acceptodds
Under review as a conference paper at ICLR 2027

Lapras: Latent Reasoning for Time Series Language Models

11 views

Abstract

Time Series Language Models (TSLMs) offer a promising path toward time series understanding by reasoning over temporal signals and producing natural language answers and explanations. A common approach is Chain-of-Thought (CoT), which generates step-by-step rationales describing relevant signal patterns and connecting them to final answers. Although these models learn from reference CoT traces during post-training, generating faithful descriptions of input time series at inference time remains challenging. The model is required to express high-dimensional and continuous temporal representations in discrete language tokens, which may lead to neglecting task-relevant patterns or describing them inaccurately. Because subsequent reasoning steps build on these descriptions, early errors can propagate through the reasoning chain, ultimately leading to incorrect answers accompanied by plausible explanations that are inconsistent with the input signal. To address this, we propose Lapras (Latent Post-trained Reasoning Across Series), a post-training framework that equips TSLMs with latent reasoning. A model trained with Lapras reasons through a sequence of continuous thoughts in the joint time series-language space, producing text only for the final answer. It learns this through teacher-student self-distillation, where a teacher trained on CoT reference traces reasons explicitly through text. The student is trained to align its hidden states with the teacher's at the answer stage, transferring the teacher's reasoning ability into the student's latent computation. We evaluate Lapras across four TSLM backbones on five diverse time series question answering benchmarks. Lapras improves average F1 by up to 10.79% over explicit CoT while generating 23.9 times fewer tokens. Importantly, the continuous thoughts generated by Lapras can be decoded into readable reasoning traces through standard language decoding, preserving the capability to provide textual explanations. Together, these results highlight Lapras as a promising post-training paradigm for TSLMs that enables efficient and effective reasoning at inference while supporting interpretability.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Language Models, Time series, Reasoning, Latent Reasoning, Wearables

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zPPQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Lapras, a post-training framework that equips time-series language models with latent reasoning instead of textual reasoning. Lapras uses a shared-parameter teacher–student self-distillation setup. Across 4 TSLM backbones and 5 time-series QA benchmarks, Lapras improves macro-F1 over explicit CoT by up to 10.79% while generating far fewer tokens, and its continuous thoughts can be decoded back into readable traces.

Critical strengths
  1. The paper is well-written and easy to follow.
  2. The problem this paper tackles is interesting and validates the feasibility of latent reasoning in time-series reasoning.
  3. The experiments on multiple benchmarks validate the effectiveness of the proposed method over other latent reasoning baselines.
Critical weaknesses
  1. The novelty is somewhat limited, since, compared to text reasoning, this work conditions on the time series as input; however, there's no specific design toward learning the time series. It is similar to Coconut [1] with multimodal input.
  2. There's a lack of comparison between textual and visualized time series representations as inputs for reasoning. Although latent reasoning shows improvements, it cannot justify its advantage over textual reasoning based on models like Qwen3.5 and Qwen3-VL.
  3. Figure 6 shows Lapras incurs substantially more wall-time than No-CoT (e.g., 171 ms vs 68 ms on ECG), contradicting the claim of “less than 8 ms overhead compared to No-CoT.
  4. Benchmarks are all multiple-choice. It remains unclear whether the approach scales to open-ended benchmarks.
Questions for authors

See Weaknesses above.

1 comment

We thank the reviewer for the thorough feedback. We have added the relevant experiments to the revised manuscript, and below we summarize these results along with our clarifications.

W1 Novelty Concern. The novelty of our work lies not only in "what" we propose but also in "why" and "how" current standard CoT-SFT fails in time series reasoning tasks, which directly motivates latent reasoning as a promising alternative paradigm for TSLM. As our primary contribution, we are the first to reveal a fundamental limitation of CoT-SFT in TSLMs, which we term the verbalization bottleneck: when the model translates a continuous time series into text, it discards task-relevant temporal evidence, and the resulting errors propagate through the reasoning chain (Section 3.1). We further support this through our systematic evaluation across four backbones and five tasks, which indicates that explicit CoT underperforms No-CoT in average F1 in most cases while incurring a substantially higher inference cost. Given this diagnosis, we adopt latent reasoning to keep reasoning in the joint time series and text embedding space, and for simplicity we only add time series masking to prevent shortcut learning. To our knowledge, Lapras is the first to show that latent reasoning can operate in a joint time series and text embedding space, and achieves significant performance improvement over the conventional CoT-SFT models.

