acceptodds
Under review as a conference paper at ICLR 2027

Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution

Abstract

Skills, reusable procedural guidance added at inference, can substantially improve LLM downstream performance (Li et al., 2026). Prior work retrieves skills from a bank by semantic relevance, then uses them as inference-time patches or for model distillation. The individual utility of each skill, however, is largely neglected. We first show that, in on-policy distillation where skill-conditioned policies serve as teachers, fewer than % of retrieved skills provide useful distillation signals. We then propose **SGUID**, a method for selecting a compact subset of skills for distillation. SGUID retains a skill only if it consistently yields effective learning signals during training. The selected skills are then distilled to produce a better model. Our results show that *not all* skills are worth distilling. Across four models from the Olmo and Qwen families, distilling 6 selected skills matches or exceeds full-bank distillation in mean avg@12 on three of the four models, and on all four after a second round that distills 3 newly selected skills, while the full banks are up to larger. Importantly, SGUID supports stable model-skill co-evolution: after a distillation round, a new candidate bank is curated from the updated model's rollouts, and SGUID selects which skills to internalize next. In the second round, this loop selects 3 new skills and improves Qwen3-8B from % to %. The selection step is essential for stability: on Qwen3-4B, naively updating the model with unfiltered skills degrades performance, including a percentage point drop on HMMT25, whereas SGUID improves HMMT25 by points after the first round and points after the second. These results identify skill selection as the key mechanism for stable model-skill co-evolution.

Primary area:
foundation or frontier models, including LLMs
Keywords:
model-skill co-evolution, self-distillation, reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_9JMkOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies skill-conditioned on-policy self-distillation (building on SGSD), where retrieved natural-language "skills" serve as privileged context for the teacher. It observes that only a small fraction of skill retrievals pass SGSD's confidence gate, and proposes SGUID, which runs one full-bank distillation, counts each skill's verifier-aligned positive signals in early/late phases, filters skills by threshold rules (Eq. 11), and re-distills from the same initialization using only the top-K (K=6, then K=3) skills. Repeating this curate–select–distill loop is termed "model–skill co-evolution". Experiments use Olmo-3-7B and Qwen3-1.7B/4B/8B trained on DAPO-Math-17K with LoRA, evaluated on AIME24/25 and HMMT25 (avg@12).

Critical strengths

The observation that most skill retrievals provide no active distillation signal (Fig. 2) is a useful diagnostic for skill-conditioned distillation, and selecting a small bank (6 vs. 30–71 skills) can match full-bank SGSD at lower teacher cost.

Critical weaknesses
  1. The empirical evidence does not reliably support the claims. (a) Checkpoints are selected by average performance on the test benchmarks themselves (Table 1 caption: "best checkpoint according to the average performance on AIME24, AIME25, and HMMT25 across 200 training steps"); there is no validation set. (b) All results are single runs without variance, on benchmarks of 30 problems each. (c) Under these conditions, the gains are within noise: R1 vs. SGSD differs by −0.7 to +0.9 mean points, and on Qwen3-8B SGUID-R1 (64.3) is below answer-only OPSD (65.3). (d) The headline 8B improvement appears only in Round 2, which uses roughly 4× the training budget of the 200-step baselines (a selection run plus a distillation run per round). Compute-matched baselines (OPSD/GRPO/SGSD trained longer) are not reported, and the "w/o selection" co-evolution comparison (Fig. 3) is given only for Qwen3-4B. (e) In Table 4, top-6 (42.3) is close to random-6 (41.2±0.9) and worst-6 (41.5), and all three are below full-bank SGSD (43.0). This weakens the claim that the selection criterion identifies the right skills.

  2. Limited novelty. SGUID is a threshold-based filtering heuristic (with hand-set h, τ and early/late split) applied on top of SGSD's existing Verify·Gate coefficient, followed by a restart. "Co-evolution" amounts to iterating this pipeline twice; skills are regenerated rather than evolved, and there is no analysis beyond two rounds. The "skills" themselves (Appendix B) are generic math hints such as "Simplify Expressions" and "Validate Solutions", so the framing (skill bank, internalization, co-evolution) overstates what is essentially selecting prompt hints for a privileged teacher.

Questions for authors

Clarifications:

  1. Can you report results with checkpoints selected on a held-out validation set, and with multiple seeds / confidence intervals?
  2. What are the results of OPSD, GRPO and SGSD trained with the same total number of steps/compute as SGUID-R2, and of full-bank (no selection) R2 on all four models?
  3. Table 4 states that the selection baselines were trained for 100 steps, yet the top-6 row matches the 200-step R1 result in Table 1. Were the variants trained with the same budget?
  4. The abstract says "fewer than 25% of retrieved skills provide useful signals", but Fig. 2 counts retrieval events passing the gate, not distinct skills. How many distinct skills are ever active, and how sensitive is this fraction to κ? Would materially affect my assessment:
  5. The paper's own data suggest that skill benefits shrink with scale (6.4% active retrievals on 8B; SGSD below OPSD on 8B). Does SGUID still help on larger models (e.g., 14B/32B) or with full fine-tuning rather than LoRA?
  6. Section 2.3 states that skill-guided teachers provide better signals than answer-guided ones, but Table 1 shows OPSD ≥ SGSD on Qwen3-4B/8B. How do you reconcile this?
6 comments
AuthorsOpenReview ↗

Author Rebuttal (Part 1/5)

We are delighted that the reviewer finds the absence of active distillation signal in most skill retrievals is a useful diagnostic for skill-conditioned distillation, and that a small bank of 6 instead of 30–71 skills can match full-bank SGSD at lower teacher cost. We also value the reviewer's specific concerns about the reliability of our empirical evidence and the scope of our claims, which led to the following new results: final-checkpoint results with no checkpoint selection, compute-matched comparisons with significance tests, co-evolution ablations on Qwen3-8B, additional skill-selection ablations, and sensitive analysis over threshold κ. We hope the evidence below, including our answer to Q4 in Part 4, resolves the reviewer's main concerns.

1(a). Checkpoints selected on the test benchmarks (Q1)

Checkpoints are selected by average performance on the test benchmarks themselves; there is no validation set.

For comparability, we follow the evaluation protocol of [1] and [2]: checkpoints every 25 steps (step 0 excluded), selected by the mean over AIME24, AIME25, and HMMT25, with the same rule for all methods. Baselines are selected over 400 steps and each SGUID stage over 200, so SGUID gets no extra selection opportunities.

We additionally report the final checkpoint after 400 steps (mean avg@12, three seeds):

ModelGRPOOPSDSGSDSGUID R1
Qwen3-1.7B33.15±0.1633.15 \pm 0.1640.77±1.5840.77 \pm 1.5841.30±1.4941.30 \pm 1.4942.66±0.59\mathbf{42.66 \pm 0.59}
Qwen3-4B61.29±1.2761.29 \pm 1.2758.49±1.6358.49 \pm 1.6362.78±0.7262.78 \pm 0.7263.61±0.28\mathbf{63.61 \pm 0.28}
Qwen3-8B63.12±2.0663.12 \pm 2.0664.20±0.7264.20 \pm 0.7263.67±0.7563.67 \pm 0.7564.32±0.44\mathbf{64.32 \pm 0.44}

SGUID R1 has the best mean on all three models, including over OPSD on Qwen3-8B, and leads SGSD by +1.36, +0.83, and +0.65. We will add this table in the revised paper.

References

[1] Zhao, Siyan, et al. "Self-distilled reasoner: On-policy self-distillation for large language models." arXiv preprint arXiv:2601.18734 (2026).

[2] Huang, Jiazhen, et al. "Skill-Conditioned Gated Self-Distillation for LLM Reasoning." arXiv preprint arXiv:2605.28791 (2026).

AuthorsOpenReview ↗

Author Rebuttal (Part 2/5)

1(b, c). Compute-matched comparisons with variances reported

All results are single runs without variance ... the gains are within noise.

