Jarrodbarnes/latent-mining

Latent Mining

Latent Mining is a benchmark-construction method for scientific-agent tasks where useful public evidence diverges from a withheld verifier-backed outcome. The resulting tasks test whether agents can make calibrated scientific triage decisions under incomplete information.

This dataset contains the first public biology subset: 165 cross-locus regulatory-edit triage tasks. Each task asks an agent to choose among candidate noncoding edits for a specified assay and cell context using public evidence such as sequence-model priors, motif and PWM effects, edit geometry, locus metadata, assay family, chromatin context, and disagreement across public signals.

Biological foundation models are beginning to expose useful signals about sequence, structure, function, and regulation. But useful signal is not the same as verified outcome. Context is sparse, measurements are noisy, and ground truth is often expensive, indirect, or unavailable at benchmark scale. Latent Mining builds evaluation tasks around that gap.

In this biology release, correctness is defined by a hidden AlphaGenome-backed regulatory-effect objective for the requested assay and cell context. The verifier-backed outcome is used for task construction and grading and is not exposed in the public task package. It is not wet-lab truth, a universal biological oracle, or evidence that selected edits would validate experimentally.

License and availability

This public dataset release is distributed under the MIT license. It is intended for open benchmark use, audit, reproduction of the public task surface, and scientific-agent evaluation research.

Files

• data/train.parquet contains one viewer-safe index row per public biology task.
• latent_mining/biology/tasks/ contains the full release-safe task packages.
• latent_mining/biology/task_index.jsonl mirrors the viewer index in JSONL form.
• latent_mining/biology/manifest/ contains release-safe manifests and audit summaries.
• dataset_manifest.json records row count, hash, and source-bundle classification for this upload.
• LICENSE contains the MIT license for this dataset release.

Biology release

The biology subset covers 165 regulatory-triage tasks across 11 non-APOE loci, 9 chromosomes, 11 biology systems, and 4 assay families: ChIP, RNA-seq, ATAC, and CAGE.

Each task includes:

• task instructions
• candidate regulatory edits
• public evidence features
• answer schema
• public validator surface
• taskbench metadata

The Dataset Viewer table is intentionally lightweight. It contains task metadata, assay and locus context, public-evidence channel summaries, admission-gate flags, and file paths into the full task package. Candidate-level public features live in the per-task package files.

Each task excludes:

• hidden verifier scores
• hidden thresholds
• hidden ranks
• cache IDs
• known solutions
• raw solver traces
• private local paths

Task design

Latent Mining builds scientific-agent tasks around constrained decisions, not open-ended prompts. Each task begins with a workflow bottleneck where a scientist must choose among plausible interventions under incomplete evidence. In this release, the decision is regulatory-edit triage for a specified assay and cell context.

The candidate set is generated from domain models and retrieval procedures. The public surface exposes the kinds of evidence a functional genomics scientist could weigh: sequence-model priors, motif and PWM effects, edit geometry, locus metadata, assay family, chromatin context, and disagreement across public signals.

Correctness is defined by a separate hidden verifier-backed outcome. The benchmark asks whether an agent can reason from useful but incomplete public evidence toward a withheld scientific outcome.

A task is admitted only when the panel preserves four properties:

1. Workflow validity: the task corresponds to a real scientific decision.
2. Public traction: the public evidence gives the agent meaningful reasons to compare candidates.
3. Hidden and public separation: no single public score, source model, motif heuristic, edit family, or generic rule should recover the answer by itself.
4. Contrastive difficulty: wrong candidates are attractive under plausible public evidence, so success requires calibrated comparison among credible interventions.

Every released task is audited for leakage, shortcut recovery, interface validity, grader replay, wrong-answer controls, calibration behavior, and trace interpretability. Latent Mining converts scientific model blind spots into verifier-backed agent tasks by contrasting plausible public evidence against withheld domain outcomes.

Benchmark context

Six complete frontier-model evaluations span the 165 released tasks. Every ranked attempt passes the public interface and the grading interface, with no missing task-agent pairs, duplicate pairs, runner exits, timeouts, or malformed submissions. Under strict hidden verification, the aggregate solve rate is 11.3%, and no evaluated model solves more than 15% of the benchmark.

Process diagnostics and a paired expert-method rescue analysis accompany the release. Rescue is a diagnostic intervention, not a second leaderboard and not a replacement for unaided benchmark results. The completed paired core contains 132 paired model-task units across three model rows, where the workflow checklist improves process metrics such as assay orientation, decoy control, evidence comparison, and contrast more than it improves strict solves.

Agent trace behavior

Frontier agents often identify relevant regulatory evidence, including motif shifts, sequence-model priors, locus context, assay constraints, chromatin evidence, or model disagreement. The failure happens when those observations do not control the final intervention choice.

Two behaviors recur in qualitative audits:

• Noticing without control: the agent mentions a relevant public signal, but the final answer is driven by a more salient proxy, such as a motif change, a single model score, or the source that made one candidate look strongest.
• Explanation instead of triage: the agent writes a plausible regulatory story for one candidate, but does not compare that candidate against the alternatives under the requested assay and cell context.

Loading

from datasets import load_dataset

ds = load_dataset("Jarrodbarnes/latent-mining")
print(ds["train"][0]["task_id"])

The full task package for each row is listed in the package_path field. Candidate edits and candidate-level public features are stored in the corresponding candidates.json file.

Release scope

This dataset contains release-safe task packages and task metadata. Raw solver traces, interactive qualitative reports, hidden verifier scores, thresholds, ranks, cache identifiers, known solutions, and private grader artifacts are not included. This is the first biology release of Latent Mining; other domains are deferred to separate releases.

Citation

Cite Latent Mining as:

@misc{barnes2026latentmining,
  title        = {Latent Mining for Verifiable Scientific Task Generation},
  author       = {Barnes, Jarrod},
  year         = {2026},
  howpublished = {Preprint and public dataset},
  url          = {https://huggingface.co/datasets/Jarrodbarnes/latent-mining}
}