acceptodds
Under review as a conference paper at ICLR 2027

Does Time-Series Question Answering Need a Large Language Model to Reason?

Abstract

Does time-series question answering need a large language model (LLM) as its reasoning engine? Many existing methods rely on an LLM both to interpret the signal and to reason over it. This is computationally expensive and makes it difficult to determine whether better answers come from better representations or better reasoning. We introduce GLEAN (**G**rounded **L**atent **E**vidence **A**ggregation **N**etwork), which separates pretrained input encoding from a dedicated recurrent reasoning engine. Frozen text and time-series encoders process each input once, supplemented by numerical features that retain details such as signal endpoints and slopes. GLEAN then reasons over these representations through recurrent latent updates. Its current state guides each new read of the inputs, so relevant details can be revisited and combined as the answer develops. A small shared network performs this computation, and structured output heads select answers or order signal fragments without generating text. The separation makes each part measurable: frozen encoders alone already lead every time-series LLM on perception questions, while the recurrent reasoner raises relational accuracy from 58.03% to 69.16% on the same representations. With 0.46B total parameters, of which only 8.2M are trained, GLEAN achieves accuracy on TSAQA, outperforming all five evaluated fine-tuned time-series LLMs we evaluated while using – fewer estimated inference FLOPs per question. These results suggest that closed-ended temporal question answering does not require a large language model as its reasoning engine: pretrained representations coupled with a small recurrent network offer a strong and far cheaper alternative.

Primary area:
learning on time series and dynamical systems
Keywords:
time series analysis, large language models, foundation models, reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_oDwUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GLEAN, a lightweight framework for time-series question answering that replaces the LLM reasoning backbone with a recurrent latent reasoning module. It combines frozen time-series and text encoders, iterative evidence retrieval, and structured prediction heads. Experiments on TSAQA demonstrate competitive performance against time-series LLMs with substantially lower inference computation.

Critical strengths

The paper explores a practically meaningful direction by decoupling time-series perception from reasoning. The extensive ablation studies and computational efficiency analysis provide useful empirical insights into the effectiveness of lightweight alternatives to LLM-based approaches.

Critical weaknesses
  1. Limited methodological novelty. The core architecture largely combines established techniques, including frozen pretrained encoders, Perceiver-style iterative cross-attention, recurrent parameter sharing, and Sinkhorn-based structured prediction. The paper lacks a sufficiently novel mechanism or deeper insight beyond integrating these existing components.
  2. Insufficient evidence for the central claim. All experiments are conducted on TSAQA, which consists of closed-ended questions with predefined answer spaces. This setting naturally favors specialized discriminative models, and the findings cannot readily generalize to open-ended time-series reasoning. Moreover, several fine-tuned general-purpose LLMs still outperform GLEAN, weakening the broader claim that LLM-based reasoning is unnecessary.
Questions for authors
  1. What is the essential methodological difference between GLEAN's recurrent reasoning module and existing Perceiver-style iterative latent processing architectures? Could the authors provide a direct comparison?
  2. Can the authors validate GLEAN on additional benchmarks, particularly open-ended time-series QA tasks, to support the generality of their central claim?
6 comments
AuthorsOpenReview ↗

Response to Reviewer oDwU (Part 1 of 4)

We thank the reviewer for recognising that decoupling time-series perception from reasoning is a practically meaningful direction, and that the ablations and the efficiency analysis give useful empirical insight into lightweight alternatives to LLM-based approaches. We address the two concerns point by point below.


W1 / Q1: Methodological novelty, and the difference from Perceiver-style architectures

W1. Limited methodological novelty. The core architecture largely combines established techniques, including frozen pretrained encoders, Perceiver-style iterative cross-attention, recurrent parameter sharing, and Sinkhorn-based structured prediction. The paper lacks a sufficiently novel mechanism or deeper insight beyond integrating these existing components.

Q1. What is the essential methodological difference between GLEAN's recurrent reasoning module and existing Perceiver-style iterative latent processing architectures? Could the authors provide a direct comparison?

Response. Although pretrained encoders, cross-attention, and sinkhorn-based module are established, and the paper credits them as such rather than claiming them, the contribution is not their assembly.

(1) First, a measurement that published work cannot provide. Existing time-series QA systems place perception and reasoning inside one LLM, and each receives the signal differently, so no published comparison can attribute accuracy to either stage. To our knowledge this is the first work in time-series question answering to hold the perception stage fixed and compare reasoners over identical evidence, a language model included, under one FLOP convention, which is what makes the question in the title answerable at all.

(2) Second, three specific differences in how the latent state is used, each isolated by its own ablation rather than asserted. State-conditioned reads are not among them, since Perceiver's queries also come from the current latents, and Related Work states that content matching and Sinkhorn assignment "are established operators, not new contributions".

(3) Third, the accuracy this combination reaches per unit of computation: 77.33 overall with 8.2M trainable parameters, at 87 to 693 times fewer FLOPs per question than the five fine-tuned time-series LLMs.

Difference from Perceiver-style processingIsolating experiment
No decoder. The latent slots are the answer representation: content matching against the candidates C and Sinkhorn assignment over the fragments P read them out directly, with no output queries and no generated text. Perceiver-IO decodes through separate output queries, and learned-query adapters hand the latents to a generating LLM.On the 5,978 reordering questions the coupling raises per-position accuracy from 29.4 to 52.1 and exact match from 8.0 to 33.5, and cuts invalid assignments from 98.9% to 29.3%, while Hungarian decoding adds only 0.7: the ordering constraint is learned in training, not imposed at decode time.
A heterogeneous evidence bank, with the answer space held outside it. H concatenates patch, question, candidate and raw-value boundary tokens, each tagged by modality, position, segment and task; C and P never enter H. Perceiver attends over one input array with no notion of an answer space.Removing boundary tokens and segment identities costs 3.71 points (73.62), encoding numeric candidates as text rather than as signals 2.35 (74.98), and reading the series through the text encoder 7.01 (70.32).
A scheduled, not uniform, block. One parameter set runs T = 3 outer loops of n = 4 FiLM-conditioned Refine steps, each closed by one Integrate step with no bank access, over two deep-supervision segments: 30 block applications from 8.2M parameters.Removing per-step reads costs 2.83 (74.50), slot interaction 1.52 (75.81), FiLM 1.19 (76.14) and Integrate 0.65 (76.68). Access beats depth: four blocks that re-read the bank (75.35) beat 30 that do not (74.50).

(continued in Part 2 of 4)

AuthorsOpenReview ↗

Response to Reviewer oDwU (Part 2 of 4)

W1 / Q1 (continued): the direct comparison with Perceiver-style processing

Direct comparison (Q1). We replaced the core while holding the frozen encoders, the bank, the K = 12 slots and the readouts fixed: an unshared Perceiver-IO-style stack at matched depth and wiring, a plain transformer over the bank and the slots swept over depth, and a small language model in the reasoner's place.

Reasoner over the same frozen evidence bankTrainable paramsReasoner GFLOPsTSAQA overall
none, pooled readout (Table 2)n/an/a70.79
Qwen3-0.6B + LoRALoRA on 0.6Bn/a74.65
plain transformer, L = 27.32M3.85 (0.50×)75.30
plain transformer, L = 410.87M7.17 (0.93×)76.90
plain transformer, L = 817.97M13.82 (1.80×)77.03
plain transformer, L = 1225.07M20.46 (2.67×)76.79
unshared core, Perceiver-IO style41.07M (5.0×)7.67 (1.00×)76.60
GLEAN shared recurrent core8.21M7.67 (1.00×)77.33

TSAQA test split, 42,000 questions; reasoner GFLOPs are the reasoning module's own forward cost under the paper's FLOP counter.

Analysis. Untying the weights is the comparison asked for most directly: identical FLOPs, 5.0× the trainable parameters, 0.73 points lower accuracy, so sharing is the better configuration at this budget rather than a compression compromise. Dropping our wiring does not recover the result either: the plain transformer reaches 75.30 at matched parameters and 77.03 at its best depth, needing 2.2× the parameters and 1.80× the reasoner FLOPs to land 0.30 short, and four further layers reduce accuracy. A language model in the same position reaches 74.65. Neither depth, nor parameters, nor a language model reproduces what the shared recurrent wiring does.

Read with the pooled-readout row, the same table also locates where reasoning capacity pays, which is the deeper insight we would claim: perception in the frozen encoders is already saturated (the pooled readout scores 90.83 there, above every time-series LLM, against 58.03 on relational questions), the reasoner is worth +11.13 on relational questions and only +0.24 on classification, and what still fails sits in the representation rather than in the reasoner, which Appendix J measures and the second corpus reproduces.

Modification. §4 Related Work states that we do not claim iterative latent cross-attention or cross-step weight sharing, and names the three differences above. New Appendix M ("Reasoner Comparisons at Fixed Evidence") with Table 10 and per-cell results in Table 11. §3.3 names the evidence-source interventions as such, so the architecture controls are unambiguous.

(continued in Part 3 of 4)

AuthorsOpenReview ↗

Response to Reviewer oDwU (Part 3 of 4)

W2 / Q2: Evidence for the central claim, and additional benchmarks

W2. Insufficient evidence for the central claim. All experiments are conducted on TSAQA, which consists of closed-ended questions with predefined answer spaces. This setting naturally favors specialized discriminative models, and the findings cannot readily generalize to open-ended time-series reasoning. Moreover, several fine-tuned general-purpose LLMs still outperform GLEAN, weakening the broader claim that LLM-based reasoning is unnecessary.

Q2. Can the authors validate GLEAN on additional benchmarks, particularly open-ended time-series QA tasks, to support the generality of their central claim?

Response. The paper does not make the broader claim. §1 states the opposite fact: "General-purpose LLMs fine-tuned on TSAQA remain more accurate overall, at 51-733× the computation, and we report where and why they lead (§3.2)", and the conclusion adds that "the answer spaces we study are closed. These findings therefore do not establish sufficiency for unrestricted language-based reasoning." The abstract scopes the claim the same way, to whether a large language model is needed as the reasoning engine. Concretely, three of the six fine-tuned general-purpose LLMs score above GLEAN, Qwen3-0.6B (78.32), Qwen3-8B (84.29) and LLaMA3.1-8B (85.26), at 51×, 733× and 576× our FLOPs per question, and three score below (Ministral-8B 74.74, LLaMA3.2-1B 73.48, Gemma3-1B 69.70); Table 5 and Figure 1c report this.

On the setting favouring discriminative models: it is the same setting for every system in the comparison, all fine-tuned on the same training split under the same protocol and scored by the same parser. The control is the pooled readout, which keeps our frozen encoders and structured readouts but has no reasoner: it reaches 70.79, 4.27 points below TimeOmni-1-7B. A discriminative readout over a predefined answer space does not by itself produce the result.

Further benchmark, retrained under the identical protocol (Q2). We retrained every system on Time-MQA, converted into the two answer formats of our readouts, 90,278 questions split 72,222 / 4,514 / 13,542, with three seeds for GLEAN. We report by subset, because the test mix is dominated by two perception-style subsets.

Time-MQAanomaly det. (5,974)classification (5,550)characterization (1,917)Overall (13,542)
GLEAN (ours), 3 seeds71.15 ± 0.6283.20 ± 0.1762.60 ± 0.6074.71 ± 0.37
ChatTime-1-7B73.8276.9952.1171.76
ChatTS-8B62.7169.4265.3665.76
TimeOmni-1-7B64.3166.5266.2565.34
PATRA-7B ¹32.1118.4020.1424.74
TimeMQA-Qwen2.5-7B ²69.4470.7264.9569.19

¹ PATRA-7B produced no parseable answer in 43.4% of its responses, so its overall figure is largely a compliance rate. ² in-domain: pre-trained on this corpus, so not a like-for-like row. Three further subsets (101 questions, 0.75% of the split) are reported as counts.

Analysis. GLEAN leads overall by +2.95 over the strongest comparable baseline, from classification (+6.21), and is behind on the other two large subsets, by 2.67 on anomaly detection and 3.65 on characterization. What bears on generality is not the margin but that the same profile reappears: strong on perception-style subsets, behind on language-heavy characterization, which is the split Table 1 reports on TSAQA. That supports reading the limitation as a property of the frozen representations rather than of one benchmark's construction.

(continued in Part 4 of 4)

Reviewer_TUMhOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper aims to answer a key research question in time-series QA task: does time-series question answering need a large language model as its reasoning engine? Through the extensive experiments, the authors reveal that time-sereis preception and relational tasks do not need large language model as the reasoner, while language-heavy tasks still require LLM reasoner to achieve good performance. The proposed method with lightweight architecture outperforms time-series reasoning methods over perception and relational tasks and achieves the competitive performance on language-heavy tasks without relying on an LLM as the reasoner.

Critical strengths
  1. The paper has a strong motivation and the central research question is very interesting.

  2. The proposed method separates the representation learning and the reasoning component, and provide the useful insights into the capabilities and limitations of frozen pretrined encoders.

  3. The paper conducts extensive experiments to demonstrate the effectivenss of the proposed method and answers the key research question raised in the introduction.

