Traceable Multi-Agent System for Knowledge-Based Forecasting

Hyeokjun Choe; Junhyeok Kang; Sangjun Han; Soonyoung Lee

REVIEW 3 major objections 5 minor 16 references

Traceable Multi-Agent System for Knowledge-Based Forecasting

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read TraceMAS organizes multi-agent forecasting around two causal-loop diagrams so every forecast revision can be traced back to the document, data, and model choice that caused it.

desk verdict A nice demo idea with a genuinely new traceability representation, but the load-bearing claim that the Data-Grounded CLD actually drives the forecast is unverified. read the letter →

arxiv 2608.03339 v1 pith:56A2U7YA submitted 2026-08-04 cs.AI

Junhyeok Kang , Sangjun Han , Hyeokjun Choe , Soonyoung Lee This is my paper

classification cs.AI

keywords traceabilitymulti-agentforecastingcausalloopdiagramsLLMagentstime-seriescrudeoilpriceprovenanceinteractivedemo

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TraceMAS is a demo system for enterprise forecasting with autonomous agents. Its claim is that the reason a forecast changed—which evidence, which data choice, which feature, which model revision—can be kept inspectable by routing all agent outputs through two causal-loop representations: an Ideal Causal Loop Diagram (CLD) extracted from domain documents, and a Data-Grounded CLD that maps each factor to a real data series or a documented proxy. The Data-Grounded CLD then guides feature construction and model design, so the diagram is a working artifact rather than a post-hoc explanation. The demo applies this to crude oil price forecasting, letting users compare iterations and trace a factor from source document to mapped variable to feature to model component. If the approach works, practitioners would be able to audit autonomous forecasting pipelines instead of trusting them blindly.

What carries the argument

The central machinery is the Data-Grounded CLD, a directed graph $G_D=(V_D,E_D,M)$ whose nodes and edges are the data-available version of the Ideal CLD, and whose mapping function $M$ attaches each node to an internal series, external series, or documented proxy. It carries the argument because it is the versioned, shared artifact to which agent outputs, features, model components, and reviewer warnings are attached; iterating on it lets users compare which evidence, data choice, or modeling decision changed the forecast.

What would settle it

Run a new batch of market reports through TraceMAS, then have an independent analyst pick three factors at random and trace each from its source document through the Data-Grounded CLD mapping to the feature and model component; if the analyst cannot recover the evidence-to-forecast path for any one factor, the core traceability claim fails.

Watch

Extended reading notes

Core claim

TraceMAS claims that traceability in multi-agent forecasting can be achieved by maintaining two causal-loop diagrams. The Ideal CLD, a directed graph $G_I=(V_I,E_I)$, records key factors and document-supported directional hypotheses with time delays. The Data-Grounded CLD, $G_D=(V_D,E_D,M)$, is derived from the Ideal CLD by selecting, merging, replacing, or extending factors according to data availability; its mapping function $M$ links each node to an internal time series, an external time series, or a documented proxy. The Data-Grounded CLD guides feature construction and model design, is versioned across iterations, and serves as the shared working artifact. The claim is that this makes e

Load-bearing premise

The system's traceability is only as reliable as the LLM agents' extraction of factors and directional hypotheses from documents, and the mapping function $M$'s faithfulness in linking each CLD node to real data or a documented proxy; the paper does not test extraction fidelity, mapping accuracy, or whether users can actually recover the evidence-to-forecast path in practice.

Editorial extensions

If this is right

Practitioners can distinguish forecast revisions caused by new textual evidence from those caused by data mapping decisions, feature revisions, or reviewer-flagged risks.
The Ideal-CLD versus Data-Grounded-CLD split makes the gap between what documents suggest and what data can support visible and comparable across iterations.
Because the Data-Grounded CLD is versioned, users can inspect which factors were added, which proxies were replaced, and which relation directions or lag assumptions were revised.
Reviewer warnings (e.g., insufficient proxy coverage, misaligned timestamps) become first-class, traceable artifacts linked to the CLD instead of buried log entries.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next step is to make the mapping function $M$ an editable control, so users can swap a proxy and watch the forecast revise in real time—turning the CLD from an audit artifact into a what-if tool. The paper does not claim this.
The two-CLD structure suggests a testable hypothesis: users can recover the provenance of a random forecast revision faster and more accurately from versioned CLDs than from unstructured agent logs; a user study could settle it.
Because edge direction and time delay are treated as model priors, one could test whether an intentionally reversed relation (say, price→supply instead of supply→price) is caught by the Risk Reviewer or silently distorts forecasts—probing whether the CLD is genuinely load-bearing.
The same two-CLD pattern could transfer to any domain where documents make causal claims (macroeconomics, energy, supply chains), though the paper only demonstrates crude oil.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

