves the accuracy and efficiency on Sudoku and Maze.

Critical strengths
  1. This paper treats update magnitute as an independent variable for recurrent inference, which fills the gap of the research of architecture and loop count.
  2. TAPS can be appiled to pretrained models without training.
  3. Theoretical analysis provides the necessary condition for TAPS.
  4. The experimental results include Sudoku and Maze on multiple variants of recurrent language models. The comparisons of different controller variants and inference methods are also conducted
Critical weaknesses
  1. Update consistency does really reflect task progress. TAPS may amplify the error if the model keeps on updating on wrong direction. There is a gap that could be filled with an experimental study of proxy alignment.
  2. Why looped transformers? It seems like this method can be applied to traditional transformers. Is there any specialty on looped transformers with TAPS?
  3. There is a lack of significance analysis, such as Maze-Hard in table 1 where the improvement on accuracy is tiny.
Questions for authors

See Weaknesses.

5 comments

We sincerely thank the reviewer for the helpful comments and for recognizing Clear and easy to follow, fills the gap of the research of architecture and loop count, be appiled to pretrained models without training, Theoretical analysis provides the necessary condition for TAPS, The inclusion of ... further broadens the empirical evaluation. Our point-by-point responses are as follows.

1. Validation of the Progress–Fluctuation Principle

Thank you for this insightful suggestion. We address this concern from three complementary perspectives: (1) empirically examining the alignment between the progress–fluctuation score and realized task progress; (2) demonstrating the practical utility of trajectory stability for adaptive inference; and (3) clarifying why correcting persistently misdirected updates requires additional task-dependent information beyond step-size control.

(1) Empirical alignment with realized task progress To examine whether the progress–fluctuation score reflects actual task progress, we evaluate TAPS on 1,000 Sudoku puzzles. For each puzzle ii and recurrent step hh, we measure the realized one-step loss reduction as Gi,h=Li,h−1−Li,hG_{i,h}=\mathcal L_{i,h-1}-\mathcal L_{i,h}, where positive values indicate progress. We then compute the Pearson correlation between B^i,h\widehat B_{i,h} and Gi,hG_{i,h} over the corresponding puzzle–step observations. The results show strong positive correlations:

Step2–162–96
Pearson r(B^,G)r(\widehat B,G)0.968 [0.955, 0.975]0.943 [0.922, 0.953]

Therefore, the progress–fluctuation score can be strongly associated with realized task progress across recurrent steps.

(2) Practical utility of trajectory stability We have also demonstrated the practical value of trajectory-based inference control in the original manuscript. Our TAPS-based adaptive-exit rule stops only when predictions remain unchanged and step-size-adjusted residual activity stays below a threshold for consecutive loops (Appendix C.3). As shown in Fig. 2, it allocates more computation to harder Sudoku instances while maintaining accuracy comparable to the unit-step baseline across difficulty levels. Notably, TAPS achieves 91.2% accuracy with only 316.5 effective updates, compared with 91.1% using 1,497 updates for FPRM, yielding a 4.7× reduction in computation. This demonstrates the usefulness of observable trajectory signals for adaptive recurrent inference.

(3) Why correcting update direction is not feasible from the information perspective? From an information perspective, determining whether an update is task-improving requires task-dependent information, such as the ground-truth target, the corresponding loss gradient, or external verifier feedback. Such information is unavailable during standard inference. As discussed in Appendix B.3.1, this is precisely why our theory distinguishes task-dependent oracle quantities from observable proxies constructed from recurrent updates.

TAPS instead builds on the premise that the balance between persistent motion and fluctuations provides a useful signal for regulating the magnitude of learned updates, without requiring every persistent update to be beneficial. The effectiveness of this principle is supported by our extensive experiments and the additional proxy-alignment analysis above. Moreover, Theorems 3–5 characterize sufficient conditions under which TAPS identifies loss-improving adjustment directions, reduces expected terminal loss, and reaches target performance with fewer recurrent steps. In the revised manuscript, we will clarify this distinction, together with the motivation and scope of TAPS.