W2 Time Series Representation. We have included the experiment that the reviewer suggested in Appendix A.4. In summary, we follow the standard protocol of ChatTS, which also compares time series encoder inputs against serialized text and rendered plot inputs. Specifically, we give each series to Qwen3-8B as serialized numeric text and to Qwen3-VL-8B as a rendered plot, and we evaluate both models zero-shot and after CoT-SFT on the same training data for the same number of steps as Lapras. As the results show, both textual and visualized time series representations perform poorly in the zero-shot setting. After CoT-SFT, the average accuracy rises from 30.5% to 56.7% with the text representation and from 34.7% to 59.8% with the plot representation. Both still underperform our time series encoder CoT-SFT baseline (61.8%) and fall well short of ChatTS-Lapras (71.2%). Overall, the results are consistent with the conclusion of ChatTS that text and plots are not effective representations of time series, which justifies our use of a time series encoder.

Table A.4. Time series encoder vs. text and plot inputs. Accuracy (%) on the five tasks. Bold marks the best per column.

InputModelFinetune MethodECGSleepHARTSREngineAvg.
TextQwen3-8BZero-shot53.5022.1027.7026.1023.3030.54
TextQwen3-8BCoT-SFT51.5069.9070.6061.4030.0656.69
PlotQwen3-VL-8BZero-shot51.3222.8625.5951.0022.5934.67
PlotQwen3-VL-8BCoT-SFT53.2081.7073.3060.8030.1059.82
TS encoderChatTS-8BCoT-SFT59.2579.6372.7961.9035.2361.76
TS encoderChatTS-8BLapras (Ours)85.2388.1977.4369.3235.7571.18

W3 Typo. Thanks for catching this typo. The intended statement is that Lapras adds less than 87 ms of overhead on average relative to No-CoT, which remains a favorable tradeoff given its performance gains. We have corrected this in the revised manuscript

W4 Evaluation Protocol. We clarify that evaluation uses open-ended generation, following the protocol of OpenTSLM [ICML 2026] and ST-Reasoner [ACL 2026 Oral]. The model reasons and writes its final answer as free text. Each question comes with its own candidate answers, which makes the setting look like multiple choice, but these candidates exist only so that rule-based answer extraction can score the free-text output reliably.

Above all, we hope we have clarified that Lapras addresses a fundamental limitation that goes beyond its training recipe. This work offers two insights that prior work has not shown. First and most importantly, we identify and empirically characterize the verbalization bottleneck, in which errors that arise while describing the signal in text propagate through the reasoning chain. Second, we show that latent reasoning can operate in the joint time series–language space rather than in text embeddings alone. If our responses have addressed your concerns, we would be grateful if you would consider raising your score.

Reviewer_Ru7WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates latent reasoning for time series language models. It identifies a potential verbalization bottleneck, where inaccurate textual descriptions of temporal signals can affect subsequent reasoning. Lapras adapts teacher–student self-distillation to this setting: a teacher processes reference textual rationales, while a student performs continuous reasoning steps and aligns its answer-stage hidden states with the teacher. Experiments cover four backbones and five temporal question-answering tasks. Lapras achieves the highest average macro-F1 on three backbones and substantially reduces inference cost compared with explicit CoT. The paper also provides component ablations and exploratory analyses of latent-state decoding, answer probabilities, and attention

Critical strengths

A useful investigation of the verbalization bottleneck in TSLMs. The paper explores how converting temporal information into discrete language tokens can affect reasoning accuracy. Teacher–student self-distillation provides a practical way to transfer information from explicit reasoning traces into continuous representations, allowing the student to reason without generating intermediate textual descriptions. The early-answering analysis in Section 3.1 provides useful motivation by showing that the accuracy gap between predicted and reference CoT traces emerges during signal description and persists through subsequent reasoning.

Extensive experiments supporting the approach’s effectiveness. The evaluation covers four backbones and five tasks, with comparisons against direct answering, explicit CoT, and existing latent reasoning methods. Lapras achieves the highest average macro-F1 on three backbones, and the reported inference times show substantial savings over explicit CoT. The component ablations and latent-step analysis provide additional evidence about the method’s behavior.

Critical weaknesses
  • The methodological novelty beyond CODI is limited. The core teacher–student self-distillation and answer-stage hidden-state alignment closely follow CODI, with relatively few additional components. The main contribution appears to be the adaptation and empirical investigation of this framework for time series. A more explicit description of the time-series-specific changes would help readers better assess the methodological contribution.

  • Some presentation details and a qualitative example need correction. TSR is used in the introduction before being clearly defined, and Figure 1 contains the typo “Explicit Reasong (CoT).” More importantly, the reference CoT in Figure 11 discusses a heating system over a 40-day period, whereas the question concerns apple sales and inventory decisions. This mismatch prevents the example from supporting the claimed consistency between the decoded latent thoughts and the reference rationale. Please verify and revise the relevant text accordingly.

Questions for authors
  • Efficiency reporting. Section 5.3 states that Lapras adds less than 8 ms of overhead over No-CoT, but the latency differences shown in Figure 6 are substantially larger. For example, ECG takes 171 ms with Lapras versus 68.3 ms with No-CoT, and the other datasets also show differences exceeding 8 ms. Could you clarify this discrepancy?

  • Source of the performance gains both Lapras and the explicit-CoT baseline learn from reference CoT traces. Could you explain what drives Lapras’s improvements over explicit CoT? It would be helpful to provide examples where explicit CoT misdescribes the input signal and gives an incorrect answer, while Lapras (with its decoded reasoning traces shown) answers the same question correctly.

  • Reliability of decoded reasoning traces. The decoded examples are interesting. Can similarly coherent and signal-grounded reasoning traces be recovered consistently across the evaluated questions, including cases where the model answers incorrectly?

1 comment

We thank the reviewer for the close reading of our manuscript and the constructive feedback. We have revised the manuscript accordingly and respond to each point below.

W1 Novelty Concern. The novelty of our work lies not only in "what" we propose but also in "why" and "how" current standard CoT-SFT fails in time series reasoning tasks, which directly motivates latent reasoning as a promising alternative paradigm for TSLM. As our primary contribution, we are the first to reveal a fundamental limitation of CoT-SFT in TSLMs, which we term the verbalization bottleneck: when the model translates a continuous time series into text, it discards task-relevant temporal evidence, and the resulting errors propagate through the reasoning chain (Section 3.1). We further support this through our systematic evaluation across four backbones and five tasks, which indicates that explicit CoT underperforms No-CoT in average F1 in most cases while incurring a substantially higher inference cost. Given this diagnosis, we adopt latent reasoning to keep reasoning in the joint time series and text embedding space, and for simplicity we only add time series masking to prevent shortcut learning. To our knowledge, Lapras is the first to show that latent reasoning can operate in a joint time series and text embedding space, and achieves significant performance improvement over the conventional CoT-SFT models.

W2 Typo. Thanks for catching the typos. We have done another iteration of proofreading and made the corresponding corrections to improve the readability of the manuscript.

Q1 Inference Speed. The value of 8 ms was a typo. We intended to state that Lapras adds less than 87ms of overhead on average compared with No-CoT. The conclusion that Lapras is both efficient and effective remains unchanged.

Q2 & Q3 Reasoning Trace Comparison. Both CoT-SFT and Lapras learn from the same reasoning traces, but explicit CoT must regenerate a textual description of the signal at inference and commit to it, so description errors propagate to the answer (the verbalization bottleneck, Section 3.1). Lapras bypasses this by reasoning in continuous thoughts and only outputs textual answers at the end. To better illustrate this, we have now included the explicit-CoT trace in Figure 4 and Figure 11 that indicates Lapras' continuous thoughts decoded to the right description while CoT failed to do so. In addition, we have included a failure example in Figure 12 showing that when Lapras answers incorrectly, which the error comes from incorrect inference rather than sensor perception.

We hope our revision makes it clearer that this work reveals a previously unrecognized yet fundamental limitation of CoT reasoning in time series language models and addresses it with Lapras as a principled solution. If these clarifications address the remaining concerns, we kindly ask the reviewer to consider raising the score.

Reviewer_Bo3POpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Lapras, a post-training framework that applies latent reasoning to time series language models. It adapts teacher-student self-distillation to learn from explicit reference rationales while generating intermediate reasoning through continuous hidden states. Experiments across four backbones and five tasks show improvements in final-answer prediction and substantial inference speedups over explicit CoT. The paper also argues that its approach addresses a verbalization bottleneck and preserves interpretability.