This is a demo paper, and that is the right frame. What is genuinely new is the dual-CLD representation: an Ideal CLD built purely from documents and a Data-Grounded CLD that ties each node to an actual data series or proxy. Separating 'what the documents say' from 'what data we could find' is a clean way to expose why a forecast changed—was it new evidence, a proxy swap, or a model revision? The versioned artifact design is a real improvement over raw agent logs, and the paper is honest that it is not claiming benchmark-level accuracy. The system description is clear and the figures look plausible.

The problem is that the central claim—that the Data-Grounded CLD guides feature construction and model design, and is 'not a post-hoc explanation'—rests entirely on assertion. There is no demo link, no code, no user study, and no check that the mapping function M is faithful. The stress-test concern is exactly right: if the CLD is generated after the fact and overlaid on model outputs, then TraceMAS is a nice visualization dashboard but not a traceable forecasting system. Table 1 and Figures 3–4 are curated examples, not evidence of a causal chain. To make the contribution stick, we need at least a demo where you can see a new document change a CLD node, then see that node feed a feature, then see the forecast change. Without that, the traceability benefit is plausible but unverified.

Given the demo track, I would treat the missing artifacts as a condition, not a rejection. The idea is good enough to deserve a serious referee, but the camera-ready version absolutely should include a working demo link and a clearer description of the software architecture connecting CLD to model code. Even a small user study—can users trace a change to its source?—would help. The citation pattern is fine; self-citations appear for supporting claims and do not affect the core idea.

So: worth engaging, worth citing as related work, and worth sending to peer review with the requirement that the demo be available. It is a solid 'maybe' for a reading group—likely to generate a good argument about what traceability should mean in agentic forecasting.

Referee Report

3 major / 5 minor

Summary. TraceMAS is an interactive demo system for multi-agent forecasting that introduces two causal-loop representations: an Ideal CLD, whose nodes and directed hypotheses are extracted from domain documents, and a Data-Grounded CLD, which links those factors to internal variables, external data, or documented proxies through a mapping function M. The paper claims that the Data-Grounded CLD is not a post-hoc explanation but a shared working artifact that guides feature construction, lag selection, interactions, and model architecture, and that versioned CLDs allow practitioners to inspect why a forecast changed and which evidence supported the change. The system is demonstrated on crude oil price forecasting through a described interface, a representative table, and initial CLD figures in the appendix. The paper explicitly disclaims any claim to benchmark-level forecasting accuracy and focuses on process traceability.

Significance. If the central traceability claim were substantiated, the two-level CLD design would be a genuinely useful organizing principle for agentic forecasting: it would give practitioners a structured, comparable artifact linking textual evidence, data choices, and model revisions, going beyond raw agent logs or opaque text-embedding methods. The paper is honest about its scope, as it does not claim forecast accuracy improvements. However, the contribution is currently at the level of a system proposal and interface description: there is no live demo, no code, no run trace, no validation of extraction fidelity, and no user study. The central claim that the Data-Grounded CLD actually guides model construction rather than being a post-hoc overlay is presented as a design statement but not demonstrated. Thus the significance is real but conditional on the missing evidence.

major comments (3)

[Sections 2.1, 2.2, and 3.2] The load-bearing claim is that the Data-Grounded CLD is a working artifact that guides feature construction and model design, not a post-hoc explanation. The manuscript states this in Section 2.2 and describes the Model Engineer's behavior in Section 2.1, but provides no mechanism, code, demo link, or run transcript showing that the CLD's nodes and edges are causally upstream of the constructed features and model components. Table 1 and Figures 3-4 are curated examples and initial artifacts; they do not establish a causal chain. To support the claim, the authors should provide an explicit mapping from CLD edges to feature templates and model structural choices, plus a logged run of the actual system showing that this mapping was executed before the forecast was produced.
[Definition 2, Section 2.2] The mapping function M is the crux of the Data-Grounded CLD, but it is only described as linking each node to an internal series, external series, or documented proxy. No formal semantics are given: there is no definition of what makes a proxy 'documented', no requirement that the chosen series actually correspond to the factor, and no stated check for semantic fidelity. If M is chosen arbitrarily by an LLM agent, the traceability claim becomes an illusion. The paper should specify the structure of M (e.g., node to source evidence, selected series, coverage, units, transformation) and show how the Risk Reviewer verifies each mapping in practice.
[Section 3.3 and Appendix A] The paper claims users can compare iterations and trace the evidence-to-forecast path, but it provides no user evaluation, no task-based inspection study, and no actual example of an iteration-level comparison. Appendix A shows only initial CLDs, not versioned differences. Since the central contribution is inspectability for practitioners, the claim is not established by interface screenshots and curated examples alone. The authors should include a worked iteration comparison from the demo and, ideally, a small user study measuring whether practitioners can recover why a forecast changed.

