Datasets:

wakeupmh
/

repro-when-to-trust-the-cheap-check-traces

Modalities:

Formats:

Size:

Tags:

format:agent-traces

Libraries:

Dataset card Data Studio Files Files and versions

Claude Code traces preview

Wed, Jul 22

# 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 #29949 — When to Trust the Cheap Check: Weak and Strong Verification for Reasoning (OpenReview id: dfN4HMZbc9). Here are the major claims that you should verify: - Claim 1: The Selective Strong Verification (SSV) algorithm adaptively maintains two thresholds to decide whether to accept, reject, or defer a response to strong verification (Section 5, Algorithm 1). - Claim 2: Theorem 5.1 proves finite-time, distribution-free error control for SSV: Type-I error stays below alpha plus a confidence term and Type-II error stays below beta plus a confidence term, without assumptions on the query stream, the language model, or the weak verifier (Section 5, Theorem 5.1). - Claim 3: Theorem 4.2 shows that under calibration assumptions the optimal verification policy has a two-threshold structure: responses scoring below the lower threshold are rejected, those above the upper threshold are accepted, and those in between are deferred to strong verification (Section 4, Theorem 4.2). - Claim 4: On the Sudoku step-by-step verification task at error tolerances alpha=beta=0.01, SSV attains 43.1% accuracy using only 2.87 strong verifier calls per puzzle, versus 5.32 calls needed by an oracle baseline for comparable accuracy, a roughly 46% reduction in strong-verifier calls (Section 6, Table 1). - Claim 5: On MATH difficulty level 5, SSV reaches 60% accuracy using approximately 2 strong verifier calls per problem, compared to an oracle policy that needs 2.8 calls to reach 63.5% accuracy (Section 6). 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-when-to-trust-the-cheap-check-weak-and-strong-verification-for-reasoning && trackio logbook publish /repro-when-to-trust-the-cheap-check-weak-and-strong-verification-for-reasoning Print the published logbook link here when done. 11 messages 171 tools

Agent traces

Agent sessions published from a Trackio Logbook.

Downloads last month: 44

Number of rows:

Total file size: