ReTool: Learn when to use tools during reasoning¶
Core mechanism in a deterministic mini-suite, not the full benchmark.
Paper information¶
| Field | Details |
|---|---|
| Paper | ReTool (arXiv 2504.11536) |
| Institution | ByteDance Seed. |
| First public release | 2025-04-15 (arXiv v1) |
| Author code | No official author repository found or released. |
| Local adapter or method key | retool |
| Local reproduction code | src/auto_research/agent_research/ |
Paper method and evidence¶
Alternate natural-language reasoning and executable tool actions. Use execution feedback to learn tool invocation, correction and stopping.
flowchart LR
R[Natural-language reasoning] --> T[Choose and execute tool]
T --> O[Observe execution feedback]
O --> R
O --> S[Learn correction or stopping]Original paper figure¶
The original figure shows reasoning interleaved with tool calls. From the original paper; copyright remains with the authors.
Core equation¶
\[
\tau=(z_1,a_1,o_1,\ldots),\quad R=R_{answer}-c\sum_t\mathbf1[a_t=tool],\quad\max_\pi\mathbb E[R].
\]
Reported results¶
The paper reports gains over no-tool and fixed-call baselines on math and tool-augmented tasks. No production A/B.
Local reproduction¶
PlanBench mini-suite, 120 episodes, seed 42: joint success 1.0000, mean cost 1.0500, and nonzero method-specific telemetry.
auto-research agent-eval --method retool --benchmark planbench-mini --episodes 120 --seed 42
auto-research evolve --model agent --dataset planbench-mini --direction "Combine retool with installed paper operators" --generations 2 --population 4
Fixed metrics. Validates method-specific objective, state updates and matched budgets only, not large models, multi-GPU RL, private environments, real web or full benchmarks.