Pith. sign in

REVIEW 4 major objections 6 minor 47 references

TSROUTER claims that joint modality–model routing on a heterogeneous graph is the right way to answer time-series reasoning queries, reporting 51.33% accuracy on TSRBench.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 07:46 UTC pith:FXKNWDE3

load-bearing objection Joint modality-model routing is a real advance and the in-distribution results are solid, but the zero-shot new-model claim rests on a single pair of models and one aggregate number. the 4 major comments →

arxiv 2607.08940 v2 pith:FXKNWDE3 submitted 2026-07-09 cs.LG

TSRouter: Dynamic Modality-Model Selection for Time Series Reasoning

Fangxu Yu , Tao Feng , Dehai Min , Lu Cheng , Ge Liu , Tianyi Zhou This is my paper
classification cs.LG
keywords time series reasoningmodel routingmodality selectionheterogeneous graphLLM routingVLM routingcost-aware routingzero-shot generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TSROUTER is a routing framework that decides, per time-series query, which input modality—textual numbers, visual chart, or both—and which off-the-shelf language or vision-language model should produce the answer. The paper's premise is that no single modality or model dominates: text preserves numerical precision but misses global shape, vision captures patterns but loses fine detail, and smaller and larger models succeed on largely disjoint sets of queries. To exploit this complementarity, TSROUTER builds a heterogeneous graph whose nodes are tasks, queries, modalities, and models, and scores each modality–model candidate against a user-weighted performance–cost objective. On the four-task TSRBench suite, it reports 51.33% overall accuracy versus 44.07% for the best prior router, 16–46% relative gains over all baselines, and zero-shot generalization to unseen models and two held-out reasoning tasks.

Core claim

On the paper's own terms, the central discovery is that joint modality-model selection for time series reasoning can be framed as candidate scoring over a heterogeneous graph, and that this framing transfers to models and tasks never seen in training. The load-bearing comparison is Table 3: TSROUTER reaches 51.33% overall accuracy on TSRBench while the best competing router reaches 44.07%, at comparable API cost. The authors attribute the gap to the graph's ability to propagate contextual signals across task, query, modality, and model nodes, and to soft-target training over the full ranking of candidates rather than a single best label.

What carries the argument

The central object is the heterogeneous graph G=(V,E) with four node types—task, query, modality, and model—and five edge types: task–query, query–modality, modality–model, query–model, and query–query k-nearest-neighbor edges. Node features are text embeddings of natural-language profiles: task descriptions, question text concatenated with time-series statistics, modality property descriptions, and model capability-and-cost profiles. A two-layer Heterogeneous Graph Transformer (a graph network that updates each node's embedding by attending over type-specific neighbors) propagates information across these edges. A candidate's score for a query is the sum of the modality and model embeddings

Load-bearing premise

The load-bearing premise is that the graph network can estimate a completely unseen model's per-query effectiveness from its text description and graph-neighbor messages alone, without ever observing that model's correctness; if descriptions and message passing cannot transfer to genuinely new models, the zero-shot plug-and-play claim collapses.

What would settle it

Insert a deliberately misdescribed unseen model into the graph—for example, a weak vision model profiled as a strong numerical reasoner—and check whether it receives the same share of queries as an accurately profiled model with identical true capabilities. If routing accuracy collapses when descriptions are corrupted, the result depends on the text prior; if it does not, the graph structure rather than the profiles is doing the work, and a second test (removing all query-query edges) would show whether the claimed generalization survives without query similarity.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim is right, systems can route each time-series query to the best modality-model pair and obtain 16–46% relative accuracy improvements over existing routers at comparable or lower cost.
  • Newly released models can be inserted into the router by adding a node and a text profile; the paper reports that adding two unseen models raised overall accuracy from 51.3% to 53.5%, while several baselines degraded.
  • The router transfers to unseen task types: on correlation prediction it reaches 31.38% accuracy versus 29.59% for the best baseline, and on imputation it reports lower MSE (0.56 vs. 0.60) and MAE (0.43 vs. 0.46).
  • The explicit cost coefficient α gives operators a tunable accuracy-cost frontier; the paper shows that at the cost budget of the best baseline, TSROUTER still exceeds all baselines in accuracy.
  • The ablations show the heterogeneous structure, query-query edges, modality-model edges, and the MLP scoring head each contribute; removing any one drops overall accuracy by roughly 2.8 to 5 points.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the same heterogeneous scoring recipe should transfer to other settings where two complementary input views and a zoo of models coexist—for example, code tasks (text vs. execution traces) or medical charts (numeric tables vs. images).
  • Beyond the paper: because query-query edges are built from embedding similarity, the router's ceiling likely depends on the embedding model; a testable extension is to replace the fixed text embedder with a task-finetuned similarity and measure whether routing accuracy rises, especially at low training-data fractions.
  • Beyond the paper: the zero-shot new-model result implies the text profile is doing real work; deliberately corrupting a new model's description (e.g., swapping capability claims) would reveal whether the GNN relies on semantic priors or on structural position in the graph.
  • Beyond the paper: the cost term is a single scalar normalized across candidates; using latency, energy, or a nonlinear cost in the same objective should yield a different Pareto front, which would show whether the framework's cost-accuracy control is robust to the choice of cost definition.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes TSRouter, a heterogeneous-graph router that jointly selects an input modality (text, vision, or both) and a foundation model (LLM or VLM) for each time-series reasoning query. Node features are initialized from LLM-generated text profiles of tasks, queries, modalities, and models; an HGT backbone performs message passing over five relation types; and candidate scoring is trained with soft targets derived from oracle effectiveness (Eq. 1). Experiments on TSRBench report 51.33% overall accuracy versus 44.07% for the best baseline (Table 3), plus claimed zero-shot generalization to two unseen models (Figure 3) and two held-out tasks (Table 4), with additional ablations, cost-preference sweeps, and efficiency measurements.

Significance. If the empirical claims hold, TSRouter makes a useful contribution: it is the first routing framework I am aware of that jointly optimizes modality and model selection for time-series reasoning, and the graph formulation provides a natural way to inject new models and tasks at test time. The paper is clearly written, the problem formulation is sound, and the authors include code, hyperparameter sensitivity experiments, component ablations, and an efficiency analysis. The central risk is not in the architecture but in the validation: the headline comparisons and the zero-shot generalization claims rest on single runs, an unspecified correctness criterion, and very thin evidence for the unseen-model setting.

major comments (4)
  1. [§3.1, Table 3, Figure 3, Table 4] No error bars or significance tests are reported for any of the main empirical claims. Table 7 shows that TSRouter itself has standard deviations of 0.2–0.5 accuracy points across 3 runs, so differences such as Perception 67.63 vs. 62.58 or Decision 42.45 vs. 39.62 are not obviously outside run-to-run noise. Please report means±std over multiple runs (or seeds) for the key tables and figures, state the API decoding temperature/sampling, and run an appropriate significance test. Without this, the central claim of 'substantially outperforms' is not quantitatively supported.
  2. [§3.1 Evaluation Metrics] The correctness criterion for TSRBench answers is never defined. Eq. (1) and the soft training targets in Eq. (5) depend on the oracle binary correctness I(q,c) for every candidate, so the router's supervision is only as reliable as this labeling. Specify how responses are scored (exact string match, rubric, multiple-choice extraction, partial credit, handling of abstentions). For the imputation task, which is evaluated with MSE/MAE, also state how correctness is defined if the router is trained on this task, or clarify that the router is applied zero-shot with an objective different from the evaluation metric.
  3. [§3.1 Baselines] Baseline implementations and hyperparameter tuning are not described. Adapting methods such as GraphRouter, which was designed for query–model edge prediction, to joint modality–model candidate selection is nontrivial, and treating each (modality, model) pair as an independent candidate may not be the intended use of those baselines. Without reporting search ranges, final hyperparameters, or using the authors' released implementations, the 16–46% relative improvements may partly reflect undertuned baselines. Please provide this information.
  4. [§2.3, Figure 3] The zero-shot plug-and-play claim for unseen models is under-validated. Only two unseen models (one LLM, one VLM) are tested, and the only evidence is the aggregate overall accuracy moving from 51.3% to 53.5%. For these new model nodes, Eq. (7) provides no correctness supervision, so the method's success depends on a transferable mapping from text profiles plus graph-neighbor messages to per-query effectiveness. A single aggregate number does not rule out that these two particular models are easy to rank (e.g., because their scale and cost make them almost always beneficial) or that the result depends on their specific text profiles. Please add per-task results for the unseen models, test with additional unseen models of varying distinctiveness, and include a diagnostic such as the rank correlation between predicted and actual per-query effectiveness for the unseen models.
minor comments (6)
  1. [Abstract] The '16% to 46% relative improvements' are not tied to specific baselines; the 16% figure matches the overall comparison to Hybrid LLM, while 46% matches the comparison to the largest-LLM rule. Please state the reference point.
  2. [Eq. (4)] The notation 'h_q d (h_m + h_l)' contains a stray 'd'; clarify that denotes element-wise product or another defined operator.
  3. [§2.1, Figure 1(b)] The terms 'smallest' and 'largest' model are used without specifying which models they refer to among the candidate set; clarify for each modality.
  4. [Table 3 and elsewhere] The paper alternates between 'TSRouter' (figures, Table 5) and 'TSROUTER' (text, tables); use one consistent spelling throughout.
  5. [Table 3 cost] Cost is reported in total USD per test set; clarify whether this includes routing overhead (e.g., GNN inference, embedding calls) and whether it is averaged over the test set, since per-task costs in the table sum to the overall cost.
  6. [Appendix A.3, Table 7] This is the only table with standard deviations; consider moving this practice to the main tables or explicitly stating the number of runs for the main experiments.

Circularity Check

0 steps flagged

No circularity: TSRouter trains on oracle effectiveness labels and is evaluated on held-out accuracy; the zero-shot claims are genuine extrapolation, not fitted predictions.

full rationale

TSROUTER's derivation chain is self-contained supervised learning. Eq. (1) defines an effectiveness score from oracle correctness I(q,c) and normalized cost; Eq. (5) converts these scores into soft targets; Eq. (7) minimizes KL divergence against the router's predicted candidate distribution; Eq. (8) selects the argmax at test time. Evaluating the trained router on held-out accuracy is standard supervised evaluation, not circular: the training objective and the evaluation metric are intentionally aligned, and the evaluation uses query/candidate data not seen in training. The zero-shot claims in §2.3 and §3.3 are also not circular. New model nodes (Qwen3.5-397B-A17B, Kimi-K2.5) are inserted with text-profile features and graph edges but no correctness labels in the training objective; the GNN must transfer learned query-model relations to these unseen nodes. That is a genuine extrapolation, even if the evidence base (one LLM and one VLM) is thin. No self-citation is load-bearing: TSRBench is used as an external evaluation benchmark and GraphRouter as a baseline; neither is invoked to define away alternatives or to justify the core routing mechanism. There is no imported uniqueness theorem and no ansatz smuggled in through citation. The paper's limitations are empirical validation concerns (e.g., small set of unseen models, single benchmark) rather than circular reasoning.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The free parameters are standard hyperparameters and the cost-preference coefficient. The key domain assumptions are deterministic labels, sufficiency of text profiles, and the completeness of the chosen edge types.

free parameters (5)
  • alpha (cost preference) = 1.0 main; varied 0.2-1.0 in Fig. 5
    User-defined trade-off between accuracy and cost in Eq. (1); chosen per experiment, not learned.
  • k (nearest neighbors) = 60
    Selected on validation set via hyperparameter search (§3.1).
  • embedding_dim = 64
    Selected on validation set (§3.1, Figure 7b).
  • num_GNN_layers = 2
    Selected on validation set (§3.1, Figure 7a).
  • learning_rate = 1e-3
    Adam optimizer; chosen via validation (§3.1).
axioms (4)
  • domain assumption Candidate correctness is a deterministic binary label for each query-candidate pair
    Eq. (1) treats I(q,c) as a fixed 0/1, ignoring LLM/VLM response stochasticity; no repeated sampling is reported (§2.2).
  • ad hoc to paper LLM-generated text descriptions (e.g., GPT-5) encode sufficient task/model/modality information for routing without benchmark leakage
    Node features are initialized from these descriptions (§2.3, Appendix B); the paper asserts no data leakage but provides no verification.
  • ad hoc to paper The five edge types (E_TQ, E_QM, E_MM, E_QL, E_QQ) capture the interactions required for optimal routing
    Edge set defined in §2.3 without theoretical or empirical justification for completeness.
  • domain assumption A statistical summary (mean, std, min/max, trend direction) of the raw time series preserves enough information for query-level modality/model selection
    Query node features use these summary statistics, discarding the raw series (§2.3, Appendix C).

pith-pipeline@v1.3.0-alltime-deepseek · 50 in / 13044 out tokens · 123260 ms · 2026-08-02T07:46:17.308522+00:00 · methodology

0 comments
read the original abstract

Time series reasoning is essential for real-world problem-solving. While both Large Language Models (LLMs) and Vision-Language Models (VLMs) can reason about time-series data, their capabilities are complementary: LLMs process time series as text sequences and thus preserve exact numerical understanding, but struggle with global patterns, whereas VLMs efficiently capture these patterns by visualizing time series but may lose fine-grained details. Moreover, models vary significantly in task-specific expertise and inference costs. Dynamically selecting the most suitable modality and model for each query is therefore crucial, yet challenging because it requires modeling the complex interactions among tasks, queries, modalities, and models, which carry rich contextual signals. To this end, we introduce TSRouter, a graph-based dynamic routing framework. TSRouter constructs a heterogeneous graph of task, query, modality, and model nodes to contextualize the interactions among query characteristics, modality attributes, and model capabilities. TSRouter formulates routing as a candidate scoring problem, where each modality-model pair is evaluated based on user-defined performance-cost preferences to select the optimal candidate. Comprehensive evaluations on 4 distinct time series reasoning tasks reveal that TSRouter substantially outperforms diverse baselines with 16\% to 46\% relative improvements. Furthermore, TSRouter demonstrates robust zero-shot plug-and-play generalization to unseen models and novel tasks and preserves high performance while reducing computational overhead through cost-aware optimization. Our code is available at https://github.com/tianyi-lab/TSRouter.

