{
  "abstract": "Sparse expert supervision can improve reinforcement learning, but deciding which\n\nstates are worth annotating remains difficult. Uncertainty identifies where a learner\n\nis unsure, but does not directly identify where supervision will most strongly\n\nchange that learner. We introduce ALBERTA, a response-aware acquisition frame-\n\nwork that measures local policy and value changes produced by controlled expert\n\ninterventions, predicts those responses from pre-annotation signals, and selects\n\ncandidates in the resulting two-dimensional response space. A calibration set of\n\n200 expert-labeled probes is used to learn response from policy entropy, historical\n\ncritic disagreement, and their interaction; Pareto selection then allocates a separate\n\n121-state downstream budget. The response-aware branch therefore uses 321 ex-\n\npert labels in total, compared with 121 for the Random baseline. In a controlled\n\nsingle-seed Atomic Chess experiment from a common RL10 branch point, the\n\nPareto branch achieves a higher mean game score than PPO-only at all four tested\n\nevaluation temperatures, with smaller margins at higher temperatures, and shows\n\nits largest margins over Random at lower temperatures. These results provide\n\nevidence that learner response can be a useful acquisition target while exposing\n\nsubstantial differences between policy and value responsiveness.",
  "abstract_shown": true,
  "affiliations": [
    "Arts et Métiers"
  ],
  "article_type": null,
  "arxiv": null,
  "author_total": 1,
  "authors": [
    {
      "affs": [
        0
      ],
      "name": "Tom Koza"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/e1z1s0qncn9kc.bib",
  "cited_by": 0,
  "doi": "10.5281/zenodo.23416966",
  "fulltext_url": null,
  "id": "e1z1s0qncn9kc",
  "issns": [],
  "landing_url": "https://zenodo.org/records/23416966",
  "language": "eng",
  "licence": "CC-BY-4.0",
  "licence_label": "CC BY 4.0",
  "licence_source": "article",
  "licence_url": "https://creativecommons.org/licenses/by/4.0/",
  "metadata_licence": null,
  "native_subjects": [],
  "openalex": null,
  "pdf_url": "https://zenodo.org/api/records/23416966/files/preprint_atomic_chess.pdf/content",
  "pmcid": null,
  "pmid": null,
  "published": "2026-10-16",
  "publisher": null,
  "ris": "https://research.draconitic.com/paper/e1z1s0qncn9kc.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T12:17:21Z",
      "native_id": "23416966",
      "source": "zenodo",
      "url": "https://zenodo.org/records/23416966"
    }
  ],
  "subjects": [],
  "title": "Learning from the Learner's Response: Bi-Objective Acquisition for Sparse Expert Supervision in Reinforcement Learning",
  "title_text": "Learning from the Learner's Response: Bi-Objective Acquisition for Sparse Expert Supervision in Reinforcement Learning",
  "type": "preprint",
  "updated_at": "2026-10-30T12:33:19Z",
  "url": "https://research.draconitic.com/paper/e1z1s0qncn9kc",
  "venue": "Zenodo",
  "versions": []
}