acceptodds
Under review as a conference paper at ICLR 2027

TimeBraid: Unifying Time Series and Language for Understanding and Forecasting

11 views

Abstract

We present TimeBraid, a series of unified time-series and language models that align pretrained language models and pretrained time-series foundation models through interleaved global residual attention layers. Each model inherits knowledge, instruction following, and reasoning from one side and continuous-signal perception and zero-shot forecasting from the other, fusing both in a shared representation space for understanding and generation. We study the design choices enabling unified modeling: where to align the two representation spaces, how to ground language in temporal structure, how to balance understanding with generation, and how to stabilize joint optimization. The resulting recipe combines a unified prompting scheme for diverse time-series and text tasks, stabilized joint training, and supervision from 2.2M curated series–text pairs and 4.9M instruction-tuning samples. Across benchmarks spanning time-series perception, understanding, reasoning, and both context-aided and unimodal forecasting, TimeBraid remains competitive with far larger general-purpose models and task-specific counterparts.

Primary area:
foundation or frontier models, including LLMs
Keywords:
unified time-series and language models, time-series analysis, multi-modal forecasting

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_REpXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents TimeBraid, a unified framework for time-series understanding and forecasting. It integrates pretrained language and time-series foundation models through a Mixture-of-Transformers architecture, using global residual attention for cross-modal interaction. The authors further introduce a two-stage training pipeline with curated time-series–text pairs and instruction-tuning data, alongside techniques for stabilizing joint optimization.

The main contribution is an architectural and training framework that enables textual understanding and numerical forecasting within a single model. Experiments across multiple benchmarks report competitive results on time-series understanding and text-conditioned forecasting. However, the improvements are not consistent across tasks, and the unified model exhibits degradation relative to its pretrained backbones in unimodal forecasting and general language understanding. Overall, the work demonstrates the feasibility of combining existing foundation models, although the benefits of the proposed integration and the extent of cross-task generalization remain insufficiently established.

Critical strengths

We consider the cross-modal integration presented in this paper promising. The paper proposes a potentially effective approach to integrating pretrained LLMs and time-series foundation models while preserving their respective modeling architectures. The results demonstrate competitive performance on several understanding and forecasting tasks compared with existing models that can also utilize both modalities. Furthermore, the visualizations in the appendix show that different textual instructions can meaningfully influence the generated time series, providing qualitative evidence of effective cross-modal interaction.

The paper provides valuable practical findings. For example, it investigates the optimization challenges of jointly training language and time-series models, particularly the instability caused by large regression losses on strongly non-stationary sequences. The ablation studies support the proposed robust normalization and loss capping techniques, showing more stable optimization for both modalities. These findings provide useful practical guidance for training unified time-series and language models.

Critical weaknesses
  1. The benefits of the unified architecture are not sufficiently demonstrated. Although TimeBraid achieves competitive performance on several benchmarks, its results are not consistently superior to existing state-of-the-art models. For example, TimeBraid-6.7B achieves 80.65% accuracy on TSAQA, below the 85.26% of the LLaMA3.1-8B; its Numeric Fidelity on CaTS-Bench is 0.684, compared with 0.802 for Gemini-2.5-Flash; and its RCRPS on CiK is 0.294, substantially worse than 0.110 for Gemini-2.5-Flash. On Time-MMD, TimeBraid also fails to consistently outperform existing forecasting foundation models across domains. More importantly, joint training leads to degradation in the original capabilities of both pretrained backbones, as demonstrated by the decline in MMLU accuracy from 60.30% to 50.68% and the inferior GIFT-Eval MASE of 0.763 compared with 0.705 for TimesFM-2.5. These results raise concerns about whether the proposed unified architecture provides sufficient benefits to justify its additional complexity and training costs. 
  2. The generalizability and sequence-length capabilities of the proposed framework are insufficiently evaluated. All three model variants are built upon the Qwen3 and TimesFM-2.5 families, without experiments involving alternative language or time-series foundation models. This limits the evidence supporting the general applicability of the proposed integration strategy. Furthermore, although the model supports long-horizon unimodal forecasting, its language-conditioned forecasting capabilities are mainly demonstrated on relatively short horizons, with many alignment samples restricted to horizons of at most 32 steps. The paper lacks systematic evaluations of cross-modal understanding and text-conditioned forecasting under longer input sequences and forecasting horizons. The authors also acknowledge limitations in accurately interpreting long and complex time series.
Questions for authors
  1. Regarding Table 3: Controlled forecasting, we believe that the author should supplement the performance of Qwen3 as a single modal comparison between the fusion language and the time series model, as the TimeFM2.5 experiment in GIFT Eval shows that the author's fusion leads to performance degradation.

  2. Referring to the open source content provided by TimeFM, we believe it is necessary for the author to compare fev bench, TIME Benchmark, etc., as the performance degradation of GIFT-EVAL needs to be determined whether it is an isolated case.

  3. The content of Figures 2, 3, and 5 needs to be reformatted, and some of the text indicates that the font size is too small, which makes it difficult to understand the model structure and performance.

7 comments
AuthorsOpenReview ↗

Official Response to Reviewer REpX (Part 1)

Dear Reviewer REpX,

Thank you for the detailed and balanced review. We appreciate your recognition that the cross-modal integration is promising, that the visualizations show text meaningfully influencing the generated time series, and that our findings on stabilizing joint training provide practical guidance. We report the requested Qwen3 comparison and fev-bench and TIME results below, and will improve figure readability in the revision. Our responses are as follows.

W1: Benefits of the unified architecture are not sufficiently demonstrated

TL;DR: The benefit of the unified architecture is that the two modalities help each other inside one model: text lowers its forecasting error below that of its time-series backbone, its understanding accuracy depends on the time-series input, and joint training is ahead of single-task training on 6 of 8 benchmarks. Each comparison cited in W1 is against a different strongest model, and the TSAQA and MMLU gaps can be reduced: one epoch of fine-tuning on TSAQA brings TimeBraid-1.2B to 85.89%, above LLaMA3.1-8B SFT (85.26%), and 500 steps of distillation bring MMLU from 50.68 to 54.61.

