Agentic RL Research

Use the public service at https://hoyant-su-agentic-rl.hf.space through the self-contained Python standard-library client in scripts/service.py. Run commands relative to this skill directory. Access is anonymous.

python3 scripts/service.py schema
python3 scripts/service.py describe search_evidence
python3 scripts/service.py call list_sources --args '{}'
python3 scripts/service.py call search_evidence --args '{"query":"credit assignment","limit":3}'

For method and benchmark comparisons, read the research_comparisons field returned by list_sources, or fetch https://hoyant-su-agentic-rl.hf.space/topics/comparisons/index.json. Each topic supplies JSON, CSV and BibTeX. The comparison collection covers credit assignment mechanisms, memory and context selection, and terminal benchmark design. Every property links to its original paper section. Keep task conditions and cited source versions attached to comparisons.

For research-condition queries, call list_method_facets, then filter_methods with the exact categorical values returned. Supplied filters combine with AND; omitted filters are unrestricted. Results include the source support for every attribute and each original paper's BibTeX. Critic status unknown is separate from no. Supervision values describe signals used by the specified method; matching one value does not exclude its other listed signals. Evaluation settings report where a method was studied, without claiming transfer to every task in that setting. The HTTP equivalents are /methods/facets and /methods with query parameters.

For original passages, inspect list_sources for retrieval coverage, versions and licenses. The passage collection declares 13 papers; 8 have indexed original body text and 5 expose bibliographic information. Use English keywords: plain terms imply AND; quoted phrases and OR are supported. Consult additional sources when a literature question extends beyond this coverage.

Use returned evidence_id values with fetch_evidence to inspect complete blocks. Preserve source URLs, versions, section anchors, authors, formulas, table cells, and extraction limitations when using evidence. Cite an original source only when it supports the claim; the service's inclusion and ranking do not validate scientific claims.

For benchmark inspection, call dataset_overview, then search_tasks, then get_task with a returned task_id and partition. Task search is a literal substring match; use next_offset for additional pages. Keep published split membership, overlap and source links intact. The tools return task instructions, reward specifications and workspace assets for inspection.

schema returns the live API schema; describe ENDPOINT returns its definition. call ENDPOINT --args JSON prints the exact single server result after transport decoding. --timeout sets the network I/O timeout and --max-wait bounds SSE waiting, checked between received lines. Errors terminate without partial results or automatic retries.