acceptodds
Under review as a conference paper at ICLR 2027

LoLa: Looped Latent Reasoning via Hidden State Feedback

Abstract

Large reasoning models (LRMs) have achieved strong performance on reasoning tasks through chain-of-thought (CoT) generation, yet the resulting trajectories are often lengthy and redundant, leading to substantial inference overhead. Recent work has explored latent reasoning to improve efficiency, but existing methods typically either generate latent representations in a single forward pass without iterative refinement or build recurrence into the model during large scale pretraining. We propose looped latent reasoning (LoLa), a framework that combines a lightweight latent reasoner with a frozen base large language model (LLM). The reasoner maintains a latent state and iteratively updates it using hidden state feedback from the base LLM. The final state is projected into latent thought tokens that condition answer generation. We train the reasoner using supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO). We evaluate LoLa with two base LLMs on five reasoning benchmarks. With DeepSeek-R1-Distill-Qwen-1.5B as the base model, LoLa outperforms existing efficient reasoning methods. With Qwen3-4B, LoLa improves pass@1 over its non-thinking mode from 54.07% to 58.64% and pass@4 from 65.78% to 72.36%, while also reducing inference latency. Once trained for a given base LLM, LoLa provides a plug-and-play reasoning mode that can be enabled on demand without modifying the base model, while explicit CoT remains available for harder problems. Code is available at https://anonymous.4open.science/r/loopedlatent_iclr2027.

Primary area:
foundation or frontier models, including LLMs
Keywords:
latent reasoning, large language models, efficient reasoning, chain-of-thought reasoning, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_4g9eOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LoLa, a framework for latent reasoning tuning with a recurrent refinement mechanism. A lightweight latent reasoner is looped with the frozen base LLM: the latent tokens are passed through the LLM, and the LLM’s hidden states are fed back to update the reasoner’s latent state. The final latent tokens condition the answer generation. The reasoner is trained with SFT followed by GRPO. Since the LLM backbone is frozen, LoLa can perform latent reasoning without modifying the target reasoning LLM. Experiments on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-4B show that it outperforms LRT.

Critical strengths
  1. The idea is clear and has the potential for enhancing and extending the mainstream latent reasoning LLMs.

  2. The method is modular and easy to implement. The base LLM is untouched, so explicit CoT is still available.

  3. It achieves better performance than LRT on both backbones.

  4. The paper is clearly written, and the appendix gives enough details of the architecture and training.

Critical weaknesses
  1. Almost everything in this paper is adopted from LRT except the recurrent mechanism, including the reasoning network, the learnable reasoning tokens, the training pipeline and the evaluation protocol. The novelty is therefore limited, and the feedback from the base LLM only brings about 2 points in the ablation (Table 3).

  2. The comparison across loop depths is unfair. As stated in Appendix B, LoLa is trained with T=4, but it is evaluated with different depths, which are out of distribution for the trained model. So the gain from T=1 to T=2 in Table 5 does not show the benefit of recurrence.

  3. The performance of LoLa (T=1) on Qwen3-4B is even worse than LRT.

Questions for authors

Why is LoLa (T=1) much worse than LRT, even though it is the closest to a single-pass reasoner?

1 comment
AuthorsOpenReview ↗

Depth-matched training and evaluation

We sincerely thank the reviewer for the thoughtful and constructive comments, and for recognizing LoLa’s potential and modular design. We appreciate the concerns regarding its methodological contribution and the comparison across loop depths. To address these points, we conduct additional controlled training experiments.

1. Methodological contribution.

Unlike LRT, which generates latent representations in a single forward pass, LoLa introduces iterative interaction between the latent Reasoner and the frozen base LLM. At each iteration, the current latent thought is processed by the base LLM, and its hidden states are projected into feedback memory and integrated by the reasoner through gated cross-attention to update the latent state. KV-cache reuse avoids recomputing prompt representations across iterations. This enables iterative latent refinement without modifying the base LLM or requiring recurrent pretraining. In the original feedback ablation, this mechanism improves average accuracy by 2.05 percentage points across five benchmarks, with gains on all five.

2. Depth-matched training and evaluation.

In the original experiments, Table 5 evaluates the same trained model at different inference depths. Its T=1 result does not represent an independently trained single-update model.

