This paper proposes Proactive Thinking, a framework that lets large language models pre-compute possible response elements during conversational downtime instead of beginning reasoning only after the user replies. It introduces a training-free baseline based on speculative continual thinking, which anticipates possible future states while balancing computation and response quality. The authors adapt three benchmarks with different complexity levels into time-aware environments that simulate real-time dialogue. According to the abstract, proactive thinking improves interaction efficiency without compromising performance, suggesting a shift from purely reactive inference toward anticipatory reasoning for conversational systems.
No heat snapshots are available in the last 24 hours.