Critical strengths
  • Improving the accuracy and efficiency of time series LLMs is an important problem.
  • The method is technically sound.
  • The evaluation covers several model sizes, sensor-fusion architectures, and application domains.
  • The performance gains are meaningful, including improvements over direct answering on several backbones.
Critical weaknesses
  • The methodological novelty and contributions are very limited. The central components of Lapras are inherited from CODI, and the paper primarily applies an existing latent-reasoning framework to the time series domain.
  • The evidence for the proposed verbalization bottleneck is not that convincing. The comparison between generated rationales and reference rationales constructed using gold information does not isolate verbalization as the cause of the errors. The observed effects could instead reflect poor signal perception, weak reasoning capabilities, or limitations of the rationale supervision.
  • The experiments are not sufficiently representative of stronger contemporary time series reasoning models. Although ChatTS-7B is included, much of the detailed analysis uses a relatively weak 1B model. It is unclear whether the observed failure modes and benefits would persist in stronger state-of-the-art time series LLMs or at substantially larger scales (e.g., 20B+ parameters). This limits confidence in the generality and significance of the findings.
  • The paper lacks comparisons with RLVR-trained reasoning models. Reinforcement learning with verifiable rewards has become an important approach for improving time series reasoning. Comparing primarily against CoT-SFT and earlier latent-reasoning methods does not establish how valuable Lapras is relative to stronger established post-training approaches. In particular, it is unclear whether the claimed limitations of explicit reasoning would persist with stronger RL-trained models.
  • The experimental controls do not adequately explain the improvements. Additional rationale supervision, hidden-state alignment, masking, and recurrent computation can all contribute to performance, and their effects are not sufficiently isolated.
  • The evaluation does not directly substantiate several of the central reasoning claims. All tasks are evaluated as classification or multiple-choice answering, including a reformulation of the ECG reasoning benchmark into binary diagnosis. Better final-answer prediction does not necessarily demonstrate better multistep reasoning or improved grounding in the original signal.
  • The interpretability claims are insufficiently supported. Decoding readable text from latent states does not establish that these explanations faithfully represent the model's computation. The selected qualitative examples and attention/probability analyses provide limited evidence for the broader claims.
Questions for authors
  • Do the reported observations and performance improvements persist with stronger time series LLMs or general LLMs?
  • How does Lapras compare with time series LLMs trained using RLVR under comparable model and training settings?
  • What evidence distinguishes the proposed verbalization bottleneck from limitations in model capabilities, signal perception, or rationale supervision?
  • How much of the improvement comes from latent reasoning itself versus the additional supervision and regularization?
  • What evidence establishes that the decoded explanations faithfully represent reasoning grounded in the input signal?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Bo3P (1/2)

We thank the reviewer for the detailed feedback. The revised manuscript includes the requested experiments, and we summarize them below together with our clarifications.

W1 Novelty Concern. The novelty of our work lies not only in "what" we propose but also in "why" and "how" current standard CoT-SFT fails in time series reasoning tasks, which directly motivates latent reasoning as a promising alternative paradigm for TSLM. As our primary contribution, we are the first to reveal a fundamental limitation of CoT-SFT in TSLMs, which we term the verbalization bottleneck: when the model translates a continuous time series into text, it discards task-relevant temporal evidence, and the resulting errors propagate through the reasoning chain (Section 3.1). We further support this through our systematic evaluation across four backbones and five tasks, which indicates that explicit CoT underperforms No-CoT in average F1 in most cases while incurring a substantially higher inference cost. Given this diagnosis, we adopt latent reasoning to keep reasoning in the joint time series and text embedding space, and for simplicity we only add time series masking to prevent shortcut learning. To our knowledge, Lapras is the first to show that latent reasoning can operate in a joint time series and text embedding space, and achieves significant performance improvement over the conventional CoT-SFT models.

W2&Q3 Verbalization bottleneck diagnosis. All four backbones we used have already demonstrated strong sensor perception in their original papers, and our early-answering analysis holds reasoning capacity and supervision fixed by using the same frozen reasoning model. We consistently observe that the accuracy gap between the predicted and reference CoT traces are large during the describe stage, before any reason step is added, and persists through the subsequent reason stage (Figure 9).Together with our main results (Table 1), this evidence supports the verbalization bottleneck we identify in Section 3.1.

W3&Q1 Larger TSLM. Most existing TSLMs are small models ranging from 0.5B to 8B, therefore we conduct experiments on these models. In terms of reasoning capability with its model size, among existing TSLMs, we observe the increasing trend of using latent reasoning compared with textual CoT when model size increases. Specifically, when TSLM size increased from 0.5B to 8B, Lapras improved average F1 by 24.49% (from 44.03% to 68.52%), while CoT improved it by 13.58% (from 44.15% to 57.73%). We also agree that scaling to larger models would allow a broader comparison. Since few publicly released TSLMs exist at the 20B+ scale, we have added a discussion of this to Section 6.