Critical weaknesses
  1. Based on the results in ablation study (Table 6), the improvement of the proposed method over series only variant is somewhat marginal across two perception tasks and some of relational tasks (e.g., TR TF and TR PZ). Although the improvement on language-heavy tasks is substantial (around +13 percents), the series-only variant does not receive textual information, which means that the comparsion is asymmetric. Therefore, the large improvement on language-heavy tasks is expected, and it remains unclear to what extent GLEAN's superior performance over existing time-series reasoning models can be attributed to the strong representations learned by the frozen encoders rather than the proposed recurrent reasoning architecture. Due to the relative small improvement on perception tasks and some of rerlation tasks, it's unclear to what extent GLEAN benefits from jointly reasoning over textual questions and time-series data. Comparisone between GLEAN with a stronger multimodal fusion (MOMENT-1-large and BGE-base + MLP or a lightweight transformer as the substitution of shared block recurrent block) might isolate the contribution of the proposed architecture from the beneift of simply providing both modalities.

  2. The experiments are conducted only on a single benchmark dataset (TSAQA), which is hard to show the generalizability of the proposed method to other time-series QA datasets. In addition, many results are evaluated only once. It's better to report the results with standard deviation after multiple runs.

  3. The proposed method takes the numerical boundary features (Equation 2) as the additional information, while the time-series LLM baselines do not appear to receive them as additional input. According to the results in Table 2, removing the numerical boundary features reduce the performance to 73.62%, which is lower than the best competitor TimeOmni-1-7B (75.08%). It seems that the additional handcrafted features provide GLEAN with additional advantage on relational and language-heavy tasks. Thus, it remains unclear whether GLEAN's superior performance over time-series LLMs is primarily attributable to the proposed recurrent reasoning architecture or to the explicitly engineered numerical features.

Questions for authors

See Weakness.

9 comments
AuthorsOpenReview ↗

Response to Reviewer TUMh (Part 1 of 5)

We thank the reviewer for recognising the motivation and the central research question, and for finding that the separation of representation from reasoning yields useful insight into what frozen pretrained encoders can and cannot do. We address the three concerns point by point below.


W1: The asymmetric modality comparison, and the fusion baselines named

Based on the results in ablation study (Table 6), the improvement of the proposed method over series only variant is somewhat marginal across two perception tasks and some of relational tasks (e.g., TR TF and TR PZ). Although the improvement on language-heavy tasks is substantial (around +13 percents), the series-only variant does not receive textual information, which means that the comparsion is asymmetric. Therefore, the large improvement on language-heavy tasks is expected, and it remains unclear to what extent GLEAN's superior performance over existing time-series reasoning models can be attributed to the strong representations learned by the frozen encoders rather than the proposed recurrent reasoning architecture. Due to the relative small improvement on perception tasks and some of rerlation tasks, it's unclear to what extent GLEAN benefits from jointly reasoning over textual questions and time-series data. Comparisone between GLEAN with a stronger multimodal fusion (MOMENT-1-large and BGE-base + MLP or a lightweight transformer as the substitution of shared block recurrent block) might isolate the contribution of the proposed architecture from the beneift of simply providing both modalities.

Response. The asymmetry is real: text only (51.48) and series only (71.93) intervene on the evidence sources rather than on the architecture, and §3.3 now says so. The first baseline the review names, MOMENT-1-large and BGE-base + MLP, is the pooled-readout row: the same frozen encoders, boundary features and readouts, with the core skipped and the two modalities fused by pooling rather than by a trained module. It reaches 70.79, 4.27 points below TimeOmni-1-7B (75.06), and because it holds the cached features fixed, comparing it with GLEAN measures what the architecture adds.

Cellnpooled readoutGLEANΔ from the core
TR PZ7,97842.9657.87+14.91
DT TF3,00266.0678.45+12.39
TR TF2,05666.9379.18+12.25
TR MC1,96666.0274.52+8.50
AD TF6,00088.7793.40+4.63
DT MC2,99878.7579.49+0.74
CLS MC6,00092.8893.12+0.24
Overall42,00070.7977.33+6.54

TSAQA test split, 42,000 questions, same protocol and same answer parser as every system in the paper.

Analysis. On fixed features the architecture is worth +6.54 overall and +11.13 on the relational group. Where the frozen encoder is already near its ceiling it adds almost nothing (CLS MC +0.24), so the perception advantage is the encoders', as the review suspects. What the architecture contributes is the relational reasoning, including the two cells the review calls marginal, TR PZ (+14.91) and TR TF (+12.25).

(continued in Part 2 of 5)

AuthorsOpenReview ↗

Response to Reviewer TUMh (Part 2 of 5)

W1 (continued): the lightweight-transformer and unshared-core controls

Further experiment. We added the lightweight transformer the review names: L pre-LN self-attention blocks over [bank ; slots], with the bank, the K = 12 slots and the readouts unchanged and L swept rather than fixed. We also untied the shared weights into an unshared stack of the same depth.