Figures

Figures reproduced from arXiv: 2607.08940 by Dehai Min, Fangxu Yu, Ge Liu, Lu Cheng, Tao Feng, Tianyi Zhou.

Figure 1
Figure 1. Figure 1: Motivation analysis results. (a) Distribution of best-performing modality across [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TSROUTER. Each query is paired with all modality–model combina￾tions to collect performance and cost, forming the training data. Then, TSROUTER builds a heterogeneous graph from task, query, modality, and model descriptions, learns node rep￾resentations, and scores candidates to route queries to the optimal modality–model pair. features and stabilize optimization: h pℓ`1q v “ h pℓq v ` HGTConv´… view at source ↗
Figure 3
Figure 3. Figure 3: Results on generalization to new mod￾els. Scenario Correlation Prediction Imputation Accuracy MSE MAE LLM (largest) 22.24 0.67 0.46 VLM (largest) 29.20 0.89 0.46 EloRouter 25.42 1.07 0.65 MFRouter 25.22 1.07 0.65 KNNRouter 24.93 0.99 0.59 GraphRouter 27.51 0.61 0.49 Hybrid LLM 29.59 0.60 0.46 RouterDC 28.60 0.74 0.47 CausalLM 27.81 0.89 0.47 Router-R1 29.59 0.70 0.48 TSROUTER 31.38 0.56 0.43 [PITH_FULL_IM… view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy–cost trade-off under varying cost preference [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Modality routing distribution across tasks of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Hyperparameter sensitivity analysis. (a) Number of HGT layers; (b) Embed [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance of TSROUTER us￾ing different amounts of training data. Effect of the Amount of Training Data. We further examine how TSROUTER performs un￾der varying amounts of training data, compar￾ing against the strongest baseline. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance with different embed￾ding models for TSROUTER. A Additional Experimental Results A.1 Performance under Different GNNs. To investigate the performance with different GNN backbones, we replace the HGT with two widely used alternatives, HAN and HeteroGAT, while keeping all other components unchanged. As shown in [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Case study from the Perception task. Model (Modality) Reasoning Process Qwen3-8B (Text) Ans: (B) — The time series data shows an initial steady increase (linear trend), followed by a more rapid growth (exponential trend), and eventually a leveling off or slower growth (logarithmic trend). The data starts with a gradual rise, then accelerates, and later appears to stabilize. . . ✗ Qwen3-32B (Text) Ans: (B) … view at source ↗
Figure 10
Figure 10. Figure 10: Description of the Perception task. Reasoning Task Description Reasoning tasks require models to go beyond observation and construct logical explanations or conclusions from time series data. This category broadly covers inferring causes, discovering relationships, applying rules, and performing quantitative analysis, demanding strong analyti￾cal and multi-step thinking capabilities [PITH_FULL_IMAGE:figu… view at source ↗
Figure 11
Figure 11. Figure 11: Description of the Reasoning task. Prediction Task Description Prediction tasks ask models to forecast future values or outcomes based on historical time se￾ries. Models need to capture temporal dependencies and extrapolate patterns, requiring both quantitative estimation and event anticipation skills [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Description of the Prediction task. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Description of the Decision-Making task. [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Description of the Text modality. Visual Modality Description Visual modality represents time series as rendered line chart images. It emphasizes spatial per￾ception and visual pattern recognition, requiring models to interpret curves, axes, and chart structures without access to raw numbers [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 16
Figure 16. Figure 16: Description of using both textual and visual modality. [PITH_FULL_IMAGE:figures/full_fig_p020_16.png] view at source ↗
Figure 18
Figure 18. Figure 18: Qwen3-32B description. LLaMA-3.3-70B-Turbo Description A large text-only language model with 70 billion parameters from Meta. Its scale provides advanced capabilities for complex multi-step logical inference and nuanced language under￾standing, at a higher cost (input: $0.25/M tokens, output: $0.75/M tokens). It is best suited for demanding tasks where smaller models fall short [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 19
Figure 19. Figure 19: LLaMA-3.3-70B-Turbo description. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Qwen3-VL-8B-Instruct description. Qwen3-VL-32B-Instruct Description A powerful vision-language model with 32 billion parameters from Alibaba, supporting both visual and multimodal inputs. It combines strong visual perception with solid analytical rea￾soning capabilities, at a moderate cost (input: $0.16/M tokens, output: $0.64/M tokens). It is well-suited for tasks requiring sophisticated interpretation o… view at source ↗
Figure 21
Figure 21. Figure 21: Qwen3-VL-32B-Instruct description. GLM-4.5V Description A large mixture-of-experts vision-language model with 106 billion parameters from Zhipu AI, supporting both visual and multimodal inputs. Its MoE architecture enables efficient scaling with strong visual and domain-specific capabilities, at a higher cost (input: $0.20/M tokens, output: $1.10/M tokens). It is designed for tasks demanding nuanced visua… view at source ↗
Figure 22
Figure 22. Figure 22: GLM-4.5V description. Qwen3.5-397B-A17B Description A very large mixture-of-experts text-only language model with 397 billion parameters from Al￾ibaba, with 397 billion total parameters and 17 billion active parameters per forward pass. Its MoE architecture delivers strong reasoning capabilities at a relatively moderate active compute cost, though pricing reflects its scale (input: $0.65/M tokens, output:… view at source ↗
Figure 23
Figure 23. Figure 23: Qwen3.5-397B-A17B description. Kimi-K2.5 Description A large vision-language model from Moonshot AI with built-in chain-of-thought reasoning ca￾pabilities, supporting both visual and multimodal inputs. It combines strong visual perception with multi-step logical reasoning, at a moderate cost (input: $0.5/M tokens, output: $2.80/M tokens). It is suited for tasks requiring integrated visual understanding an… view at source ↗
Figure 24
Figure 24. Figure 24: Kimi-K2.5 description. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_24.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 33 linked inside Pith