To address this issue, we independently train LoLa-1U, which performs one reasoner update throughout both SFT and GRPO. We also retrain LRT using its original single-pass architecture. All three models use the same frozen Qwen3-4B backbone, training data, random seeds, effective batch size, and training recipe: two SFT epochs (learning rate 1e-4), followed by 50 GRPO updates (learning rate 3e-5). Full LoLa uses T=4 during SFT and T=3 during GRPO and inference.

Before GRPO, the SFT-only Pass@1 scores are 44.55% for LRT, 44.62% for LoLa-1U, and 51.83% for LoLa.

Pass@1 and Pass@4 on Qwen3-4B (%).

MethodAMCMATH-500GSM8KLSATGPQAAverage
LRT@145.7873.4091.5838.7038.8957.67
LoLa-1U@146.9973.4091.5139.5738.8958.07
LoLa@148.1975.0092.3441.3041.4159.65
LRT@455.4281.2095.6058.7059.6070.10
LoLa-1U@456.6380.4095.8363.4859.6071.19
LoLa@461.4583.0096.1364.7862.6373.60

Under matched SFT+GRPO, LoLa-1U improves over LRT by 0.40 and 1.09 percentage points in average Pass@1 and Pass@4, respectively. Full LoLa further improves over LoLa-1U by 1.58 and 2.41 points, respectively.

These results clarify the original T=1 comparison: independently training a single-update LoLa model yields better performance than LRT, while iterative refinement further improves both Pass@1 and Pass@4.

We sincerely thank the reviewer again for these valuable suggestions. They motivate us to conduct additional controlled training experiments and clarify the depth-matched comparison. We hope these results address the reviewer’s concerns and provide a clearer understanding of LoLa’s contribution. We will incorporate these results and analyses into the revised manuscript.

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

The paper studies efficient latent reasoning and proposes LoLa, a framework that combines a frozen base LLM with a trainable latent reasoner. LoLa iteratively updates a fixed-length latent state using hidden-state feedback from the base LLM, and uses the final state to condition answer generation. The reasoner is trained through SFT followed by GRPO. Experiments with two backbones across different reasoning benchmarks report improved average accuracy over existing efficient reasoning methods and lower inference latency.

Critical strengths
  1. The motivation is intuitive and clear. Based on the limited opportunity for iterative refinement in single-pass latent reasoning, the paper explores whether hidden-state feedback from a frozen LLM can improve latent representations through repeated updates. This provides a practical direction for introducing iterative computation without modifying the base model or requiring recurrent pretraining.
Critical weaknesses
  1. The methodological contribution is incremental. LoLa extends an existing modular latent reasoning framework with looped hidden-state feedback, making the contribution primarily an empirical investigation of looped computation for reasoning tasks. The training pipeline also follows the standard combination of SFT and GRPO techniques.

  2. The comparison with prior methods is not fully controlled, raising concerns about its fairness. The paper does not clearly establish whether LRT receives equivalent SFT and GRPO training. This is consequential because GRPO improves LoLa’s average accuracy from 50.60% to 58.64% in Table 4. Without matched training conditions, it remains difficult to disentangle whether the reported gains come from the proposed looped-feedback mechanism or from differences in training.

Questions for authors
  1. Could the authors clarify whether LRT receives equivalent SFT and GRPO training, including the training data and optimization budget?

  2. Could the authors provide an accuracy-latency comparison with LRT, including the overhead of latent-state refinement, to establish the practical benefit of looped computation?

4 comments
AuthorsOpenReview ↗

Matched SFT+GRPO training and accuracy–latency analysis

We sincerely thank the reviewer for the thoughtful and constructive comments, and for recognizing the practical motivation of LoLa’s iterative hidden-state refinement.

1. Methodological contribution.

Unlike LRT, which generates latent representations in a single forward pass, LoLa introduces iterative interaction between the latent reasoner and the frozen base LLM. At each iteration, the current latent thought is processed by the base LLM, and its hidden states are projected into feedback memory and integrated by the reasoner through gated cross-attention to update the latent state. KV-cache reuse avoids recomputing prompt representations across iterations. This enables iterative latent refinement without modifying the base LLM or requiring recurrent pretraining.