AuthorsOpenReview ↗

Rebuttal (continued1)

2. Why Looped Transformers?

Thanks for this insightful question. We focus on looped transformers because their recurrent structure makes adaptive update-scale control particularly natural. Specifically, loop transformers repeatedly apply a shared operator to an evolving hidden state: Xk+1=Φθ(Xk;u)=Xk+Δk.X_{k+1}=\Phi_\theta(X_k;u)=X_k+\Delta_k. This shared operator produces a sequence of recurrent updates, allowing TAPS to track their persistence and fluctuations and adapt their magnitudes accordingly: Xk+1=Xk+ηkΔk.X_{k+1}=X_k+\eta_k\Delta_k. In contrast, standard transformers typically apply different operators across layers and do not exhibit the same shared-operator recurrent dynamics. Moreover, looped transformers have emerged as an important direction for scaling test-time reasoning by reusing model parameters across recurrent steps. Improving their inference efficiency is therefore a timely problem for the community.

We also appreciate your observation that TAPS can extend beyond models with native recurrence. Indeed, as shown in Table 4, we have already applied TAPS to a frozen Qwen3-4B-Instruct model with intermediate-layer recurrencem and results demonstrate the applicability of TAPS beyond models originally designed with native recurrence. Since this observation suggests that TAPS can have a broader impact, we will highlight this broader applicability in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal (continued2)

3. Significance of the Imrovement Brought by TAPS

Thank you for raising this concern. To demonstrate that the improvement brought by TAPS is significant, we first provide significance tests on the evaluation metrics including Acc. and Speedup. Then we show in which situations TAPS can bring greater advantages.

(1) Significance tests on Acc. and Speedup

To show the improvement is statistically significant, we add significance tests on Maze-Hard using 1,000 paired puzzles at H=16H=16, where Accuracy p-values are computed by one-sided paired McNemar tests and Timing p-values are based on seven CUDA timing repeats.

MethodAcc. (%)Acc. p-valueGPU SpeedupTiming p-value
Standard78.8—1.000×—
TAPS (GD)79.80.0551.518×<0.001
TAPS (P/S-Sign)79.90.0681.566×<0.001
TAPS (Momentum)79.80.0551.517×<0.001
TAPS (RMSProp)79.70.1211.518×<0.001
TAPS (Adam)79.80.0971.498×<0.001
TAPS (BB)79.60.1281.476×<0.001

Generally speaking, the accuracy improvements of most variants of TAPS are significant under the nominal level 10%. Most importantly, All TAPS variants achieve substantially larger gains in inference efficiency (1.46–1.56× GPU speedups) and these improvements are all significant (p-value <0.001). These results are sufficient to support the advantage of TAPS, since our primary objective is to improve the accuracy–compute trade-off of recurrent inference.

(2) When TAPS would have more significant advantages

To reveal the greater potential of TAPS in several important situations, we add experiments of TAPS (RMSProp) on 1,000 Sudoku problems. Briefly speaking, we find that TAPS achieves more gains when we have more loop budgets or tasks become more challenging.

  • TAPS delivers larger efficiency gains with more recurrent loops. We extend the evaluation in Table 1 to larger loop budgets under the same setup. TAPS consistently improves accuracy under the same budgets, and its GPU-time speedup at matched accuracy increases from 1.15× to 1.34× as the loop budget grows from 32 to 96. Such observed trend indicates that the performance improvement may grow given larger loop budgets.
Budget HHAccuracy (%) (Base → TAPS)Matched-Acc. Loops (Base → TAPS)GPU Speedup
3290.4 → 91.132 → 271.15×
4891.7 → 92.448 → 371.26×
6492.6 → 93.264 → 511.22×
9693.6 → 94.596 → 701.34×
  • TAPS achieves more speedups when tasks are more challenging. We further evaluate inference efficiency at fixed accuracy targets. At 95% accuracy, TAPS achieves 1.96× and 1.80× GPU-time speedups on easy and medium Sudoku problems. On hard problems, TAPS reaches 90% accuracy in 400 loops, compared to 565 loops for the baseline, yielding a 1.38× speedup. These results show that the benefits of TAPS become more pronounced at higher accuracy targets and on more challenging problems.
Target Acc.DifficultyLoops (Base → TAPS)GPU Speedup
90%Easy3 → 30.98×
90%Medium33 → 301.08×
90%Hard565 → 4001.38×
95%Easy12 → 61.96×
95%Medium350 → 1901.80×
95%HardNot reached—

