AREX-2 Advances Self-Improving Agents via Long-Horizon Reflective Tasks: Scoring 81.8 on MLE-bench and 92.2 on GAIA

核心背景与行业痛点

Autonomous LLM agents struggle when deployed on open-ended tasks requiring sustained trial-and-error across extended horizons, such as machine learning experimentation or deep research. This fragility stems from two bottlenecks in test-time self-improvement: inadequate reflection (the inability to propose hypotheses strictly superior to current iterations) and fragile long-horizon execution (the failure to maintain goal coherence over dozens of turns without context collapse or looping).

架构亮点与底层机制

Researchers from Renmin University, Microsoft Research, and USTC introduce AREX-2 (arXiv:2609.38288):

1. Domain-Agnostic Dual Capabilities: Demonstrates that reflective critique and long-horizon execution represent generalized metacognitive capacities that transfer across modalities.
2. Verifiable Long-Horizon Data Synthesis: Curates iterative trajectory data from machine learning engineering (MLE-bench) and algorithmic programming, domains providing unambiguous execution traces and metric feedback.
3. Continuous Refinement Distillation: Internalizes complete problem-solving lifecycles—from buggy prototypes and performance diagnosis to hardened solutions—directly into model parameters.
4. Monotonic Test-Time Compute Scaling: Empowers the agent to dynamically allocate reflection iterations during inference, maintaining positive scaling across extended search budgets.

权威 Benchmark 与实测跑分对比

Evaluated on demanding autonomous workflows:

1. 81.8 on MLE-bench Lite: A Qwen3.8-27B agent post-trained with AREX-2 achieves 81.8 on MLE-bench Lite and 70.7 on Frontier-CS.
2. Strong Zero-Shot Transfer to Deep Research: Generalizes to knowledge-intensive benchmarks with 84.0 on BrowseComp, 52.6 on Humanity's Last Exam (HLE), 92.2 on GAIA, and 93.8 on DeepSearchQA.
3. Monotonic Budget Scaling: Task success rates scale monotonically as allowable reflection rounds expand from 1 to 8, reversing conventional multi-turn decay.

开发者实战落地与开箱指南

AREX-2 training recipes and checkpoints are openly available. AI engineering teams developing autonomous software engineers or deep research copilots can leverage verifiable coding environments to teach agents iterative reflection, deploying models capable of sustained test-time self-evolution.