AREX-2: BAAI's Model That Bakes the Agent Retry Loop Into Its Weights
BAAI's AREX-2, a Qwen3.8-27B finetune, trains try-check-revise loops into the model itself and beats GPT-5.6 Soul on MLE-Bench Light.
What is AREX-2?
AREX-2 is an open-weights language model built by the Beijing Academy of Artificial Intelligence (BAAI), along with two university partners, as a finetune of Qwen3.8-27B. It’s released under an Apache 2.0 license. What sets it apart isn’t raw scale (27B parameters is modest by current standards) but what it was trained to do: run its own try-check-revise loop across a long time budget, without external code telling it when to retry, what to keep, or when to stop. On MLE-Bench Light, an ML engineering benchmark, it scores 81.8, ahead of GPT-5.6 Soul’s 72.7. On Frontier CS, a set of open-ended algorithmic problems, it scores 70.7, trailing GPT-5.6 Soul by about six points but leading other open-weights models.
TL;DR
- AREX-2 moves retry logic from the harness into the model, so a single model call can try a solution, check it, revise it, and decide when to stop, rather than relying on external orchestration code.
- The architecture rests on two combined skills, reflection (judging whether a new attempt is actually better) and long-horizon execution (remembering past attempts and recovering from setbacks over many hours).
- Training data was built from real GitHub repositories turned into scored environments in machine learning engineering and algorithmic programming, chosen because success there is unambiguous (a score goes up or it doesn’t).
- Failed attempts were deliberately kept in the training trajectories but the loss was only computed on the steps that moved the task forward, teaching the model to recover from bad rounds instead of just mimicking clean successes.
- The method is supervised finetuning, not reinforcement learning, despite BAAI describing the approach with language that sounds like RL in its launch materials.
- On MLE-Bench Light, the score climbs steadily across a 5-hour budget (54 after one hour, 66 after two, 70.7 after five), while a comparison model, DeepSeek V3.2 Pro, flattens out around 45 after two hours and stops improving.
- An ablation study shows task-specific skills in context drive most of the gain, with the finetuning itself adding roughly 14 points on top of what skills alone deliver.
How does the training process actually work?
Most agent training data is built the simple way: give the model a task, let it attempt an answer, run a verifier, and keep the passing runs. That approach teaches a model what a correct answer looks like, but it throws away everything about how an answer got better, the failed tries, the feedback, the fixes along the way. BAAI’s pipeline was designed specifically to preserve that messy middle.
The process has three steps. First, a teacher model converts GitHub repositories into structured environments, each with a defined task and a scoring function. Second, a teacher agent works each environment for hours, sometimes hundreds of tool calls, reading code, searching papers, and running small experiments to see what actually improves the score. Third, the full trajectories are kept, not just the final successful ones, and failures are deliberately left in context.
The key training trick: the model sees the entire messy run, including the dead ends, but the loss is only computed on steps that represent genuine forward progress. That teaches the model to respond to its own failures rather than pretend they didn’t happen. BAAI calls this “reinforcing” in its launch post, which reads like reinforcement learning, but the transcript’s creator confirms it’s supervised finetuning on these curated trajectories, mixed in with deep research data carried over from an earlier AREX model.
Why focus on machine learning engineering and algorithmic programming?
BAAI picked these two domains for a practical reason: feedback is unambiguous. A validation score either rises or it doesn’t. A test suite either passes or it doesn’t. That removes the ambiguity that plagues domains like open-ended writing or general reasoning, where “better” is subjective and hard to score automatically.
Both domains also have long improvement runways. ML engineering tasks can run from a weak baseline all the way to a tuned ensemble. Algorithmic programming tasks can run from a brute-force solution to something close to optimal. That gives a long-horizon training signal room to actually matter, rather than converging in a handful of steps where a short-horizon model would perform just as well.
What do the benchmark numbers actually show?
The headline numbers: 81.8 on MLE-Bench Light (ahead of GPT-5.6 Soul’s 72.7) and 70.7 on Frontier CS (behind GPT-5.6 Soul by roughly six points, but ahead of other open-weights models). Given AREX-2 is a 27B parameter model and GPT-5.6 Soul is almost certainly far larger, the MLE-Bench Light result is a notable efficiency story even setting aside the long-horizon claims.
