Research map / Time series / Foundation models for time series

LLMs can construct powerful representations and streamline sample-efficient supervised learning

NeurIPS 2026 · Time series › Foundation models for time series · approach: Time-series language reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes rubric representation learning: an LLM inspects a small diverse cohort of text-serialized EHRs to synthesize a global rubric (or writes per-example local summaries) that standardizes inputs for sample-efficient downstream classifiers on 15 EHRSHOT tasks.

Problem

Supervised learning on complex heterogeneous data (e.g., EHRs with text, codes, labs) is bottlenecked by input representation design, which requires domain expertise and may discard or bury signal, especially when labeled data is scarce.

Approach

An LLM analyzes a small label-balanced, k-means-medoid cohort of text-serialized inputs in-context to synthesize a task-specific global rubric that defines what to extract and how; the rubric is applied to all inputs (by LLM, by generated parser script, or by tabularization script). Local rubrics are per-example task-conditioned LLM summaries.

Key results

• At n=10 per class, Local-Rubric reaches 0.701 AUROC vs 0.609 for NaiveText and 0.597 for CLMBR-T; Global-Rubric reaches 0.649 AUROC.
• At n=All, Global-Rubric obtains 0.763 AUROC (0.459 AUPRC) and Local-Rubric 0.772 AUROC, versus CLMBR-T at 0.725 AUROC.
• Global-Rubric-Auto (deterministic parser, no per-example LLM calls) gives up about 0.02 AUPRC at n=All (0.437 vs 0.459).

Surprising findings

• Stripping LLM interpretation from local rubrics (NoInterp) costs noticeable performance at small n but almost nothing at n=All.
• Zero-shot GPT5-Mini and Qwen3-8B CoT perform near chance on most tasks (about 0.61-0.64 AUROC overall).

Limitations

• Evaluation restricted to a single benchmark (EHRSHOT) without clinical notes or images.
• Rubric synthesis bounded by context length (40 patients per global rubric); only a single global rubric per task reported.
• Chest X-ray task shows near-chance performance for all methods at n=All.

Experimental setup

Datasets: EHRSHOT

Benchmarks: EHRSHOT

Metrics: AUROC AUPRC

Baselines: Count-GBM CLMBR-T NaiveText Qwen3-8B zero-shot CoT GPT5-Mini zero-shot CoT Global-Rubric-Blind

Builds on: EHRSHOT CLMBR-T chain-of-thought prompting XGBoost GEPA

Closest papers by meaning

• Small Models are LLM Knowledge Triggers for Medical Tabular Prediction (ICLR 2025 · similarity 0.87)
• MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making (NeurIPS 2024 · similarity 0.84)
• Format Matters: The Robustness of Multimodal LLMs in Reviewing Evidence from Tables and Charts (AAAI 2026 · similarity 0.84)
• HC-LLM: Historical-Constrained Large Language Models for Radiology Report Generation (AAAI 2025 · similarity 0.84)
• OctoMed: Data Recipes for State-of-the-Art Multimodal Medical Reasoning (CVPR 2026 · similarity 0.83)
• PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments (NeurIPS 2026 · similarity 0.83)