Real execution traces from this repository (real tool-execution mode) on three GTA tasks:
step through each rollout — candidate drafts, pre-evaluation, execution with real tool outputs,
post-evaluation, pruning, and Q/N backpropagation.
Provenance: every candidate draft, score (r_pre / r_post), Q/N value, tool output, and final
answer below is taken verbatim from two logged runs of this repository — 5 GTA tasks (Cases 1–2)
and 4 GTA tasks with 3–4-step reference plans (Case 3) — both with
tool_execution: real (seed 42, τpre 0.3, τpost 0.4,
c 1.4, max 40 rollouts). Long tool outputs are truncated to 300 characters and flagged. Failed tool
calls (e.g. a tool without a real backend, or a malformed argument draft) are shown verbatim —
post-evaluation prunes those branches; one error message that echoed a local filesystem path has
that path masked and is flagged. Task images are referenced by their dataset-relative paths and are
not embedded here.
Walkthrough data failed to load. This page needs static/demo_data.js next to it
(shipped in the repo under docs/static/). If you are viewing a copy of this file alone,
grab the full docs/ directory from the
GitHub repo.
Search tree of this case (reconstructed from the logged run): solid boxes are executed tool
calls with their pre-/post-evaluation scores, dashed boxes are pruned or duplicate (skipped)
candidate drafts, and the bold green path is the final highest-reward plan.
