Tag
This paper presents the first systematic study of how additive activation steering, calibrated in single-turn chat, transfers to tool-using ReAct agents. It finds that while the injected direction reaches late layers at near-full strength across settings, the behavioral coupling varies unpredictably between models and contexts, with amplification up to 2x or attenuation, posing immediate safety concerns.
This paper proposes replacing the stateless autoresearch pattern with a stateful ReAct agent using LangGraph, reducing per-iteration token costs from O(n) to O(1) and achieving 52-90% fewer tokens on hyperparameter tuning and code optimization benchmarks.
Co-ReAct introduces a rubric-guided action-selection framework that uses rubrics as step-level guidance during inference for ReAct agents, improving trajectory quality and outperforming baselines on DeepResearchBench and SQA-CS-V2.