ReAct gives the answer a next move
ReAct alternates reasoning with actions that gather information. A useful research result—and a sharper question about what we want models to do.
A fluent answer can still need to look something up. ReAct’s researchers explore a way for language models to alternate reasoning traces with task-specific actions. Those actions gather information from external sources; the traces help update plans. The interesting shift is from producing an answer to choosing another step. I find that an encouraging expansion of the job description. [1]
The authors report that using a Wikipedia API helps address hallucination and error propagation on HotpotQA question answering and Fever fact verification. They also report more interpretable task-solving trajectories than baselines without reasoning traces. These are claims about the study’s evaluations. They do not establish that an explanation is always reliable, or that consulting a source settles every question. An account of the steps deserves scrutiny too. [1]
On ALFWorld, the paper reports better results than the compared imitation-learning methods; on WebShop, it reports better results than the compared reinforcement-learning methods. These experiments used one or two examples. Keep those claims attached to their settings. My larger interest is the invitation: if gathering information can change the next action, the boundaries of those actions matter. Before asking a machine to continue, it is worth deciding what continuing permits. [1]
