Adaptive Stopping of Streamed Reasoning for Language Model Agents
We introduce StreamStop, which learns model-specific stopping policies offline from prefix-level action agreement.
ProofPaper ↗
Key points
- Reasoning-enabled language model agents can incur substantial computation before acting, although the reasoning needed varies across decisions.
- A frozen prefix reader encodes available reasoning text for a lightweight stop head, allowing stopping without access to the reasoning model's internal signals.
- We construct and publicly release a dataset of agent reasoning trajectories with prefix-level action annotations.
- These annotations compare each prefix action with the full-reasoning reference action, enabling offline comparison of stopping rules and exact expected-return optimization of a discounted action-agreement objective.
Sources (1)
- [1]Adaptive Stopping of Streamed Reasoning for Language Model AgentsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 10:24 AM
We introduce StreamStop, which learns model-specific stopping policies offline from prefix-level action agreement.
Reasoning-enabled language model agents can incur substantial computation before acting, although the reasoning needed varies across decisions.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026SAPD: Learning from Agent Experience without Environment Interaction
- Oct 20, 2026Self-Evolving Agents through Privileged Retrospective Distillation
- Oct 20, 2026Evolving Agent Interfaces
- Oct 19, 2026Procedural Exposure Bias in Skill-Augmented Language Model Agents
- Oct 19, 2026Claim Laundering: When Agent Decision Models Trust Records Over Reality
- Oct 19, 2026Zephyrus-R1: Agentic Reinforcement Learning for Weather Reporting