AREX-2: Self-Improving LLM Agents via Long-Horizon Reflection

Researchers introduced AREX-2, a Qwen3.8-27B-based agent trained on verifiable self-improvement trajectories in ML engineering and programming. It scored 81.8 on MLE-bench Lite, 92.2 on GAIA and 93.8 on DeepSearchQA, with performance still improving after up to 30 refinement rounds.

DEV Community

Open source

AREX-2: Self-Improving LLM Agents via Long-Horizon Reflection

Built on a Qwen3.8-27B backbone trained on synthesized verifiable improvement trajectories
Scores: 81.8 on MLE-bench Lite, 92.2 on GAIA, 93.8 on DeepSearchQA
Accuracy keeps rising up to 30 reflection rounds, while GPT-4o plateaus after 5–10
Reflective-Self-Correction loop distills state updates to limit context contamination

#AREX-2#Qwen

AREX-2: Self-Improving LLM Agents via Long-Horizon Reflection

REAL-Q uses dynamic gradient descent to fix LLM quantization

Nokia open-sources AnyJev: a training-free layer that turns any open LLM into a calibrated decision model

PrismML hopes its tiny LLM could change how we all use AI

Intel squeezed a 1.58-bit LLM down to 1.485 bits without changing a single weight