Micro1 says expert-written STEM problems improved a 748B model’s reasoning scores

Micro1 says post-training a 748-billion-parameter model on 1,600 expert-written STEM problems improved its scores across several reasoning tests. The results suggest carefully designed training data can matter even when the model is already large, though this is the company’s own study, not an independent replication.

Micro1 Watch analysis

What happened

In a research post published on 27 October, Micro1 describes post-training DeepSeek-V4.1-Flash with a rank-32 LoRA reinforcement-learning setup for 25 steps. Its training set contained 1,600 distinct STEM problems written and reviewed by PhD experts, with reference answers and grading criteria. Micro1 calls its approach an adaptive curriculum: it shifts training effort between problems the model usually solves, problems it sometimes solves and harder problems.

Micro1 reports that the model’s score on Humanity’s Last Exam rose from 36.40% to 39.84%. It also reports gains on FrontierScience’s Olympiad and Research splits, and a rise from 80% to 85% on the AA-LCR v1.1 benchmark. The company says 11 of 44 held-out STEM questions that had received no accepted answer in eight baseline attempts received at least one after post-training.

Key findings

• Humanity’s Last Exam: +3.44 percentage points
  Micro1 reports a rise from 36.40% to 39.84% across 2,500 questions, with one answer per question.
• Different STEM tests also improved
  Reported gains were 1.95 points on FrontierScience Olympiad and 2.28 on its Research split.
• Some previously missed questions became solvable
  Of 44 held-out questions with no accepted baseline answer across eight attempts, 11 got at least one accepted answer across eight post-training attempts.
• The training set stayed fixed
  The adaptive system changed how it allocated generation effort among the 1,600 problems, rather than adding new problems during the run.

Why it matters

The practical point is not simply that a bigger model scored better. Micro1’s experiment centres on the quality and selection of training problems: expert-reviewed questions, dependable grading and an adaptive allocation of practice. If the approach generalises, model developers may be able to get more from a relatively small, carefully prepared dataset than from indiscriminate extra training.

There are important limits to the comparison. Micro1 reports that the STEMQA evaluation used different execution paths for the base model and trained checkpoints, making that result descriptive rather than a clean like-for-like test. The study also combines several elements, including its curriculum and reward shaping, so it does not isolate the effect of any single one. The score gains are promising evidence for the overall recipe, not a verdict on which ingredient did the heavy lifting.

Our read

This is a useful argument for treating training data as an engineered resource, not just a pile of examples. The strongest takeaway is the measured improvement across several tests, alongside a method that tries to spend training effort where it can still help. The next question is whether outside researchers can reproduce the gains, and whether the recipe holds up beyond this model and these benchmarks.

What to watch

• Whether Micro1 releases the full experimental materials and enough detail for independent replication.
• Whether further tests use matched execution settings across base and post-trained models.
• Whether the same approach improves other models or tasks without relying on the same benchmarks.

Discussion spark: Would you put more value on a small, expert-reviewed training set with adaptive practice, or on scaling up the amount of training data?

not affiliated with or endorsed by Micro1