Note: Difficulty is defined by the number of empty cells: Easy (46–54), Medium (55–58), and Hard (59–64).

The above results together demonstrate that TAPS can indeed achieve better accuracy-efficiency tradeoff, where the improvements are statistically significant. These improvements are even more pronounced in several important scenarios.

Reviewer_2kvWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates adaptive update-scale control for recurrent reasoning in looped Transformers. The authors propose Trajectory Adaptive Progress–Fluctuation Scheduler (TAPS), which dynamically adjusts the magnitude of recurrent hidden-state updates based on the persistence and fluctuation of recent updates. The method can be applied to pretrained recurrent models without modifying their parameters. The paper further provides a theoretical analysis connecting trajectory statistics to terminal task loss under certain conditions and introduces a training–inference co-design strategy to encourage more stable recurrent dynamics. Experiments on Sudoku, Maze, recurrent language models, and intermediate-layer recurrence demonstrate improvements in accuracy and inference efficiency across several settings.

Critical strengths

S1. An interesting and lightweight approach to recurrent inference control.

The paper explores update-scale adaptation as an additional dimension of recurrent reasoning, complementing existing approaches that primarily focus on recurrent architectures or computation depth. TAPS adjusts hidden-state update magnitudes using observable trajectory statistics without requiring changes to pretrained model parameters. This provides a simple and potentially practical mechanism for controlling recurrent computation at inference time.

S2. Experimental evaluation across different recurrent architectures.

The authors evaluate TAPS on multiple settings, including TRM-based Sudoku and Maze solving, recurrent language models such as Ouro and Huginn, and intermediate-layer recurrence in Qwen3-4B. The results demonstrate performance improvements in several settings, particularly on GSM8K with Ouro-1.4B. The inclusion of inference-efficiency measurements and training–inference co-design experiments further broadens the empirical evaluation.

Critical weaknesses

W1. Insufficient justification for trajectory-adaptive scheduling.

The fixed-step baselines in Tables 1–3 only consider a few selected values (e.g., 0.8, 1.0, and 1.2). Without comparisons against properly tuned constant scales or predefined schedules, it remains unclear whether the improvements genuinely come from trajectory adaptivity.

W2. Limited validation of the progress–fluctuation principle.

The method assumes that persistent updates indicate beneficial progress, while fluctuations should be suppressed. However, consistent updates may also lead toward incorrect solutions, and fluctuations may reflect necessary corrections. The paper lacks direct empirical evidence connecting these trajectory statistics to actual task improvements, weakening the justification for its core mechanism.

W3. Modest and inconsistent empirical gains.

The improvements are relatively small on several benchmarks (e.g., at most +0.52 on Sudoku and +0.15 on Qwen3-4B average accuracy). Moreover, the gains on Ouro-1.4B are largely concentrated on GSM8K (Table 3), with negligible or negative changes on other tasks. This raises concerns about the robustness and generalizability of the reported improvements.

Questions for authors

Q1. Would TAPS still outperform properly tuned constant step sizes or predefined schedules, rather than only the selected fixed-step baselines?

Q2. Can the authors provide empirical evidence that the proposed progress–fluctuation statistics reliably predict beneficial step-size adjustments?

Q3. Why are the performance gains concentrated on certain tasks (e.g., GSM8K), while remaining marginal or negative on others? Can the authors provide evidence of their statistical reliability?

7 comments

We sincerely thank the reviewer for the helpful comments and for recognizing An interesting and lightweight approach to recurrent inference control, complementing existing approaches, provides a simple and potentially practical mechanism, results demonstrate performance improvements in several settings, The inclusion of ... further broadens the empirical evaluation. Our point-by-point responses are as follows.

1. Comparison with Tuned Constant Step Sizes and Predefined Schedules (W1 & Q1)

Thank you for raising this important concern. We clarify that the novelty of TAPS lies in adapting recurrent update scales online according to the evolving trajectory. To the best of our knowledge, there are no works to address this problem, and therefore, there is no established off-the-shelf schedule for this objective. Moreover, a single constant step size can be fundamentally inadequate even in simplified recurrent systems; see Proposition 7 in Appendix B.2. To further address this concern, we provide both a dynamical motivation for trajectory adaptivity and additional experiments on a broader range of constant and predefined schedules.

(1) Why are constant step sizes insufficient? As illustrated in Figure 1(b), recurrent updates initially become more aligned but gradually exhibit conflicting directions as iterations proceed. A constant step size cannot adapt to such evolving dynamics, motivating TAPS to adjust the update magnitude based on the progress and stability of the recurrent trajectory. More fundamentally, Appendix B.2 (Proposition 7) provides an example in which no constant step size can achieve the optimal terminal loss, while a nonconstant schedule can eliminate all error components.

(2) Additional empirical evidence To further address this concern, we evaluate five constant step sizes η∈{0.8,0.9,1.0,1.1,1.2}\eta\in\lbrace0.8,0.9,1.0,1.1,1.2\rbrace and two predefined schedules (linear and cosine), under the same experimental setup. The results show that TAPS consistently outperforms constant step sizes and predefined schedules across both loop budgets, achieving higher accuracy and greater inference efficiency. This is intuitive, as recurrent update dynamics evolve throughout inference, and TAPS adapts to these changes rather than following a fixed or predetermined schedule.

MethodAcc@32Speedup@32Acc@96Speedup@96
Constant 0.890.3%N/R94.1%1.157×
Constant 0.989.7%N/R93.1%N/R
Standard (1.0)90.4%1.000×93.6%1.000×
Constant 1.188.5%N/R92.1%N/R
Constant 1.286.1%N/R91.4%N/R
Linear (1.2→0.8)89.7%N/R93.2%N/R
Cosine (1.2→0.8)90.1%N/R93.7%1.021×
TAPS (RMSProp)91.1%1.145×94.5%1.326×

2. Validation of the progress–fluctuation principle (W2 & Q2)

Thank you for this insightful suggestion. We agree that update consistency does not necessarily imply task progress, as a model may consistently update in the wrong direction. However, our extensive experiments and theoretical analyses have shown that TAPS is effective in most cases. To further examine the progress–fluctuation principle, we measure the actual task progress on 1,000 Sudoku puzzles. Specifically, we define the loss reduction at step hh as Gi,h=Li,h−1−Li,hG_{i,h}=\mathcal L_{i,h-1}-\mathcal L_{i,h}. A positive Gi,hG_{i,h} indicates progress, while a negative value indicates that the update worsens the task loss. We then average the progress–fluctuation score B^i,h\widehat B_{i,h} and loss reduction Gi,hG_{i,h} across puzzles at each step, and correlate the resulting curves.

Step2–162–96
Pearson r(B^,G)r(\widehat B,G)0.968 [0.955, 0.975]0.943 [0.922, 0.953]

We observe strong positive correlations both during early inference ((r=0.968)) and across the full trajectory ((r=0.943)). These results provide empirical evidence that the proposed score is strongly aligned with realized task progress at the trajectory level.