Reasoner over the same bankTrainable paramsReasoner GFLOPsOverall
plain transformer, L = 27.32M3.85 (0.50×)75.30
plain transformer, L = 410.87M7.17 (0.93×)76.90
plain transformer, L = 817.97M13.82 (1.80×)77.03
plain transformer, L = 1225.07M20.46 (2.67×)76.79
unshared core, Perceiver-IO style41.07M (5.0×)7.67 (1.00×)76.60
GLEAN shared recurrent core8.21M7.67 (1.00×)77.33

Reasoner GFLOPs are the reasoning module's own forward cost, measured with the paper's FLOP counter.

Analysis. A trained fusion module is a stronger baseline than pooling, as the review expects, but it does not match the recurrent core: at matched parameters it reaches 75.30, 2.03 below, and its best depth (L = 8) needs 2.2× the trainable parameters and 1.80× the reasoner FLOPs to land 0.30 short, while four further layers reduce accuracy. It also attends jointly over [bank ; slots], so its cost is quadratic in bank length where our cross-attention reads are linear. Untying the shared weights gives identical FLOPs, 5.0× the trainable parameters and 0.73 points less accuracy. A LoRA-adapted Qwen3-0.6B in the reasoner's place reaches 74.65, 2.68 below the core and lower on all three capability groups (details in our reply to Reviewer FLNp, W1/Q1).

Modification. §3.3 states that text only and series only intervene on the evidence sources rather than on the architecture, and names the symmetric architecture controls: the pooled readout and the plain transformer over the same bank. New Appendix M ("Reasoner Comparisons at Fixed Evidence") with Table 10 and per-cell results in Table 11. §3.2 carries the fixed-feature decomposition in the main text.

(continued in Part 3 of 5)

AuthorsOpenReview ↗

Response to Reviewer TUMh (Part 3 of 5)

W2: A single benchmark, and single runs

The experiments are conducted only on a single benchmark dataset (TSAQA), which is hard to show the generalizability of the proposed method to other time-series QA datasets. In addition, many results are evaluated only once. It's better to report the results with standard deviation after multiple runs.

Response. We retrained GLEAN and all five time-series LLMs with three seeds each under the same protocol. Seed 1 is the number reported in Table 1 in every row.

TSAQA overall (n = 42,000)seed 1seed 2seed 3mean ± sd
GLEAN (ours)77.3377.2377.2477.27 ± 0.06
TimeOmni-1-7B75.0675.0474.8975.00 ± 0.09
TimeMQA-Qwen2.5-7B72.7472.8172.7772.77 ± 0.04
ChatTS-8B72.1371.9171.8771.97 ± 0.14
ChatTime-1-7B56.8357.0456.6556.84 ± 0.20
PATRA-7B54.8756.5454.0855.16 ± 1.26

Analysis. Each margin is an order of magnitude larger than the seed variation of the two systems it compares: +2.27 over TimeOmni-1-7B (sd 0.06 and 0.09), +4.49 over TimeMQA-Qwen2.5-7B (sd 0.04) and +5.30 over ChatTS-8B (sd 0.14). Standard deviations are reported as descriptive statistics; we do not attach significance tests to three runs.

Further benchmark, retrained under the identical protocol. We converted Time-MQA into the two answer formats of our readouts, giving 90,278 questions split 72,222 / 4,514 / 13,542 (stratified by subset and question type), and trained every system with the recipe used on TSAQA; GLEAN has three seeds here as well. We report by subset, because the test mix is dominated by two perception-style subsets.

Time-MQAanomaly det. (5,974)classification (5,550)characterization (1,917)Overall (13,542)
GLEAN (ours), 3 seeds71.15 ± 0.6283.20 ± 0.1762.60 ± 0.6074.71 ± 0.37
ChatTime-1-7B73.8276.9952.1171.76
ChatTS-8B62.7169.4265.3665.76
TimeOmni-1-7B64.3166.5266.2565.34
PATRA-7B ¹32.1118.4020.1424.74
TimeMQA-Qwen2.5-7B ²69.4470.7264.9569.19

¹ PATRA-7B produced no parseable answer in 43.4% of its responses; its format compliance is 56.6% and its accuracy among parseable answers 43.7, so its overall figure is largely a compliance rate. ² in-domain: pre-trained on this corpus, so not a like-for-like row. Three further subsets (comparison 53, temporal relationship 42, data transformation 6, together 0.75% of the split) are reported as counts.

