AI Paper Atlas

Research map / Time series / Foundation models for time series

TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models

Tong Guan, Zijie Meng, Dianqi Li, Shiyu Wang, Chao-Han Huck Yang, Qingsong Wen, Zuozhu Liu, Sabato Marco Siniscalchi, Ming Jin, Shirui Pan

ICLR 2026 · Time series › Foundation models for time series · approach: Time-series language reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CoT-SFT with reasoning traces instead of answer-only SFT (ablation)TimeOmni-1 multi-task · Causality Discovery (ID)accuracy: 30.5 -> 67.7Improved
add Stage 1 SFT on reasoning traces to the base model (ablation)Qwen2.5-Instruct-7B · Task 2 Causality Discovery (ID)accuracy: +46.1%Improved
add human-guided templates to the zero-shot prompt (ablation)GPT-4.1 · Task 2 Causality Discoveryconsistency accuracy: 28.7 -> 71.1Improved
apply Stage 2 RL directly to the base model without Stage 1 SFT (ablation)Qwen2.5-Instruct-7B · Task 4 Decision Makingaccuracy: 5.3% dropHurt
add RL (Stage 2) after CoT-SFT (ablation)TimeOmni-1 multi-task · Scenario Understanding (ID)accuracy: 89.8 -> 90.7Improved
add RL after CoT-SFT on decision making (ablation)TimeOmni-1 multi-task · Decision Making (ID)accuracy: 40.9 -> 47.9Improved
use multi-task joint training instead of single-task training (ablation)CoT-SFT+RL · TSR-Suite ID testbed, four tasksaccuracy / MAE: +8.2%, +1.8%, 2.46 (MAE), +7.0%Improved
train decision-making jointly with perception and extrapolation tasks (ablation)TimeOmni-1 · Decision Making (ID)accuracy: 40.9 -> 45.7 -> 47.9Improved
zero-shot transfer from perception and extrapolation training to decision making (ablation)base model · Decision Making (ID)accuracy: 25.5 -> 26.2 -> 31.3Improved
add the counting bonus reward for sequence lengthRL reward for event-aware forecasting · Task 3 Event-aware Forecastingvalid sequence length: Stage I checkpoint 55.7% success on sequence lengthImproved
time series specialization training (SFT+RL)TimeOmni-1 · DROP, GPQA, ReClor averageaccuracy: +16.5% over base, +1.3% over Stage 1Improved
multi-task joint training for ANS-SFT on causality discovery (ablation)ANS-SFT · Causality Discovery (ID)accuracy: 35.7 -> 30.5Hurt

Later papers that compared with this method

2 later accepted papers report TimeOmni-1 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy