DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory
Xincheng Wei, Yifan Ding, Yoshua Li, Dongsheng Ma, Rongxiang Weng, Xunliang Cai, Wenjian Ding, Yao Zhang
cs.AI
2026-09-01
DiagEvo turns solver failures into the next self-play curriculum. Qwen3-8B reaches 72.3% math average, 4.5 points above R-Zero, and beats DARC overall without external data.
Self-play lets a challenger write questions and a solver train on them, round after round, with no human labels. Unguided runs often grow longer questions while the solver stalls. Difficulty, learnability, and diversity keep items hard and varied; they do not say which recurring mistake the next round should hit. Guided methods borrow that direction from external banks, document corpora, or difficulty tags, so the curriculum sits outside the loop.
DiagEvo treats the solver's own failures as the missing signal. The LongCat team at Meituan extracts transferable error causes, stores them in a hierarchical memory, and uses that memory to steer the next round of question writing.
Each round has four steps. The challenger mixes free exploration with cause-targeted generation aimed at Active error causes, and can stitch an Active cause with a Mastered sibling under the same skill node. More Active failures raise the targeted share. Once a cause is marked Mastered, its count resets and probability flows back to free exploration.
The solver then samples a construction set. Double-confidence filtering keeps questions of intermediate self-consistency whose leading answer clearly outvotes the runner-up; the default ratio is τ=1.6. Retained items get a second GRPO sample for optimization. Failed trajectories go to a diagnostician: one response that disagrees with the pseudo-label is compared with one that agrees, and the earliest reasoning fork becomes a transferable cause.
Memory has two levels. Causes carry an Active/Mastered flag and a recurrence count for the current episode. Skill nodes group related causes. The default diagnostician is a frozen Qwen3-4B-Instruct-2507, used for extraction, deduplication, and assignment. The solver never sees the memory, only questions and pseudo-labels.
With the default 4B diagnostician, DiagEvo leads every baseline on the nine-benchmark mean for all three solvers. On Qwen3-8B, the five-task math average is 72.3%, 4.5 points above label-free R-Zero and 1.2 above DARC, which uses external resources. The four-task general average is 38.8%, up 2.6 and 1.0 points. The nine-task overall mean is 57.4%, 1.1 above DARC. Qwen3-4B and OctoThinker-8B reach 53.5% and 41.9% overall, each 1.3 above DARC.
Scaling the diagnostician from 4B to 235B-A22B adds about 1.0 to 1.2 math points and almost nothing on general reasoning. The 4B default already clears the field.
Ablations on Qwen3-8B are sharp. Freezing the challenger drops math to 68.5, 3.8 below the full 72.3, the largest cut. Pure free exploration scores 69.5; pure targeted generation scores 70.1. Dropping the relative-confidence constraint leaves 70.9; no filtering leaves 69.4. Same-skill stitching beats random pairing by 1.0 math point.
The run is not monotonic. Math rises for five rounds to 72.3, then slips to 72.0 and 71.4, so the paper reports the round-5 checkpoint. Unguided exploration peaks at 69.5 in round 3. Memory grows from 151 to 244 causes, with skill nodes settling near 36. Oracle agreement on pseudo-labels falls from 83% to 65% with only the absolute constraint; memory lifts round 5 to 72%; the relative constraint reaches 75%.
Self-play often needs a better target, not a harder surface form. DiagEvo turns failure history into that target without an external corpus, and still beats DARC on the overall mean. A 4B diagnostician is enough, so the cost is not in diagnostician scale.
The solver trains only on math questions. General-reasoning gains are transfer, not a second curriculum.
Both filters score agreement among solver samples, not ground truth. A shared error with a strong majority still gets through. The round count is fixed in advance; performance falls after round 5, and there is no stop rule. The curriculum is built from math. Gains on general suites do not show that the same loop would work if diagnosis started in those domains. Some baseline numbers are taken from DARC rather than fully reproduced.