Datasets:

bsenst
/

hearts-reproduction-traces

Modalities:

Formats:

Size:

Tags:

Libraries:

Dataset card Data Studio Files Files and versions

Pi traces preview

Sat, Jul 25

# First, read the challenge instructions: curl -sL https://huggingface.co/datasets/ICML-2026-agent-repro/challenge/resolve/main/README.md Your job is to reproduce the ICML 2026 paper #12789 — HEARTS: Benchmarking LLM Reasoning on Health Time Series (OpenReview id: qj4EnIvNU4). Here are the major claims that you should verify: - Claim 1: HEARTS integrates 16 real-world datasets spanning 12 health domains and 20 distinct signal modalities, with temporal lengths ranging from 60 to over 1 million steps and sampling frequencies from daily to 48kHz. - Claim 2: HEARTS organizes 110 tasks into four hierarchical reasoning categories: Perception (10 tasks), Inference (32 tasks), Generation (34 tasks), and Deduction (34 tasks). - Claim 3: Across 14 state-of-the-art LLMs evaluated on more than 20,226 test cases using the CodeAct framework, evaluated LLMs substantially underperform specialized models, with specialized models concentrated in a high-performance regime while LLMs are shifted substantially lower. - Claim 4: LLM performance on HEARTS correlates only weakly with general reasoning benchmark scores, indicating domain-specific health time-series reasoning skill is not explained by raw model capability. - Claim 5: LLMs predominantly rely on simple heuristics such as copy-pasting with noise, linear interpolation, and statistical averaging rather than genuine multi-step temporal reasoning, with performance degrading further as input sequence length and sampling frequency increase. Logbook structure (fixed order): - Index: title + Pages table only - Executive summary: pinned summary + Scope & cost table + pinned Chenruishuo/posterly poster (`poster_embed.html`) - Claim 1 … Claim N: one page per claim - Conclusion: overall findings and reproducibility notes Try to verify the claims as much as you can locally or using Hugging Face Jobs. Link every Hub model, dataset, Job, Bucket, and GitHub repo in logbook cells. # Last step — validate, then publish: curl -sL https://huggingface.co/spaces/ICML-2026-agent-repro/challenge/raw/main/scripts/validate_icml_logbook.py | python3 - --space /repro-hearts-benchmarking-llm-reasoning-on-health-time-series && trackio logbook publish /repro-hearts-benchmarking-llm-reasoning-on-health-time-series Print the published logbook link here when done. 4 messages 166 tools

Agent traces

Agent sessions published from a Trackio Logbook.

Downloads last month: 36

Number of rows:

Total file size: