The idea in one paragraph.
Interleaving reasoning traces with environment actions lets each correct the other - thinking plans the next action, and observations correct the thinking.
THE MACHINERY
Thought-Action-Observation loop
Alternate free-form reasoning with a tool call and its observation.
Acting as grounding
Retrieved observations correct hallucinated reasoning mid-trajectory.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
On ALFWorld and WebShop, outperforms imitation and reinforcement learning methods by an absolute success rate of 34% and 10% respectively, while prompted with only one or two examples.
Inspect a source excerpt
outperforms imitation and reinforcement learning methods by an absolute success rate of 34% and 10% respectively, while being prompted with…Read the passage in context on arXiv ↗
Reasoning alone hallucinates; acting alone cannot plan. The combination beats either.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- The loop is fixed by the prompt; nothing accumulates between episodes.
- No verifier - the trajectory ends when the model says it is done.
Why it belongs in this atlas.
The canonical agent loop, and still the default shape of almost every harness. Note what it lacks: no memory across runs, no verification, no budget the agent can reason about.