---
title: "AREX: Towards a Recursively Self-Improving Agent for Deep Research"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.21461v1"
permalink: "https://feed.astralor.com/items/itm_2af7e9352c3787b41a0661d5"
published: "2026-07-23T16:05:46.000Z"
score: 81
channel: "ai"
tags: ["AREX","arXiv","Deep research agent","Mixture-of-Experts","Recursive Self-Improvement"]
---

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

## AI commentary

> This paper demonstrates how to break through the bottleneck of deep research via recursive self-improvement, especially valuable for developers building AI agents capable of autonomous complex research. Its context compression tool addresses the core challenge of long-horizon interaction, and open-sourcing models and training data will accelerate progress in this field.

## At a glance

AREX is a recursively self-improving deep research agent framework that alternates between an inner research loop for gathering evidence and constructing a provisional answer, and an outer self-improvement loop that audits the answer constraint-wise, identifies unresolved claims, and launches targeted follow-up research. It learns an autonomous context-update tool to compress growing interaction history into a compact improvement state without relying on an external model. Trained on a 4B dense model and a 122B-A10B Mixture-of-Experts model, AREX substantially outperforms comparable-scale baselines on BrowseComp, WideSearch, DeepSearchQA, Humanity's Last Exam, and other benchmarks, remaining competitive with models using substantially more activated parameters.

## Key points

- AREX proposes a recursively self-improving (RSI) paradigm with inner and outer loops to iteratively refine answers
- Learns an autonomous context-update tool that compresses interaction history into a compact improvement state, eliminating external model dependency
- Outperforms comparable-scale baselines on multiple deep research and reasoning benchmarks, and remains competitive with models using more activated parameters
- Supports scaling from 4B dense model to 122B-A10B Mixture-of-Experts model

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_2af7e9352c3787b41a0661d5