0

Wikispeedia

Fresh

Navigate Wikipedia by following links to reach a target article.

Type
RL Env
Runtime
tool-use
License
unknown
Size
v0.2.2
Published
Apr 2026
Updated
Sep 2026

Cite

Notes

Only stored in your browser.

wikispeedia

Navigate Wikipedia by clicking links to reach a target article. Each task gives a (source, target) pair sampled from the Wikispeedia link graph; the agent uses a stateful wiki_click_link tool to move between articles, and the reward is 1.0 if the target is reached. Two zero-weight rewards track solution quality in [0, 1]: path_efficiency (shortest path over the net path walked) and click_efficiency (shortest path over all clicks spent).

Taskset

  • Source: Stanford SNAP Wikispeedia article graph; (source, target) pairs sampled from the graph within a distance band
  • Size: procedurally generated; a lazy stream that enumerates every in-band (source, target) pair exactly once, in a seeded pseudorandom order. The full graph has 4,604 articles, ~120k links, and ~15.5M pairs in the default 3–8 distance band. The taskset is declared infinite: training streams it; eval runs must be bounded with -n.

Train/eval split

--env.taskset.task.tools.split {train,eval} restricts the taskset and the navigation tools to one side of a fixed vertex partition: 15% of articles hash into the eval partition (a seeded name hash, so the split is identical everywhere), and each side plays on its induced subgraph. Train and eval share no pages and no links, so eval on the held-out side measures generalization to unseen graph structure rather than memorized link menus. Task distances are recomputed on the induced subgraph (BFS), not the full-graph distance matrix. The eval side has ~700 pages and ~186k pairs in the 3–8 band; the train side keeps ~11M.

Turn caps are not part of the taskset; set them on the agent (--env.agent.max-turns).

Changelog

  • 2026-09-03: Restore default solver network access by reverting the network_allow=[] default-deny policy introduced in #780; training rollouts need outbound network.
  • 2026-08-12: Added the fixed train/eval vertex partition (split); made the taskset a lazy generator over all in-band pairs (num_tasks removed, bound evals with -n); removed the taskset-level max_turns (use --env.agent.max-turns); added zero-weight path_efficiency and click_efficiency rewards.
  • 2026-07-31: Migrated the per-rollout wiki tools to the explicit toolsets(config) API required by verifiers>=0.2.2.dev65.
  • 2026-06-23: Initial v1 taskset.