  1. [1]

    Let’s sample step by step: Adaptive- consistency for efficient reasoning and coding with llms.arXiv preprint arXiv:2305.11860,

    Pranjal Aggarwal, Aman Madaan, Yiming Yang, et al. Let’s sample step by step: Adaptive- consistency for efficient reasoning and coding with llms.arXiv preprint arXiv:2305.11860,

    Pith/arXiv arXiv

  2. [3]

    ✓ Qwen3-VL-8B(Text+Vision) Ans: (D)— The data shows a clear progression through multiple trend phases

    gradually decreasing slope (logarithmic). ✓ Qwen3-VL-8B(Text+Vision) Ans: (D)— The data shows a clear progression through multiple trend phases. Initially (to„40), rapid accelerating growth characteristic of an exponential trend. After index 40, growth slows down significantly. . . ✓ Qwen3-VL-32B(Text+Vision) Ans: (B)— The data starts with a slow increase...

    2026

  3. [5]

    Tempo: Prompt-based generative pre-trained transformer for time series forecasting

    Defu Cao, Furong Jia, Sercan O Arik, Tomas Pfister, Yixiang Zheng, Wen Ye, and Yan Liu. Tempo: Prompt-based generative pre-trained transformer for time series forecasting. arXiv preprint arXiv:2310.04948,

    Pith/arXiv arXiv

  4. [6]

    Mtbench: A multimodal time series benchmark for temporal reasoning and question answering.arXiv preprint arXiv:2503.16858,

    Jialin Chen, Aosong Feng, Ziyu Zhao, Juan Garza, Gaukhar Nurbek, Cheng Qin, Ali Maatouk, Leandros Tassiulas, Yifeng Gao, and Rex Ying. Mtbench: A multimodal time series benchmark for temporal reasoning and question answering.arXiv preprint arXiv:2503.16858,

    arXiv

  5. [7]

    Visionts: Visual masked autoencoders are free-lunch zero-shot time series forecasters

    Mouxiang Chen, Lefei Shen, Zhuo Li, Xiaoyun Joy Wang, Jianling Sun, and Chenghao Liu. Visionts: Visual masked autoencoders are free-lunch zero-shot time series forecasters. arXiv preprint arXiv:2408.17253, 2024a. Shuhao Chen, Weisen Jiang, Baijiong Lin, James Kwok, and Yu Zhang. Routerdc: Query- based router by dual contrastive learning for assembling lar...

    Pith/arXiv arXiv

  6. [9]

    Hybrid llm: Cost-efficient and quality-aware query routing.arXiv preprint arXiv:2404.14618, 2024a

    Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Ruhle, Laks VS Lakshmanan, and Ahmed Hassan Awadallah. Hybrid llm: Cost-efficient and quality-aware query routing.arXiv preprint arXiv:2404.14618, 2024a. Ruomeng Ding, Chaoyun Zhang, Lu Wang, Yong Xu, Minghua Ma, Wei Zhang, Si Qin, Saravan Rajmohan, Qingwei Lin, and Dongmei Zha...