The more revealing chart tracks score against time budget. AREX-2 doesn’t lead at the start but climbs from 54 at one hour to 66 at two hours to 70.7 at five hours. A comparison model, DeepSeek V3.2 Pro, plateaus around 45 after two hours and doesn’t move meaningfully after that. That continued climb across a multi-hour budget is the core claim of the paper: give the model more time and it keeps finding real improvements, rather than stalling once the easy gains are gone. AREX-2 also outperforms earlier AREX models across four research benchmarks, including a 122-billion-parameter predecessor, reportedly because of this long-horizon training rather than added scale.
How much of the improvement comes from skills versus training?
BAAI published an ablation that’s worth taking seriously because it complicates the simple “training alone did this” narrative. The base Qwen3.8-27B model with no skills in context and no finetuning scores 28.8. Adding general machine learning practice skills pushes that to 41.2. Adding task-specific skills (generated by searching around the specific problem) along with more rounds pushes the untrained base model to 68.2. Only when you swap in the actual AREX-2 finetune, using the same skills and the same number of rounds, does the score reach 81.8.
In other words, skills in context account for most of the lift, moving the base model from 28.8 to 68.2. The finetuning adds roughly another 14 points on top. That’s a real, measurable contribution, but it’s smaller than the context-engineering lever. It lines up with something practitioners have noticed building their own agents: well-built skills and context often matter more than people expect, and model-level training is an additive refinement rather than the whole story.
Is this actually recursive self-improvement?
BAAI frames AREX-2 using the language of recursive self-improvement (RSI), but it’s worth being precise about what’s demonstrated. This is self-improvement at test time: within a single task, across a single budget of hours, the model gets progressively better at that one problem by trying, checking, and revising. The weights themselves never change during this process.
A useful way to think about it: it’s like giving an intern a week to attempt a problem twenty different ways, having them identify the best approach, and then wiping their memory before next week’s task so they repeat all twenty attempts again. True RSI, in the stronger sense researchers usually mean, would involve the model’s own successful trajectories feeding back into the next round of training, so future attempts need fewer tries to reach the same quality. To its credit, BAAI’s own paper reportedly states this feedback loop into training data is the intended next step, not something already achieved.
What does this mean for people building agent harnesses?
If more models get trained this way, hand-built harness logic doesn’t disappear, but its job shifts. Instead of managing when to retry, what to keep, and when to stop, a harness increasingly needs to focus on three things: supplying a genuine feedback signal (real scores, real logs, real error messages), supplying the operational knowledge the model needs (the skills layer that did most of the heavy lifting in the ablation), and setting a sensible time or round budget for the model to pace itself against.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
There’s a real risk of friction here too. A harness with its own hardcoded reflect-and-retry nodes, layered on top of a model that’s been trained to pace itself across one continuous long session, can end up with two competing loops. Resetting context every few steps, a common harness pattern, could actively hurt a model built for sustained, memory-preserving execution rather than help it. This tension is likely to become a bigger design question as infrastructure catches up: cloud providers are already building toward multi-day continuous agent runs, which only matters if the model underneath can actually use that time productively.
Frequently Asked Questions
What base model is AREX-2 built on?
AREX-2 is a finetune of Qwen3.8-27B, released as open weights under an Apache 2.0 license by BAAI and two university research partners in China.
Does AREX-2 use reinforcement learning?
No. Despite launch materials describing the training as “reinforcing,” the actual method is supervised finetuning on curated agent trajectories, combined with deep research data from an earlier AREX model.
How does AREX-2 compare to GPT-5.6 Soul?
AREX-2 scores higher on MLE-Bench Light (81.8 versus 72.7) but trails by about six points on Frontier CS algorithmic benchmarks, despite being a much smaller model at 27B parameters.
Is AREX-2 true recursive self-improvement?
Not yet in the strongest sense. It demonstrates self-improvement within a single task during one time budget, but its weights don’t update from that experience. The researchers reportedly describe feeding successful trajectories back into future training as the intended next step.
What matters more for performance, the finetuning or the skills in context?
Based on BAAI’s published ablation, skills provided in context account for most of the improvement over the untrained base model, while the finetuning itself adds roughly 14 additional points on top of that.