Benefits of the unified architecture. The main benefit of unification is the mutual benefit between the two modalities, and each part of it is measured:

  • Text improves forecasting: with the text context, TimeBraid-6.7B has lower error than its time-series backbone TimesFM-2.5 on CAF (nCRPS 0.225 vs. 0.281), CGTSF (macro MSE 0.413 vs. 0.535) and Ctrl-F (MSE 6.232 vs. 8.784), and lower error than its language backbone Qwen3-4B on all three benchmarks in Q1 (see also our response to Reviewer MyWN, W1).
  • The time-series input improves understanding: removing it lowers TSAQA accuracy from 80.65 to 48.68 (Reviewer 5qHY, W4 & Q4), and TimeBraid-6.7B is 16 to 25 points ahead of Qwen3 reading the series as text at every TSAQA input length (W2).
  • Joint training helps both sides: against understanding-only and forecasting-only variants with the same architecture and training budget, the jointly trained model is ahead on 6 of 8 benchmarks and level on the other two (Reviewer 5qHY, W1 & Q1).

One set of weights provides all of this, and its full training takes 120–194 GPU-hours (Reviewer MyWN, W1). Our claim is that a single model is competitive with much larger general-purpose models and with task-specific models. We do not claim that it outperforms every task-specific model on each benchmark, and none of the baselines cited in W1 covers all the tasks.

Below, we respond to the individual results and then to the degradation of the backbones.

AuthorsOpenReview ↗

Official Response to Reviewer REpX (Part 2)

TSAQA (80.65% vs. 85.26% for LLaMA3.1-8B SFT). LLaMA3.1-8B is fine-tuned only on TSAQA with an 8B language model. TimeBraid-6.7B has a 4B language tower and is trained on a diverse mixture in which TSAQA accounts for about 3%. To check whether this gap comes from the training mixture or from the architecture, we fine-tuned the trained TimeBraid-1.2B (0.6B language tower) on TSAQA alone for one epoch. It reaches 85.89%, above LLaMA3.1-8B SFT (85.26%), so the gap comes from TSAQA's small share in our mixture and not from the capacity of the unified architecture.

CaTS-Bench numeric fidelity (0.684 vs. 0.802 for Gemini-2.5-Flash). Gemini reads the time series as text digits, which keeps the exact values. TimeBraid uses patch embeddings that compress 32 values into one token. This loses some local numeric detail, but it gives better semantic alignment (DeBERTa F1 of 0.712, higher than all closed-source models) and potentially supports much longer contexts. Patch tokenization is a standard compression trade-off in time-series models, and fine-grained numeric detail can be incorporated in future work via multi-scale patching or continuous number encodings. We admit it as a limitation of the time-series tokenizer and are actively working on methods introducing more fine-grained perception into the model.

CiK (0.294 vs. 0.110 for Gemini-2.5-Flash). CiK requires multi-step deductive reasoning over quantitative constraints (e.g., inferring capacity bounds or scheduled outages from prose). On this benchmark TimeBraid improves over TimesFM (0.295) on Future (0.359 vs. 0.373), Causal (0.437 vs. 0.474), and Covariate (0.243 vs. 0.251) subsets (Table 13), and is much better than 7B–8B open LLMs (Qwen-2.5-7B 0.344, Llama-3-8B 0.476) and other unified models (ChatTime 0.344, TimeOmni-VL 0.655). Consistent with CiK's finding that larger models benefit much more from context than smaller ones, and with the follow-up finding that small models fail mainly at the reasoning stage [1], our 0.6–4B towers after non-reasoning SFT do not train chain-of-thought forecasting-centric reasoning; adding reasoning supervision or RL post-training is an important direction for future work.

Time-MMD. The three TimeBraid sizes have the best three average ranks over the nine domains (4.50 to 5.83 vs. 6.33 across all 20 models in Table 15). TimeBraid has the lowest MSE in four domains and is in the top two in five domains. As our analysis in the Fig. 8 showed, the text effect on time-series in different domains can be very different and some domains' context are bringing negative effect, which is a general challenge for multimodal forecasting.

Degradation of the backbones. We acknowledge the degradation and will revise the Introduction to say "largely retains". We would like to note that the degradation is moderate and can be reduced.

  • Language: MMLU drops from 60.30 to 50.68. Table 21 shows that mixing unimodal data into the alignment stage improves MMLU from 40.88 to 53.38, and the final result depends on the unimodal share (12%) in the fine-tuning mixture. This is consistent with unified vision–language models, where fine-tuning the language model on multimodal data reduces its text-only accuracy [2][3] unless text-only data is blended back in [2] or the language modules are frozen [2][4][5]. Moreover, we found that a large part of the current remaining gap comes from the different multiple-choice formats in our fine-tuning data and MMLU. With a light calibration of 500 steps by distilling the base model, TimeBraid-2.5B recovers MMLU from 50.68 to 54.61.
  • Forecasting: please see Table 20 (where zero-shot TimeBraid-2.5B achieves lower average MSE than TimesFM-2.5 on all five datasets) and our response to Q2. We acknowledge that at the same context length TimeBraid-2.5B is 4–7% behind TimesFM-2.5 in MASE on GIFT-Eval, fev-bench and TIME, while it stays on par with or ahead of dedicated forecasting models such as Moirai-2.0 and Chronos-Bolt on fev-bench and TIME. This gap comes with our current fine-tuning data mixture (unimodal forecasting is 7.6% of the training mixture, Table 6), and we expect it can be reduced with the same established remedies as on the language side, i.e., more unimodal replay and distillation from the frozen backbone; this is ongoing work.
AuthorsOpenReview ↗

Official Response to Reviewer REpX (Part 3)

W2: Generality across backbones, and longer sequences and horizons

TL;DR: Variants trained with other backbones (Gemma-3-1B, TimesFM-3.0) show that the integration is not tied to Qwen3 and TimesFM-2.5, and 56% of the alignment forecasting samples have horizons up to 512 steps (only the LLM-rewritten samples are capped at 32). Results by length show gains over TimesFM-2.5 at every history length tested (up to about 500 points) and at horizons up to 64 steps, and on understanding TimeBraid stays ahead of Qwen3 reading the series as text at every input length (up to about 2,048 points); long-horizon text-conditioned forecasting (480–900 steps) remains open.

Backbones. We trained variants with a cross-family language model (Gemma3-1B) and a newer time-series model (TimesFM-3) under the same training budget. Please see our response to Reviewer 5qHY (W2 & Q2) for the results, which demonstrates the general applicability of our integration strategy and findings on base models.

Horizons in the training data. We would like to clarify that only the LLM-rewritten samples are limited to horizons of at most 32. About 56% of the forecasting samples in the alignment corpus (867k of 1.56M) have horizons up to 512 (Table 5).

Evaluation at different lengths. For unimodal and multivariate long sequences, Table 20 already evaluates 2,880-step inputs and horizons up to 720 across ETT and 21-variate Weather, where TimeBraid-2.5B has MSE rank 2.48 vs. 3.83 for TimesFM-2.5. Below, we report our text-conditioned results by horizon and history length, and will include the new GIFT-Eval-CTX [6] evaluation in the revision.

Text-conditioned forecasting by horizon (TimeBraid-6.7B vs. TimesFM-2.5, which sees the history only; lower is better):

Benchmark (metric)Horizon (steps)TimeBraid-6.7BTimesFM-2.5Diff
CAF (nCRPS ↓)≤ 320.1830.230−0.047
33–640.2940.364−0.070
CiK (RCRPS ↓)6–70.1630.166−0.002
15–240.3390.352−0.013
29–990.3090.306+0.003
GIFT-Eval-CTX (rel. MASE ↓)30–600.6850.764−0.079
480–6000.8720.812+0.060
720–9000.8640.851+0.013

Text-conditioned forecasting by history length (same comparison; lower is better):

Benchmark (metric)History lengthTimeBraid-6.7BTimesFM-2.5Diff
CGTSF (MSE ↓, mean of 3 datasets; Diff: mean per-dataset change)48-1920.4260.563−28.4%
72-2880.4150.553−29.3%
96-3840.4050.515−26.1%
120-4800.4050.510−25.4%
CAF (nCRPS ↓)≤ 320.1980.260−0.062
33–1280.2220.272−0.050
129–5000.2660.322−0.056

Understanding by input length (accuracy, higher is better):

Benchmark (accuracy ↑)Input lengthTimeBraid-6.7BQwen3-4B (text input)Diff
TimeSeriesExam1 series (about 1,024 points)64.147.4+16.7
2 series (about 2,048 points)61.253.2+8.0
TSAQA (all 42,000)1–3254.9–
33–6477.6–
65–12886.5–
129–25687.6–
257–51284.4–

Text usually provides information about the near future, and the horizons in Time-MMD, CiK and CAF are relatively short (Time-MMD 6–336, CiK 6–99, CAF 5–64 steps), as in other multimodal forecasting benchmarks: CGTSF forecasts one day ahead (24–96 steps) and From News to Forecast 7–48 steps [7]. Our training data therefore focuses more on short and medium horizons, where TimeBraid performs best. Its performance decreases at long horizons. In the tables above, the gain from text over TimesFM-2.5 holds at every history length (CGTSF: 25–29% lower MSE at all four history lengths; CAF: 0.05–0.06 lower nCRPS for histories from under 32 to 500 points) and at horizons up to 64 steps, while on GIFT-Eval-CTX it is limited to the 30–60-step horizons (0.685 vs. 0.764) and TimeBraid is behind TimesFM-2.5 at 480–900 steps. On understanding, accuracy stays high with longer inputs: TSAQA is at 84–88% from 65 to 512 points (77.6% at 33–64), and the margin over Qwen3-4B reading the series as text is +16.0 at 1–64 points and +24 to +25 beyond. We will add these length-stratified results and a detailed discussion to the revision. We agree that long-horizon text-conditioned forecasting is not fully solved in this work, and we consider it important future work.

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

The paper proposed TimeBraid, a joint time-series-language model that aligns pretrained LLM and TSFM by a set of fine-tuned fusion layers. The framework combines diverse time-series and text tasks, such as reasoning and forecasting in a unified way. To show the effectiveness of the proposed method, the paper conducts experiments on diverse tasks and settings.

Critical strengths
  1. The paper is well written and easy to follow.

  2. The target problem of the paper, aligning time series and language modalities is essential.

  3. The strategy that the paper proposed to align the two modalities is reasonable.

Critical weaknesses
  1. The improvement of the proposed method is limit, especially by substantially increasing model size and jointly fine-tuning both pretrained backbones. I assume the training costs of the proposed method is large, which also requires some discussions in the paper. However, TimeBraid-2.5B even underperforms its TimesFM-2.5 backbone in the forecasting task, which is not satisfied as people usually expect to get better performance for the original task by introducing new information (language) and parameters (Qwen and extra fusion module). On out-of-domain contextual forecasting, TimeBraid-6.7B performs similarly to the pretrained TimesFM baseline. Although the CGTSF and CAF results are promising, I still concerns the overall benefits with the additional model complexity and training investment.

  2. I didn't see enough ablation study in the main paper to demonstrate whether global residual attention improves reasoning and forecasting over simpler fusion mechanisms under comparable parameter and compute budgets.

  3. Since the proposed method utilizes both time series and language information, the compared baselines are also expected to utilize both, such as the time series agent methods [1, 2, 3], which is not well discussed and compared in the experiments.

  4. As TSFM advances very fast recently, TimesFM-2.5 is not good enough anymore. I'm wondering is it possible to replace or upgrade the TSFM while preserving the language and fusion components? This is related to the long-term costs of the proposed method.

[1] TS-Agent: Understanding and Reasoning Over Raw Time Series via Iterative Insight Gathering [2] TimeCopilot [3] TemporalBench: A Benchmark for Evaluating LLM-Based Agents on Contextual and Event-Informed Time Series Tasks

Questions for authors

Please check the weaknesses.

7 comments
AuthorsOpenReview ↗

Official Response to Reviewer MyWN (Part 1)

Dear Reviewer MyWN,

Thank you for the thoughtful review. We are glad that you found the paper well written and easy to follow, the problem of aligning time series and language essential, and our alignment strategy reasonable. Our responses are as follows.

W1: Limited improvement given the model size and training cost

TL;DR: When text context is available, TimeBraid consistently improves forecasting over TimesFM-2.5 across controlled and real-world benchmarks while unlocking time-series QA and captioning that TimesFM-2.5 cannot support; when no text exists at test time (GIFT-Eval, fev-bench, TIME), language parameters add no new signal and TimeBraid largely retains backbone accuracy (within about 4–7% MASE). Because both towers start from pretrained checkpoints, full alignment and fine-tuning require only about 120–194 GPU-hours (15–24 hours on an 8-GPU node).

We discuss forecasting with text, forecasting without text, and the training cost separately.

Forecasting with text (and understanding). When text context is available, introducing language and fusion parameters consistently improves forecasting over the TimesFM-2.5 backbone across both in-domain and out-of-domain benchmarks:

  • Ctrl-F: Top-1 improves from 33.33% to 43.33% (MSE 8.784 → 6.232). TimesFM-2.5 cannot use the text condition, so it stays at the chance level.
  • CAF: nCRPS improves from 0.281 to 0.225 (about 20% lower), outperforming GPT-5.4 (0.233).
  • CGTSF: every TimeBraid size outperforms every baseline on all three datasets, with up to 19% lower MSE than the best baseline; against TimesFM-2.5 the macro MSE is 0.413 vs. 0.535 (Table 16).
  • Time-MMD: the three sizes achieve the best three average MSE ranks (4.11 to 5.33 vs. 5.56 for TimesFM-2.5 in Table 2; 4.50 to 5.83 vs. 6.33 across all 20 models in Table 15).
  • GIFT-Eval-CTX [1] (new, OOD): TimeBraid-6.7B achieves relative MASE 0.7733 vs. 0.7975 for TimesFM-2.5 (and 0.8234 for TimeBraid without text), on par with TimesFM-3 (0.7796) and Claude Opus 4.8 (0.7807).
  • TemporalBench OOD forecasting (Table 14b): TimeBraid-6.7B ranks 1st on two of the four domains, PSML (0.197 vs. 0.231 for TimesFM-2.5 and 0.243 for GPT-5.4) and MIMIC (8.972 vs. 9.223 for TimesFM-2.5).
  • CiK (OOD, Table 13): TimeBraid improves over TimesFM on explicit Future (0.359 vs. 0.373), Causal (0.437 vs. 0.474), and Covariate (0.243 vs. 0.251) subsets while beating 7B–8B open LLMs (0.344–0.476); overall parity (0.294 vs. 0.295) comes from tasks that require multi-step quantitative reasoning over the context [2] (e.g., scaling the forecast by a stated factor within a stated time window), which TimeBraid has not been trained on.

Moreover, our controlled ablation (Reviewer 5qHY, W1 & Q1) shows that joint training with understanding tasks further lowers text-conditioned forecasting error over a forecasting-only variant, while unlocking time-series question answering, captioning, and reasoning (Table 1, surpassing most task-specific fine-tuned models and larger general-purpose models) that TimesFM-2.5 does not support.

Forecasting without text. GIFT-Eval has no text input, so no language information is introduced at test time, and prior work shows that LLM parameters alone do not improve purely numerical forecasting without text [3]. Our goal on unimodal forecasting is therefore to largely retain the backbone's ability, similar to retaining text-only MMLU in unified vision–language models [4][5]. Across GIFT-Eval, fev-bench, and TIME, TimeBraid-2.5B stays within about 4–7% MASE of TimesFM-2.5; on fev-bench and TIME it is on par with Moirai-2.0 and ahead of Chronos-Bolt (see Reviewer REpX, Q2 for the table), and it achieves lower zero-shot average MSE than TimesFM-2.5 on all five ETT and Weather datasets (Table 20).

Training cost. We will add the training cost to Appendix C.

VariantAlignment (GPU-hours)Fine-tuning (GPU-hours)
TimeBraid-1.2B36.7 H100-h83.0 H100-h
TimeBraid-2.5B41.7 H100-h93.2 H100-h
TimeBraid-6.7B59.1 H200-h135.0 H200-h

Because TimeBraid initializes from pretrained backbones, full alignment and fine-tuning take about 120 H100 GPU-hours for 1.2B, 135 H100 GPU-hours for 2.5B, and 194 H200 GPU-hours for 6.7B (15–24 hours on an 8-GPU node), which is a modest training cost relative to the added text-conditioned forecasting gains and understanding capabilities, while also acquiring two robust unimodal abilities.

AuthorsOpenReview ↗

Official Response to Reviewer MyWN (Part 2)

W2: Ablation of global residual attention against simpler fusion

TL;DR: Under an identical training budget and matched fusion parameter count (377.5M vs. 377.4M), layer-wise global residual attention is ahead of a bidirectional input/output MLP connector on 6 of 8 downstream benchmarks, including all four understanding benchmarks (up to +3.1 points on TSAQA) and two text-conditioned forecasting benchmarks (CGTSF −4.0% MSE, CAF −1.5% CRPS).

To support this, we will include the additional ablation reported below in the revision.

For this ablation, we evaluated a simpler fusion method that uses a pair of MLP connectors at the two ends of the towers: one MLP writes the TimesFM patch states into the language stream, and a second MLP adds the language model's outputs at the patch positions back to the TimesFM states before the forecasting head. The two towers then interact only at the input and the output, while global residual attention lets them interact at every paired layer.

We trained the bidirectional MLP connector above with a matched number of fusion parameters (377.4M vs. 377.5M) and evaluated it on downstream tasks. Both models use the towers of Qwen3-0.6B and TimesFM-2.5 as towers with the same training settings (10k alignment steps + 10k SFT steps).

BenchmarkMetricMLP connectorGlobal residual attentionDifference
Understanding
TSAQAacc ↑64.867.9+3.1
TemporalBench MCQacc ↑30.631.2+0.6
TimeSeriesExamacc ↑52.052.8+0.8
CaTS-BenchDeBERTa F1 ↑70.070.3+0.2
Text-conditioned forecasting
CGTSFmean MSE ↓0.4590.440−4.0%
CAFCRPS ↓0.2380.235−1.5%
CiKRCRPS ↓0.2890.299+3.3%
Time-MMDdomains won (of 9) ↑544 : 5
Fusion parameters377.4M377.5M

W3: Baselines that use both modalities, and agent methods

TL;DR: All multimodal and LLM baselines in our text-conditioned experiments already receive both time series and text, and TemporalBench is already evaluated zero-shot in Tables 1 and 14. Under the official TemporalBench leaderboard scoring, single-pass TimeBraid-6.7B (37.15) is ahead of four of the five agent frameworks listed below (MetaGPT, TimeCopilot, CAMEL, AgentScope) and first among all leaderboard entries on event-conditioned forecasting (T4, 46.07 vs. 43.63 for TimeClaw); it can also serve as the time-series component of an agent system.

We would like to clarify two points and add a new comparison with agent methods on TemporalBench.

  • We have already evaluated on TemporalBench as an out-of-domain benchmark for both QA and forecasting (Tables 1 and 14).
  • The multimodal baselines in our text-conditioned experiments use both modalities, including ChatTS, TimeOmni-1, ChatTime, TimeOmni-VL, Aurora, MIGAS-1.5, Time-LLM, GPT4MTS, Time-VLM, and DoubleCast. The frontier LLMs also receive both the time series and the text.
  • Following your suggestion, we compare with the agent baselines reported on the official TemporalBench leaderboard.
ModelTemporalBench (leaderboard scoring) ↑T4 (event-conditioned forecasting) ↑Backbone
TimeClaw42.7343.63deepseek-v4-pro
TimeBraid-6.7B37.1546.07one 6.7B model, one forward pass
MetaGPT36.5934.56deepseek-chat
TimeCopilot36.5228.35deepseek-chat
CAMEL36.4835.08deepseek-chat
AgentScope36.2135.08deepseek-chat

T4 is the event-conditioned forecasting tier, where the answer depends on the series together with the event text; TimeBraid-6.7B is first among all leaderboard entries (46.07 vs. 43.63 for TimeClaw). The 36.58 in Table 1 is the unweighted mean over the 16 tasks, a different metric.

We will add TS-Agent and TimeCopilot to the Related Work. We consider agent methods complementary to our work. Agents use a frontier LLM with tools over multiple calls, while TimeBraid is a unified model with one generation pass, and it can serve as the time-series component in an agent system. We will discuss this direction in Appendix I.

AuthorsOpenReview ↗

Official Response to Reviewer MyWN (Part 3)

W4: Replacing or upgrading the time-series backbone

TL;DR: Replacing TimesFM-2.5 with TimesFM-3.0 in a trained TimeBraid-1.2B, with both towers frozen and only the global residual attention layers tuned for 2,000 steps, recovers most of the drop right after the swap on TSAQA, CAF and CGTSF and is already ahead of TimeBraid-1.2B on TemporalBench (36.50 vs. 34.17) and on 5 of 9 Time-MMD domains.

We ran this experiment on a trained TimeBraid-1.2B. We replaced its TimesFM-2.5 tower with TimesFM-3.0 and froze both the language tower and the new time-series tower.

Right after the swap and before any tuning, the performance drops, as expected for a tower the fusion layers were not trained with (e.g., TSAQA 76.61 → 55.87, CiK 0.301 → 0.526). Tuning only the global residual attention module for 2,000 steps recovers most of this drop on TSAQA (72.93 vs. 76.61), CAF (0.250 vs. 0.235) and CGTSF (0.430 vs. 0.417), and brings CaTS within 1% (70.16 vs. 70.6 DeBERTa F1). The architecture supports this: the two towers interact only through the shared attention layers, where each tower has its own projections, so a new time-series tower can be connected by adapting these layers alone while both towers stay intact.

Time-series towerFusion layers tunedTSAQA ↑TSExam ↑CaTS DeBERTa F1 ↑TemporalBench ↑CAF CRPS ↓CiK RCRPS ↓Time-MMD domains won ↑ (vs. TimeBraid-1.2B)CGTSF mean MSE ↓
TimesFM-2.5 (TimeBraid-1.2B)–76.6150.1370.634.170.2350.301–0.417
TimesFM-3.00 steps55.8743.5769.7132.750.30490.52642 of 90.576
TimesFM-3.02,000 steps72.9345.8470.1636.500.25030.36195 of 90.430

We wish to clarify that TimeBraid is a jointly trained unified model and not a plug-in adapter. Similar to unified vision–language models, replacing a backbone requires some re-alignment. The result above shows that this cost is small, so we believe the long-term cost is acceptable as time-series foundation models improve.

We hope our responses address your concerns. If so, we would appreciate it if you could consider raising your score. Please let us know if you have further questions.

References

[1] Xu Liu et al. GIFT-Eval-CTX: A Benchmark For Contextual Forecasting Model Evaluation. arXiv:2610.17794, 2026.

[2] Arjun Ashok et al. Beyond Naïve Prompting: Strategies for Improved Context-aided Forecasting with LLMs. TMLR, 2026.

[3] Mingtian Tan et al. Are Language Models Actually Useful for Time Series Forecasting? NeurIPS, 2024.

[4] Ji Lin et al. VILA: On Pre-training for Visual Language Models. CVPR, 2024.

[5] Sicheng Mo et al. X-Fusion: Introducing New Modality to Frozen Large Language Models. ICCV, 2025.

Reviewer_5qHYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TimeBraid, a unified framework that aligns pretrained language models and pretrained time-series foundation models through interleaved global residual attention layers. The model aims to handle both time-series understanding (generating text outputs) and time-series forecasting (generating continuous temporal sequences) within a single set of model weights, trained on a large-scale curated dataset of series-text pairs and instruction-tuning samples.

Critical strengths
  1. Clean Architecture and Clear Objectives: The overall architectural design is concise and well-structured, with a clearly defined goal of bridging text and time-series modalities.
  2. Comprehensive Baselines and Sub-task Evaluations: The empirical evaluation covers a wide range of downstream sub-tasks across perception, reasoning, and forecasting, and compares the proposed model against a broad set of competitive baselines.
Critical weaknesses
  1. Unclear Motivation and Lack of Mutual Benefit for Joint Modeling: Time-series understanding (text generation) and forecasting (time-series generation) target fundamentally different output modalities. The paper fails to convincingly justify why these two distinct capabilities must be integrated into a single set of shared weights. Without strong empirical evidence demonstrating mutual enhancement—i.e., showing that understanding explicitly improves forecasting or vice versa—combining them into one model appears largely incremental compared to using separate task-specific models.
  2. Insufficient Justification for Expert Selection: The model adopts a classical Mixture-of-Experts (MoE) style architecture, yet the specific choices for the underlying expert backbones are poorly justified. While selecting a general LLM for reasoning and a time-series foundation model for forecasting is intuitive, it remains unclear how the perception expert is selected and why this specific choice is optimal. Moreover, the paper does not explore whether upgrading to newer or stronger backbone models directly translates to base-level performance gains.
  3. Unfair Evaluation Setup and Lack of True Out-of-Distribution (OOD) Benchmarks: Since TimeBraid claims to be a general-purpose model capable of answering diverse time-series queries, it should be rigorously evaluated on held-out, third-party OOD benchmarks with frozen weights. Currently, the comparisons (especially in Table 1) are not entirely fair: TimeBraid and a few fine-tuned baselines are explicitly trained on the curated in-domain datasets, whereas other strong baselines (such as proprietary LLMs and existing time-series language models) are evaluated zero-shot without seeing the same training data. This discrepancy naturally favors TimeBraid and undermines the validity of the comparison.
  4. Unclear Contribution of the Perception Expert vs. Text Shortcut Learning: It is unclear how much temporal information is genuinely extracted and retained by the perception expert. In multimodal time-series QA tasks, models often rely heavily on textual prompts to guess the answers without truly utilizing the time-series signal. The paper lacks a strict ablation removing the perception expert (or the time-series input) to quantify the actual performance gain brought by continuous-signal perception.
Questions for authors
  1. Synergy Between Understanding and Forecasting: Can you provide rigorous ablation studies comparing the unified TimeBraid model against two separately trained single-task variants (one trained purely for understanding and one purely for forecasting under the same architecture and data budget)? Does joint optimization actually yield positive transfer between the two modalities?
  2. Selection and Scaling of Expert Backbones: Could you clarify the exact criteria for selecting the perception expert? Additionally, how sensitive is TimeBraid to the choice of backbones—does replacing the current experts with newer or stronger pretrained backbones yield consistent improvements?
  3. Fair Zero-Shot OOD Evaluation: To ensure a fair comparison against zero-shot proprietary models and existing time-series LLMs, can you evaluate the frozen TimeBraid model on entirely unseen, third-party benchmarks that have no distribution overlap with your 2.2M/4.9M training corpora?
  4. Ablation on the Perception Expert: What happens if the perception expert is completely removed during evaluation (i.e., evaluating the model using text prompts alone)? How much of the current performance is retained solely from textual priors, and how much is genuinely contributed by the perception expert capturing temporal dynamics?
8 comments
AuthorsOpenReview ↗

Official Response to Reviewer 5qHY (Part 1)

Dear Reviewer 5qHY,

Thank you for the careful review. We appreciate that you found the architecture concise and well-structured, the goal of bridging text and time series clear, and our baselines and sub-task evaluations comprehensive. We have conducted new experiments for each of your four questions and will add them to the revised paper. Our responses are as follows.

W1 & Q1: Unclear motivation and lack of mutual benefit for joint modeling

TL;DR: TimeBraid unifies two modalities, time series and language, and our motivation is their mutual benefit inside one set of weights (1× vs. 2× weights compared with two separate multimodal models): language guides forecasting, and the continuous time-series input supports language understanding. The controlled ablation you suggested shows an additional benefit of joint modeling: under identical architectures and training budgets, it outperforms the single-task variants on all four text-conditioned forecasting benchmarks and on two of the four understanding benchmarks, and matches them on the other two.

Below, we explain the motivation, report the ablation, and share a pilot finding.

Motivation. TimeBraid unifies two modalities, and our main claim is the mutual benefit between the two modalities: language guides forecasting, and the continuous time-series input supports language understanding, with one set of weights. With text, Ctrl-F Top-1 improves from the chance level to 43.33% and CAF nCRPS improves from 0.281 to 0.225 compared with the time-series backbone. Without the time-series input, TSAQA accuracy drops from 80.65 to 48.68 (your Q4). The positive transfer between the two tasks in the ablation below is an additional benefit. Regarding separate task-specific models, unimodal models cannot do text-conditioned forecasting (a numerical-only forecaster cannot read text, and an understanding-only model cannot generate time series), while maintaining two separate multimodal models, one for understanding and one for forecasting, doubles parameter memory (about 13.4 GB vs. 26.8 GB in BF16 at 6.7B) and training/deployment cost. A single TimeBraid model outperforms existing task-specific open-source models on most metrics in Tables 1–3 and achieves lower zero-shot MSE than TimesFM-2.5 on ETT and Weather (Table 20). Connecting an independently pretrained LLM and time-series foundation model in their native representations for both understanding and generation is itself non-trivial (Section 4.3).

Ablation (Q1). Following your suggestion, we trained a unified TimeBraid-1.2B, an understanding-only variant and a forecasting-only variant with the same architecture, initialization, and target data budget, and compared the two variants with the jointly trained model. Due to resource limitations, we trained them with 10k alignment steps and 10k SFT steps.

VariantTemporalBench MCQ acc ↑TimeSeriesExam acc ↑TSAQA acc ↑CaTS-Bench DeBERTa F1 ↑CGTSF mean MSE ↓CiK RCRPS ↓CAF CRPS ↓Time-MMD domains won (of 9) ↑
Understanding-only25.150.068.270.4––––
Forecasting-only––––0.4600.3120.2433
Joint modeling31.252.867.970.30.4400.2990.2356
Joint − single-objective+6.1+2.8−0.2−0.1−4.2%−4.1%−3.4%6 : 3

The joint modeling of understanding and forecasting tasks enables positive transfers between two tasks. The understanding clearly helps the model consistently use text better for forecasting on different benchmarks: CGTSF 0.440 vs 0.460 (−4.2%), CiK 0.299 vs 0.312 (−4.1%), CAF 0.235 vs 0.243 (−3.4%; joint ahead at SFT 4k / 6k / 8k / 10k), Time-MMD 6 of 9 domains won. The forecasting helps the model achieve either better or comparable performance on understanding tasks: TemporalBench 31.2 vs 25.1 (+6.1), TimeSeriesExam 52.8 vs 50.0 (+2.8); TSAQA 67.9 vs 68.2 (−0.2) and CaTS DeBERTa F1 70.3 vs 70.4 (−0.1).

We would also like to share an interesting finding during the pilot evaluation in the alignment stage. We took ~3500 unseen controlled forecasting samples and evaluated how the forecasting-only variant and joint modeling work on blank text and the sample with context. We see that joint modeling is much better than the forecasting-only variant with text (0.321 vs 0.571), and we infer that this comes from the better time-series and language alignment from joint modeling, while they achieve similar performance on unimodal forecasting with blank text (1.358 vs. 1.311).

Held-out samples in the alignment stageForecasting-onlyJointJoint − forecasting-only
The sample's own text, MSE ↓0.5710.321−44%
Blank text, MSE ↓1.3111.358+3.6%
Gain from the text (blank − own) ↑0.7401.037+0.297
AuthorsOpenReview ↗

Official Response to Reviewer 5qHY (Part 2)

W2 & Q2: Insufficient justification for expert selection

TL;DR: TimeBraid is a modality-routed Mixture-of-Transformers (not a gated MoE) in which perception and forecasting share one causal TimesFM-2.5 tower, because adding a separate perception tower brings no gain (Fig. 3E). We chose TimesFM-2.5 for its zero-shot accuracy, shared KV-cache decoding, and permissive Apache-2.0 license. Variants trained with other backbones (TimesFM-3.0, Gemma-3-1B), together with our 0.6B → 1.7B → 4B scaling, show that upgrading the time-series tower lowers the error on all four forecasting benchmarks while stronger language towers improve understanding.

We explain the selection below and will make this clear in the revision.

Perception expert. The perception expert uses the same time-series model as the forecasting expert. In our ablation (RQ2, Fig. 3E, App. G.1), using separate time-series models for perception and forecasting did not bring any benefit, so we let the two experts share one time-series tower and only differ in input preprocessing. We would also like to clarify that TimeBraid is a Mixture-of-Transformers and not a classical MoE: tokens are routed by modality, and there is no learned gating.

Forecasting expert. We selected TimesFM-2.5 based on three considerations.

  1. Forecasting performance. When we designed the model, TimesFM-2.5 was among the strongest zero-shot forecasting models on unimodal forecasting benchmarks (e.g., GIFT-Eval).
  2. Decoder-only architecture. A causal time-series tower allows text tokens and time-series patches to share one KV cache at inference. Strong encoder-based models such as Chronos-2 do not support this.
  3. License. TimesFM-2.5 is released under Apache 2.0, so we can release the checkpoints without restrictions for further studies in the community.

Based on these criteria, we consider TimesFM-2.5 a strong choice, while exploring heterogeneous perception and forecasting backbones remains an interesting direction for future work.

Stronger backbones. Following your suggestion, we trained TimeBraid variants with a cross-family language model (Gemma3-1B) and a newer time-series model (TimesFM-3). All variants use the same budget (10k alignment steps and 10k fine-tuning steps).

Language towerTime-series towerTSAQA ↑TSExam ↑CaTS DeBERTa F1 ↑TemporalBench ↑CAF CRPS ↓CiK RCRPS ↓Time-MMD avg. rank ↓ (among the 3)CGTSF mean MSE ↓
Qwen3-0.6BTimesFM-2.567.9352.8270.2631.160.23470.29892.220.440
Qwen3-0.6BTimesFM-3.068.2252.6870.3129.760.23380.29071.780.437
Gemma-3-1B-itTimesFM-2.559.9543.3069.7828.420.23260.29582.000.447

On the time-series side, replacing TimesFM-2.5 with the newer TimesFM-3.0 lowers the error on all four forecasting benchmarks (CiK RCRPS 0.2989 → 0.2907, CGTSF MSE 0.440 → 0.437, CAF CRPS 0.2347 → 0.2338, Time-MMD average rank 2.22 → 1.78), while understanding changes little (TSAQA 67.93 → 68.22, CaTS 70.26 → 70.31, TimeSeriesExam 52.82 → 52.68, TemporalBench 31.16 → 29.76). The margins are small because the two TimesFM versions are close to each other. Qwen3-0.6B has better general performance than Gemma 3-1B, and leads to better unified model performance on the understanding side but not consistently on the forecasting side. We also wish to raise the point that we have shown the benefits of updating stronger LLM backbones in the paper: TimeBraid-1.2B, 2.5B and 6.7B use Qwen3-0.6B, Qwen3-1.7B and Qwen3-4B, and every understanding benchmark in Table 1 improves with the language tower (TSAQA 76.61 → 78.31 → 80.65, TimeSeriesExam 50.13 → 60.05 → 63.14, TemporalBench 34.17 → 35.74 → 36.58, CaTS-Bench DeBERTa F1 0.706 → 0.708 → 0.712), while the forecasting metrics in Table 3 change less and not always in one direction (CAF 0.235 → 0.235 → 0.225, CiK 0.301 → 0.289 → 0.294).

AuthorsOpenReview ↗

Official Response to Reviewer 5qHY (Part 3)

W3 & Q3: Unfair evaluation setup and lack of true OOD benchmarks

TL;DR: We respectfully disagree that the submission lacks a fair evaluation setup: it already includes three out-of-domain benchmarks evaluated with frozen weights (TemporalBench, CiK, and zero-shot ETT and Weather; Table 7). For the in-domain benchmarks (Tables 1 and 2), our primary comparison is against benchmark-specific fine-tuned or supervised baselines, following common practice in time-series language modeling; zero-shot models are listed for reference. We further add five new out-of-domain benchmarks. Evaluated zero-shot with frozen weights, TimeBraid-6.7B is ahead of its language backbone Qwen3-4B on every benchmark in the tables below, ahead of the listed time-series language models and unified models on CiK, MTSFBench-300, TemporalBench and BEDTime recognition, and ahead of GPT-4o on TemporalBench, GPT-5.1 on FactoryBench-Lite and DeepSeek-R1 on MTSFBench-300.

We would like to point out that the submission already includes three out-of-domain benchmarks evaluated with frozen weights (Table 7): (1) TemporalBench, where TimeBraid-6.7B reaches 36.58% vs. 38.50% for GPT-5.4 on QA (Tables 1, 14a) and ranks 1st overall on PSML and MIMIC forecasting ahead of TimesFM-2.5 and GPT-5.4 (Table 14b); (2) CiK (Table 13); and (3) zero-shot long-horizon forecasting on ETT and Weather, where TimeBraid-2.5B achieves lower average MSE than TimesFM-2.5 on all five datasets (Table 20). Following your request, we additionally evaluated TimeBraid on several out-of-domain benchmarks:

Forecasting benchmark (metric ↓)TimeBraid-6.7BBaselines
CiK (RCRPS)0.294Qwen3-4B 0.361TimesFM-2.5 0.295GPT-4o 0.257Chronos-2 0.295ChatTime 0.344TimeOmni-VL 0.655
GIFT-Eval-CTX [1] (relative MASE)0.773Qwen3-4B 1.502TimesFM-2.5 0.797Claude Opus 4.8 0.781TimesFM-3 0.780Moirai-2 0.782Chronos-2 0.795
MTSFBench-300 [2] (mean MSE)0.711Qwen3-4B 1.471TimesFM-2.5 0.801DeepSeek-R1 0.899QwQ-32B 0.961TimesFM-2.0 0.819ChatTime-7B 1.175
Understanding benchmark (accuracy % ↑)TimeBraid-6.7BBaselines
TemporalBench36.6Qwen3-4B 33.1Claude-Sonnet-4 36.1GPT-4o 32.3TimeOmni-VL 32.3ChatTime-7B 28.5TimeOmni-1-7B 25.7
FactoryBench-Lite [3] L1 / L2 / L3 (chance-corrected, 0 = chance)+4.3 / +7.6 / +29.4Qwen3-4B +0.6 / −4.2 / +5.8GPT-5.1 +3.2 / +6.9 / +27.5Mistral Large 3 +5.9 / +5.3 / +19.7Qwen3-235B +3.6 / +4.9 / +20.6DeepSeek V3.2 −0.3 / +5.8 / +17.1Simple Baseline (non-LLM) +4.3 / −0.2 / +6.2
TS-Instruct QA Gold [4]76.6Qwen3-4B 73.6GPT-4o 78.5GPT-4o-mini 74.6Phi-3-medium 64.6Ministral-8B 58.7Llama-3.1-8B 54.2
BEDTime recognition [5]63.1Qwen3-4B 56.6GPT-4o 69.6Qwen2.5-7B 57.8Llama-3.1-8B 50.0ChatTS-14B 62.2ChatTime-7B 26.0
BEDTime differentiation [5]57.4Qwen3-4B 41.0GPT-4o 71.0Qwen2.5-7B 45.4Llama-3.1-8B 34.4ChatTS-14B 61.4ChatTime-7B 25.8

