- 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
- 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.
- 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
-
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.
-
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.
-
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
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.
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-6.7B | TimesFM-2.5 | Diff |
|---|---|---|---|---|
| CAF (nCRPS ↓) | ≤ 32 | 0.183 | 0.230 | −0.047 |
| 33–64 | 0.294 | 0.364 | −0.070 | |
| CiK (RCRPS ↓) | 6–7 | 0.163 | 0.166 | −0.002 |
| 15–24 | 0.339 | 0.352 | −0.013 | |
| 29–99 | 0.309 | 0.306 | +0.003 | |
| GIFT-Eval-CTX (rel. MASE ↓) | 30–60 | 0.685 | 0.764 | −0.079 |
| 480–600 | 0.872 | 0.812 | +0.060 | |
| 720–900 | 0.864 | 0.851 | +0.013 |
Text-conditioned forecasting by history length (same comparison; lower is better):
| Benchmark (metric) | History length | TimeBraid-6.7B | TimesFM-2.5 | Diff |
|---|---|---|---|---|
| CGTSF (MSE ↓, mean of 3 datasets; Diff: mean per-dataset change) | 48-192 | 0.426 | 0.563 | −28.4% |
| 72-288 | 0.415 | 0.553 | −29.3% | |
| 96-384 | 0.405 | 0.515 | −26.1% | |
| 120-480 | 0.405 | 0.510 | −25.4% | |
| CAF (nCRPS ↓) | ≤ 32 | 0.198 | 0.260 | −0.062 |
| 33–128 | 0.222 | 0.272 | −0.050 | |
| 129–500 | 0.266 | 0.322 | −0.056 |
Understanding by input length (accuracy, higher is better):
| Benchmark (accuracy ↑) | Input length | TimeBraid-6.7B | Qwen3-4B (text input) | Diff |
|---|---|---|---|---|
| TimeSeriesExam | 1 series (about 1,024 points) | 64.1 | 47.4 | +16.7 |
| 2 series (about 2,048 points) | 61.2 | 53.2 | +8.0 | |
| TSAQA (all 42,000) | 1–32 | 54.9 | – | |
| 33–64 | 77.6 | – | ||
| 65–128 | 86.5 | – | ||
| 129–256 | 87.6 | – | ||
| 257–512 | 84.4 | – |
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.
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:
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.