2. Matched SFT+GRPO training.

We retrain LRT, and independently train LoLa-1U and full LoLa on Qwen3-4B using the same frozen backbone, training data, random seeds, effective batch sizes, and SFT+GRPO recipe. SFT runs for two epochs (learning rate 1e-4), followed by 50 GRPO updates (learning rate 3e-5). LRT uses its original single-pass architecture; LoLa-1U performs one reasoner update; full LoLa uses four refinement steps during SFT and three during GRPO and inference.

Before GRPO, the SFT-only Pass@1 scores are 44.55% for LRT, 44.62% for LoLa-1U, and 51.83% for LoLa.

Pass@1 and Pass@4 on Qwen3-4B (%).

MethodAMCMATH-500GSM8KLSATGPQAAverage
LRT@145.7873.4091.5838.7038.8957.67
LoLa-1U@146.9973.4091.5139.5738.8958.07
LoLa@148.1975.0092.3441.3041.4159.65
LRT@455.4281.2095.6058.7059.6070.10
LoLa-1U@456.6380.4095.8363.4859.6071.19
LoLa@461.4583.0096.1364.7862.6373.60

Under matched SFT+GRPO, LoLa-1U improves over LRT by 0.40 and 1.09 percentage points in average Pass@1 and Pass@4, respectively. Full LoLa further improves over LoLa-1U by 1.58 and 2.41 points, and over LRT by 1.98 and 3.50 points.

3. Latency comparison and refinement overhead.

We evaluate all methods on the same 64 MATH-500 questions using Qwen3-4B and a 1,024-token generation budget.

Inference efficiency on Qwen3-4B.

MethodMean end-to-end latency (s)Mean output tokensLatent preparation (ms)
Base (non-thinking)16.15629.20—
LRT23.70902.0946.08
LoLa-1U16.03620.3886.39
LoLa17.12653.94205.35

Compared with LRT, LoLa reduces mean end-to-end latency by 27.76% and mean output tokens by 27.51%, despite its additional latent preparation.

We further profile the refinement overhead:

Latent refinement breakdown (ms).

MethodPrompt encodingFeedbackProcessorUpdate/Norm.Projection
LoLa-1U26.1634.5824.550.230.48
LoLa26.35103.5172.900.671.12

Most refinement time is spent on base LLM feedback and reasoner processing, while state updates and projection account for little overhead.

4. Evaluation on a different and larger backbone.

We also train and evaluate LoLa on DeepSeek-R1-Distill-Llama-8B using the same 1,024-token generation budget.

Pass@1 and Pass@4 on DeepSeek-R1-Distill-Llama-8B (%).

MethodAMCMATH-500GSM8KLSATGPQAAverage
Base@136.1461.4080.5922.6136.3647.42
LoLa@143.3764.6081.6528.2636.3650.85
Base@451.8172.6093.8636.0951.0161.07
LoLa@461.4579.0094.0144.7864.6568.78

LoLa improves average Pass@1 and Pass@4 by 3.43 and 7.70 percentage points, respectively.

We sincerely thank the reviewer again for these thoughtful and valuable suggestions. They motivate us to conduct additional controlled training and evaluation experiments. We hope these results provide a clearer understanding of LoLa’s methodological contribution and practical benefits. We will incorporate these results and analyses into the revised manuscript.

Dear Reviewer XDSG,

Thank you again for your thoughtful review. We have posted detailed responses to your comments and uploaded a substantially revised version of the manuscript. As the discussion period is scheduled to end on November 18, we wanted to kindly check whether you have had a chance to review our responses and whether any further clarification would be helpful. We would be happy to address any remaining questions.

Thank you for your time and consideration.

Best regards, The Authors

Thank you for the detailed rebuttal and the additional experiments. My concerns are partially resolved, particularly regarding the matched SFT+GRPO comparisons. After considering the rebuttal and the other reviewers’ comments, the following concerns remain:

  • The methodological contribution is incremental. LoLa adds iterative hidden-state feedback to LRT while retaining its modular design, frozen backbone, and explicit CoT capability. The training pipeline also follows the standard combination of SFT and GRPO techniques, without introducing a new training objective or optimization method. The rebuttal clarifies the architectural differences, but does not substantially change my assessment, consistent with the concern in Reviewer KoH7’s review that the method primarily integrates and refines existing approaches.

  • The empirical gains do not establish a compelling advantage of iterative refinement. On the 1.5B backbone, LoLa improves average accuracy over LRT by only 0.84 percentage points, while MATH-500 accuracy decreases from 68.50% to 67.00%. On Qwen3-4B, the matched improvements over LoLa-1U are 1.58 and 2.41 points in Pass@1 and Pass@4, respectively, while the additional 8B experiment lacks a single-pass baseline and therefore does not establish whether this advantage generalizes. Given the incremental methodological contribution, these modest gains and the limited comparative evidence across backbones do not sufficiently establish the significance of feedback-driven iteration.

I appreciate the authors’ efforts to strengthen the evaluation. However, addressing these issues would require substantial methodological and empirical reworking beyond the scope of a rebuttal. At the current stage, I remain unconvinced that the methodological and empirical contributions are sufficient for acceptance and cannot give a positive recommendation.

Thank you for taking the time to carefully review our rebuttal and additional experiments. We understand and respect your assessment.

We sincerely appreciate your thoughtful and constructive feedback throughout the review process. Your suggestions have motivated us to conduct additional controlled experiments, strengthen our evaluation, and improve the quality of our revised manuscript.

Thank you again for your time and consideration.

Reviewer_h2VcOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Long CoT is expensive, so latent reasoning tries to move the intermediate computation into continuous space. The closest prior work, LRT, uses a small trainable reasoner next to a frozen LLM, but it produces its latent thoughts in a single forward pass with no refinement. Looped models (Ouro, Huginn) do refine iteratively, but they bake the recurrence into the backbone during pretraining. LoLa sits between the two: a 0.6B reasoner keeps a latent state, sends it through the frozen base LLM, reads back the base model’s hidden states as feedback, and updates the state for T rounds. The final state becomes latent thought tokens that condition answer generation. The reasoner is trained with SFT and then GRPO. On DeepSeek-R1-Distill-1.5B it gets the best average among efficient-reasoning baselines.

Critical strengths
  • It fills a real gap with a clean, modular design. Iterative refinement with base-LLM feedback, without touching the backbone, is a sensible middle ground between LRT and pretraining-level looped models. Because the base model stays frozen, explicit CoT remains available as a fallback.
  • The architecture is carefully engineered for stability. Gated cross-attention initialized near zero, RMS-clipped residual state updates, per-loop embeddings, and KV-cache reuse across loops are all sensible choices that keep the loop cheap and trainable.
Critical weaknesses
  • What’s driving the gains isn’t isolated. GRPO on GSM8K/MATH alone adds 8pp (GSM8K +12.7, LSAT +12.6), whereas the base-LLM feedback loop, the paper’s actual contribution, adds only about 2pp in a single run (Table 3). The comparison against non-thinking mode is therefore largely “task-specific RL vs. no training.” It’s also unclear whether LRT on Qwen3-4B received the same SFT+GRPO recipe

overall the idea is good and the design is clean. However, the gains over prior work are within noise on DeepSeek, and on Qwen they are mostly explained by GRPO training rather than by the looped feedback.

Questions for authors

Can you train a single-pass reasoner (no feedback, T=1 throughout training) and LRT on Qwen3-4B with exactly the same SFT+GRPO recipe? That comparison would isolate how much the loop itself contributes.

3 comments
AuthorsOpenReview ↗

LRT, LoLa-1U and LoLa under the same SFT+GRPO recipe

We sincerely thank the reviewer for the thoughtful comments and for recognizing LoLa’s modular design. To better isolate the contribution of iterative refinement, we conduct matched training comparisons on Qwen3-4B, including retrained LRT, independently trained LoLa-1U, and a newly trained full LoLa model. We also add inference-efficiency and 8B evaluations.

1. Controlled training experiments.

LRT uses its original single-pass architecture. LoLa-1U uses a single reasoner update in both SFT and GRPO, without further iterations. Full LoLa uses iterative refinement.

All three use the same frozen Qwen3-4B backbone, training data, random seeds, effective batch size, and learning rates. Training consists of two SFT epochs followed by 50 GRPO updates, with learning rates of 1e-4 and 3e-5, respectively. Before GRPO, the SFT-only Pass@1 scores are 44.55% for LRT, 44.62% for LoLa-1U, and 51.83% for LoLa, showing that LoLa’s advantage is already present before GRPO.

Pass@1 and Pass@4 on Qwen3-4B (%).

MethodAMCMATH-500GSM8KLSATGPQAAverage
LRT@145.7873.4091.5838.7038.8957.67
LoLa-1U@146.9973.4091.5139.5738.8958.07
LoLa@148.1975.0092.3441.3041.4159.65
LRT@455.4281.2095.6058.7059.6070.10
LoLa-1U@456.6380.4095.8363.4859.6071.19
LoLa@461.4583.0096.1364.7862.6373.60

Under matched SFT+GRPO, LoLa-1U improves over LRT by 0.40 and 1.09 percentage points in average Pass@1 and Pass@4, respectively. Full LoLa further improves over LoLa-1U by 1.58 and 2.41 points, and over LRT by 1.98 and 3.50 points.

2. Inference efficiency.

We evaluate all methods on the same 64 MATH-500 questions using Qwen3-4B and a 1,024-token generation budget.

MethodMean end-to-end latency (s)Mean output tokensLatent preparation (ms)
Base (non-thinking)16.15629.20—
LRT23.70902.0946.08
LoLa-1U16.03620.3886.39
LoLa17.12653.94205.35

Compared with LRT, LoLa reduces mean end-to-end latency by 27.76% and mean output tokens by 27.51%, despite its additional latent preparation.

3. Evaluation on a different and larger backbone.

We also train and evaluate LoLa on DeepSeek-R1-Distill-Llama-8B using the same 1,024-token generation budget.

Pass@1 and Pass@4 on DeepSeek-R1-Distill-Llama-8B (%).

MethodAMCMATH-500GSM8KLSATGPQAAverage
Base@136.1461.4080.5922.6136.3647.42
LoLa@143.3764.6081.6528.2636.3650.85
Base@451.8172.6093.8636.0951.0161.07
LoLa@461.4579.0094.0144.7864.6568.78

LoLa improves average Pass@1 and Pass@4 by 3.43 and 7.70 percentage points, respectively.

We sincerely thank the reviewer again for these thoughtful and valuable suggestions. They motivate us to conduct additional controlled training and evaluation experiments. We hope these results help clarify LoLa’s contribution and practical benefits, and will incorporate them into the revised manuscript.

Dear Reviewer h2Vc,

Thank you again for your thoughtful review. We have posted detailed responses to your comments and uploaded a substantially revised version of the manuscript. As the discussion period is scheduled to end on November 18, we wanted to kindly check whether you have had a chance to review our responses and whether any further clarification would be helpful. We would be happy to address any remaining questions.

Thank you for your time and consideration.

Best regards, The Authors

Area_Chair_M1EbOpenReview ↗

Reminder: Rebuttal Discussion

Dear Reviewer,

The discussion deadline is only a few hours away. Could you please respond to the authors' rebuttal before the deadline?

Best regards, Area Chair

Reviewer_KoH7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LoLa, a framework that aims to reduce the cost of LLM reasoning using a small external reasoner. While keeping the base LLM frozen, the reasoner iteratively refines a fixed set of latent vectors using feedback from the LLM’s hidden states. These vectors then guide answer generation. The method is trained through supervised fine-tuning followed by reinforcement learning. Experiments on two models across five reasoning benchmarks evaluate its performance, with additional ablations and measurements of inference latency and memory usage.

Critical strengths
  1. The paper addresses a valuable problem. Explicit chain-of-thought (CoT) reasoning often involves generating lengthy reasoning traces, which increases inference time, computational cost, and context usage. These are practical concerns. Exploring whether a relatively small external module can reduce these costs while keeping the base LLM frozen is a worthwhile direction.
  2. The paper is logically coherent, and the method is clearly presented. The mathematical notation and equations help readers understand the method of paper.
  3. The experimental evaluation is comprehensive. Ablations cover several key design choices and measurements of end-to-end latency and memory usage. The authors have released the code for the core model and training pipeline, making it convenient for others to verify and use this work.
Critical weaknesses
  1. The novelty is limited. LRT already uses a small external model to generate latent vectors that support LLM reasoning. PCCoT explores iterative, parallel updates of latent vectors. This work LoLa feeds the base model’s hidden states back to a separate reasoner to update the latent state. However, the paper does not sufficiently explain what new capabilities these modifications enable or why this particular architecture is needed. At present, I view the method primarily as an integration and refinement of existing approaches.
  2. The results show improvements, but they are inconsistent across models and tasks. The advantages are modest on several tasks, with performance even declining on some tasks. In addition, the five repeated evaluations are reported only for LoLa, without corresponding evaluations of the baselines. This makes it difficult to assess how stable those improvements are.
  3. The range of models evaluated is narrow. The paper considers only a 1.5B model and a 4B model, both from the Qwen family. These experiments do not establish whether similar improvements would hold for larger models or other model families. The paper is motivated by reducing the inference cost of long reasoning traces. However, small models face limitations in supporting long and complex reasoning traces due to their capacity and context-window constraints, while the time and computational costs of such traces should be more pronounced in larger models.
  4. The accuracy and latency evaluations are not well aligned. The main results use a maximum generation length of 1,024 tokens, whereas the latency evaluation uses a limit of 8,192 tokens on 64 problems from MATH500, without reporting accuracy under the same setting. The latency table also omits LRT, the most directly relevant baseline. Therefore, although the paper reports end-to-end latency, it remains unclear how much time LoLa saves at comparable answer quality and whether it retains an efficiency advantage over LRT and other baselines.
Questions for authors
  1. Why does the SFT stage use four loop iterations, while RL and final inference use three? Have the authors compared this setup with one that uses three iterations in both SFT and RL?
  2. The number of loop iterations is evaluated only from one to four. The high similarity of the latent vectors after three iterations suggests that the state changes very little under the current setting. Could this also be related to the model size and the number of iterations used during training?
  3. Why is all the training data drawn from mathematics? How to explain the improvement on LSAT and the decline on GPQA? Could this be related to the training data distribution?
  4. Could the authors report the average number of output tokens for both methods, along with a breakdown of LoLa’s latency between latent generation and answer decoding?
2 comments
AuthorsOpenReview ↗

Additional experiments on robustness, generalization, and inference efficiency

We sincerely thank the reviewer for the detailed and constructive comments. In response, we conduct additional experiments on Qwen3-4B and DeepSeek-R1-Distill-Llama-8B, together with latent-state analyses across the 1.5B, 4B, and 8B backbones.

1. Methodological contribution.

LRT generates fixed-length latent representations through a single forward pass of an external reasoner. PCCoT adopts Jacobi-style iterations to update latent thought vectors in parallel. In contrast, LoLa introduces an external feedback loop: the frozen base LLM processes the latent thoughts and returns hidden states, which the external reasoner uses to iteratively update its latent state. This enables feedback-based latent refinement without modifying the base LLM.

2. Evaluation robustness and model generalization.

In the original manuscript, the base and LRT results are taken from the LRT paper. For this rebuttal, we re-evaluate the non-thinking base and retrain LRT using the same SFT+GRPO recipe as LoLa, including the same frozen Qwen3-4B backbone, training data, random seeds, effective batch size, two SFT epochs, and 50 GRPO updates.

Pass@1 on Qwen3-4B (mean ± std, %).

MethodAMCMATH-500GSM8KLSATGPQAAverage
Base47.59 ± 3.1972.20 ± 0.6489.61 ± 0.4637.39 ± 3.2037.88 ± 0.8756.93
LRT45.78 ± 3.0273.40 ± 0.4391.58 ± 0.3138.70 ± 1.9338.89 ± 1.5357.67
LoLa48.19 ± 1.9875.00 ± 0.7792.34 ± 0.2841.30 ± 1.0041.41 ± 2.6659.65

LoLa outperforms both base and retrained LRT across all five benchmarks in Pass@1 and Pass@4. Their average Pass@4 scores are 69.97%, 70.10%, and 73.60%, respectively.

We also train and evaluate LoLa on DeepSeek-R1-Distill-Llama-8B using the same 1,024-token generation budget.

Pass@1 and Pass@4 on DeepSeek-R1-Distill-Llama-8B (%).

MethodAMCMATH-500GSM8KLSATGPQAAverage
Base@136.1461.4080.5922.6136.3647.42
LoLa@143.3764.6081.6528.2636.3650.85
Base@451.8172.6093.8636.0951.0161.07
LoLa@461.4579.0094.0144.7864.6568.78

LoLa improves average Pass@1 and Pass@4 by 3.43 and 7.70 percentage points, respectively, demonstrating its applicability to a larger backbone from a different model family.