We retrained GRPO, OPSD, and SGSD for 400 steps over three seeds, matching SGUID R1's budget (200 selection + 200 distillation steps):

ModelMethodAIME24AIME25HMMT25Mean
Qwen3-1.7BGRPO50.93±1.5850.93 \pm 1.5836.30±1.1636.30 \pm 1.1622.41±1.1222.41 \pm 1.1236.54±1.1936.54 \pm 1.19
Qwen3-1.7BOPSD55.09±0.8555.09 \pm 0.8542.59±1.1342.59 \pm 1.1328.33±1.47‾\underline{28.33 \pm 1.47}42.01±0.3242.01 \pm 0.32
Qwen3-1.7BSGSD56.39±1.21‾\underline{56.39 \pm 1.21}44.54±0.42‾\underline{44.54 \pm 0.42}26.30±1.3126.30 \pm 1.3142.41±0.3442.41 \pm 0.34
Qwen3-1.7BSGUID R155.74±0.8955.74 \pm 0.8944.07±0.5844.07 \pm 0.5827.96±1.5427.96 \pm 1.5442.59±0.90‾\underline{42.59 \pm 0.90}
Qwen3-1.7BSGUID R256.94±2.82\mathbf{56.94 \pm 2.82}46.67±1.27\mathbf{46.67 \pm 1.27}28.89±1.11\mathbf{28.89 \pm 1.11}44.17±0.73∗∗\mathbf{44.17 \pm 0.73}^{\ast\ast}
Qwen3-8BGRPO78.15±0.32‾\underline{78.15 \pm 0.32}68.89±0.5668.89 \pm 0.5644.81±1.1644.81 \pm 1.1663.95±0.4763.95 \pm 0.47
Qwen3-8BOPSD78.89±0.28\mathbf{78.89 \pm 0.28}70.28±1.11‾\underline{70.28 \pm 1.11}46.67±1.39‾\underline{46.67 \pm 1.39}65.28±0.33‾\underline{65.28 \pm 0.33}
Qwen3-8BSGSD77.04±0.5877.04 \pm 0.5869.81±1.1669.81 \pm 1.1646.20±1.4346.20 \pm 1.4364.35±0.2564.35 \pm 0.25
Qwen3-8BSGUID R176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
Qwen3-8BSGUID R277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗\mathbf{72.31 \pm 1.43}^{\ast\ast\ast}48.06±2.40\mathbf{48.06 \pm 2.40}65.93±0.56∗\mathbf{65.93 \pm 0.56}^{\ast}

n = 3 seeds; ± is the standard deviation. Bold / underline: best / second-best. Stars: SGUID R2 vs. SGUID R1 within the same model, one-sided paired t-test (each R2 run continues from the matching R1 run; df = 2; ∗p≤0.05^{\ast}p \le 0.05, ∗∗p≤0.01^{\ast\ast}p \le 0.01, ∗∗∗p≤0.001^{\ast\ast\ast}p \le 0.001).

  • We find the conclusions are consistent with Table 1. R1 matches SGSD with a compact bank (42.59 vs. 42.41 on Qwen3-1.7B; 64.14 vs. 64.35 on Qwen3-8B) while distilling 6 instead of 34 and 71 skills.
  • R2 gains are significant: R2 improves over R1 (p = 0.0018 and 0.017) and over SGSD (Welch p = 0.018 and 0.012) on Qwen3-1.7B and Qwen3-8B.

We will add the revised paper results for all four models.

1(d). Round-2 budget and compute-matched baselines (Q2)

The headline 8B improvement appears only in Round 2, which uses roughly 4× the training budget of the 200-step baselines.

R1 + R2 totals 800 steps, four times the 200-step baselines in Table 1. However:

  • R1 is step-matched (400 steps, table above) and matches SGSD; we do not present R2 as a compute-matched comparison.
  • The baselines have plateaued: going from 200 steps (Table 1) to 400 steps changes their means by at most 0.6 points (e.g., Qwen3-8B OPSD 65.3 → 65.28).
  • R1 + R2 costs about the same as 800-step SGSD in GPU-hours (22.96 vs. ≈ 23.2 on Qwen3-4B), since compact-bank distillation needs fewer teacher passes.

We will add 800-step baselines on all four models and the GPU-hour breakdown (Appendix A) in the revision. The requested full-bank R2 is in Part 3.

AuthorsOpenReview ↗

Author Rebuttal (Part 3/5)

Co-evolution ablations on Qwen3-8B (1d, Q2)

The "w/o selection" co-evolution comparison (Fig. 3) is given only for Qwen3-4B.

We ran it on Qwen3-8B over three seeds, plus an ablation that continues to reuse the R1 bank for another round:

Qwen3-8BRoundAIME24AIME25HMMT25Mean
Base (Table 1)–77.5077.5067.2267.2243.8943.8962.8762.87
w/o selection (full bank)R176.20±0.8976.20 \pm 0.8970.83±1.2770.83 \pm 1.2747.04±0.5847.04 \pm 0.5864.69±0.3864.69 \pm 0.38
w/o selection (full bank)R277.41±0.5877.41 \pm 0.5869.91±0.3269.91 \pm 0.3246.39±1.4746.39 \pm 1.4764.57±0.6864.57 \pm 0.68
w/ R1 bankR276.85±0.5876.85 \pm 0.5870.37±1.1270.37 \pm 1.1245.74±1.7645.74 \pm 1.7664.32±0.1964.32 \pm 0.19
w/ selection (SGUID)R176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
w/ selection (SGUID)R277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗72.31 \pm 1.43^{\ast\ast\ast}48.06±2.4048.06 \pm 2.4065.93±0.56∗\mathbf{65.93 \pm 0.56}^{\ast}

Stars: SGUID R2 vs. SGUID R1 (same test as above).

All second rounds use the same budget. Consistent with our claims in the paper, re-curating and re-selecting skills is crucial: full-bank R2 (64.57) and R1-bank R2 (64.32) give no gain, while SGUID R2 beats both (Welch p = 0.029 and 0.014), so the 8B gain is not explained by extra budget. We will add these results to Figure 3.

1(e). Clarifications on Table 4, and evidence that the criterion finds the right skills (Q3)

Top-6 (42.3) is close to random-6 (41.2 ± 0.9) and worst-6 (41.5), and all three are below full-bank SGSD (43.0). ... Were the variants trained with the same budget?

Budget. All K = 6 variants are trained for 100 steps; top-6 equals Table 1's R1 because that run peaks at step 100. The SGSD reference of 43.0 is its best over 200 steps to show an upper bound. Within the same 100 steps, SGSD's performance is 41.57:

Qwen3-1.7B, 100 stepsAIME24AIME25HMMT25Mean
Base50.650.635.635.623.123.136.436.4
SGSD (full bank, 34 skills)56.3956.3941.1141.1127.2227.2241.5741.57
top-6 (ours)55.355.343.1\mathbf{43.1}28.6\mathbf{28.6}42.3\mathbf{42.3}
random-6 (3 selection seeds)54.1±0.854.1 \pm 0.842.6±1.242.6 \pm 1.226.9±0.926.9 \pm 0.941.2±0.941.2 \pm 0.9
worst-655.6\mathbf{55.6}42.842.826.126.141.541.5

Bold: best among the three K = 6 variants.

Criterion. As the margins on Qwen3-1.7B are small, we additionally compare Qwen3-4B, which further supports the effectiveness of our skill selection. The reviewer might be interested to see additional frequency-6 (the most-retrieved skills) and fast-decay-6 (the fastest-decaying skills) baselines we added. All rows use 100 steps:

Qwen3-4B, 100 stepsAIME24AIME25HMMT25Mean
Base73.973.965.865.846.746.762.162.1
SGSD (full bank, 30 skills)75.075.069.72\mathbf{69.72}45.8345.8363.5263.52
top-6 (ours)77.2\mathbf{77.2}67.267.248.1\mathbf{48.1}64.2\mathbf{64.2}
random-6 (3 selection seeds)74.9±0.774.9 \pm 0.768.5±0.768.5 \pm 0.745.3±0.645.3 \pm 0.662.9±0.262.9 \pm 0.2
frequency-674.774.766.966.945.345.362.362.3
worst-674.474.467.867.844.444.462.262.2
fast-decay-674.774.768.368.343.643.662.262.2

Top-6 is the only 6-skill bank that outperforms the full bank (64.2 vs. 63.52); all alternatives are 1.3–2.0 points lower, and frequency-6 is no better than worst-6. With the persistence ablation in Table 3 (65.1 → 62.3), this supports Reviewer MgrQ's observation that these controls "show that the choice of skills matters, not just the number of skills." We will add this table to Section 5 and the 100-step SGSD row to Table 4.

