Thinking longer cannot reveal information hidden behind a link the agent has never clicked. The useful scaling question is therefore not just how many tokens to spend before an action, but how much opportunity the agent gets to explore, backtrack, and revise its plan from fresh observations.
Key Findings
- On a controlled 62-task WebArena subset, prompting an agent to check its work again raised success from roughly 23% to at least 28%; the authors report smaller gains from spending comparable compute on longer per-step reasoning or best-of-n selection. These experiments establish a mechanism rather than a full-benchmark ranking.
- The TTI curriculum gradually increases rollout horizons during online reinforcement learning. Learning useful short trajectories first helps prevent extra steps from becoming aimless exploration.
- Experiments with Gemma 3 12B on WebVoyager and WebArena support interaction length as a complementary scaling axis. The trained policy can explore on harder tasks while taking more direct paths on easier ones.
Who Should Read It
Great fit if you train browser agents or need to decide whether extra inference budget should buy deeper reasoning or additional environment feedback. Look elsewhere if you need evidence that longer interaction always helps: repeated re-checking can overturn correct answers, and the reported evaluations concern web navigation rather than arbitrary agent domains.