Many real‑world tasks require LLM agents to interact repeatedly with their environments, but current environments suffer from three major issues: scattered information, mixed‑in misleading evidence, and time‑varying noise. These factors can cause state‑of‑the‑art agents to drop from 83.9% to 57.6% accuracy.
Env‑Rethink, built on a 27B post‑trained model, offers three capabilities. First, it adaptively builds Collection Maps to group related files and Event Logs to capture cross‑data contextual relationships, supplementing missing context. Second, offline trajectory learning lets the model pinpoint underlying noise in the environment. Third, it evolves environments by injecting virtual event histories that alter states and evidence links, creating harder scenarios for recursive self‑improvement.
Experiments on 30 tasks across nine models show a 15.1% absolute increase in rubric pass rate, confirming the system’s effectiveness in boosting downstream performance.
Review