Reviewer_MgrQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SGUID improves language models by identifying which reusable skills are most useful to learn from and incorporating them directly into the model through self-distillation. It creates skills from past problem-solving experiences, selects those that consistently provide useful feedback during training, and repeats the process as the model improves. Across mathematical reasoning tasks, the authors show that a small set of carefully selected skills can perform as well as or better than much larger skill collections.

Critical strengths

SGUID tackles an interesting problem: figuring out which skills are actually useful for distillation. Instead of selecting skills based only on how relevant they are to a problem, the method looks at whether they continue to provide useful training signals as the model improves. The results are encouraging, as using just six selected skills gives similar or better performance than the full skill bank on three of the four models, despite the much smaller bank.

The paper also includes some helpful ablations to understand why the method works. For example, removing the persistence constraint drops Qwen3-4B's accuracy from 65.1% to 62.3%, suggesting that keeping skills that remain useful throughout training is important. The comparisons with random and bottom-ranked skills are useful as well, since they help show that the choice of skills matters, not just the number of skills being used.

Critical weaknesses

One concern is that it is hard to tell how reliable the reported improvements are. The authors choose the best checkpoint based on performance on the same three benchmarks used for the final evaluation. This could make the results look better than they would on unseen data. The paper also does not report variation across training runs, even though some of the gains over SGSD are less than one percentage point. Running the main experiments with multiple seeds would make the results more convincing.

I also think the comparison with SGSD could be fairer. SGSD uses eight teachers per prompt, while SGUID uses six in the first round. It would be useful to compare SGUID against SGSD using the same number of teachers, to see whether the improvement really comes from selecting better skills. Another issue is the training cost. SGUID first runs 200 training steps to select skills, then starts training again with the selected bank. Since this adds a substantial amount of computation, I would like to see a comparison where both methods have the same overall training budget.

Finally, I am not fully convinced by the co-evolution results. The paper only evaluates two rounds, and the direct comparison with unfiltered co-evolution is limited to Qwen3-4B, where the improvement between rounds is quite small. I think an important missing ablation is to run the second round without generating a new skill bank, using the original skills instead. This would help show whether the improvement comes from creating new skills or simply from doing another round of distillation. As it stands, it is difficult to tell how much the co-evolution step itself is helping.

Questions for authors

.

5 comments
AuthorsOpenReview ↗

Author Rebuttal (Part 1/3)

We are glad the reviewer found our research question, which skills are actually useful for distillation, interesting, and our results encouraging: six selected skills perform similar or better than the full skill bank on three of the four models. We appreciate that the reviewer pointed to the persistence ablation (65.1 → 62.3 on Qwen3-4B) and the other baselines as evidence that the choice of skills matters.

The review raises three questions: are the gains robust to checkpoint selection and seed variance, is the comparison with SGSD fair in teacher count and training budget, and does the co-evolution step itself contribute? Taking these to heart, we ran new experiments for each: compute-matched comparisons with reported variances, co-evolution ablation on Qwen3-8B, and a fair SGSD baseline with 6 teachers. The details follow.

1. Checkpoint selection on the evaluation benchmarks

The authors choose the best checkpoint based on performance on the same three benchmarks used for the final evaluation.

To stay comparable with prior baselines, we follow the protocol of [1] Self-Distilled Reasoner(ICML) and [2] Skill-Conditioned Gated Self-Distillation(EMNLP). Checkpoints are saved every 25 steps (step 0 excluded) and selected by the mean over the three benchmarks.

To check that our conclusions do not depend on this selection, we report the final checkpoint after the full 400-step budget, with no checkpoint selection at all (mean avg@12, three seeds):

ModelGRPOOPSDSGSDSGUID R1
Qwen3-1.7B33.15±0.1633.15 \pm 0.1640.77±1.5840.77 \pm 1.5841.30±1.4941.30 \pm 1.4942.66±0.59\mathbf{42.66 \pm 0.59}
Qwen3-4B61.29±1.2761.29 \pm 1.2758.49±1.6358.49 \pm 1.6362.78±0.7262.78 \pm 0.7263.61±0.28\mathbf{63.61 \pm 0.28}
Qwen3-8B63.12±2.0663.12 \pm 2.0664.20±0.7264.20 \pm 0.7263.67±0.7563.67 \pm 0.7564.32±0.44\mathbf{64.32 \pm 0.44}

Without checkpoint selection, SGUID R1 has the highest mean on all three models and stays ahead of full-bank SGSD (+1.36, +0.83, +0.65). Best-checkpoint selection therefore does not favor SGUID. We will add this table to the revision and state the selection protocol explicitly in the Table 1 caption.

References

[1] Zhao, Siyan, et al. "Self-distilled reasoner: On-policy self-distillation for large language models." arXiv preprint arXiv:2601.18734 (2026).

[2] Huang, Jiazhen, et al. "Skill-Conditioned Gated Self-Distillation for LLM Reasoning." arXiv preprint arXiv:2605.28791 (2026).

AuthorsOpenReview ↗

Author Rebuttal (Part 2/3)

2. Compute-matched comparisons and variance across seeds

The paper also does not report variation across training runs, even though some of the gains over SGSD are less than one percentage point.

Since this adds a substantial amount of computation, I would like to see a comparison where both methods have the same overall training budget.

We retrained GRPO, OPSD, and SGSD for 400 steps, matching SGUID R1's total step budget (200 selection + 200 distillation steps), over three seeds. In GPU-hours, SGUID R1 costs no more than SGSD at this budget (11.54 vs. ≈ 11.8 on Qwen3-4B), since distilling the compact bank uses fewer teacher passes per prompt.

ModelMethodAIME24AIME25HMMT25Mean
Qwen3-1.7BGRPO50.93±1.5850.93 \pm 1.5836.30±1.1636.30 \pm 1.1622.41±1.1222.41 \pm 1.1236.54±1.1936.54 \pm 1.19
Qwen3-1.7BOPSD55.09±0.8555.09 \pm 0.8542.59±1.1342.59 \pm 1.1328.33±1.47‾\underline{28.33 \pm 1.47}42.01±0.3242.01 \pm 0.32
Qwen3-1.7BSGSD56.39±1.21‾\underline{56.39 \pm 1.21}44.54±0.42‾\underline{44.54 \pm 0.42}26.30±1.3126.30 \pm 1.3142.41±0.3442.41 \pm 0.34
Qwen3-1.7BSGUID R155.74±0.8955.74 \pm 0.8944.07±0.5844.07 \pm 0.5827.96±1.5427.96 \pm 1.5442.59±0.90‾\underline{42.59 \pm 0.90}
Qwen3-1.7BSGUID R256.94±2.82\mathbf{56.94 \pm 2.82}46.67±1.27\mathbf{46.67 \pm 1.27}28.89±1.11\mathbf{28.89 \pm 1.11}44.17±0.73∗∗\mathbf{44.17 \pm 0.73}^{\ast\ast}
Qwen3-8BGRPO78.15±0.32‾\underline{78.15 \pm 0.32}68.89±0.5668.89 \pm 0.5644.81±1.1644.81 \pm 1.1663.95±0.4763.95 \pm 0.47
Qwen3-8BOPSD78.89±0.28\mathbf{78.89 \pm 0.28}70.28±1.11‾\underline{70.28 \pm 1.11}46.67±1.39‾\underline{46.67 \pm 1.39}65.28±0.33‾\underline{65.28 \pm 0.33}
Qwen3-8BSGSD77.04±0.5877.04 \pm 0.5869.81±1.1669.81 \pm 1.1646.20±1.4346.20 \pm 1.4364.35±0.2564.35 \pm 0.25
Qwen3-8BSGUID R176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
Qwen3-8BSGUID R277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗\mathbf{72.31 \pm 1.43}^{\ast\ast\ast}48.06±2.40\mathbf{48.06 \pm 2.40}65.93±0.56∗\mathbf{65.93 \pm 0.56}^{\ast}

n = 3 seeds; ± is the standard deviation. Bold / underline: best / second-best mean. Stars: SGUID R2 vs. SGUID R1 within the same model, one-sided paired t-test over the three seeds (each R2 run continues from the matching R1 run; df = 2; ∗p≤0.05^{\ast}p \le 0.05, ∗∗p≤0.01^{\ast\ast}p \le 0.01, ∗∗∗p≤0.001^{\ast\ast\ast}p \le 0.001).

