Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

A 4-billion-parameter multimodal LLM can reason over irregular ICU vital-sign trajectories using only 16 time-series tokens and 0.15 seconds per question.

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

2026-08-01 01:00 UTC pith:SLWDZMZY

load-bearing objection Careful, internally consistent engineering, but the headline SOTA sits on a same-group benchmark with only ID-level de-overlap — read 49.83% as unverified until CLIR-Bench and code are released. the 3 major comments →

arxiv 2607.25947 v1 pith:SLWDZMZY submitted 2026-07-28 cs.AI cs.CL

A Cost-Effective Multimodal LLM Reasoning Framework for Question Answering over Irregular Clinical Time Series

Frank Nie , Ethan B Liu , Yuan Zhu , Wei Fan , Jindong Han This is my paper
classification cs.AI cs.CL
keywords irregular clinical time seriesmultimodal LLMquestion answeringtime-series tokensmulti-scale encodertemporal evidence distillerprogressive alignmentCLIR-Bench
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.

The paper proposes ClinPRISM, a framework for question answering over irregular clinical time series—records where vital signs are measured at nonuniform, variable-specific times. The central claim is that sparse, asynchronous ICU trajectories can be compressed into as few as 16 LLM-compatible tokens while preserving enough multi-scale evidence to answer clinical questions, and that this can be done with a 4-billion-parameter language model at 0.15 seconds per question. The framework's core is an irregularity-aware multi-scale encoder that treats global trends, episode-level dynamics, and fine-grained measurements as one evidence hierarchy, followed by a temporal evidence distiller that fuses and compresses them, and a three-stage progressive alignment strategy that first pretrains the encoders, then aligns them with hierarchical captions, then adapts them to QA. On a held-out clinical QA benchmark the model attains the highest macro-average accuracy among open-source systems. The sympathetic reader should care because it suggests that reasoning over raw, irregular clinical signals need not be expensive, either in tokens or compute.

Core claim

ClinPRISM's central claim is that irregular clinical time series can be represented as a small set of continuous tokens that an LLM can reason over directly, without grid resampling or text serialization. The encoder converts every observed measurement into a timestamp-aware event, then aggregates under three scale conditions—macro (whole trajectory), meso (learnable soft windows), and micro (reference times)—and each scale reports not only a local state but the degree of observation support behind it, so weakly supported estimates are masked out. A temporal evidence distiller fuses the three scales, using meso as a bridge between macro context and micro detail, and a learned-query resampler

What carries the argument

The central object is the irregularity-aware multi-scale encoder plus the temporal evidence distiller. The encoder defines a shared timestamp-aware event schema: each valid measurement is a combination of value, normalized time, and observation gap, so missing values never enter. Aggregation at the macro, meso, and micro scales produces an evidence hierarchy, and a support-aware aggregation rule weights observations by relevance to a temporal location while appending log(1+support), so sparse local estimates are downweighted and masked by a threshold. The temporal evidence distiller projects the three scales into the LLM hidden space, refines meso tokens with macro and micro context, and a l

Load-bearing premise

The load-bearing premise is that the evaluation benchmark is genuinely held-out: although no patient, admission, or ICU-stay identifiers overlap with training, the benchmark shares the same 11 task schemas, evidence rules, and automatic rewriting style used to build the training corpus, so distributional independence is not established.

What would settle it

A concrete check: build a validation set with independently authored task schemas and question styles (for example, clinician-written questions or tasks not in the training taxonomy) over trajectories from other hospitals; if accuracy drops to near the level of general LLMs, the 16-token multi-scale interface is exploiting benchmark-specific regularities rather than transferable clinical reasoning.

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

If this is right

  • If the reported numbers hold, a 4B-parameter LLM with 16 temporal tokens can outperform much larger general LLMs and dedicated time-series LLMs on irregular clinical QA.
  • The 0.15-second latency and compact token interface suggest interactive clinical question answering is feasible on modest hardware, enabling bedside or triage use.
  • The ablation shows the joint LoRA–distiller adaptation is load-bearing: removing it drops accuracy from 49.83 to 36.35, so co-adapting the token interface with the LLM is essential.
  • Support-aware aggregation masks weakly supported local estimates, meaning answers are tied to where observations actually exist—a property that matters for trust in sparse electronic health records.

Where Pith is reading between the lines

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

  • Because the training QA corpus and the evaluation benchmark share the same 11-task taxonomy and the same automatic rewriting style, part of the measured gain may reflect in-distribution familiarity rather than general reasoning over irregular clinical series; a true test would use unseen tasks or a differently sourced benchmark.
  • The learned-query resampler is question-independent; conditioning it on the question could allocate the 16-token budget to question-relevant scales, potentially improving tasks like cross-variable reasoning and intervention response where the margins are smaller.
  • The 16-token optimum (performance falls at 8 and at 32–128) hints that the token budget acts as a compression regularizer; a testable extension would be to vary the budget per task or per question complexity.
  • The same encoder–distiller–alignment recipe could transfer to other irregular time-series QA domains—wearable sensor streams, industrial telemetry, or financial event data—where observation gaps themselves carry information.

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

3 major / 5 minor

Summary. The paper proposes ClinPRISM, a multimodal LLM framework for question answering over irregular clinical time series. The method combines an irregularity-aware multi-scale encoder (macro, meso, micro), a temporal evidence distiller that compresses the fused representation into 16 LLM-compatible tokens, and a three-stage progressive training procedure: self-supervised multi-scale pretraining, hierarchical caption alignment, and QA-oriented distiller/LoRA adaptation. Training resources are constructed from MIMIC-IV: 30,000 ICU trajectories with hierarchical captions and approximately 41,000 four-way multiple-choice QA instances across 11 tasks. On CLIR-Bench, the paper reports a macro-average accuracy of 49.83% with a 4B-parameter backbone, 16 temporal tokens, and 0.15 s inference latency. Ablations show consistent gains from each scale and training stage, and the robustness/efficiency analyses support the paper's practical claims.

Significance. If the results hold, ClinPRISM would be a meaningful efficiency/performance point for irregular clinical time-series QA: a 4B-parameter model with a very compact temporal-token interface is competitive with much larger open-source time-series LLMs and approaches closed-source GPT-5.4 mini. The architecture and training pipeline are described in enough detail to be reproducible in principle, and the ablations are internally consistent. The main significance, however, is conditional on CLIR-Bench being a fair independently valid held-out benchmark. Because CLIR-Bench is authored by the same group from the same MIMIC-IV data and uses the same 11-task taxonomy as the training corpus, the external validity of the state-of-the-art claim is not yet established. The paper provides no code, data, confidence intervals, or significance tests, which limits independent verification of the reported margins.

major comments (3)
  1. [Section 4.2 / Appendix B.1] The central SOTA claim rests entirely on CLIR-Bench [18], a benchmark co-authored by the same group and built from the same MIMIC-IV source as the training corpus. Appendix B.1 reports only that no subject_id, hadm_id, or stay_id overlaps with CLIR-Bench. This does not establish distributional independence: the training QA corpus (Section 3.3) and CLIR-Bench share the same 11-task taxonomy and, as far as the text shows, similar deterministic program schemas and GPT-5.5-based rewriting. If the task schemas, evidence rules, question templates, and rewriting style are shared, the benchmark is an in-distribution split rather than an independent test. The paper offers no evidence to the contrary. Please quantify the overlap in schemas/evidence rules or release the benchmark construction prompts and generation scripts, and/or evaluate on an externally authored benchmark.
  2. [Table 1 / Section 4.2] The reported margins are fragile and no uncertainty quantification is provided. ClinPRISM leads KiMi-2.6 by 0.74 points and trails GPT-5.4 mini by 0.32 points. With 11 tasks and roughly 60 items per task, the 0.74-point lead would be about 5 questions; if the benchmark is smaller or task sizes are heterogeneous, the margin could be noise. No confidence intervals, per-task sample sizes, or significance tests are reported. Please report the number of items per task, bootstrap confidence intervals, and a paired test (e.g., McNemar) to support the claim of being the best open-source system.
  3. [Abstract / Section 4.2] The abstract states that ClinPRISM "achieves state-of-the-art performance on the held-out evaluation benchmark" without qualification. However, Table 1 shows GPT-5.4 mini at 50.15%, above ClinPRISM's 49.83%. The main text correctly qualifies the result as the highest among open-source systems, but the unqualified abstract claim is not supported by the reported numbers and should be revised for consistency.
minor comments (5)
  1. [Table 1] The table formatting is garbled in several rows (e.g., the Gemma3-27B-it row shows values such as "21.6731.6720.50" run together). The caption also mentions gray shading, but no shading is visible in the text.
  2. [Appendix C.1] The notation for the temporal-token budget is inconsistent: Section 2.4 uses K, while Table 6 and the appendix use T. This should be aligned.
  3. [Figure 1(b)] The motivation figure is extremely dense and lacks a legend. Grouping of model families (Qwen, ITFormer, ChatTS, etc.) is not visually distinguishable, making it hard to verify the claimed capability-level patterns.
  4. [Section 5.1] The related-work text says "EngineMT-QA [23]", but reference [23] is titled "ITFormer: Bridging Time Series and Natural Language for Multi-Modal QA..." — the attribution appears to be a typo or mis-citation.
  5. [General] No data or code availability statement is included. Given that the evaluation benchmark is central and also not described in this paper, a release plan or a link to the companion CLIR-Bench paper is needed.

Circularity Check

1 steps flagged

SOTA claim is evaluated on CLIR-Bench, a same-author, same-MIMIC-IV, same-11-task benchmark with only ID-level de-overlap; the held-out result is in-distribution.

specific steps
  1. self citation load bearing [Sec. 4.1 (Datasets and metrics); Sec. 3.3 / Table 5; Appendix B.1; Ref. [18]]
    "We train ClinPRISM on the QA corpus constructed in Section 3 and evaluate it on CLIR-Bench [18], a held-out benchmark for irregular multivariate time-series QA. CLIR-Bench contains 11 tasks grouped into four capability dimensions ... To ensure separation from CLIR-Bench, we compare the selected corpus with the benchmark using all three identifiers. None of the selected trajectories shares a subject_id, hadm_id, or stay_id with any CLIR-Bench trajectory."

    The central state-of-the-art claim rests entirely on CLIR-Bench [18], whose authors (Nie, Liu, Zhu, Yan, Fan, Han) overlap with this paper and which is derived from the same MIMIC-IV ICU collection as the training corpus. Sec. 3.3/Table 5 define the training QA corpus with the identical 11-task taxonomy (TG, ASR, TPR, MA, TSS, TF, NIF, CVR, IR, IID, MED) that Sec. 4.1 attributes to CLIR-Bench. Appendix B.1 reports only subject/hadm/stay-ID de-overlap, not independence of task schemas, evidence rules, question templates, or generation style. Thus evaluating on CLIR-Bench is an in-distribution test of the same task-generation process rather than an externally validated held-out benchmark; the SOTA claim is supported by a same-author citation without demonstrated distributional separation.

full rationale

The model components themselves are not circular: the multi-scale encoder, temporal evidence distiller, and progressive alignment stages are trained with self-supervised, caption-alignment, and QA objectives on corpora described in the paper, and no fitted parameter is renamed as a prediction. The circularity is in the evaluation chain. The headline SOTA is measured solely on CLIR-Bench [18], a benchmark by the same research group, built from the same MIMIC-IV source, using the same 11-task taxonomy as the training QA corpus. Appendix B.1's de-overlap is limited to subject/admission/stay IDs and does not establish that CLIR-Bench was generated with different task schemas, evidence rules, or paraphrase pipeline. Under the stated evidence, the held-out evaluation reduces to an in-distribution evaluation of the same construction process. The efficiency and ablation findings are independent and not implicated, so the overall circularity is partial rather than total.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claim rests on a stack of domain assumptions: one dataset (MIMIC-IV), one benchmark from the same authors, programmatic gold labels, and an internal 16-token compression. Fitted design choices reported are K=16 and loss weights; P, R, and support thresholds are unspecified. No physical entity is introduced; the main invented component is the learned query-token bank, which has no out-of-sample handle.

free parameters (4)
  • Temporal token budget K = 16
    Selected as best in sensitivity sweep K=8..128 (Table 6) on CLIR-Bench; used in all main results.
  • Support thresholds eta_s (branch-specific) = not reported
    Equation (1) masks weakly supported locations with eta_s; values are not specified and appear hand-chosen.
  • Loss weights lambda_scale, lambda_text = lambda_scale=0.05 (Stage 2); lambda_text=0.1, lambda_scale=0.01 (Stage 3 Step 1)
    Chosen by hand; no sensitivity analysis is reported for these weights.
  • Numbers of meso windows P and micro reference times R = not reported
    Architecture hyperparameters that determine token counts; not specified in main text or appendix.
axioms (5)
  • domain assumption MIMIC-IV ICU stays are a sufficient and representative source for irregular clinical time-series QA training and evaluation.
    The whole pipeline and benchmark are built from this one dataset; no external corpus is used.
  • domain assumption Deterministic clinical summarization plus GPT-5.5 verbalization preserves the clinical facts needed for QA, and human verification catches errors.
    Caption corpus (B.2) and QA rewriting (B.3) depend on this; no quantitative measure of caption fidelity is given.
  • domain assumption The 11-task taxonomy (Table 5) is the right target, and gold labels computed from observations are correct.
    Gold labels are computed programmatically; this assumes observation records are ground truth and multiple-choice format captures clinical reasoning.
  • ad hoc to paper CLIR-Bench is a fair held-out benchmark despite shared authorship, source data, and task taxonomy.
    Central evaluation uses ref [18]; no independent third-party benchmark is used.
  • domain assumption Masked cross-attention over 16 compressed tokens retains sufficient clinical evidence for QA.
    Information-theoretic assumption behind token compression; not proven beyond ablations.
invented entities (1)
  • Learned query-token bank Q (temporal evidence distiller) no independent evidence
    purpose: Compress multi-scale irregular trajectory evidence into a fixed 16-token sequence for the LLM.
    Internal design component with no external falsifiable handle; its adequacy is established only by CLIR-Bench performance.

pith-pipeline@v1.3.0-alltime-deepseek · 18165 in / 14176 out tokens · 122978 ms · 2026-08-01T01:00:49.422411+00:00 · methodology

0 comments
read the original abstract

Question answering (QA) over irregular clinical time series (ICTS) plays a pivotal role in a wide range of healthcare applications. Although recent multimodal time-series large language models (LLMs) have shown considerable promise in general-purpose time-series QA, they remain poorly equipped to model the sparsity, asynchrony, and irregular sampling patterns of clinical observations. To fill this gap, we propose ClinPRISM, a cost-effective multimodal LLM reasoning framework for question answering over ICTS data. First, we devise an irregularity-aware multi-scale encoder to capture sparse clinical evidence at diverse temporal scales. Then, we propose a temporal evidence distiller to integrate representations across these scales and compress them into a small number of LLM-compatible tokens. Moreover, we introduce a progressive alignment strategy that sequentially aligns the irregular trajectories with the LLM's textual embedding space. To facilitate training, we construct 30,000 clinical time series paired with multi-scale descriptions, together with 41,000 instruction-tuning instances spanning 11 tasks. Using a 4-billion-parameter LLM backbone, ClinPRISM achieves state-of-the-art performance on the held-out evaluation benchmark while using only 16 time-series tokens and achieving an average inference latency of 0.15 seconds per question.

Figures

Figures reproduced from arXiv: 2607.25947 by Ethan B Liu, Frank Nie, Jindong Han, Wei Fan, Yuan Zhu.

Figure 1
Figure 1. Figure 1: Motivation and empirical results for ClinPRISM. (a) Clinical questions require evidence at different temporal scales. (b) Existing model families exhibit limited performance across 11 clinical temporal reasoning tasks, suggesting that they fail to consistently capture the sparse and asynchronous nature of irregular clinical time series. identify and preserve complementary evidence across scales while avoid… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ClinPRISM. 2.2 Framework Overview [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Capability-level reliance on hierarchical temporal representations. Panels show bootstrapped accuracy-point con [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy under different observation-density [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy-efficiency trade-off. per question. Under the same 16-token budget, t-PatchGNN is 0.03 seconds faster but achieves 11.18 percentage points lower accuracy. ClinPRISM also substantially outperforms ITFormer and ChatTS while using comparable or lower inference latency and fewer tem￾poral tokens. The text-serialized Qwen baselines require approx￾imately 21K input tokens and 6.6–40 seconds per question… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond Frame Selection: Generative Latent Evidence Aggregation for Long-Video Understanding

    cs.CV 2026-07 conditional novelty 6.0

    A query-conditioned latent evidence aggregator after frozen frame selection improves long-video QA by up to +5.2 average / +10.1 LVBench with 0.11–0.40% token overhead.

Reference graph

Works this paper leans on

35 extracted references · 14 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Defu Cao, Furong Jia, Sercan O Arik, Tomas Pfister, Yixiang Zheng, Wen Ye, and Yan Liu. 2024. TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting. arXiv:2310.04948 [cs.LG] https://arxiv.org/abs/2310. 04948

    Pith/arXiv arXiv 2024

  2. [2]

    Jialin Chen, Aosong Feng, Ziyu Zhao, Juan Garza, Gaukhar Nurbek, Cheng Qin, Ali Maatouk, Leandros Tassiulas, Yifeng Gao, and Rex Ying. 2026. MTBench: A Multimodal Time Series Benchmark for Temporal Reasoning and Question Answering. arXiv:2503.16858 [cs.CL] https://arxiv.org/abs/2503.16858

    arXiv 2026

  3. [3]

    Nguyen, Hao Xue, Imran Razzak, Flora D

    Felix Divo, Maurice Kraus, Anh Q. Nguyen, Hao Xue, Imran Razzak, Flora D. Salim, Kristian Kersting, and Devendra Singh Dhami. 2025. QuAnTS: Question Answering on Time Series. arXiv:2511.05124 [cs.LG] https://arxiv.org/abs/2511. 05124

    arXiv 2025

  4. [4]

    Tong Guan, Zijie Meng, Dianqi Li, Shiyu Wang, Chao-Han Huck Yang, Qing- song Wen, Zuozhu Liu, Sabato Marco Siniscalchi, Ming Jin, and Shirui Pan. 2026. TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Lan- guage Models. arXiv:2509.24803 [cs.LG] https://arxiv.org/abs/2509.24803

    Pith/arXiv arXiv 2026

  5. [5]

    Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. 2024. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. arXiv:2310.01728 [cs.LG] https://arxiv.org/abs/2310.01728

    Pith/arXiv arXiv 2024

  6. [6]

    Baoyu Jing, Sanhorn Chen, Lecheng Zheng, Boyu Liu, Zihao Li, Jiaru Zou, Tianxin Wei, Zhining Liu, Zhichen Zeng, Ruizhong Qiu, Xiao Lin, Yuchen Yan, Dongqi Fu, Jingchao Ni, Jingrui He, and Hanghang Tong. 2026. TSAQA: Time Series Analysis Question And Answering Benchmark. arXiv:2601.23204 [cs.AI] https: //arxiv.org/abs/2601.23204

    Pith/arXiv arXiv 2026

  7. [7]

    Alistair Johnson, Lucas Bulgarelli, Tom Pollard, Brian Gow, Benjamin Moody, Steven Horng, Leo Anthony Celi, and Roger Mark. 2024. MIMIC-IV.PhysioNet (Oct. 2024). Version 3.1. doi:10.13026/kpb9-mt58

    doi:10.13026/kpb9-mt58 2024

  8. [8]

    Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al

  9. [9]

    Yaxuan Kong, Yiyuan Yang, Yoontae Hwang, Wenjie Du, Stefan Zohren, Zhangyang Wang, Ming Jin, and Qingsong Wen. 2025. Time-mqa: Time se- ries multi-task question answering with context enhancement. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 29736–29753

    2025

  10. [10]

    Boyuan Li, Yicheng Luo, Zhen Liu, Junhao Zheng, Jianming Lv, and Qianli Ma

  11. [11]

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. 2024. Foundation models for time series analysis: A tutorial and survey. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 6555–6565

    2024

  12. [12]

    Chenxi Liu, Shaowen Zhou, Qianxiong Xu, Hao Miao, Cheng Long, Ziyue Li, and Rui Zhao. 2025. Towards Cross-Modality Modeling for Time Series Analytics: A Survey in the LLM Era. arXiv:2505.02583 [cs.LG] https://arxiv.org/abs/2505.02583

    Pith/arXiv arXiv 2025

  13. [13]

    Haoxin Liu, Harshavardhan Kamarthi, Zhiyuan Zhao, Shangqing Xu, Shiyu Wang, Qingsong Wen, Tom Hartvigsen, Fei Wang, and B Aditya Prakash. 2025. How can time series analysis benefit from multiple modalities? a survey and outlook. arXiv preprint arXiv:2503.11835(2025)

    arXiv 2025

  14. [14]

    Xu Liu, Junfeng Hu, Yuan Li, Shizhe Diao, Yuxuan Liang, Bryan Hooi, and Roger Zimmermann. 2024. Unitime: A language-empowered unified model for cross- domain time series forecasting. InProceedings of the ACM Web Conference 2024. 4095–4106

    2024

  15. [15]

    Xvyuan Liu, Xiangfei Qiu, Hanyin Cheng, Xingjian Wu, Chenjuan Guo, Bin Yang, and Jilin Hu. 2025. Astgi: Adaptive spatio-temporal graph interactions for irregular multivariate time series forecasting.arXiv preprint arXiv:2509.23313 (2025)

    arXiv 2025

  16. [16]

    Xvyuan Liu, Xiangfei Qiu, Xingjian Wu, Zhengyu Li, Chenjuan Guo, Jilin Hu, and Bin Yang. 2026. Rethinking irregular time series forecasting: A simple yet effective baseline. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 23873–23881

    2026

  17. [17]

    Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, and Mingsheng Long

  18. [18]

    Liu, Yuan Zhu, Loe Yan, Wei Fan, and Jindong Han

    Frank Nie, Ethan B. Liu, Yuan Zhu, Loe Yan, Wei Fan, and Jindong Han. 2026. CLIR-Bench: Benchmarking Multimodal Question Answering over Irregular Clinical Time Series. arXiv:2607.09880 [cs.CL] https://arxiv.org/abs/2607.09880

    Pith/arXiv arXiv 2026

  19. [19]

    Qwen Team. 2026. Qwen3.6-Plus: Towards Real World Agents. https://qwen.ai/ blog?id=qwen3.6

    2026

  20. [20]

    Satya Narayan Shukla and Benjamin M. Marlin. 2024. Heteroscedastic Temporal Variational Autoencoder For Irregular Time Series. arXiv:2107.11350 [cs.LG] https://arxiv.org/abs/2107.11350

    Pith/arXiv arXiv 2024

  21. [21]

    Qwen Team. 2026. Qwen3.5: Accelerating Productivity with Native Multimodal Agents. https://qwen.ai/blog?id=qwen3.5

    2026

  22. [22]

    Chengsen Wang, Qi Qi, Jingyu Wang, Haifeng Sun, Zirui Zhuang, Jinming Wu, Lei Zhang, and Jianxin Liao. 2024. ChatTime: A Unified Multimodal Time Series Foun- dation Model Bridging Numerical and Textual Data. arXiv:2412.11376 [cs.CL] https://arxiv.org/abs/2412.11376

    Pith/arXiv arXiv 2024

  23. [23]

    Yilin Wang, Peixuan Lei, Jie Song, Yuzhe Hao, Tao Chen, Yuxuan Zhang, Lei Jia, Yuanxiang Li, and Zhongyu Wei. 2025. ITFormer: Bridging Time Series and Natural Language for Multi-Modal QA with Large-Scale Multitask Dataset. In International Conference on Machine Learning. PMLR, 63324–63344

    2025

  24. [24]

    Stephan Xie, Ben Cohen, Mononito Goswami, Junhong Shen, Emaad Khwaja, Chenghao Liu, David Asker, Othmane Abou-Amal, and Ameet Talwalkar. 2026. ARFBench: Benchmarking Time Series Question Answering Ability for Software Incident Response. arXiv:2604.21199 [cs.LG] https://arxiv.org/abs/2604.21199

    Pith/arXiv arXiv 2026

  25. [25]

    Zhe Xie, Zeyan Li, Xiao He, Longlong Xu, Xidao Wen, Tieying Zhang, Jianjun Chen, Rui Shi, and Dan Pei. 2025. ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning.Proceedings of the VLDB Endowment18, 8 (2025), 2385–2398

    2025

  26. [26]

    Hao Xue and Flora D Salim. 2023. Promptcast: A new prompt-based learning paradigm for time series forecasting.IEEE Transactions on Knowledge and Data Engineering36, 11 (2023), 6851–6864

    2023

  27. [27]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    Pith/arXiv arXiv 2025

  28. [28]

    Fangxu Yu, Hongyu Zhao, and Tianyi Zhou. 2025. TS-Reasoner: Aligning Time Series Foundation Models with LLM Reasoning. arXiv:2510.03519 [cs.CL] https: //arxiv.org/abs/2510.03519

    arXiv 2025

  29. [29]

    Jiawen Zhang, Shun Zheng, Wei Cao, Jiang Bian, and Jia Li. 2023. Warpformer: A multi-scale modeling approach for irregular clinical time series. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3273–3285

    2023

  30. [30]

    Weijia Zhang, Chenlong Yin, Hao Liu, Xiaofang Zhou, and Hui Xiong. 2024. Irregular multivariate time series forecasting: A transformable patching graph neural networks approach. InForty-first International Conference on Machine Learning

    2024

  31. [31]

    Gupta, and Jingbo Shang

    Xiyuan Zhang, Ranak Roy Chowdhury, Rajesh K. Gupta, and Jingbo Shang. 2024. Large Language Models for Time Series: A Survey. arXiv:2402.01801 [cs.LG] https://arxiv.org/abs/2402.01801

    Pith/arXiv arXiv 2024

  32. [32]

    Tian Zhou, PeiSong Niu, Xue Wang, Liang Sun, and Rong Jin. 2023. One Fits All:Power General Time Series Analysis by Pretrained LM. arXiv:2302.11939 [cs.LG] https://arxiv.org/abs/2302.11939 Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al. A Ethics and Privacy Statement This study uses de-identified MIMIC-IV ICU records and involves no...

    Pith/arXiv arXiv 2023

  33. [2023]

    MIMIC-IV, a freely accessible electronic health record dataset.Scientific data10, 1 (2023), 1

    2023

  34. [2024]

    arXiv:2402.02370 [cs.LG] https://arxiv.org/abs/2402.02370

    AutoTimes: Autoregressive Time Series Forecasters via Large Language Models. arXiv:2402.02370 [cs.LG] https://arxiv.org/abs/2402.02370

    Pith/arXiv arXiv

  35. [2025]

    arXiv:2505.17431 [cs.LG] https://arxiv.org/abs/2505.17431

    HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting. arXiv:2505.17431 [cs.LG] https://arxiv.org/abs/2505.17431

    Pith/arXiv arXiv