AuthorsOpenReview ↗

Rebuttal (continued)

3. Modest and task-dependent empirical gains (W3 & Q3)

Thanks for this helpful question. We clarify that TAPS is primarily designed to improve the accuracy–compute trade-off of recurrent inference, rather than just to improve accuracy across tasks. Therefore, the improvements brought by TAPS should be evaluated from two aspects including both the accuracy and the speedup. Several results (e.g., Table 1) in the original paper have already demonstrated the advantages of TAPS. To fully address your concern, in what follows, we (1) show when TAPS can have more significant advantages, (2) answer Q3 by giving possible reasons about why gains concentrated on GSM8K, and (3) provide results to show that the improvements brought by TAPS are statistical significant.

(1) When TAPS would have more significant advantages?

To reveal the greater potential of TAPS in several important situations, we add experiments of TAPS (RMSProp) on 1,000 Sudoku problems. Briefly speaking, we find that TAPS achieves more gains when we have more loop budgets or tasks become more challenging.

  • TAPS delivers larger efficiency gains with more recurrent loops. We extend the evaluation in Table 1 to larger loop budgets under the same setup. TAPS consistently improves accuracy under the same budgets, and its GPU-time speedup at matched accuracy increases from 1.15× to 1.34× as the loop budget grows from 32 to 96. Such observed trend indicates that the performance improvement may grow given larger loop budgets.
Budget HHAccuracy (%) (Base → TAPS)Matched-Acc. Loops (Base → TAPS)GPU Speedup
3290.4 → 91.132 → 271.15×
4891.7 → 92.448 → 371.26×
6492.6 → 93.264 → 511.22×
9693.6 → 94.596 → 701.34×
  • TAPS achieves more speedups when tasks are more challenging. We further evaluate inference efficiency at fixed accuracy targets. At 95% accuracy, TAPS achieves 1.96× and 1.80× GPU-time speedups on easy and medium Sudoku problems. On hard problems, TAPS reaches 90% accuracy in 400 loops, compared to 565 loops for the baseline, yielding a 1.38× speedup. These results show that the benefits of TAPS become more pronounced at higher accuracy targets and on more challenging problems.
Target Acc.DifficultyLoops (Base → TAPS)GPU Speedup
90%Easy3 → 30.98×
90%Medium33 → 301.08×
90%Hard565 → 4001.38×
95%Easy12 → 61.96×
95%Medium350 → 1901.80×
95%HardNot reached—

Note: Difficulty is defined by the number of empty cells: Easy (46–54), Medium (55–58), and Hard (59–64).

(2) Why are the accuracy gains concentrated on GSM8K? (Q3)

Larger gains on GSM8K with Ouro-1.4B may partly reflect the nature of the task. Unlike MMLU, ARC-C, and HellaSwag, which use zero-shot label likelihoods, GSM8K requires multi-token chain-of-thought generation, potentially offering more opportunities for recurrent refinement. This is also consistent with Ouro-1.4B's strong performance on GSM8K in the original Ouro paper [1].

(3) Are the improvement statistically significant?

To show the improvement is statistically significant, we add significance tests on Sudoku using 1,000 paired Sudoku at H=96H=96, where Accuracy p-values are computed by one-sided paired McNemar tests and Timing p-values are based on seven CUDA timing repeats.

MethodAcc. (%)Acc. p-valueGPU SpeedupTiming p-value
Standard93.6—1.000×—
TAPS (RMSProp)94.50.0611.340×0.00391

We see that improvements of TAPS on both Acc. and Speedup are statistical significant under the nominal level 10%. We also conduct statistical tests on more results; see our response in 3. Significance of the Imrovement Brought by TAPS (1) Significance tests on Acc. and Speedup for Reviewer rW2d. These results together demonstrate that the improvements brought by TAPS are significant.

Reference [1] Zhu, Rui-Jie, et al. "Scaling latent reasoning via looped language models." arXiv preprint arXiv:2510.25741 (2025).

Thank you for the additional experiments, which partly address my concerns. I am not requesting further experiments, but would appreciate clarification on:

  1. Given the small accuracy gap over the best constant baseline (94.5% vs. 94.1%), and that the reported significance test uses the unit-step baseline, how strong is the current evidence for an advantage attributable to adaptivity?

  2. Please confirm whether the reported GPU timings include all trajectory-estimation and controller overhead.

  3. Do the FPRM comparisons use its original FPOpt adaptive damping rule? How do the existing comparisons distinguish the contribution of the scheduler from differences in the model and halting mechanism?

Reviewer_g7a9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary
  • The paper proposes adaptive update scales for recurrent reasoning models by incorporating information from the history of previous updates. It evaluates several update rules inspired by well-known gradient-based optimizers, including an Adam-like rule. It also experiments with adding a loss term during training to suppress fluctuations in the updates.
  • Experiments are conducted on reasoning benchmarks (Sudoku and Maze) with TRM, as well as LLM benchmarks such as GSM8K and MMLU.
Critical strengths
  • It is interesting that the proposed update rules can be applied without additional training. However, as noted in the weaknesses, the improvements are relatively small.
  • Also, the paper shows that training models with the proposed update rules and penalizing fluctuations in the updates helps them reach the target accuracy faster during inference.
Critical weaknesses

Overall, the improvements are marginal, except for the results on GSM8K. Basically, in most benchmarks, the improvement is within 1 pt. The sign of the improvement is consistent across benchmarks, so I do not think the improvements are just noise. However, in terms of the significance of the method, it remains unclear whether these modest gains justify the additional complexity and hyperparameter tuning.

Questions for authors
  • How are rho and gamma selected? The selection procedure is described for Qwen, but not for the other models.
  • How important is the fluctuation loss? The paper only shows experimental results with both the proposed update rule and the loss, so we cannot isolate the efficacy of the proposed fluctuation loss. Can the authors show SFT results with the proposed update rule but without the fluctuation loss?
5 comments

We sincerely thank the reviewer for the insightful comments and for recognizing interesting, applied without additional training, helps them reach the target accuracy faster during inference. Below, we address your concerns point by point.

1. Gains of TAPS v.s. Additional Complexity and Hyperparameter Tuning

Thank you for raising this practical concern.

(1) Additional complexity and hyperparameter tuning. We clarify that TAPS is deliberately lightweight: it introduces no learned parameters and additional model evaluations, which only performs scalar updates using recurrent updates that are already computed by the model. The metric Speedup is measured end-to-end, which have already included the additional time needed by TAPS. Regarding hyperparameter tuning, we would like to emphasize that each model locks in these hyperparameters after the tuning process, without the need for individual setting adjustments. Moreover, the tuning process is quite straightforward. Therefore, TAPS does not introduce excessive complexity, and the performance improvement is attributed to the design of the method itself.

(2) TAPS can be more powerful. We further explore when TAPS would have a more significant advantage. Briefly speaking, we find that TAPS achieves more gains when we have more loop budgets or tasks become more challenging. To show this, we provide additional experimental results of TAPS (RMSProp) on 1,000 Sudoku problems.

  • TAPS delivers larger efficiency gains with more recurrent loops. We extend the evaluation in Table 1 to larger loop budgets under the same setup. TAPS consistently improves accuracy under the same budgets, and its GPU-time speedup at matched accuracy increases from 1.15× to 1.34× as the loop budget grows from 32 to 96. Such observed trend indicates that the performance improvement may grow given larger loop budgets.
Budget HHAccuracy (%) (Base → TAPS)Matched-Acc. Loops (Base → TAPS)GPU Speedup
3290.4 → 91.132 → 271.15×
4891.7 → 92.448 → 371.26×
6492.6 → 93.264 → 511.22×
9693.6 → 94.596 → 701.34×

Note: Speedup is measured by GPU time at matched baseline accuracy on 1,000 Sudoku problems.