We find the conclusions are consistent with Table 1. At the matched budget, SGUID R1 is on par with SGSD (42.59 vs. 42.41 on Qwen3-1.7B; 64.14 vs. 64.35 on Qwen3-8B) while distilling 6 instead of 34 and 71 skills, consistent with the paper's claim that a compact bank matches or exceeds the full bank. At the final checkpoint (point 1), R1 is ahead of SGSD on all three models.

We will replace the single-run Table 1 with these three-seed results, and add the results for Qwen3-4B and Olmo-3-7B in the revised draft.

3. Matching the number of teachers

It would be useful to compare SGUID against SGSD using the same number of teachers.

We reproduce SGSD with its original configuration (top-8 retrieval [2]). SGUID follows this configuration and changes only the skill bank, so its teacher count equals its bank size K, which means fewer teacher passes per prompt than SGSD (6 vs. 8 in R1, 3 in R2). We reran SGSD with 6 teachers on Qwen3-4B ( everything else as in Table 1):

Qwen3-4BTeachers / promptAIME24AIME25HMMT25Mean
Base—73.973.965.865.846.746.762.162.1
SGSD (original, Table 1)8875.675.670.370.346.746.764.2064.20
SGSD6674.474.468.968.946.946.963.4363.43
SGUID R16676.776.771.471.447.247.265.10\mathbf{65.10}

SGSD w/ 6 teachers has a lower mean performance by 0.77 points. With the same number of teachers, SGUID R1 outperforms SGSD by 1.67 points.The gain therefore comes from which skills condition the teachers, not from how many teachers there are. We will add this row to the revision.

AuthorsOpenReview ↗

Author Rebuttal (Part 3/3)

4. Does the co-evolution step itself help?

I think an important missing ablation is to run the second round without generating a new skill bank, using the original skills instead.

We ran this ablation on Qwen3-8B over three seeds, which also extends the without-selection comparison of Figure 3 (Qwen3-4B ) to a second model. The settings are:

  • w/o selection: each round distills the full candidate bank, and the round-2 bank is curated from the round-1 model's rollouts.
  • w/ R1 bank (the requested control): round 2 continues from the round-1 model but distills the same 6 skills selected in round 1, with no new bank.
  • w/ selection (SGUID): round 2 curates a new bank from the round-1 model and selects 3 skills from it.
Qwen3-8BRoundAIME24AIME25HMMT25Mean
Base (Table 1)–77.5077.5067.2267.2243.8943.8962.8762.87
w/o selectionR176.20±0.8976.20 \pm 0.8970.83±1.2770.83 \pm 1.2747.04±0.5847.04 \pm 0.5864.69±0.3864.69 \pm 0.38
w/o selectionR277.41±0.5877.41 \pm 0.5869.91±0.3269.91 \pm 0.3246.39±1.4746.39 \pm 1.4764.57±0.6864.57 \pm 0.68
w/ R1 bankR276.85±0.5876.85 \pm 0.5870.37±1.1270.37 \pm 1.1245.74±1.7645.74 \pm 1.7664.32±0.1964.32 \pm 0.19
w/ selectionR176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
w/ selectionR277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗72.31 \pm 1.43^{\ast\ast\ast}48.06±2.4048.06 \pm 2.4065.93±0.56∗\mathbf{65.93 \pm 0.56}^{\ast}

Stars: w/ selection R2 vs. w/ selection R1 (same test as in point 2).

he benefit of a second round depends on updating and selecting the skill bank. Reusing the six skills from round 1 does not improve performance (64.32 vs. 64.14 for R1), and simply re-curating a bank and distilling all candidate skills also gives no gain (64.69 → 64.57). In contrast, SGUID's second round improves over both baselines by 1.61 points (p=0.014p = 0.014) and 1.36 points (p=0.029p = 0.029), respectively, using one-sided Welch's tt-tests. It also improves significantly over its own round-1 result (p=0.017p = 0.017). These results suggest that the gain comes from selecting new skills based on the updated model, rather than from performing another distillation round alone. We will extend Figure 3 with the Qwen3-8B results.

The paper only evaluates two rounds.

The paper only evaluates two rounds.

Our claim is limited to stability from one round to the next, rather than to indefinite co-evolution. Even two rounds reveal a clear difference between the three settings. Without selection, the gain decreases on Qwen3-4B (+2.1 → +1.6, with a negative change on HMMT25; Figure 3) and disappears on Qwen3-8B (64.69 → 64.57). With selection, performance instead improves from 64.14 to 65.93. Thus, the results support stable co-evolution across the two evaluated rounds, but do not establish what happens over many more rounds. We will state the number of rounds explicitly in the abstract and discuss longer-horizon co-evolution as a limitation.

We hope these additions address the reviewer's concerns, and we welcome any further questions.

Reviewer_jxdKOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper builds on an on-policy self-distillation method (SGSD) where a teacher that shares the student's weights is conditioned on a retrieved natural-language skill. The paper observes that, on Qwen models, only 6–21% of skill retrievals pass a magnitude gate. The paper therefore proposes SGUID, which trains once with the full skill bank, keeps the skills that give useful teacher signals both early and late in training, and retrains from the original checkpoint using only the top six. A second round repeats this with a new bank written by the updated model, keeping three skills. Across four models evaluated on math benchmarks, the proposed method matches or beats SGSD's mean score on three models after round 1 and on all four after round 2.

Critical strengths

The diagnostic analysis that only a small proportion of retrieved skills give a non-zero teacher signal is interesting. The proposed method gets matching or better performance on math benchmarks than the baseline.

Critical weaknesses
  • SGUID's gains cannot be attributed to skill selection, because the comparisons are not controlled. SGUID uses 2× SGSD's training in round 1 and around 4 times by round 2 but is compared to single-run baselines. Batch size and teachers per prompt also differ across methods, and in Table 4, top-6 appears to be trained for 200 steps vs. 100 for random/worst. There is no retrieval-frequency baseline, so the ranking may simply favor frequently retrieved generic skills, and the pipeline's hand-set choices are not ablated. This is consequential because the paper's contribution is the selection rule, and none of the experiments isolate its effect. 

  • The evaluation protocol does not support the claimed improvements. The reported checkpoint is the best of 200 steps on the evaluation benchmarks themselves, which inflates all methods, and the co-evolution stability claim rests on differences of only −0.3 to +1.1 points with no test of statistical significance. This is consequential because the claimed advantages over SGSD are of similar size, so it is unclear whether any of them matter.

Questions for authors

Clarifications:

  • How many checkpoints were evaluated for best-checkpoint selection? What are the results at the final checkpoint, or with checkpoints selected on a held-out validation set?
  • In Table 4, were random-6 and worst-6 trained for 100 steps while top-6 (identical to Table 1) was trained for 200?
  • What is the total training cost (e.g., GPU-hours) of each method, including SGUID's selection runs and bank curation in each round?
  • Why do batch size and teachers per prompt differ across methods? Was SGSD run with its original configuration?
  • How were h, τ, the early/late split and K chosen, and why does K change from 6 to 3 between rounds? What are the round-2 values, which Table 5 does not list?

