Xin Eric Wang

Evolving Agent Interfaces

Saaket Agashe, Yuzhe Yang, Gonzalo Gonzalez-Pumariega, et al.

LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents

Qingni Wang, Changdae Oh, Yuzhe Yang, et al.

Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants

Deepak Nathani, Cheng Zhang, Chang Huan, et al.

Context Bootstrapped Reinforcement Learning

Saaket Agashe, Jayanth Srinivasa, Gaowen Liu, et al.

This paper introduces Context Bootstrapped Reinforcement Learning (CBRL), a technique that helps AI models learn better by occasionally providing example demonstrations during training. The method starts by frequently showing examples to help the model explore and learn initial patterns, then gradually removes these examples so the model must eventually succeed on its own. The approach is tested on reasoning tasks and a specialized programming language, showing consistent improvements in success rates and learning efficiency.