TimeSage-MT - L4 Full Decision Path: leaderboard
TimeSage-MT (2026) evaluates LLM agents in multi-turn time-series analysis dialogues with verifiable, code-derived answers, here under its Code-Enabled Reasoning setting in which the model writes Python that a sandbox executes. The board averages the outcome score in percent over the 60 L4 tasks, long dialogues that end in a structured recommendation with rationale, confidence and risks; higher is better.
Metric: Outcome score (%) on the 60 L4 full-decision tasks (12-20 turns ending in a structured decision): per-task mean of the applicable outcome dimensions (numerical accuracy, factual verification, GPT-5.4-judged analytical quality and code correctness, plus decision making at L4); Code-Enabled Reasoning setting (the model writes Python, a sandbox runs it, the model answers from stdout; 240 multi-turn time-series analysis tasks over 8 domains), temperature 0, reasoning effort high where supported and medium for GPT-5.3-Codex; higher is better. Source: arxiv.org. Saturation forecast: Around May 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.1 | 38.87 | #179 |
| 2 | Claude Sonnet 4.6 | 37.96 | #161 |
| 3 | GPT-5.3 Codex | 37.69 | #116 |
| 4 | Qwen 3.5 122B A10B | 36.13 | #331 |
| 5 | MiniMax-M2.7 | 35.44 | #518 |
| 6 | Gemini 3 Flash (Preview) | 33.45 | #144 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=timesage-mt-l4-full-decision-path · How It Works · Data refreshed daily, snapshot 2026-11-13.