Requested experiments that could affect my assessment:

  • Ablations of the pipeline itself, namely restarting from θ_r vs. continuing training with the pruned bank, the selection run's own checkpoint as a baseline, and sensitivity to h, τ and the early/late split.
  • Compute-matched baselines, i.e. SGSD and OPSD trained for the same total budget as SGUID R1 and R2, and the without-selection round-2 control on all four models.
  • A retrieval-frequency baseline that selects the top-K skills by retrieval count, to show that the ranking captures usefulness rather than how often a skill is retrieved. 9.
  • Test co-evolution results (quite small improvements with the proposed method) for statistical significance.
5 comments
AuthorsOpenReview ↗

Author Rebuttal (Part 1/4)

We thank the reviewer for the detailed and constructive review. We are glad that the reviewer found our analysis interesting: only a few retrieved skills give a non-zero teacher signal. The two main concerns are (i) whether SGUID's gains can be validated under controlled budgets, and (ii) justification of the evaluation protocol. During the rebuttal period, we ran: compute-matched baselines with variance reported, an ablation of co-evolution gains with significance test, a retrieval-frequency baseline to ablate the skill selection, and an SGSD baseline with 6 teachers for fair comparison. We respond point-by-point below.

1. Compute-matched comparison (W1, W2, Q7, Q9)

SGUID R1 uses 2× the training steps of the 200-step baselines in Table 1 (a 200-step selection run plus a 200-step distillation run). We therefore retrained GRPO, OPSD, and SGSD for 400 steps, matching SGUID R1's total step budget, over three seeds.

ModelMethodAIME24AIME25HMMT25Mean
Qwen3-1.7BGRPO50.93±1.5850.93 \pm 1.5836.30±1.1636.30 \pm 1.1622.41±1.1222.41 \pm 1.1236.54±1.1936.54 \pm 1.19
Qwen3-1.7BOPSD55.09±0.8555.09 \pm 0.8542.59±1.1342.59 \pm 1.1328.33±1.47‾\underline{28.33 \pm 1.47}42.01±0.3242.01 \pm 0.32
Qwen3-1.7BSGSD56.39±1.21‾\underline{56.39 \pm 1.21}44.54±0.42‾\underline{44.54 \pm 0.42}26.30±1.3126.30 \pm 1.3142.41±0.3442.41 \pm 0.34
Qwen3-1.7BSGUID R155.74±0.8955.74 \pm 0.8944.07±0.5844.07 \pm 0.5827.96±1.5427.96 \pm 1.5442.59±0.90‾\underline{42.59 \pm 0.90}
Qwen3-1.7BSGUID R256.94±2.82\mathbf{56.94 \pm 2.82}46.67±1.27\mathbf{46.67 \pm 1.27}28.89±1.11\mathbf{28.89 \pm 1.11}44.17±0.73∗∗\mathbf{44.17 \pm 0.73}^{\ast\ast}
Qwen3-8BGRPO78.15±0.32‾\underline{78.15 \pm 0.32}68.89±0.5668.89 \pm 0.5644.81±1.1644.81 \pm 1.1663.95±0.4763.95 \pm 0.47
Qwen3-8BOPSD78.89±0.28\mathbf{78.89 \pm 0.28}70.28±1.11‾\underline{70.28 \pm 1.11}46.67±1.39‾\underline{46.67 \pm 1.39}65.28±0.33‾\underline{65.28 \pm 0.33}
Qwen3-8BSGSD77.04±0.5877.04 \pm 0.5869.81±1.1669.81 \pm 1.1646.20±1.4346.20 \pm 1.4364.35±0.2564.35 \pm 0.25
Qwen3-8BSGUID R176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
Qwen3-8BSGUID R277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗\mathbf{72.31 \pm 1.43}^{\ast\ast\ast}48.06±2.40\mathbf{48.06 \pm 2.40}65.93±0.56∗\mathbf{65.93 \pm 0.56}^{\ast}

n = 3 seeds; ± is the standard deviation. Bold / underline: best / second-best mean. Stars: SGUID R2 vs. SGUID R1 within the same model, one-sided paired t-test over the three seeds (each R2 run continues from the matching R1 run; df = 2; ∗p≤0.05^{\ast}p \le 0.05, ∗∗p≤0.01^{\ast\ast}p \le 0.01, ∗∗∗p≤0.001^{\ast\ast\ast}p \le 0.001).

At a matched 400-step budget, SGUID R1 has the best mean on Qwen3-1.7B while distilling 6 instead of 34 skills. On Qwen3-8B, R1 trails OPSD (64.14 vs. 65.28), consistent with Table 1;R2 significantly improves over R1 on the mean on both models (p = 0.0018 on Qwen3-1.7B, p = 0.017 on Qwen3-8B), which directly tests the co-evolution gain. We will add the results for Qwen3-4B and Olmo-3-7B in the revised version.

2. Without-selection round-2 control on Qwen3-8B (W2, Q7)

In addition to Figure 3 (Qwen-4B), we additionally provide evidence with Qwen3-8B that a second round helps only when skills are re-selected from the bank curated by the updated model.* Re-distilling the full bank (64.69 → 64.57) or reusing the R1 bank (64.32) gives no gain, whereas re-selecting improves the model from 64.14 to 65.93 (R2 vs. R1, p = 0.017).

SettingRoundAIME24AIME25HMMT25Mean
Base (Table 1)–77.5077.5067.2267.2243.8943.8962.8762.87
w/o selectionR176.20±0.8976.20 \pm 0.8970.83±1.2770.83 \pm 1.2747.04±0.5847.04 \pm 0.5864.69±0.3864.69 \pm 0.38
w/o selectionR277.41±0.5877.41 \pm 0.5869.91±0.3269.91 \pm 0.3246.39±1.4746.39 \pm 1.4764.57±0.6864.57 \pm 0.68
w/ R1 bankR276.85±0.5876.85 \pm 0.5870.37±1.1270.37 \pm 1.1245.74±1.7645.74 \pm 1.7664.32±0.1964.32 \pm 0.19
w/ selectionR176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
w/ selectionR277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗72.31 \pm 1.43^{\ast\ast\ast}48.06±2.4048.06 \pm 2.4065.93±0.56∗65.93 \pm 0.56^{\ast}

Stars: w/ selection R2 vs. w/ selection R1 (same test as in point 1).

AuthorsOpenReview ↗

Author Rebuttal (Part 2/4)

3. Evaluation protocol (W2, Q1)

The reported checkpoint is the best of 200 steps on the evaluation benchmarks themselves.

To stay comparable with prior baselines, we follow the protocol of [1] Self-Distilled Reasoner(ICML) and [2] Skill-Conditioned Gated Self-Distillation(EMNLP). Checkpoints are saved every 25 steps (step 0 excluded) and selected by the mean over the three benchmarks. GRPO, OPSD, and SGSD are selected over their full 400 steps.

In addition, we report the final checkpoint after the full 400-step budget, with no checkpoint selection at all, over three seeds:

ModelGRPOOPSDSGSDSGUID R1
Qwen3-1.7B33.15±0.1633.15 \pm 0.1640.77±1.5840.77 \pm 1.5841.30±1.4941.30 \pm 1.4942.66±0.59\mathbf{42.66 \pm 0.59}
Qwen3-4B61.29±1.2761.29 \pm 1.2758.49±1.6358.49 \pm 1.6362.78±0.7262.78 \pm 0.7263.61±0.28\mathbf{63.61 \pm 0.28}
Qwen3-8B63.12±2.0663.12 \pm 2.0664.20±0.7264.20 \pm 0.7263.67±0.7563.67 \pm 0.7564.32±0.44\mathbf{64.32 \pm 0.44}

SGUID R1 has the highest final-checkpoint mean on all three models and keeps its advantage over full-bank SGSD (+1.36, +0.83, +0.65). On Qwen3-8B, OPSD is ahead under best-checkpoint selection but comparable at the final checkpoint (64.20 vs. 64.32). We will add this table to the revision.