3. Loop depth and latent-state analysis.

To examine the choice of loop depth, we evaluate the Qwen3-4B SFT checkpoint at inference depths from T=1 to T=4.

SFT results on Qwen3-4B.

Loop depthAverage Pass@1 (%)Average output tokens
T=131.82947.65
T=228.18833.66
T=351.83515.33
T=443.45417.20

T=3 achieves the highest average accuracy, while T=4 generates shorter responses but with lower accuracy. We therefore use T=3 for GRPO and final inference.

We further analyze latent-state trajectories across all three backbones. The following table reports the average cosine similarity and relative displacement between consecutive projected latent thoughts across the five benchmarks.

ModelUpdate 1Update 2Update 3Update 4
1.5B0.851 / 54.62%0.922 / 39.46%0.891 / 46.66%0.917 / 40.66%
4B0.781 / 66.18%0.917 / 40.76%0.974 / 22.72%0.991 / 13.19%
8B0.808 / 62.03%0.813 / 61.15%0.883 / 48.45%0.962 / 27.64%

The 1.5B model exhibits a notable change at the third update. In contrast, both the 4B and 8B models show progressively increasing similarity and decreasing relative displacement across updates, indicating increasingly stable latent states.

4. Training data and out-of-domain generalization.

We use mathematical reasoning data for SFT and GRPO because it provides structured multi-step reasoning supervision and verifiable answers. LSAT and GPQA serve as out-of-domain benchmarks for evaluating logical and scientific reasoning.

With the updated base results, LoLa improves LSAT and GPQA Pass@1 by 3.91 and 3.53 percentage points, respectively, and Pass@4 by 2.61 and 10.10 points. Both benchmarks show improvements over base.

5. Inference efficiency and output-length analysis.

We evaluate all methods on the same 64 MATH-500 questions using Qwen3-4B and a 1,024-token generation limit.

Inference efficiency on Qwen3-4B.

MethodMean end-to-end latency (s)Mean output tokensLatent preparation (ms)
Base (non-thinking)16.15629.20—
LRT23.70902.0946.08
LoLa-1U16.03620.3886.39
LoLa17.12653.94205.35

Compared with LRT, LoLa reduces mean end-to-end latency by 27.76% and output tokens by 27.51%.

We hope these results address the reviewer’s concerns and provide a clearer understanding of LoLa’s contribution. We will incorporate these additional results and analyses into the revised manuscript.

Dear Reviewer KoH7,

Thank you again for your thoughtful review. We have posted detailed responses to your comments and uploaded a substantially revised version of the manuscript. As the discussion period is scheduled to end on November 18, we wanted to kindly check whether you have had a chance to review our responses and whether any further clarification would be helpful. We would be happy to address any remaining questions.

Thank you for your time and consideration.

Best regards, The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised Paper and Summary of Changes

We sincerely thank the Area Chair and all reviewers for their time, careful evaluation, and valuable comments. Their feedback has helped us substantially improve the clarity, presentation, and experimental validation of our work.

We have uploaded a thoroughly revised version of the paper. The main changes are summarized below:

  1. Methodological clarification (Appendix C.4): We have added a discussion comparing LoLa with LRT and PCCoT to clarify the differences in their architectures, information flow, and iterative refinement mechanisms.

  2. Revised experimental evaluation (Section 3): We have substantially rewritten the entire experimental section, redesigned the figures and tables, and revised the corresponding analyses. This includes additional comparisons among LRT, LoLa-1U, and LoLa under the same training settings, as well as an extended evaluation on DeepSeek-R1-Distill-Llama-8B.

  3. Further analysis (Section 3.3): We have revised and expanded the analysis of LoLa's latent reasoning mechanism. We have also improved the analyses of the ablation and inference efficiency results.

  4. Additional experiments and appendix revisions (Appendix C): We have reorganized and substantially rewritten the appendix, adding analyses of inference loop depth, performance variation across three independent runs, latent preparation time breakdown, and comparisons with other latent reasoning methods.

  5. Revised presentation: We have rewritten the abstract, introduction (including the contributions), and conclusion to better reflect the method, experimental findings, and contributions of the work.

We hope that these revisions adequately address the reviewers' concerns and welcome any further feedback.

Dear Reviewers,

I hope you're doing well.

Please respond to the authors' rebuttal if you have not already done so.

Thank you for your valuable time and engagement.

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.