9B Model as Strategist, DeepSeek V4 Writes Code: AIME Up 22 Points After Ten Rounds

While running recursive self-improvement (RSI) experiments, the HKUDS team hit a somewhat counterintuitive result: a 9B model can direct the much larger DeepSeek V4 better.

The division of labor is clean. The 9B model writes not a single line of code; it reads the task description and previous attempt records, then decides several things: which version to continue modifying, what algorithm or mechanism to try in the next round, how the coding agent should implement and verify it, and where to spend the remaining budget. DeepSeek V4 turns these strategies into runnable code and hands the execution results back.

Strategy → execution → evaluation → feedback → next-round strategy. This loop ran for ten rounds.

Testing covered two directions: AIME tests mathematical reasoning and problem solving, while Terminal-Bench 2.1 tests coding, debugging, tool use, and long-horizon execution.

After ten rounds, AIME went from 55.6% to 77.8%, and Terminal-Bench from 44.9% to 69.4%.

The control group had DeepSeek V4 serve as both strategy proposer and coder. In their setup, V4 still handled code writing; it just offloaded the algorithm-proposal work. Besides the scores, there are two details: to reach performance comparable to the baseline, it used 9% to 23% fewer tokens; and the model proposed not only bug fixes but also new mechanisms.

TJ's summary under the tweet captures it well: the 9B handles planning, V4 handles typing—small brain, big hands and feet.

jimerrors noticed another point: the key to this split is not capability boundaries but budget allocation.

A common assumption in RSI is that a model improves itself, and the larger the parameters, the better it improves. This experiment separates proposal from execution, which is tantamount to saying the speed of improvement is bottlenecked by strategy quality, not by the typing speed of writing code. Under this split, the 9B model actually benefits: it does not need to memorize all implementation details; it only needs to learn from failures what to try next.

Code is open on GitHub: https://github.com/HKUDS/AgentPilot

One interesting part is the rhetorical question: the best algorithm proposer may not need to be the largest model at all. It only needs to become a better strategist after each round of experiments.

If strategy quality is scarcer than coding ability, then the direction of scaling up parameters may need to be reordered. The converse also holds: how far can a model go if it can only write code and cannot propose strategies?