AREX: Towards a Recursively Self-Improving Agent for Deep Research

2026-07-23

AI Summary

This work addresses the inefficiency of multi-constraint answer discovery in deep research by proposing AREX, a recursively self-improving agent framework. AREX features an inner research loop that generates preliminary answers and an outer refinement loop that continuously verifies constraints, identifies unresolved claims, and initiates targeted investigations to iteratively enhance answer quality. A novel autonomous context updating mechanism compresses interaction history into a compact state representation, enabling long-term self-optimization without reliance on external models. The framework integrates in-agent training, long-horizon reinforcement learning, and dense reward signals to prioritize critical evidence acquisition and error correction. Evaluated on benchmarks including BrowseComp, WideSearch, DeepSearchQA, and HLE, AREX significantly outperforms same-scale baselines and achieves performance comparable to substantially larger models.

Abstract

Deep research requires agents to find answers that jointly satisfy multiple constraints. Discovering such answers is costly, whereas verifying a candidate can often be decomposed into tractable constraint-wise checks. This discovery--verification asymmetry suggests that a research agent should do more than simply search longer: it should recursively improve its current answer by verifying intermediate results and using the partially verified state to guide subsequent refinement. We introduce AREX, a family of Recursively Self-Improving (RSI) deep research agents. AREX alternates between an inner research loop that gathers evidence and constructs a provisional answer, and an outer self-improvement loop that audits the answer constraint-wise, identifies unresolved claims, and launches targeted follow-up research. To sustain RSI over long horizons, AREX learns an autonomous context-update tool that compresses growing interaction history into a compact improvement state preserving verified evidence and unresolved constraints, without relying on an external model. We train AREX on verified synthetic tasks and high-quality trajectories through agentic mid-training and long-horizon reinforcement learning. To mitigate sparse final rewards during long horizon learning, we emphasize key steps where decisive evidence is acquired or erroneous research directions are corrected. We instantiate a dense 4B model and a 122B-A10B Mixture-of-Experts model. Across BrowseComp, WideSearch, DeepSearchQA, Humanity's Last Exam (HLE), and other reasoning and tool-use benchmarks, AREX substantially outperforms comparable-scale baselines and remains competitive with models using substantially more activated parameters.

Problem

• deep research
• recursively self-improving agent
• constraint satisfaction
• answer verification
• long-horizon reasoning

Innovation

• Recursively Self-Improving
• Deep Research Agent
• Constraint-wise Verification
• Autonomous Context Update
• Long-horizon Reinforcement Learning