4. Does the selection rule itself matter? (W1, Q8)

None of the experiments isolate its effect.

The submission already includes two ablation studies regarding the selection rule. Dropping the persistence constraint (Table 3, Qwen3-4B) lowers the mean from 65.1 to 62.3. Replacing the top-6 with random-6 or worst-6 (Table 4, Qwen3-1.7B) also lowers it. Reviewer MgrQ noted that these comparisons "help show that the choice of skills matters, not just the number of skills."

Following the reviewer's suggestion, we added the retrieval-frequency baseline (the 6 most-retrieved skills) and a fast-decay-6 baseline (the 6 skills whose signal decays fastest) on Qwen3-4B. All rows, including full-bank SGSD, are trained for 100 steps:

Qwen3-4BAIME24AIME25HMMT25Mean
Base73.973.965.865.846.746.762.162.1
SGSD (full bank, 30 skills)75.075.069.72\mathbf{69.72}45.8345.8363.5263.52
top-6 (ours)77.2\mathbf{77.2}67.267.248.1\mathbf{48.1}64.2\mathbf{64.2}
frequency-674.774.766.966.945.345.362.362.3
random-6 (3 seeds)74.9±0.774.9 \pm 0.768.5±0.768.5 \pm 0.745.3±0.645.3 \pm 0.662.9±0.262.9 \pm 0.2
worst-674.474.467.867.844.444.462.262.2
fast-decay-674.774.768.368.343.643.662.262.2

Note that Frequency-6 is 1.9 points below top-6 and no better than worst-6, so our ranking captures usefulness rather than how often a skill is retrieved. Top-6 is also the only 6-skill bank that beats the 30-skill full bank at the same 100-step budget (64.2 vs. 63.52). We will add this table and the frequency baseline to Section 5.

References

[1] Zhao, Siyan, et al. "Self-distilled reasoner: On-policy self-distillation for large language models." arXiv preprint arXiv:2601.18734 (2026).

[2] Huang, Jiazhen, et al. "Skill-Conditioned Gated Self-Distillation for LLM Reasoning." arXiv preprint arXiv:2605.28791 (2026).

AuthorsOpenReview ↗

Author Rebuttal (Part 3/4)

5. Training budget in Table 4 (W1, Q2)

All three K = 6 variants in Table 4 (top-6, random-6, worst-6) are trained for 100 steps, as mentioned in the figure title. The top-6 row equals the Table 1 R1 row because the best checkpoint of that R1 run on Qwen3-1.7B is at step 100. We use the 200-step SGSD result to show the best achievable performance. We additionally report the table with SGSD also limited to 100 steps, so that every row has the same budget:

Qwen3-1.7B, 100 stepsAIME24AIME25HMMT25Mean
Base50.635.623.136.4
SGSD (full bank, 34 skills)56.3941.1127.2241.57
top-6 (ours)55.343.128.642.3
random-6 (3 selection seeds)54.1 ± 0.842.6 ± 1.226.9 ± 0.941.2 ± 0.9
worst-655.642.826.141.5

We find top-6 has the best mean, ahead of other baselines. We will revise Table 4 tp add the 100-step SGSD row and state the budget in its caption.

6. Total computational cost (Q3)

We report GPU-hours (wall-clock hours × GPU count) on Qwen3-4B in Table 1, where all baseline runs are 200 steps: GRPO, OPSD, and SGSD. Each SGUID stage (skill selection, distillation) consumes 200 steps. Training uses 4× H200 and bank curation uses 1× H200.

MethodBank curationSkill selectionDistillationTotal GPU-hours
GRPO——10.0010.0010.0010.00
OPSD——3.003.003.003.00
SGSD (full bank)0.440.44—5.685.686.126.12
SGUID R10.440.445.685.685.425.4211.5411.54
SGUID R2 (extra round)0.920.925.655.654.854.8511.4211.42
SGUID R1 + R2———22.96\mathbf{22.96}

SGUID R1 costs no more than SGSD at the step-matched budget of point 1: a 400-step SGSD run costs ≈ 0.44 + 2 × 5.68 ≈ 11.8 GPU-hours, versus 11.54 for SGUID R1.

7. Batch size and teachers per prompt (W1, Q4)

Why do batch size and teachers per prompt differ across methods? Was SGSD run with its original configuration?

We reproduce SGSD with its original code and configuration. Each baseline uses its original published setup: GRPO and OPSD use effective batch 32 (Zhao et al., 2026, Table 6), and SGSD uses effective batch 16 and top-8 retrieval (Huang et al., 2026, Table 11). SGUID inherits SGSD's configuration in full and changes only the skill bank. The teacher count is determined by each method's privileged information: OPSD conditions on one ground-truth solution, and SGUID's count equals its selected bank size K. SGUID therefore uses fewer teacher passes per prompt than SGSD (6 vs. 8 in R1, 3 in R2). To rule this out explicitly, we also ran SGSD with 6 teachers:

Qwen3-4Bteachers / promptAIME24AIME25HMMT25Mean
Base—73.973.965.865.846.746.762.162.1
SGSD (original, Table 1)8875.675.670.370.346.746.764.2064.20
SGSD6674.474.468.968.946.946.963.4363.43
SGUID R16676.776.771.471.447.247.265.10\mathbf{65.10}

With the same number of teachers, SGUID R1 outperforms SGSD by 1.67 points, so the gain does not come from teacher count.

Reviewer_CLo4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the problem of how LLMs leverage external "skills" during training in On-Policy Self-Distillation (OPSD). It discovers that fewer than 25% of semantically retrieved skills provide effective learning signals, and blindly using the full skill bank often introduces noise. To address this, the authors propose an innovative method named SGUID. By monitoring training dynamics throughout the distillation process, SGUID filters and retains a compact subset of skills that consistently provide verifier-aligned teacher signals from early to late training. Experiments show that distilling just 6 skills carefully selected by SGUID matches or even outperforms full-bank distillation (which is much larger) across multiple math reasoning benchmarks. Furthermore, multi-round "model-skill co-evolution" (distilling 3 new skills in the second round) not only achieves stable performance gains (such as raising the accuracy of Qwen3-8B from 64.3% to 66.3%), but also effectively avoids the late-stage performance degradation caused by unfiltered co-evolution, proving that skill selection is a key mechanism supporting stable co-evolution.

Critical strengths
  1. The paper identifies a meaningful and insufficiently explored problem. The paper identifies an important limitation of existing skill-guided self-distillation methods: semantically relevant skills do not necessarily provide useful training signals. SGUID addresses this problem with a simple and interpretable selection strategy based on training dynamics, offering a practical way to identify skills worth distilling.

  2. The experiments cover four models from two families and demonstrate that a small set of selected skills can match or outperform full-bank distillation. The ablation studies further investigate the effects of skill-bank size, selection strategies, and signal persistence, providing useful evidence for the effectiveness of the proposed method.

Critical weaknesses
  1. Lack of Compute-Matched Comparisons. SGUID requires an additional full-bank distillation run to collect training statistics before restarting training with selected skills. This introduces substantial computational overhead compared with the baselines. Without compute-matched experiments, it is unclear whether the improvements come from better skill selection or simply from additional computational resources.

  2. Insufficient Evidence for Stable Co-Evolution. Although stable model–skill co-evolution is presented as a key contribution, the experiments only evaluate two rounds, with limited comparisons against unfiltered co-evolution. Moreover, the main results lack variance estimates across independent training runs. These limitations make it difficult to determine whether the observed improvements are statistically reliable and whether the proposed method can support stable long-term co-evolution.

Questions for authors

Q1. Computational Cost SGUID requires an additional full-bank distillation run to collect training statistics before training with selected skills. What is the total computational cost of SGUID compared with SGSD and OPSD? Does SGUID still outperform these baselines under the same overall compute budget?

Q3. Effectiveness of the Selection Criterion SGUID assumes that skills providing persistent positive training signals are more valuable for distillation. However, a skill's contribution may decrease because the student has already learned it. How do the authors distinguish this case from genuinely ineffective skills? Is there evidence that the proposed persistence metric predicts actual learning benefits?

5 comments
AuthorsOpenReview ↗

Author Rebuttal (Part 1/3)

We appreciate the reviewer's acknowledgment that semantically relevant skills do not necessarily provide useful training signals, a meaningful and insufficiently explored problem, and that SGUID is a practical way to select which skill is worth distilling. We are also encouraged that the reviewer found our experiments across four models and the ablations to be useful evidence. The two weaknesses concern compute-matched comparisons (W1, Q1) and the evidence for stable co-evolution (W2), and Q3 asks whether persistence reflects real learning benefit. In response, we add compute-matched baselines with final-checkpoint results, total computation cost in terms of GPU hours, additional co-evolution experiments on Qwen3-8B, and two new skill selection baselines (frequency-6 and fast-decay-6). We address each point in turn.

1. Compute-matched comparison (W1, Q1)

We retrained GRPO, OPSD, and SGSD for 400 steps over three random seeds, matching SGUID R1's total step budget (200 steps of full-bank selection + 200 steps of distillation).

ModelMethodAIME24AIME25HMMT25Mean
Qwen3-1.7BGRPO50.93±1.5850.93 \pm 1.5836.30±1.1636.30 \pm 1.1622.41±1.1222.41 \pm 1.1236.54±1.1936.54 \pm 1.19
Qwen3-1.7BOPSD55.09±0.8555.09 \pm 0.8542.59±1.1342.59 \pm 1.1328.33±1.47‾\underline{28.33 \pm 1.47}42.01±0.3242.01 \pm 0.32
Qwen3-1.7BSGSD56.39±1.21‾\underline{56.39 \pm 1.21}44.54±0.42‾\underline{44.54 \pm 0.42}26.30±1.3126.30 \pm 1.3142.41±0.3442.41 \pm 0.34
Qwen3-1.7BSGUID R155.74±0.8955.74 \pm 0.8944.07±0.5844.07 \pm 0.5827.96±1.5427.96 \pm 1.5442.59±0.90‾\underline{42.59 \pm 0.90}
Qwen3-1.7BSGUID R256.94±2.82\mathbf{56.94 \pm 2.82}46.67±1.27\mathbf{46.67 \pm 1.27}28.89±1.11\mathbf{28.89 \pm 1.11}44.17±0.73∗∗\mathbf{44.17 \pm 0.73}^{\ast\ast}
Qwen3-8BGRPO78.15±0.32‾\underline{78.15 \pm 0.32}68.89±0.5668.89 \pm 0.5644.81±1.1644.81 \pm 1.1663.95±0.4763.95 \pm 0.47
Qwen3-8BOPSD78.89±0.28\mathbf{78.89 \pm 0.28}70.28±1.11‾\underline{70.28 \pm 1.11}46.67±1.39‾\underline{46.67 \pm 1.39}65.28±0.33‾\underline{65.28 \pm 0.33}
Qwen3-8BSGSD77.04±0.5877.04 \pm 0.5869.81±1.1669.81 \pm 1.1646.20±1.4346.20 \pm 1.4364.35±0.2564.35 \pm 0.25
Qwen3-8BSGUID R176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
Qwen3-8BSGUID R277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗\mathbf{72.31 \pm 1.43}^{\ast\ast\ast}48.06±2.40\mathbf{48.06 \pm 2.40}65.93±0.56∗\mathbf{65.93 \pm 0.56}^{\ast}

n = 3 seeds; ± is the standard deviation. Bold / underline: best / second-best mean. Stars: SGUID R2 vs. SGUID R1 within the same model, one-sided paired t-test over the three seeds (each R2 run continues from the matching R1 run; df = 2; * p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001).

We find the conclusions are consistent with Table 1. At the matched budget, SGUID R1 matches full-bank SGSD (42.59 vs. 42.41 on Qwen3-1.7B; 64.14 vs. 64.35 on Qwen3-8B) while distilling 6 instead of 34 and 71 skills. On Qwen3-8B it remains below OPSD (65.28), consistent with Table 1. We also report the final checkpoint after 400 steps with no checkpoint selection:

ModelGRPOOPSDSGSDSGUID R1
Qwen3-1.7B33.15±0.1633.15 \pm 0.1640.77±1.5840.77 \pm 1.5841.30±1.4941.30 \pm 1.4942.66±0.59\mathbf{42.66 \pm 0.59}
Qwen3-4B61.29±1.2761.29 \pm 1.2758.49±1.6358.49 \pm 1.6362.78±0.7262.78 \pm 0.7263.61±0.28\mathbf{63.61 \pm 0.28}
Qwen3-8B63.12±2.0663.12 \pm 2.0664.20±0.7264.20 \pm 0.7263.67±0.7563.67 \pm 0.7564.32±0.44\mathbf{64.32 \pm 0.44}

At the final checkpoint, SGUID R1 has the highest mean on all three models and is ahead of SGSD (+1.36, +0.83, +0.65). Under a matched step budget, SGUID R1 therefore matches or exceeds the baselines while using a 6-skill bank. We will revise the paper to add these three-seed results and the final-checkpoint evaluation results.

AuthorsOpenReview ↗

Author Rebuttal (Part 2/3)

2. Evidence for stable co-evolution (W2)

The experiments only evaluate two rounds, with limited comparisons against unfiltered co-evolution. Moreover, the main results lack variance estimates across independent training runs.

(a) Comparison against unfiltered co-evolution. We extend the without-selection comparison of Figure 3 (Qwen3-4B) to Qwen3-8B over three seeds. w/o selection distills the full candidate bank in each round. w/ R1 bank continues round 2 with the 6 skills selected in round 1, with no new bank. w/ selection (SGUID) curates a new bank from the round-1 model and selects 3 skills.

Qwen3-8BRoundAIME24AIME25HMMT25Mean
Base (Table 1)–77.5077.5067.2267.2243.8943.8962.8762.87
w/o selectionR176.20±0.8976.20 \pm 0.8970.83±1.2770.83 \pm 1.2747.04±0.5847.04 \pm 0.5864.69±0.3864.69 \pm 0.38
w/o selectionR277.41±0.5877.41 \pm 0.5869.91±0.3269.91 \pm 0.3246.39±1.4746.39 \pm 1.4764.57±0.6864.57 \pm 0.68
w/ R1 bankR276.85±0.5876.85 \pm 0.5870.37±1.1270.37 \pm 1.1245.74±1.7645.74 \pm 1.7664.32±0.1964.32 \pm 0.19
w/ selectionR176.67±1.5476.67 \pm 1.5470.18±1.5870.18 \pm 1.5845.56±0.2845.56 \pm 0.2864.14±0.1464.14 \pm 0.14
w/ selectionR277.41±1.3777.41 \pm 1.3772.31±1.43∗∗∗72.31 \pm 1.43^{\ast\ast\ast}48.06±2.4048.06 \pm 2.4065.93±0.56∗\mathbf{65.93 \pm 0.56}^{\ast}

Stars: w/ selection R2 vs. w/ selection R1 (same test as in point 1).

Only the selected second round gives a significant gain. Re-distilling the full bank does not improve the model (64.69 → 64.57), and reusing the R1 bank barely moves it (64.14 → 64.32). SGUID's R2 improves over both (+1.36, p = 0.029; +1.61, p = 0.014; one-sided Welch's t-test). In round 1, full-bank distillation is slightly ahead (64.69 vs. 64.14) while using 71 instead of 6 skills. The advantage of selection appears in round 2, as in Figure 3 on Qwen3-4B. With this experiment, the without-selection control covers two of the four models.