On these benchmarks, every model, including TimeBraid, is evaluated zero-shot. With frozen weights, TimeBraid-6.7B is ahead of its language backbone Qwen3-4B on every row, and ahead of its time-series backbone TimesFM-2.5 on MTSFBench-300 (0.711 vs. 0.801). It is also ahead of the time-series language models and unified models listed above on CiK, MTSFBench-300, TemporalBench and BEDTime recognition (e.g., CiK RCRPS 0.294 vs. 0.344 for ChatTime, TemporalBench 36.6 vs. 32.3 for TimeOmni-VL, BEDTime recognition 63.1 vs. 62.2 for ChatTS-14B). Against the frontier LLMs, also evaluated zero-shot, it is ahead of GPT-4o on TemporalBench (36.6 vs. 32.3), of GPT-5.1 on all three FactoryBench-Lite levels, and of DeepSeek-R1 on MTSFBench-300 (0.711 vs. 0.899), and it is within two points of GPT-4o on TS-Instruct (76.6 vs. 78.5).

Clarification about the evaluation setup in the submission. Table 1 has several groups of baselines. Closed-source models are listed as references and are excluded from the ranking. For the in-domain benchmarks, the fair comparison is with the "Benchmark-Specific Fine-Tuned Models". These models are fine-tuned specifically on the corresponding training split (e.g., LLaMA3.1-8B on TSAQA and Idefics2 on CaTS-Bench), and each of them is trained on a single benchmark. Likewise, on Time-MMD and CGTSF (Table 2), the baselines include supervised models trained on each dataset (e.g., Time-LLM, GPT4MTS and Time-VLM). In-domain evaluation alongside benchmark-specific fine-tuned baselines is common practice in time-series language modeling [6][7][8], because time-series domains are heterogeneous and have different annotation rules. The zero-shot open-source models are included for reference.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

We sincerely thank all reviewers for their constructive feedback. We are encouraged that the reviewers find the problem essential and the alignment strategy reasonable (Reviewer MyWN), the architecture concise and the evaluation comprehensive (Reviewer 5qHY), the writing clear (Reviewers 5qHY and MyWN), and the cross-modal integration promising, with practical findings on joint training (Reviewer REpX).

After our responses, Reviewer 5qHY acknowledges that the new ablations "provide empirical evidence of positive transfer between understanding and forecasting", and Reviewer MyWN finds that the fusion ablation and the backbone replacement study help "clarify the contribution and flexibility of the method". Some of these points bear on other reviewers' concerns. The positive transfer acknowledged by Reviewer 5qHY is a measured benefit of the unified architecture, which Reviewer REpX asks for in W1. The flexibility that Reviewer MyWN sees in the backbone replacement study speaks to the question of other backbones raised by Reviewers 5qHY and REpX (W2). Reviewer REpX notes that our stability ablations show "more stable optimization for both modalities", which speaks to Reviewer 5qHY's question of how the two losses are balanced.

Below is a summary of our responses. Numbers refer to the revised manuscript, where new content is highlighted in blue.

Changes: The main text has two new research questions in the ablation studies, joint versus single-objective training (RQ6, Table 5) and reliance on the time-series input (RQ7, Table 6). A new Appendix G reports seven new benchmarks evaluated with frozen weights, agent frameworks, and result analysis by horizon and input length (Tables 23 to 29). Appendix H adds fusion and backbone ablations (Tables 32 to 34).

To Reviewer 5qHY: We clarified that TimeBraid unifies two modalities: it builds unified time-series and language representations by connecting an independently pretrained language model and a pretrained time-series foundation model for understanding and forecasting (RQ1 to RQ5 identify what makes this work). In the comparison you suggested, joint training is ahead of the single-objective variants on 6 of 8 benchmarks and level on the other two (RQ6, Table 5). We explained why perception and forecasting share one time-series tower and how the backbone was chosen (RQ2, Appendix C), and trained variants with TimesFM-3.0 and Gemma-3-1B (Table 33). We added five out-of-domain benchmarks evaluated with frozen weights (Tables 23 and 24) and the ablation without the time-series input, which lowers TSAQA accuracy from 80.65 to 48.68 (RQ7, Table 6). In reply to your follow-up, we positioned TimeBraid relative to SciTS (Appendix I), and explained why understanding and forecasting reinforce each other, as the two directions of one correspondence between series and language, and how the losses are balanced (Appendix H.4, Table 31; RQ3 and RQ4).

To Reviewer MyWN: With text, TimeBraid-6.7B has lower or equal error than its TimesFM-2.5 backbone on CGTSF (23% lower), CAF (20%), MTSFBench-300 (11%), GIFT-Eval-CTX (3%) and CiK (level) (Tables 3, 18 and 23). Without text, TimeBraid-2.5B stays within 4 to 7% of it in MASE (Table 26). We added a fusion ablation with matched parameter counts, where global residual attention is ahead of a bidirectional MLP connector on 6 of 8 benchmarks (Table 32). We added the comparison with agent frameworks on the TemporalBench leaderboard (Table 25) and discussed time-series agents (Appendix I). We replaced the time-series tower with TimesFM-3.0 and tuned only the global residual attention layers for 2,000 steps, which recovers most of the drop on TSAQA, CAF and CGTSF (Table 34). In reply to your follow-up, we set these gains against the training cost of 120 to 194 GPU-hours (Table 11), which is paid once for one model that covers understanding and forecasting with and without text.

To Reviewer REpX: We stated what unification provides: one model for all tasks in the paper, which none of the baselines cited in W1 covers, with lower or equal error than either backbone alone when forecasting with text (Table 3). On the results cited in W1, one epoch of fine-tuning on TSAQA brings TimeBraid-1.2B to 85.89%, above the fine-tuned LLaMA3.1-8B (Appendix H.6). We revised the claim about the backbones to "largely retains" (Introduction, Section 4.2, Conclusion) and showed that 500 steps of distillation raise MMLU from 50.68 to 54.61 (Table 30). We added variants with other backbones (Table 33), results by horizon and input length (Tables 27 to 29), the Qwen3-4B row in Table 3, and the fev-bench and TIME results (Table 26), and redrew Figures 2, 3, 4 and 6 with larger fonts. In reply to your follow-up, we listed the results per benchmark and the benefits of unification beyond complementarity, such as one generation pass for both directions and lower error from joint training (Table 5).

Please see our detailed responses to each point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.