AREX-2: Moving the Improvement Loop Into the Weights — 27B vs. 2.8T, With All Failures Kept
How do today's coding agents get better? For most, the answer lives outside the model: the harness decides when to retry, what to keep, and what to discard, while the model only produces answers round after round. The improvement loop lives in the framework; the model itself learns nothing.
AREX-2 (arXiv 2609.38288, released September 29 by BAAI) reverses this: iterative improvement itself is trained into the model as a capability. The harness is demoted from boss to messenger. At test time, the model plans what to fix next round, where the previous round went wrong, and when to stop — given more rounds, the final result keeps improving, with no saturation point.
Results
AREX-2 uses Qwen3.8-27B as its base. On MLE-bench Lite it scores 81.8 — first among all systems compared in the paper, including closed-source flagships GPT-5.6 Sol (76.4) and Claude Opus 4.8 (74.5). On Frontier-CS it scores 70.7, first among open models, though GPT-5.6 Sol (76.4) still leads there. The scale gap is striking: open-source rivals it surpasses include Kimi-K3 (2.8T parameters) and DeepSeek-V4-Pro (1.6T). 27B versus 2.8T — a 100x-plus parameter difference.
Two Capabilities, One Product
The paper decomposes self-improvement into two capabilities, with total improvement approximately a product: r̄ × T*.
- r̄ is reflectiveness: how much score a valuable iteration earns on average.
- T* is long-horizon execution: after how many rounds the model starts spinning in circles.
- Producer: Beijing Academy of Artificial Intelligence (BAAI); corresponding author Zheng Liu, author of the BGE series.
- Code on GitHub at VectorSpaceLab/AREX-2; model weights on HuggingFace under the BAAI/arex-2 collection — genuinely open.
- Base model Qwen3.8-27B — the Qwen family continues to serve as the foundation of self-improvement/RL work.
- This v1 has no standalone Limitations section; the r̄ × T* product is an approximation, not a theorem; cross-domain gains are described only as "stable improvements" without per-benchmark breakdowns.
Strong single-round improvements stall after three rounds — total score won't climb; the ability to run fifty rounds at small per-round gains is equally insufficient. The paper states total improvement ≈ r̄·T*.
This product turns self-improvement from a slogan into two separately trainable, separately measurable quantities. Giving more test-time rounds yields monotonically rising performance — on Frontier-CS's 188 problems, the budget still helped at hour five. More strikingly, BrowseComp, a deep-research task with no correctness feedback, also improved with more rounds — meaning the model learned to climb even without a score.
Data Construction: Keep Every Failure
Training data comes from two domains: machine learning engineering and algorithmic programming — chosen, the paper says plainly, because they offer executable verification feedback: whether an improvement works can be checked by running it. Where verification is cheap, supervision signal fits.
Construction has three steps:
1. Teacher models convert GitHub repositories and online-judge problems into executable environments, each a pair of "task + scoring function"; environments failing execution and validation checks are discarded. 2. Agents run many rounds in each environment, revising approaches based on feedback. 3. Full-trajectory filtering: trajectories with high final scores and compliant behavior are kept in their entirety, without requiring any single round to succeed.
Step 3 is the paper's most counterintuitive and valuable choice. Conventionally, you'd trim failed steps and keep only clean wins. AREX-2 refuses: full trajectories keep failed runs, performance regressions, and abandoned lines — deliberately, per the paper. Cut these out and the model only learns from "every step succeeds" samples, leaving nothing to learn from when it actually hits setbacks. Teaching the model to keep iterating after hitting a wall relies precisely on the wall-hitting samples.
The Precise Claim on Cross-Domain Transfer
The paper's most widely circulated claim: capabilities trained on ML and algorithmic programming directly transferred to deep research. The direction is right, but the precise scope matters.
AREX-2's training data, beyond the newly synthesized two-domain trajectories, reuses the deep-research data of the original AREX recipe (Lu et al., 2026). So "zero new data" means no new deep-research data was created — not that deep research was absent from training. The paper's actual control is cleaner than "direct transfer": the original AREX model also consumed deep-research data, and AREX-2, adding only long-horizon reflection trajectories in two new domains, improved across all four research benchmarks — BrowseComp 84.0, HLE 52.6, GAIA 92.2, DeepSearchQA 93.8 — all surpassing the original. With not a single research task in the new training trajectories, the gains can only come from the reflection meta-skill itself. This control is the real evidence for the meta-skill hypothesis.
Context and Caveats
In the lineage of self-improvement (RSI) work, AREX-2 occupies a unique position: previous "R" occurred either during training or at the harness layer; AREX-2 trains the loop itself into the model's test-time behavior — experience, for the first time, enters the weights as the process structure of recovering from setbacks. That "keep all failures" data construction effectively burns a negative-results archive into parameters.
Additional facts worth noting: