AREX-2: The Harness Moves Into the Model

🔁 The loop you wrote by hand 0:00

If you have built an agent in the last year, you have written the same code: try something, check the result, decide whether to retry, keep the best version so far, and work out when to stop. All of that logic lives in the harness. The model gives you one attempt at a time and your code does everything else.

AREX-2 asks the obvious question: what if the model did that part itself? Give it a budget — say five hours on one problem — and let it work out how to spend those hours making its answer better.

📦 What AREX-2 is 0:52

Being a 27B fine-tune is not a footnote, it is the point of the comparison: every number below is achieved by a model roughly an order of magnitude smaller than the closed systems it is measured against.

🏛 Who built it 1:38

The paper's affiliation is the Beijing Academy of Artificial Intelligence (BAAI) — a twelve-author team led by Hongjin Qian, with Zhicheng Dou among the senior authors. The video notes it was built alongside two other Chinese institutions, and that apart from the paper the main write-up is a blog post available only in Chinese, which is a fair explanation for why an English-language audience has been slow to notice it.

🎯 Two capabilities, and a diagnostic 2:24

The paper's definition of self-improvement at test time rests on exactly two things, and the framing is the most useful idea in the video because it applies to agents you already run.

Multiply them and the failure modes separate cleanly: if reflection is weak, more rounds just burn tokens; if the horizon is short, even sharp reflection plateaus early.

Budget awareness is the third ingredient: the model is told how many rounds it has, and paces itself — early rounds to measure, later rounds to act on what was measured.

🗃 Why agent data hides improvement 4:58

The paper's diagnosis of the training-data problem is blunt: the standard recipe poses a task, has the model attempt it, has a verifier mark pass or fail, and keeps the passes. That teaches a model what a correct answer looks like — not how an answer gets improved. The intermediate attempts, the feedback, the revisions and the abandoned approaches are all discarded.

Two domains were chosen because their feedback is unambiguous and their headroom is real: machine-learning engineering (a validation score goes up or it does not) and algorithmic programming (tests pass or they do not), both with a long path from a baseline to a tuned ensemble or from brute force to near-optimal.

🏗 The three-step pipeline 5:41

1. A teacher model builds environments. It reads a GitHub repository or an online-judge problem and turns it into a task plus a scoring function — a repository has code and a metric but no stated task; a judge problem has a statement and tests but no sandbox. The teacher supplies the missing half.
2. An agent works the task for hours. Hundreds of tool calls per task: reading code, searching papers, running small experiments, seeing what actually improves the score.
3. Trajectories are selected as a whole. A trajectory is kept if its final score is high and its process satisfies the requirements — no individual round is required to succeed.

Then the subtle part, and the bit worth copying if you fine-tune on your own agent traces:

⚙️ It is fine-tuning, not RL 6:54

The launch material described the tasks as being "reinforced", which sounds like reinforcement learning. The paper is explicit: AREX-2 is supervised fine-tuning on the selected trajectories, mixed with the deep-research data from the earlier AREX recipe — which is left unchanged, so any gain on research benchmarks is transfer, not new training data.

📊 The numbers 7:08

Two benchmarks, and the pattern across them is different — which the video is careful to point out.

• MLE-bench Lite: 81.8 — the highest score in the paper's comparison, above GPT-5.6 Sol's 72.7, from a model a fraction of the size.
• Frontier-CS: 70.7 — the best open-weights result, about six points behind GPT-5.6 Sol (76.4). Open-ended algorithmic problems remain harder to win outright.
• Transfer without new data: BrowseComp 84.0, HLE 52.6, GAIA 92.2, DeepSearchQA 93.8 — the deep-research scores improve over the previous recipe even though no research trajectories were added.

📈 Five hours, two flat lines 8:10

The most important chart in the paper is not a leaderboard — it is score against wall-clock budget on the 188 problems of the Frontier-CS Agent Track, five hours each.

AREX-2 does not start in front; it takes the lead as the budget grows, and the later hours stay productive — 4.8 points between the second and fifth hour, of which 2.2 arrive in the final hour. DeepSeek-V4-Flash flattens even earlier, at 39.1 after three hours.

🔬 The ablation that reframes the headline 8:35

AREX-2's MLE-bench Lite score was measured with skills in context — general ML-practice skills plus task-specific ones assembled by searching around the problem. The paper publishes the stage-wise ablation, and it changes how you should read the 81.8.