Analysis. GLEAN leads overall by +2.95 over the strongest comparable baseline, from classification (+6.21), and is behind on the other two large subsets, by 2.67 on anomaly detection and 3.65 on characterization. The ranking of the baselines does not transfer: TimeOmni-1-7B is the strongest time-series LLM on TSAQA (75.06) and the weakest here (65.34), while ChatTime-1-7B is second weakest on TSAQA (56.83) and second strongest here (71.76). What bears on generality is not the margin but the shape: strong on the perception-style subsets, behind on language-heavy characterization, which is the same profile Table 1 reports on TSAQA. That supports reading this limitation as a property of the frozen representations rather than of one benchmark.

(continued in Part 4 of 5)

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

The paper asks whether closed ended time-series question answering needs a large language model as its reasoning engine. In existing time-series LLMs, the same LLM does both the perception and the reasoning, which is computationally expensive and makes it difficult to tell whether better answers come from better representations or from better reasoning. GLEAN encodes each instance once with a frozen time-series encoder and a frozen text encoder, plus a few raw value features, into a token-level evidence bank. A small recurrent core then refines K latent answer slots by alternating refine steps with integrate steps.

Contributions:

  1. Authors separate perception from reasoning and use this separation to identify a gap in frozen representations.
  2. An evidence grounded recurrent reasoner with value sensitive inputs and structured readouts, training 8.2M parameters while keeping both encoders frozen.
  3. On TSAQA, GLEAN outperforms five time-series LLMs under a common fine-tuning protocol with 87–693× fewer FLOPs per question and the highest perception accuracy among the compared systems.

Validation: The evaluation uses the full TSQA test set and its standard scoring. Baselines are five time-series LLMs, plus zero-shot and instruction-tuned general-purpose LLMs.

Critical strengths
  1. GLEAN outperforms all five fine-tuned time-series LLMs while training only 8.2 million parameters and using far less compute. The general LLMs that perform better need much more compute. This efficiency is the main reason I lean towards acceptance.

  2. The ablations are detailed and reported per task, so it is clear what each part of the model contributes. The pooled readout baseline is especially useful, since it shows that frozen encoders alone already outperform every time-series LLM on perception.

Critical weaknesses
  1. The paper is motivated by separating perception from reasoning, yet no experiment keeps the frozen encoders and evidence bank fixed and swaps only the reasoner. Every LLM baseline also differs in how it receives the signal, so the results cannot show if an LLM would add value on top of the same evidence. A small LLM reading the same evidence bank tokens would directly test the question in the title.

  2. GLEAN's advantage over TimeOmni comes mostly from anomaly detection and classification, where even the simple pooled baseline performs well. On the relational questions, the results are mixed. GLEAN is ahead on data transformation but clearly behind on temporal relationships. This suggests the frozen encoders, not the reasoner, drive most of the overall gain.

  3. All results seem to come from single runs, but many of the gaps, including the overall margin over TimeOmni and many ablation effects, are only one or two points. The paper itself shows signs of seed sensitivity. Without results over a few seeds, it is hard to know which of the smaller differences are real.

Questions for authors
  1. With the frozen encoders and evidence bank fixed, how does a small LLM reading the evidence bank tokens compare with the recurrent core in accuracy?
  2. Can you report the mean and std over three seeds for the full model and, if possible, TimeOmni?
  3. How does a core with only the refine module, or a core without “FiLM”, perform? Since the refine module already contains self-attention and an ffn block, what is the role of the separate integrate module?
  4. How sensitive are accuracy and cost to the number of answer slots K?
6 comments
AuthorsOpenReview ↗

Response to Reviewer FLNp (Part 1 of 4)

We thank the reviewer for the careful review, for recognising the efficiency of the approach and for finding the per-task ablations and the pooled-readout control informative. We address the four concerns point by point below.


W1 / Q1: A small LLM reading the same evidence bank

W1. The paper is motivated by separating perception from reasoning, yet no experiment keeps the frozen encoders and evidence bank fixed and swaps only the reasoner. Every LLM baseline also differs in how it receives the signal, so the results cannot show if an LLM would add value on top of the same evidence. A small LLM reading the same evidence bank tokens would directly test the question in the title.

Q1. With the frozen encoders and evidence bank fixed, how does a small LLM reading the evidence bank tokens compare with the recurrent core in accuracy?

Response. We added a reasoner-swap experiment: the frozen encoders, the evidence bank, the K = 12 answer slots, the readouts and the evaluation code are held fixed, and only the reasoning module is replaced. The language model reads the bank as input embeddings and is read out by the same heads, so no text is generated; it is Qwen3-0.6B with LoRA adapters and a frozen backbone, adapted as the time-series LLM baselines are (Appendix C).

Reasoner over the same bankPerceptionRelationalLanguage-heavyOverall
none, pooled readout (Table 2)90.8358.0369.8970.79
Qwen3-0.6B + LoRA (new)90.2365.2373.2174.65
GLEAN recurrent core93.2669.1673.6677.33

TSAQA test split, 42,000 questions, same protocol and same answer parser as every system in the paper.

Analysis. The LLM reasoner recovers much of what the pooled readout loses, which confirms that the bank carries usable evidence, but it stays 2.68 points below the recurrent core while adapting a 0.6B backbone against our 8.2M trainable parameters, and it is lower on all three capability groups, including language-heavy questions (−0.45), where an LLM reasoner would be expected to help most. Given the same evidence bank and the same readout, a small LLM does not add value over the recurrent core.

Modification. New Appendix M ("Reasoner Comparisons at Fixed Evidence") with Table 10 (capability groups, trainable parameters, measured reasoner FLOPs) and Table 11 (all eleven cells). §3.2 points to it: "Appendix M isolates the reasoner by replacing it at fixed evidence, a small language model included."

(continued in Part 2 of 4)

AuthorsOpenReview ↗

Response to Reviewer FLNp (Part 2 of 4)

W2: Whether the frozen encoders or the reasoner drive the gain

GLEAN's advantage over TimeOmni comes mostly from anomaly detection and classification, where even the simple pooled baseline performs well. On the relational questions, the results are mixed. GLEAN is ahead on data transformation but clearly behind on temporal relationships. This suggests the frozen encoders, not the reasoner, drive most of the overall gain.

Response. We can answer this with the features held exactly fixed. The pooled readout has the same two frozen encoders (MOMENT-1-large and BGE-base-en-v1.5), the same boundary features and the same readouts, with no reasoning module: as the reviewer notes it already beats every time-series LLM on perception, but overall it reaches 70.79, which is 4.27 points below TimeOmni-1-7B (75.06). Comparing it with GLEAN therefore isolates what the reasoner adds on identical features.

Cellnpooled readoutGLEANΔ from the core
TR PZ7,97842.9657.87+14.91
DT TF3,00266.0678.45+12.39
TR TF2,05666.9379.18+12.25
TR MC1,96666.0274.52+8.50
AD TF6,00088.7793.40+4.63
DT MC2,99878.7579.49+0.74
CLS MC6,00092.8893.12+0.24
Overall42,00070.7977.33+6.54

Analysis. The core's gains concentrate in exactly the relational cells the review calls mixed (+11.13 on the group) and vanish where the frozen encoder is already near its ceiling (CLS MC +0.24); the four language-heavy cells each gain 3.3 to 4.6 (Table 6). The perception advantage is therefore the encoders' and the relational improvement is the core's. The review is also right about the two smaller temporal cells, where TimeOmni leads on TR TF (−15.42) and TR MC (−18.26), while TR PZ, 7,978 of the 12,000 TR samples, is a tie (+0.19); weighted by sample count the margin splits into +4.89 from the two perception cells and −2.63 from the other nine, which §3.2 now reports.

Further experiment. We also replaced the core by two architectures of comparable size over the same bank: an unshared core that unties the shared weights at matched depth and wiring, and a plain transformer over the bank and the slots, swept over depth.

Reasoner over the same bankTrainable paramsReasoner GFLOPsOverall
plain transformer, L = 27.32M3.85 (0.50×)75.30
plain transformer, L = 410.87M7.17 (0.93×)76.90
plain transformer, L = 817.97M13.82 (1.80×)77.03
plain transformer, L = 1225.07M20.46 (2.67×)76.79
unshared core, Perceiver-IO style41.07M (5.0×)7.67 (1.00×)76.60
GLEAN shared recurrent core8.21M7.67 (1.00×)77.33

Reasoner GFLOPs are the reasoning module's own forward cost, measured with the paper's FLOP counter.

Analysis. Neither more depth nor more parameters reproduces the result. At matched parameters the plain transformer reaches 75.30, 2.03 below the core; its best depth (L = 8) needs 2.2× the trainable parameters and 1.80× the reasoner FLOPs and is still 0.30 short, and four further layers reduce accuracy. Untying the shared block costs 0.73 points at 5.0× the parameters and identical FLOPs, so weight sharing is not a compression compromise at this budget. The cost comparison also moves in our favour as the evidence grows, since the plain transformer attends jointly over [bank ; slots] and is quadratic in bank length where our cross-attention reads are linear.

Modification. §3.2 carries the decomposition and the fixed-feature result in the main text (the +4.89 / −2.63 split, TR PZ +14.91, DT TF +12.39, CLS MC +0.24, and the conclusion that the perception advantage is the encoders' and the relational improvement the core's). The two architecture controls are Appendix M, Tables 10 and 11, and §3.3 names the pooled readout and the plain transformer as the symmetric architecture controls.

(continued in Part 3 of 4)

AuthorsOpenReview ↗

Response to Reviewer FLNp (Part 3 of 4)

W3 / Q2: Mean and standard deviation over three seeds

W3. All results seem to come from single runs, but many of the gaps, including the overall margin over TimeOmni and many ablation effects, are only one or two points. The paper itself shows signs of seed sensitivity. Without results over a few seeds, it is hard to know which of the smaller differences are real.

Q2. Can you report the mean and std over three seeds for the full model and, if possible, TimeOmni?

Response. We retrained GLEAN and all five time-series LLMs with three seeds each under the same protocol. Seed 1 is the number reported in Table 1 in every row.

TSAQA overall (n = 42,000)seed 1seed 2seed 3mean ± sd
GLEAN (ours)77.3377.2377.2477.27 ± 0.06
TimeOmni-1-7B75.0675.0474.8975.00 ± 0.09
TimeMQA-Qwen2.5-7B72.7472.8172.7772.77 ± 0.04
ChatTS-8B72.1371.9171.8771.97 ± 0.14
ChatTime-1-7B56.8357.0456.6556.84 ± 0.20
PATRA-7B54.8756.5454.0855.16 ± 1.26

Analysis. Each margin the comparison rests on is an order of magnitude larger than the seed variation of the two systems it compares: +2.27 over TimeOmni-1-7B (sd 0.06 and 0.09), +4.49 over TimeMQA-Qwen2.5-7B (sd 0.04) and +5.30 over ChatTS-8B (sd 0.14), so the submitted single-run margin of 2.26 is not seed noise. Seed variance is similarly small on the second benchmark, where GLEAN scores 74.31 / 75.03 / 74.79, i.e. 74.71 ± 0.37 on Time-MQA (13,542 questions, retrained under the identical protocol; see our reply to Reviewer TUMh, W2).

Modification. New Appendix O ("Seed Variance") with Table 13. §3.2 states the result in the main text: "Three training seeds leave the margin at +2.27, with seed standard deviations of 0.06 for GLEAN and 0.09 for TimeOmni-1-7B."

(continued in Part 4 of 4)

Comments on the paper

2 comments
Area_Chair_EQRnOpenReview ↗

Discussion period ending soon

Dear Reviewers,

Thank you for your reviews of this submission. The authors have posted responses addressing the concerns raised.

With limited time remaining in the discussion period, please take a moment to review the authors' responses and post a brief follow-up indicating whether your concerns have been addressed or what remains outstanding. If your assessment changes as a result, please update your score accordingly.

Active engagement at this stage is essential for a fair and well-substantiated final decision. Thank you for your time and effort in reviewing this submission.

Best regards, AC

AuthorsOpenReview ↗

Summary of the Reviews and Responses

We sincerely thank the AC and all reviewers for the time and effort devoted to our submission, and for feedback that has substantially improved our paper.


We are encouraged that the reviewers found the central research question very interesting, with a strong motivation (TUMh), and the decoupling of perception from reasoning a practically meaningful direction (oDwU) that yields useful insights into frozen pretrained encoders (TUMh). They also noted that our method outperforms all five fine-tuned time-series LLMs while using far less compute, supported by ablations detailed and reported per task (FLNp), and an efficiency analysis with useful empirical insights (oDwU).

In response to the reviewers' concerns, we have made the following key updates:

  • We have isolated the contribution of the architecture itself: with the evidence, the answer slots and the readouts held fixed, we replaced only the reasoner, with a small LLM, a Perceiver-IO-style unshared stack and a plain transformer. None of them matches our recurrent core at comparable cost (FLNp, TUMh, oDwU).
  • We have shown that the numeric boundary features are not the source of our margin: supplying them to the strongest baseline does not improve its overall accuracy, and a probe on those features alone explains where they help and where they mislead (TUMh).
  • We have added three-seed results for our model and all five baselines, retrained every system on a second benchmark under the identical protocol, and transferred to a third without retraining: the margin is neither run-to-run variation nor specific to one dataset (FLNp, TUMh, oDwU).
  • We have reported the requested ablations on each part of the recurrent core, together with the accuracy and cost of the answer-slot count: removing any single component lowers accuracy (FLNp).
  • We have clarified the scope of the central claim, so that it matches exactly the settings in which it is evaluated (oDwU).

After the discussion, all three reviewers have confirmed their concerns to be adequately addressed, with no further questions. We are grateful for their time and their extremely helpful suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.