(b) Statistical significance. All new results above use three seeds. With selection, R2 improves significantly over R1 on both models (Qwen3-1.7B: 42.59 → 44.17, p = 0.0018; Qwen3-8B: 64.14 → 65.93, p = 0.017). It also improves over full-bank SGSD (one-sided Welch's t-test: p = 0.018 and p = 0.012).

(c) Number of rounds. Our claim is that selection keeps co-evolution stable from one round to the next, and the second round is where unfiltered co-evolution already starts to degrade: without selection, the gain shrinks on Qwen3-4B (+2.1 → +1.6, turning negative on HMMT25; Figure 3) and stalls on Qwen3-8B (64.69 → 64.57), whereas with selection it grows (64.14 → 65.93). After round 1, the model has already internalized the first-round skills, so the gate saturates earlier and fewer skills keep contributing late in training, which is why round 2 retains K = 3. We will state the number of rounds explicitly in the abstract.

3. Total computational cost (Q1)

We report the GPU-hours (wall-clock hours × GPU count) on Qwen3-4B. All baseline runs are 200 steps: GRPO, OPSD, and SGSD as in Table 1, and each SGUID stage (skill selection, distillation) for 200 steps. Training uses 4× H200 and bank curation uses 1× H200.

MethodBank curationSkill selectionDistillationTotal GPU-hours
GRPO——10.0010.0010.0010.00
OPSD——3.003.003.003.00
SGSD (full bank)0.440.44—5.685.686.126.12
SGUID R10.440.445.685.685.425.4211.5411.54
SGUID R2 (extra round)0.920.925.655.654.854.8511.4211.42
SGUID R1 + R2———22.9622.96
  • Against SGSD, at the same cost: a 400-step SGSD run costs ≈ 0.44 + 2 × 5.68 ≈ 11.8 GPU-hours, versus 11.54 for SGUID R1, and R1 matches or exceeds it (point 1). Distilling the compact bank is cheaper than distilling the full bank (5.42 vs. 5.68; 4.85 in R2) because each prompt uses fewer teacher passes.
  • Against OPSD: OPSD is cheaper per step because its teacher conditions on a single ground-truth solution rather than on several skills. At the same step count, OPSD is ahead of R1 on Qwen3-8B at the best checkpoint (65.28 vs. 64.14) and comparable at the final checkpoint (64.20 vs. 64.32).
AuthorsOpenReview ↗

Author Rebuttal (Part 3/3)

4. Does persistence reflect real learning benefit? (Q3)

A skill's contribution may decrease because the student has already learned it. How do the authors distinguish this case from genuinely ineffective skills? Is there evidence that the proposed persistence metric predicts actual learning benefits?

We thank the reviewer for this insightful question.

(a) Learned vs. ineffective skills. For deciding what to distill next, SGUID does not need to tell these two cases apart. A skill the student has fully absorbed and a skill that never helped lead to the same decision: neither has signal left to distill. What the criterion must avoid is discarding skills that are still being learned. It is built to tolerate the decay that learning causes. Every useful skill shows a moderate early-to-late drop in its active-signal rate (2–4.4× on Qwen3-1.7B), and SGUID keeps such skills as long as the drop stays within τ = 5 and they keep at least h = 10 late-phase positive events. Only skills whose rate collapses beyond τ are treated as exhausted. Persistence therefore measures remaining learning signal, which is what matters for re-distillation.

(b) Evidence that persistence predicts benefit. Two experiments hold the pipeline and budget fixed and vary only the criterion. First, dropping the persistence constraint (Table 3, Qwen3-4B) lowers the mean from 65.1 to 62.3: it admits two skills whose rate collapses, and the result is barely above the base model. Second, we compare top-6 against alternative 6-skill banks on Qwen3-4B, all trained for 100 steps. The alternatives include two new baselines: frequency-6 (the most-retrieved skills) and fast-decay-6 (the skills whose signal decays fastest, i.e. those most likely already learned).

Qwen3-4BAIME24AIME25HMMT25Mean
Base73.973.965.865.846.746.762.162.1
SGSD (full bank, 30 skills)75.075.069.72\mathbf{69.72}45.8345.8363.5263.52
top-6 (ours)77.2\mathbf{77.2}67.267.248.1\mathbf{48.1}64.2\mathbf{64.2}
random-6 (3 seeds)74.9±0.774.9 \pm 0.768.5±0.768.5 \pm 0.745.3±0.645.3 \pm 0.662.9±0.262.9 \pm 0.2
worst-674.474.467.867.844.444.462.262.2
frequency-674.774.766.966.945.345.362.362.3
fast-decay-674.774.768.368.343.643.662.262.2

Top-6 is the best bank overall, 0.7 points above the 30-skill full bank at the same 100-step budget (64.2 vs. 63.52), while every alternative reaches 62.2–62.9. Fast-decay-6 does no better than worst-6, so skills that decay quickly are poor targets for re-distillation. Together with the persistence ablation, this is consistent with persistence tracking actual learning benefit. We will add the frequency-6 and fast-decay-6 baselines to Section 5 and expand the discussion of learned vs. ineffective skills.

We hope these additions address the reviewer's concerns, and we welcome any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

Dear Reviewers,

We sincerely appreciate your careful reading and thoughtful feedback, which has pushed us to strengthen our experimental evidence considerably. We appreciate the recognition regarding our proposal: semantically relevant skills do not necessarily provide useful training signals, a meaningful and insufficiently explored problem, and that SGUID is a practical way to select which skill is worth distilling. We have taken every suggestion seriously in preparing our revision. Responding directly to your feedback, we summarize the additions below, and respond to each reviewer in detail in the individual threads.

1. Compute-matched comparison

We retrained GRPO, OPSD, and SGSD for 400 steps over three random seeds, matching SGUID R1's total step budget (200 steps of full-bank selection + 200 steps of distillation).

ModelGRPOOPSDSGSDSGUID R1SGUID R2
Qwen3-1.7B36.5442.0142.4142.5944.17
Qwen3-8B63.9565.2864.3564.1465.93

At the matched budget, SGUID R1 matches full-bank SGSD while distilling 6 instead of 34 and 71 skills. On Qwen3-8B, SGUID R1 remains below OPSD (65.28). SGUID R2 significantly improves over R1 on both models (p = 0.0018 on Qwen3-1.7B; p = 0.017 on Qwen3-8B). This conclusion is consistent with Table 1 in our submission.

2. Final-checkpoint results

SGUID R1 achieves the highest final-checkpoint mean across all three model scales (1.7B/4B/8B), outperforming full-bank SGSD by 1.36/0.83/0.65 points, respectively.

3. Stronger evidence for stable co-evolution

We extend the without-selection comparison of Figure 3 (Qwen3-4B) to Qwen3-8B over three seeds. The mean results are:

SettingR1R2
Full bank64.6964.57
Fixed R1 bank64.1464.32
SGUID selection64.1465.93

The improvement comes from selecting new skills based on the updated model, rather than simply performing another round of distillation. SGUID's R2 improves over the full-bank and fixed-bank controls by 1.36 (p = 0.029) and 1.61 (p = 0.014) points, respectively, using one-sided Welch's t-tests. This directly supports our claim that skill selection is the key mechanism for stable co-evolution.

4. Does our skill selection criterion reflect real learning benefit?

We compare against baselines selecting six skills randomly (random-6), by retrieval frequency (frequency-6), or by fastest signal decay (fast-decay-6).

Top-6 (ours) achieves a mean of 64.2, outperforming frequency-6 (62.3), fast-decay-6 (62.2), and random-6 (62.9), demonstrating that our selection criterion captures training utility rather than shallow features like retrieval frequency. Together, these results support the conclusion that SGUID effectively identifies skills with greater learning utility.

5. Number of teachers per prompt

We additionally controlled for teacher count by comparing SGUID R1 and SGSD with the same six teachers per prompt. SGUID achieves 65.10 versus 63.43 for SGSD, ruling out teacher count as the explanation for the improvement and supporting the importance of skill selection.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.