So: skills and rounds take an untrained base model from 28.8 to 68.2, and the fine-tune adds roughly fourteen points on top of that. Both levers are real; the free one is bigger.

🧰 What it means for your harness 10:23

If models start carrying the loop, do harnesses go away? No — their job changes. The video's version of the new division of labour:

That is the practical risk, and it is testable: a harness that resets context every few steps and re-decides when to retry may actively damage a model trained to hold the thread across hours. His own test bench included exactly this condition — wiping history every three rounds and passing only a summary forward.

🌀 Is this recursive self-improvement? 11:30

BAAI frames AREX-2 as self-improving, and the video makes the necessary distinction: what was demonstrated is self-improvement at test time. Inside one task the model gets better over many rounds — but the weights never change.

True recursive self-improvement, in this framing, is when the model's own trajectories become the next round of training data — so the next generation needs fewer attempts, not the same 20. To their credit, that is exactly what the authors propose as future work: they list it as the path forward rather than repeating the many steps every time.

🧪 His own test bench 12:32

Rather than eyeballing chats, he built a small bench ("Round Lab") that runs AREX-2 and base Qwen3.8-27B through the identical loop: same prompt, same budget, same settings. Each round the model submits a full solution, it runs in a sandbox, and the model gets back its score, the logs and how many rounds remain.

The result was modest and reported as such: AREX-2 does slightly better, especially on harder problems, but the difference is not huge — plausibly because he used his own harness rather than BAAI's, and because TSP is unlike the training domains.

✅ Claims checked

Basis note. Built from the full 16:27 transcript, then verified against the paper's full text (arXiv HTML), the Hugging Face model card and the GitHub repository. Every benchmark figure in this article is the paper's, including two the video rounds: the 1-hour and 2-hour points are 54.4 and 65.9, and the ablation's intermediate stage of 75.8 (same skills, base rounds) is not mentioned in the video. The paper is an arXiv preprint, not a peer-reviewed publication.

🔑 Key takeaways

• The loop is moving into the model. Try, check, revise used to be harness code; AREX-2 learns to do it inside one long session, given a budget.
• Two capabilities, two failure modes. Reflection decides whether the next round is worth more; long-horizon execution decides whether round ten is still useful. Diagnose your stalls accordingly.
• Ask the diagnostic question before buying anything: bad changes mean you need a better feedback signal; repeated work means you need state and memory. Different problems, different fixes.
• Agent training data usually discards the useful part. Keeping only passes teaches what a correct answer looks like, not how an answer improves.
• Select by outcome, mask by progress. Keep failures in context, apply loss only to steps that diagnose, repair, change strategy or submit — the reusable recipe from the paper.
• It is SFT, not RL. The result came from how trajectories were selected and masked, which is within reach of a small team with a good log store.
• MLE-bench Lite 81.8 beats GPT-5.6 Sol's 72.7 — but read the ablation first: skills and extra rounds take the untrained base model from 28.8 to 68.2, and training adds ~14 points on top.
• Most of the gain is free: operational knowledge in context is the biggest single lever, and it needs no training run at all.
• The real differentiator is the tail of the curve: at a five-hour budget, AREX-2 keeps climbing while DeepSeek-V4-Pro stops at 44.7 after two hours.
• Test-time improvement is not recursive self-improvement. The weights never change; doing twenty attempts again tomorrow is the intern whose memory is wiped weekly.
• Harnesses do not disappear — they change jobs: less loop control, more feedback signal, operational skills, and a declared budget the model can plan against.
• Watch for two loops fighting: a harness that resets context and re-decides retries can damage a model trained to hold the thread for hours.

🔗 Resources

• The source video — Sam Witteveen, "AREX-2 - The Harness in the Model", 16:27, 6 November 2026
• AREX-2 paper — "Advancing Self-Improving Agents through Long-Horizon Reflective Tasks", BAAI, 29 September 2026
• BAAI/AREX-2 on Hugging Face — the model weights, Apache-2.0
• VectorSpaceLab/AREX-2 — code and inference (Apache-2.0)
• Full text (HTML) — including the ablation and round-scaling sections the video summarises

⏱ Timestamp index