    Pith/arXiv arXiv 2024

  7. [10]

    Graphrouter: A graph-based router for llm selections.arXiv preprint arXiv:2410.03834,

    Tao Feng, Yanzhen Shen, and Jiaxuan You. Graphrouter: A graph-based router for llm selections.arXiv preprint arXiv:2410.03834,

    Pith/arXiv arXiv

  8. [11]

    Tao Feng, Haozhen Zhang, Zijie Lei, Peixuan Han, and Jiaxuan You

    GitHub repository. Tao Feng, Haozhen Zhang, Zijie Lei, Peixuan Han, and Jiaxuan You. Graphplan- ner: Graph memory-augmented agentic routing for multi-agent llms.arXiv preprint arXiv:2604.23626,

    Pith/arXiv arXiv

  9. [12]

    Fast graph representation learning with pytorch geo- metric.arXiv preprint arXiv:1903.02428,

    Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geo- metric.arXiv preprint arXiv:1903.02428,

    Pith/arXiv arXiv 1903

  10. [14]

    Timeomni-1: Incen- tivizing complex reasoning with time series in large language models.arXiv preprint arXiv:2509.24803,

    Tong Guan, Zijie Meng, Dianqi Li, Shiyu Wang, Chao-Han Huck Yang, Qingsong Wen, Zuozhu Liu, Sabato Marco Siniscalchi, Ming Jin, and Shirui Pan. Timeomni-1: Incen- tivizing complex reasoning with time series in large language models.arXiv preprint arXiv:2509.24803,

    Pith/arXiv arXiv

  11. [15]

    Timeomni-vl: Unified models for time series understanding and generation

    Tong Guan, Sheng Pan, Johan Barthelemy, Zhao Li, Yujun Cai, Cesare Alippi, Ming Jin, and Shirui Pan. Timeomni-vl: Unified models for time series understanding and generation. arXiv preprint arXiv:2602.17149,

    Pith/arXiv arXiv

  12. [16]

    Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992,

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhit- ing Hu. Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992,

    Pith/arXiv arXiv

  13. [17]

    Arcmemo: Abstract reasoning composition with lifelong llm memory.arXiv preprint arXiv:2509.04439,

    Matthew Ho, Chen Si, Zhaoxiang Feng, Fangxu Yu, Zhijian Liu, Zhiting Hu, and Lianhui Qin. Arcmemo: Abstract reasoning composition with lifelong llm memory.arXiv preprint arXiv:2509.04439,

    arXiv

  14. [18]

    Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006,

    Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006,

    Pith/arXiv arXiv

  15. [19]

    Routerbench: A benchmark for multi- llm routing system.arXiv preprint arXiv:2403.12031,

    Qitian Jason Hu, Jacob Bieker, Xiuyu Li, Nan Jiang, Benjamin Keigwin, Gaurav Ranganath, Kurt Keutzer, and Shriyash Kaustubh Upadhyay. Routerbench: A benchmark for multi- llm routing system.arXiv preprint arXiv:2403.12031,

    Pith/arXiv arXiv

  16. [20]

    Heterogeneous graph trans- former

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. Heterogeneous graph trans- former. InProceedings of the web conference 2020, pp. 2704–2710,

    2020

  17. [22]

    Time-llm: Time series forecasting by reprogramming large language models.arXiv preprint arXiv:2310.01728,

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, et al. Time-llm: Time series forecasting by reprogramming large language models.arXiv preprint arXiv:2310.01728,

    Pith/arXiv arXiv

  18. [23]

    Reasoning with sampling: Your base model is smarter than you think.arXiv preprint arXiv:2510.14901,

    Aayush Karan and Yilun Du. Reasoning with sampling: Your base model is smarter than you think.arXiv preprint arXiv:2510.14901,

    Pith/arXiv arXiv

  19. [24]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Pith/arXiv arXiv

  20. [27]

    Qinghua Liu, Sam Heshmati, Zheda Mai, Zubin Abraham, John Paparrizos, and Liu Ren

    URLhttps://api.semanticscholar.org/CorpusID:273963297. Qinghua Liu, Sam Heshmati, Zheda Mai, Zubin Abraham, John Paparrizos, and Liu Ren. Mllm4ts: Leveraging vision and multimodal language models for general time-series analysis.arXiv preprint arXiv:2510.07513,

    arXiv

  21. [29]

    Lan- guage models still struggle to zero-shot reason about time series.arXiv preprint arXiv:2404.11757,

    Mike A Merrill, Mingtian Tan, Vinayak Gupta, Tom Hartvigsen, and Tim Althoff. Lan- guage models still struggle to zero-shot reason about time series.arXiv preprint arXiv:2404.11757,

    Pith/arXiv arXiv

  22. [30]

    Stop when reasoning converges: Semantic-preserving early exit for reasoning models.arXiv preprint arXiv:2605.17672,

    Dehai Min, Giovanni Vaccarino, Huiyi Chen, Yongliang Wu, Gal Yona, and Lu Cheng. Stop when reasoning converges: Semantic-preserving early exit for reasoning models.arXiv preprint arXiv:2605.17672,

    Pith/arXiv arXiv

  23. [31]

    Routellm: Learning to route llms with preference data.arXiv preprint arXiv:2406.18665,

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonza- lez, M Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data.arXiv preprint arXiv:2406.18665,

    Pith/arXiv arXiv

  24. [32]

    Time2lang: Bridging time-series foundation models and large language models for health sensing beyond prompting.arXiv preprint arXiv:2502.07608,

    Arvind Pillai, Dimitris Spathis, Subigya Nepal, Amanda C Collins, Daniel M Mackin, Michael V Heinz, Tess Z Griffin, Nicholas C Jacobson, and Andrew Campbell. Time2lang: Bridging time-series foundation models and large language models for health sensing beyond prompting.arXiv preprint arXiv:2502.07608,

    Pith/arXiv arXiv

  25. [34]

    Large language model routing with benchmark datasets.arXiv preprint arXiv:2309.15789,

    Tal Shnitzer, Anthony Ou, M ´ırian Silva, Kate Soule, Yuekai Sun, Justin Solomon, Neil Thompson, and Mikhail Yurochkin. Large language model routing with benchmark datasets.arXiv preprint arXiv:2309.15789,

    Pith/arXiv arXiv

  26. [35]

    Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, et al. Kimi k2. 5: Visual agentic intelligence.arXiv preprint arXiv:2602.02276,

    Pith/arXiv arXiv

  27. [36]

    Chatts: Aligning time series with llms via synthetic data for enhanced understanding and reasoning.arXiv preprint arXiv:2412.03104,

    Zhe Xie, Zeyan Li, Xiao He, Longlong Xu, Xidao Wen, Tieying Zhang, Jianjun Chen, Rui Shi, and Dan Pei. Chatts: Aligning time series with llms via synthetic data for enhanced understanding and reasoning.arXiv preprint arXiv:2412.03104,

    arXiv

  28. [37]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    Pith/arXiv arXiv

  29. [38]

    Self-distilled rlvr.arXiv preprint arXiv:2604.03128,

    Chenxu Yang, Chuanyu Qin, Qingyi Si, Minghui Chen, Naibin Gu, Dingyu Yao, Zheng Lin, Weiping Wang, Jiaqi Wang, and Nan Duan. Self-distilled rlvr.arXiv preprint arXiv:2604.03128,

    Pith/arXiv arXiv

  30. [40]

    Ts-reasoner: Aligning time series foundation models with llm reasoning.arXiv preprint arXiv:2510.03519, 2025a

    Fangxu Yu, Hongyu Zhao, and Tianyi Zhou. Ts-reasoner: Aligning time series foundation models with llm reasoning.arXiv preprint arXiv:2510.03519, 2025a. Fangxu Yu, Xingang Guo, Lingzhi Yuan, Haoqiang Kang, Hongyu Zhao, Lianhui Qin, Furong Huang, Bin Hu, and Tianyi Zhou. Tsrbench: A comprehensive multi-task multi-modal time series reasoning benchmark for ge...

    arXiv 2026

  31. [41]

    Time-vlm: Exploring multimodal vision-language models for augmented time series forecasting

    Siru Zhong, Weilin Ruan, Ming Jin, Huan Li, Qingsong Wen, and Yuxuan Liang. Time-vlm: Exploring multimodal vision-language models for augmented time series forecasting. arXiv preprint arXiv:2502.04395,

    Pith/arXiv arXiv

  32. [42]

    Least-to-most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625,

    Denny Zhou, Nathanael Sch ¨arli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625,

    Pith/arXiv arXiv

  33. [43]

    15 Published as a conference paper at COLM 2026 Variant Perc. Reas. Pred. Deci. Over. HGT 67.63 52.81 43.11 42.45 51.33 HAN 71.22 51.12 43.11 42.45 51.21 HeteroGAT 69.06 47.75 44.89 43.40 50.00 Table 6: Results with different GNN back- bone architectures for TSROUTER. Perception Reasoning Prediction Decision-Making Overall 35 40 45 50 55 60 65 70 75Accura...

    2026

  34. [44]

    Even without the descriptions, TSROUTERexhibits only a minor performance drop and still clearly outperforms the best baseline. This is be- cause the descriptions serve merely as an initial semantic prior for node features, whereas the routing signal is primarily produced by structural message passing over the heteroge- neous graph. Consequently, TSROUTERr...

    2026

  35. [45]

    Given a query withNchannels, we plot each channel as a separate subplot, with a shared x-axis (time/index) and individual y-axes

    Time Series Visualization.Each time series is rendered as a line chart image before being passed to VLMs. Given a query withNchannels, we plot each channel as a separate subplot, with a shared x-axis (time/index) and individual y-axes. For univariate series, a single plot is generated; for multivariate series, we stackNsubplots vertically, each labeled wi...

    2026

  36. [47]

    In contrast, all three VLMs correctly identify the trend ordering from the visual plot, where the curvature of each regime is di- rectly observable

    All three LLMs operating on textual time series arrive at the wrong answer, misidentifying the initial exponential phase as linear, likely because the early ab- solute increments appear small in raw numerical form. In contrast, all three VLMs correctly identify the trend ordering from the visual plot, where the curvature of each regime is di- rectly obser...

    2026

  37. [1990]

    Self-distillation en- ables continual learning.arXiv preprint arXiv:2601.19897,

    13 Published as a conference paper at COLM 2026 Idan Shenfeld, Mehul Damani, Jonas H ¨ubotter, and Pulkit Agrawal. Self-distillation en- ables continual learning.arXiv preprint arXiv:2601.19897,

    Pith/arXiv arXiv 2026

  38. [2014]

    Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907,

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907,

    Pith/arXiv arXiv

  39. [2016]

    Time-mqa: Time series multi-task question answering with context enhancement.arXiv preprint arXiv:2503.01875, 2025a

    Yaxuan Kong, Yiyuan Yang, Yoontae Hwang, Wenjie Du, Stefan Zohren, Zhangyang Wang, Ming Jin, and Qingsong Wen. Time-mqa: Time series multi-task question answering with context enhancement.arXiv preprint arXiv:2503.01875, 2025a. Yaxuan Kong, Yiyuan Yang, Shiyu Wang, Chenghao Liu, Yuxuan Liang, Ming Jin, Ste- fan Zohren, Dan Pei, Yan Liu, and Qingsong Wen. ...

    Pith/arXiv arXiv 2026

  40. [2019]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    11 Published as a conference paper at COLM 2026 Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Pith/arXiv arXiv 2026

  41. [2020]

    Timexl: Explainable multi-modal time series prediction with llm-in- the-loop.arXiv preprint arXiv:2503.01013,

    Yushan Jiang, Wenchao Yu, Geon Lee, Dongjin Song, Kijung Shin, Wei Cheng, Yanchi Liu, and Haifeng Chen. Timexl: Explainable multi-modal time series prediction with llm-in- the-loop.arXiv preprint arXiv:2503.01013,

    arXiv

  42. [2021]

    Time series analysis for education: Methods, applications, and fu- ture directions.arXiv preprint arXiv:2408.13960,

    Shengzhong Mao, Chaoli Zhang, Yichi Song, Jindong Wang, Xiao-Jun Zeng, Zenglin Xu, and Qingsong Wen. Time series analysis for education: Methods, applications, and fu- ture directions.arXiv preprint arXiv:2408.13960,

    Pith/arXiv arXiv

  43. [2022]

    Flow of reason- ing: Training llms for divergent reasoning with minimal examples.arXiv preprint arXiv:2406.05673,

    Fangxu Yu, Lai Jiang, Haoqiang Kang, Shibo Hao, and Lianhui Qin. Flow of reason- ing: Training llms for divergent reasoning with minimal examples.arXiv preprint arXiv:2406.05673,

    Pith/arXiv arXiv

  44. [2023]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    Pith/arXiv arXiv

  45. [2024]

    Large language monkeys: Scaling inference compute with re- peated sampling.arXiv preprint arXiv:2407.21787,

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher R ´e, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with re- peated sampling.arXiv preprint arXiv:2407.21787,

    Pith/arXiv arXiv

  46. [2025]

    Graph of thoughts: Solving elaborate problems with large language models

    10 Published as a conference paper at COLM 2026 Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. InProceed- ings of the AAAI conference on artificial intelligenc...

    2026

  47. [2026]

    Under- standing different design choices in training large time series models.arXiv preprint arXiv:2406.14045,

    Yu-Neng Chuang*, Songchen Li*, Jiayi Yuan*, Guanchu Wang*, Kwei-Herng Lai*, Leisheng Yu, Sirui Ding, Chia-Yuan Chang, Qiaoyu Tan, Daochen Zha, and Xia Hu. Under- standing different design choices in training large time series models.arXiv preprint arXiv:2406.14045,

    arXiv