Self-Certainty Guided Test-Time Scaling for Web Agents

Large language models (LLMs) are increasingly deployed in real-world interactive environments where response quality and decision efficiency are both critical. We propose a adaptive test-time scaling strategy facing response selection based on self-certainty, a confidence measure computed from the model’s own token-level log-probabilities to adaptively control branching and inference cost.

Evaluated on the VisualWebArena and WebArena benchmark, our method consistently outperforms the baselines, achieving a +3.6% improvement in average success rate while reducing prompt token usage by over 80%. These results highlight the utility of internal confidence estimation in balancing exploration and efficiency, paving the way for more selective and scalable test-time inference in LLM based agents.

Video Presentation: Video URL

Slides: SC-ATS Project Overview

Link: GitHub Repository