minor comments (5)

[Abstract] Typos: 'We presentTraceMAS,' should have a space after 'present'; the comma after 'TraceMAS' appears to be a typo from the text-formatting pipeline.
[Section 2.1] 'Data Engineersearches' is missing a space. Also, in the role list, entry (7) calls the main agent 'Coordinator' while Section 2.3 calls it 'Coordinator' with mixed capitalization; please standardize.
[Figure 2] The workflow diagram appears linear, but the text emphasizes iterative revision. Adding feedback arrows or a note about versioned loops would make the figure consistent with the described process.
[Table 1] The caption says 'Representative outputs from the TraceMAS demo.' If these are curated examples rather than direct system outputs, the caption should say so explicitly, and ideally the actual output should be available in the demo.
[Definition 2] The notation '𝑉 𝐷' and '𝐸𝐷' has spacing and font inconsistencies; also, the term 'documented proxy' is used in prose without a formal definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TraceMAS is a system/demo description with no fitted predictions; its traceability claims, though unverified, are not equivalent to their inputs by construction.

full rationale

The paper does not derive any numeric forecast from its CLD representations; it describes a pipeline in which the Ideal CLD is extracted from documents, the Data-Grounded CLD links nodes to data via a mapping function M, and the Model Engineer uses the Data-Grounded CLD for feature and model construction. Each step is an architectural design assertion rather than a mathematical derivation, so there is no equation-level reduction. Self-citations ([6], [7], [12]) appear but are not load-bearing: they support background claims about retrieval importance and forecasting applications, not the core traceability mechanism. The main weakness is evidential, not circular: the paper does not demonstrate that the CLD genuinely constrains model construction or that M is semantically faithful, and it disclaims benchmark-level performance ('rather than to claim benchmark-level forecasting performance'). These are validation gaps, not circularity. The system's self-report that the Data-Grounded CLD is 'not a post-hoc explanation' (Sec 2.2) is an unsupported assertion about implementation practice, but it is not derived from, nor identical to, its inputs. Hence score 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 2 invented entities

The paper's central contribution is architectural: it introduces two new representation objects, the Ideal CLD and the Data-Grounded CLD, and relies on the reliability of LLM-based document understanding and data mapping. No free parameters are fitted in the paper; there are no quantitative claims at all. The main burden is the unvalidated assumption that agent-generated CLDs faithfully preserve evidence-to-data-to-forecast links.

assumptions (3)

domain assumption LLM-based agents, specifically the Domain Analyst and Causal Analyst, can reliably extract key factors and directional hypotheses from domain documents with sufficient fidelity for the CLD to be a faithful representation.
Invoked in Section 2.1 and 2.3, where the Domain Analyst 'extracts key factors and document-supported directional hypotheses' and the Causal Analyst 'synthesizes them into an Ideal CLD.' No evidence of extraction accuracy is provided.
domain assumption The Data-Grounded CLD mapping function M, which links each node to internal series, external data, or documented proxies, preserves the semantic connection between the document hypothesis and the data used.
Defined in Definition 2 and used throughout Section 3. The paper asserts the mapping 'preserving the connection between textual evidence, data choices, and model revisions,' but does not test whether the mapping is semantically accurate or whether proxy substitutions weaken the connection.
domain assumption Causal loop diagrams are an adequate shared representation for capturing the reasoning structure of forecasting documents and for guiding feature construction.
Adopted in Section 2.2 as 'the main inspectable representation.' The paper assumes CLDs can represent the relevant causal structure without a comparative study against other representation formats (e.g., knowledge graphs, event chains).

invented entities (2)

Ideal Causal Loop Diagram (Ideal CLD)
purpose: A document-derived directed graph of key factors and directional hypotheses, serving as the pre-data causal structure from which the forecasting pipeline starts.
This is a novel conceptual representation proposed by the paper. It has no falsifiable handle outside the paper; it is an internal organization of agent outputs. Its value is assessed subjectively as a traceability aid.
Data-Grounded Causal Loop Diagram (Data-Grounded CLD)
purpose: A versioned directed graph that extends the Ideal CLD with node-to-data mappings, used to guide feature construction, model design, and traceable revisions.
Another internal representation. It is introduced to connect document-derived factors to actual data, but no external validation or measurement of its effectiveness is given within the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Traceable Multi-Agent System for Knowledge-Based Forecasting." pith.science (2026). https://pith.science/paper/56A2U7YA

@misc{pith2026260803339,
  author       = {Pith},
  title        = {Pith review of: Traceable Multi-Agent System for Knowledge-Based Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/56A2U7YA}},
  note         = {Machine review of arXiv:2608.03339}
}

read the original abstract

Enterprise forecasting increasingly relies on autonomous agents that interpret documents, search for data, generate code, and revise models. While this autonomy helps build adaptive forecasting pipelines, it also makes it difficult for practitioners to inspect why a forecast changed, which evidence supported the change, and how data and modeling choices were revised. We present TraceMAS, an interactive demo system for traceable multi-agent forecasting. TraceMAS organizes agent outputs around two causal-loop representations: an Ideal Causal Loop Diagram (Ideal CLD), which captures key factors and their causal relations extracted from domain documents, and a Data-Grounded Causal Loop Diagram (Data-Grounded CLD), which links those factors to internal variables, external data, or documented proxies. The Data-Grounded CLD guides feature construction and model design while preserving the connection between textual evidence, data choices, and model revisions. We demonstrate TraceMAS on crude oil price forecasting. The demo interface allows users to compare forecasting iterations, inspect agent-level revisions, explore causal maps, review feature-data mappings and model architecture, and connect scenario forecasts to market narratives. This demonstration shows how autonomous forecasting agents can retain flexibility while making the evidence-to-forecast process inspectable.

Figures

Figures reproduced from arXiv: 2608.03339 by the authors.

**Figure 1.** TraceMAS demo interface for crude oil price forecasting. The interface allows users to inspect iteration-level agent [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Overview of the TraceMAS workflow. Domain documents are used to construct an Ideal CLD. The Ideal CLD is then [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Initial Ideal CLD. The diagram summarizes document-derived factors and directional hypotheses extracted from [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Initial Data-Grounded CLD. The diagram links document-derived factors to available internal time-series variables, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages

[1]

Pierre-Daniel Arsenault, Shengrui Wang, and Jean-Marc Patenande. 2025. A Survey of Explainable Artificial Intelligence (XAI) in Financial Time Series Fore- casting.Comput. Surveys(2025)

work page 2025
[2]

Chawla, Olaf Wiest, and Xiangliang Zhang

Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V. Chawla, Olaf Wiest, and Xiangliang Zhang. 2024. Large Language Model Based Multi-Agents: A Survey of Progress and Challenges. InProceedings of the Thirty- Third International Joint Conference on Artificial Intelligence. 8048–8057

work page 2024
[3]

Niyousha Hosseinichimeh, Aritra Majumdar, Ross Williams, and Navid Ghaf- farzadegan. 2024. From text to map: A system dynamics bot for constructing causal loop diagrams.System Dynamics Review(2024), e1782

work page 2024
[4]

Furong Jia, Kevin Wang, Yixiang Zheng, Defu Cao, and Yan Liu. 2024. GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence

work page 2024
[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. InInternational Conference on Learning Representations

work page 2024
[6]

Junhyeok Kang, Jun Seo, Soyeon Park, Sangjun Han, Seohui Bae, Hyeokjun Choe, and Soonyoung Lee. 2026. Channel-Wise Retrieval for Multivariate Time Series Forecasting. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1336–1340

work page 2026
[7]

Junhyeok Kang, Yooju Shin, and Jae-Gil Lee. 2025. VarDrop: Enhancing Training Efficiency by Reducing Variate Redundancy in Periodic Time Series Forecasting. InAAAI

work page 2025
[8]

Peiyuan Liu, Hang Guo, Tao Dai, Naiqi Li, Jigang Bao, Xudong Ren, Yong Jiang, and Shu-Tao Xia. 2025. CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 18915–18923. doi:10.1609/aaai.v39i18.34082

doi:10.1609/aaai.v39i18.34082 2025

Show all 16 references

[9]

Marius Schlegel et al . 2025. Capturing end-to-end provenance for machine learning pipelines.Information Systems132 (2025), 102495

2025
[10]

Renan Souza, Amal Gueroudji, Stephen DeWitt, Daniel Rosendo, Tirthankar Ghosal, Robert Ross, Prasanna Balaprakash, and Rafael Ferreira da Silva. 2025. PROV-AGENT: Unified Provenance for Tracking AI Agent Interactions in Agentic Workflows.arXiv preprint arXiv:2508.02866(2025)

2025 arXiv
[11]

John D. Sterman. 2000.Business Dynamics: Systems Thinking and Modeling for a Complex World. Irwin/McGraw-Hill

2000
[12]

Patara Trirat, Yooju Shin, Junhyeok Kang, Youngeun Nam, Jihye Na, Minyoung Bae, Joeun Kim, Byunghyun Kim, and Jae-Gil Lee. 2024. Universal time-series representation learning: A survey.arXiv preprint arXiv:2401.03717(2024)

2024 arXiv
[13]

Xinlei Wang, Maike Feng, Jing Qiu, Jinjin Gu, and Junhua Zhao. 2024. From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection. InAdvances in Neural Information Processing Systems

2024
[14]

Andrew Robert Williams, Arjun Ashok, Étienne Marcotte, Valentina Zantedeschi, Jithendaraa Subramanian, Roland Riachi, James Requeima, Alexandre Lacoste, Irina Rish, Nicolas Chapados, and Alexandre Drouin. 2025. Context is Key: A Benchmark for Forecasting with Essential Textual...

2025
[15]

White, Doug Burger, and Chi Wang

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. 2023. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation.arXiv pre...

2023 arXiv
[16]

Zhijian Xu, Yuxuan Bian, Jianyuan Zhong, Xiangyu Wen, and Qiang Xu. 2024. Beyond Trend and Periodicity: Guiding Time Series Forecasting with Textual Cues.arXiv preprint arXiv:2405.13522(2024). KDD ’26, August 9–13, 2026, Jeju, Korea Junhyeok Kang, Sangjun Han, Hyeokjun Choe, a...

2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Pierre-Daniel Arsenault, Shengrui Wang, and Jean-Marc Patenande. 2025. A Survey of Explainable Artificial Intelligence (XAI) in Financial Time Series Fore- casting.Comput. Surveys(2025)

work page 2025

[2] [2]

Chawla, Olaf Wiest, and Xiangliang Zhang

Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V. Chawla, Olaf Wiest, and Xiangliang Zhang. 2024. Large Language Model Based Multi-Agents: A Survey of Progress and Challenges. InProceedings of the Thirty- Third International Joint Conference on Artificial Intelligence. 8048–8057

work page 2024

[3] [3]

Niyousha Hosseinichimeh, Aritra Majumdar, Ross Williams, and Navid Ghaf- farzadegan. 2024. From text to map: A system dynamics bot for constructing causal loop diagrams.System Dynamics Review(2024), e1782

work page 2024

[4] [4]

Furong Jia, Kevin Wang, Yixiang Zheng, Defu Cao, and Yan Liu. 2024. GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence

work page 2024

[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. InInternational Conference on Learning Representations

work page 2024

[6] [6]

Junhyeok Kang, Jun Seo, Soyeon Park, Sangjun Han, Seohui Bae, Hyeokjun Choe, and Soonyoung Lee. 2026. Channel-Wise Retrieval for Multivariate Time Series Forecasting. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1336–1340

work page 2026

[7] [7]

Junhyeok Kang, Yooju Shin, and Jae-Gil Lee. 2025. VarDrop: Enhancing Training Efficiency by Reducing Variate Redundancy in Periodic Time Series Forecasting. InAAAI

work page 2025

[8] [8]

Peiyuan Liu, Hang Guo, Tao Dai, Naiqi Li, Jigang Bao, Xudong Ren, Yong Jiang, and Shu-Tao Xia. 2025. CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 18915–18923. doi:10.1609/aaai.v39i18.34082

doi:10.1609/aaai.v39i18.34082 2025