W4&Q2 Comparison with RLVR. We position Lapras within the SFT paradigm. RL is typically applied on top of an SFT-initialized model, so RL-based methods are complementary to our approach rather than direct alternatives. To address the reviewer’s question more directly, we use the recent STReasoner [ACL 2026 Oral] as a testbed, because it provides a strong RL-trained time series reasoning model for fair comparison. We compared 1) CoT-SFT (which is the baseline in our paper), 2) CoT-SFT + RLVR (STreasoner) and 3). Lapras using their provided training datasets and 8B TSLM. The performance is shown in the following table, we observe that RL improves over CoT-SFT on all three tasks, which agrees with the reviewer’s expectation. However, Lapras outperforms the RL-trained model by a large margin on every task. Lapras achieves this without any RL stage or the additional data used during RL. This shows that the limitations of explicit reasoning remain even after RL post-training. The experiment details and discussion is now in Appendix A.5.

Method / DatasetEtiol.EntityCorr.
SFT89.465.479.7
SFT + RLVR (S-GRPO)95.775.787.1
Lapras99.596.493.2
AuthorsOpenReview ↗

Response to Reviewer Bo3P (2/2)

W5&Q4 Experiment Setting. The factors the reviewer lists are intrinsic to each method and cannot be ablated independently. For example, recurrent latent computation is what distinguishes latent reasoning methods from CoT-SFT. Similarly, hidden-state alignment cannot be applied to CoT-SFT because CoT-SFT is itself the teacher. Aligning it would mean aligning the model with itself, so the loss would be identically zero. Therefore, our experiments follow the standard evaluation protocol of iCoT, COCONUT, and CODI, in which the backbone and training data remain fixed within each experiment group. Under this protocol, the performance gain of our method can be attributed to its reasoning mechanism rather than to differences in backbone or training data.

W6 Evaluation Protocol. We follow the standard evaluation protocol of time series reasoning benchmarks (i.e. OpenTSLM [ICML 2026] and ST-Reasoner [ACL 2026 Oral]) to evaluate a model's reasoning ability. Tasks are posed as multiple-choice questions so that answers can be scored reliably by string matching. The model is not restricted to a fixed set of labels, where the model is required to generate its answer as free text.

W7&Q5 Interpretability. We intended the decoding analysis to show that Lapras keeps the human-readable reasoning traces CoT provides while outperforming CoT-SFT by up to 10.79% in average F1. We have revised the manuscript to make our claim more clear.

Above all, we hope our clarification resolves the misunderstanding about our experimental setting, which is following standard reasoning evaluation protocol. We also thank the reviewer for suggesting the comparison with RLVR, in which Lapras outperforms the RL-trained STReasoner model on all three tasks by up to 20.7%. This further supports our main contribution, which reveals a previously unrecognized limitation of explicit reasoning in time series language models and addresses it with Lapras as a principled solution. Since the reviewer already finds the method technically sound, the evaluation broad and the gains meaningful, we kindly ask the reviewer to raise the overall score accordingly.

Comments on the paper

2 comments

We have uploaded a revised manuscript that includes all the changes and corrections described in our responses, and we thank all reviewers for their valuable feedback.

AuthorsOpenReview ↗

General Note

We want to thank all reviewers for their time, reading and feedback.

We summarize the main additions and clarification we did in our rebuttal:

Reviewer Bo3P: We added a comparison with the RLVR-trained STReasoner under the same 8B backbone and training data. Lapras outperforms it on all three tasks by up to 20.7% without any RL stage, which shows that the limitations of explicit reasoning remain after RL post-training.

Reviewer Ru7W: We added explicit-CoT traces to Figures 4 and 11, where Lapras’s decoded latent thoughts describe the signal correctly and explicit CoT misdescribes it. We also added a failure case in Figure 12, which shows that Lapras’s errors come from inference and not from signal perception.

Reviewer zPPQ: We added the requested comparison with text and plot inputs (Appendix A.4), where Qwen3-8B with serialized text and Qwen3-VL-8B with rendered plots reach 56.7% and 59.8% average accuracy after CoT-SFT on the same data, both well below ChatTS-Lapras at 71.2%.

General note. We would like to restate that our central contribution is identifying the verbalization bottleneck, a fundamental limitation of CoT-SFT in TSLMs that prior work has overlooked and that causes explicit CoT to underperform direct answering in most of our settings. Lapras addresses it with a simple yet effective solution that keeps reasoning in the joint time series and language space, improving average F1 over CoT-SFT by up to 10.79%.

We hope reviewers will consider the new insight into a previously overlooked failure mode of CoT-SFT, the broad evaluation across four backbones and five tasks, and the consistent gains over both SFT and RL post-training baselines when assessing the paper. We would greatly appreciate it if reviewers could let us know whether our responses address their concerns, and we are happy to provide any further clarification before the discussion period ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.