  • TAPS achieves more speedups when tasks are more challenging. We further evaluate inference efficiency at fixed accuracy targets to reflect practical deployment scenarios. At 95% accuracy, TAPS achieves 1.96× and 1.80× GPU-time speedups on easy and medium Sudoku problems. On hard problems, TAPS reaches 90% accuracy in 400 loops, compared to 565 loops for the baseline, yielding a 1.38× speedup. These results show that the benefits of TAPS become more pronounced at higher accuracy targets and on more challenging problems.
Target Acc.DifficultyLoops (Base → TAPS)GPU Speedup
90%Easy3 → 30.98×
90%Medium33 → 301.08×
90%Hard565 → 4001.38×
95%Easy12 → 61.96×
95%Medium350 → 1901.80×
95%HardNot reached—

Note: Difficulty is defined by the number of empty cells: Easy (46–54), Medium (55–58), and Hard (59–64).

By the way, TAPS was originally designed to improve inference efficiency rather than accuracy. However, its consistent accuracy gains suggest a broader insight: progress–stability is a simple yet effective principle for recurrent reasoning. Our co-design experiments further validate this insight, leading to substantially larger improvements. We believe this intuitive and empirically validated principle is itself a meaningful contribution to the community.

2. Details about Hyperparameter Tuning

Thanks for this question. In all TAPS experiments, we select hyperparameters via grid search. For TRM and Ouro, we search over ρ∈{0.1,0.2,0.3}\rho \in \lbrace 0.1, 0.2, 0.3 \rbrace and γ∈{1.0,1.5,2.0}\gamma \in \lbrace1.0, 1.5, 2.0\rbrace. For Huginn, we use ρ∈{0.02,0.04,0.06,0.08,0.10,0.20}\rho \in \lbrace0.02, 0.04, 0.06, 0.08, 0.10, 0.20\rbrace and γ∈{0.5,1,2,4}\gamma \in \lbrace0.5, 1, 2, 4\rbrace. We will include these details in our revised version.

Robustness to hyperparameters. We further evaluate the sensitivity of TAPS to ρ\rho and γ\gamma. As shown below, TAPS(RMSProp) maintains accuracy and efficiency gains under multiple hyperparameter settings. At H=96H=96, three configurations outperform the baseline, with speedups of up to 1.46×. This suggests that the benefits of TAPS are not limited to a single hyperparameter choice.

ρ\rhoγ\gammaAcc@32 (%)Speedup@32Acc@96 (%)Speedup@96
Standard—90.41.00×93.61.00×
0.22.091.11.17×94.51.35×
0.12.090.51.01×94.01.19×
0.21.590.91.01×94.81.46×

Note: Bold parameter values indicate the configuration used in the paper. Speedups are measured using GPU time at matched baseline accuracy.

AuthorsOpenReview ↗

Rebuttal (continued)

3. Ablation Experiments on the Fluctuation Loss

Thanks for this important suggestion. To show the effect of the fluctuation term in the SFT, we compare the TAPS with standard method (ηk=1\eta_k=1) under SFT with and without the fluctuation loss, respectively.

SFT LossHHStandard Acc.TAPS Acc.GPU Speedup
without fluctuation3291.6%92.3%1.16×
with fluctuation3292.1%93.5%1.20×
without fluctuation9694.9%96.0%1.49×
with fluctuation9695.2%96.3%1.74×

The above table shows that, for each HH, both standard method and TAPS achieve higher accuracy when SFT includes the flucatuation term. (e.g., for TAPS, HH=32, the accuracy is improved from 92.3% to 93.5%). Furthermore, the advantage of TAPS is more significant when SFT includes the flucatuation term (e.g, H=96H=96, the speedup is improved from 1.49x to 1.74x). These results show demonstrate the effectiveness of the proposed loss.

Dear Reviewer g7a9,

We sincerely thank you for the time and effort you have dedicated to reviewing our submission. As the discussion period will end on Nov 18 AOE, we would greatly appreciate any feedback you may have regarding our response. Please let us know if there are any remaining concerns or points that would benefit from further clarification.

Thank you for your time and consideration.

Best regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.