{"id":"afdc022d-61a5-4ffa-afb8-acdae0e159fe","arxiv_id":"2608.09826","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SKALD shows that distilling skill-conditioned teacher predictions into a question-only student improves math reasoning more than GRPO alone, with gains concentrated on rollout groups where rewards are uniform.","lead":"The paper introduces SKALD, a training method that uses a skill summary shown only to a teacher branch of the same language model to create dense learning signals, then distills those signals into the student's weights. On five math benchmarks, SKALD improves average scores over GRPO by 2 to 12 points across model sizes, suggesting a cheap way to add supervision where group rewards give no signal.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SKALD's own shuffled-skill control recovers 68% of the 1.7B gain over GRPO, so generic card conditioning, not abstract skill content, may drive most of the improvement. The matched-skill increment is only +1.54, and no shuffled control is reported at 4B.","rationale":"The reader's weakest assumption and my concern overlap: both worry that gains are not specifically due to skill content. I make this concrete by using the paper's own SKALD-shuffled number. At 1.7B, shuffled cards recover 3.31 of the 4.85-point GRPO gap, leaving only 1.54 points attributable to matched skill content. That is not a speculative alternative; it is reported in §4.3. The paper's leakage audits and sanitized-card control address answer leakage directly, but they do not address the shuffled-card regularization path, and the 5.1% matched-skill answer recovery leaves a residual channel that could inflate the small matched-versus-shuffled gap. The 4B case is especially important because the claimed gain is largest there and no shuffled control is reported; its absence makes the central claim underdetermined at the scale where it matters most. I still credit the paper's strengths: three seeds with hierarchical intervals, equal-FLOP GRPO baselines, decontamination, and the zero-variance stratification are real evidence that SKALD improves mathematical reasoning numerically beyond GRPO. The concern is about interpretation, not about whether the numbers are reproducible. If the proposed shuffled-card experiment shows matched cards clearly outperforming shuffled cards by several points at 4B, the central claim would be supported. Until then, the verdict should remain conditional, with the condition being this control.","tokens_in":13661,"tokens_out":8906,"duration_ms":83554,"concrete_test":"Run the full SKALD pipeline at 1.7B and 4B under an identical protocol except that each problem receives a randomly assigned skill card from the same 1,926-skill library, keeping the same gate, tilt, seeds, and FLOP budget. Report overall avg@8, pass@8, and the zero-variance-group stratified gain for matched versus shuffled versus no-skill teacher conditions. If shuffled-card SKALD at 4B is within about one point of matched SKALD, or reproduces most of the +11.19 equal-FLOP margin, the central skill-content claim is falsified. If the shuffled condition falls back to roughly the no-skill or OPSD level (about 48-49 at 1.7B and near 52-55 at 4B), the concern is refuted. A secondary control replacing the matched card with an answer-scrambled but same-method card would separate residual leakage from genuine skill transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that abstract skills provide dense supervision where group-relative rewards are uninformative. For that claim, the skill content has to be the operative cause. The paper's own SKALD-shuffled condition (§4.3) reaches 48.83 at 1.7B, which is +3.31 over GRPO (45.52) and only -1.54 below matched SKALD (50.37). The matched-skill increment is therefore 1.54 points out of a 4.85-point total; shuffled or irrelevant cards reproduce roughly 68% of the gain. The paper acknowledges that 'generic card exposure or regularization can explain part of the gain,' but the part it can explain is a majority. If a randomly assigned card in the teacher context yields most of the improvement, the mechanism is a context-induced regularization or distribution-shift effect, not the abstract skill's content. This is not resolved by the leakage audits: answer recovery with a matched skill is 5.1% versus 2.4% question-only, and the paper concedes these audits 'cannot logically eliminate' leakage. Moreover, the 4B headline gain (+12.01; +11.19 over tuned equal-FLOP GRPO) has no shuffled-card control, so the largest claimed effect is exactly where the skill-content attribution is least tested. The numerical improvements may stand, but the paper's central interpretation, 'abstract skills provide dense supervision,' is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SKALD (Skill-Anchored Latent Distillation), a shared-parameter on-policy self-distillation method for mathematical reasoning. A teacher branch conditions the same Qwen3-Base checkpoint on an abstract, explicit-answer-filtered skill card, while a student branch sees the question only; the student's own rollout prefixes are trained with an annealed exponentially tilted cross-entropy, and a fixed empirical gate activates distillation only when the skill-conditioned branch has higher initial sampled accuracy. Across five held-out math benchmarks and three Qwen3-Base sizes (0.6B, 1.7B, 4B), SKALD reports avg@8 gains over GRPO of +2.46, +4.85, and +12.01. The paper attributes the gains to dense supervision on zero-variance rollout groups, where group-relative rewards vanish, and supports this with a zero-variance-only distillation run recovering 84.7% of the 1.7B gain. It also reports leakage audits, decontamination, equal-FLOP GRPO baselines, gate-sensitivity checks, and ordered component ablations.","tokens_in":13967,"tokens_out":10567,"duration_ms":88918,"significance":"If the central claim holds, SKALD is a compute-aware complement to RLVR that improves math reasoning without a larger online teacher or test-time privileged input. The paper's strengths are its controls: three training seeds with hierarchical intervals, equal-FLOP GRPO comparisons, an ordered component ablation, conditional leakage audits, decontamination, gate-sensitivity analysis, and group-stratified tests. The derivation of the tilted objective in Lemma 1 is correct and clearly presented, and the paper honestly acknowledges several residual limitations. The main weakness is attribution: the paper's own shuffled-skill control indicates that most of the 1.7B gain is reproduced by irrelevant skill cards, and no shuffled control is reported at 4B. The numerical improvements may stand, but the central interpretation that abstract skill content provides the dense supervision is not yet established.","major_comments":[{"comment":"The shuffled-skill control (SKALD-shuffled, 48.83) reproduces +3.31 of the +4.85 SKALD-over-GRPO gain, leaving only +1.54 (50.37 vs. 48.83) attributable to matched skill content. Thus roughly 68% of the 1.7B headline gain is explained by generic card conditioning or regularization, not by the abstract skill's content. The paper's statement that 'generic card exposure or regularization can explain part of the gain' understates this: at 1.7B it explains a majority. Since the central claim is that abstract skills provide dense supervision, this control directly challenges the causal interpretation. Moreover, no shuffled-skill control is reported at 4B, where the gain is +12.01 and the equal-FLOP margin is +11.19; the largest claimed effect is exactly where content attribution is untested. Please add shuffled/random-card controls at 4B, and if the pattern persists, revise the central interpretation to distinguish skill-conditioned distillation from generic context-conditioned distillation.","section":"Table 1, 1.7B block; §4 Setup paragraph"},{"comment":"The zero-variance-only run (49.63, +4.11 over GRPO, recovering 84.7% of the full 1.7B gain) shows that distillation helps where group-relative rewards are zero, but it does not isolate skill content because this condition still uses matched skill cards. A shuffled-card zero-variance-only condition is needed to determine whether the dense supervision comes from the skill's semantic content or from the presence of any conditioning card. Without this control, the paper's headline mechanism test supports 'dense supervision in zero-variance groups' but not 'abstract skills provide dense supervision.' The matched-vs-shuffled margin at 1.7B is too small to carry the central claim on its own.","section":"§4.2, zero-variance-only distillation"}],"minor_comments":[{"comment":"The aggregate token-category KL plot should report the number of problems, tokens, seeds, category definitions, and inter-annotator agreement; the text acknowledges that a reproducible claim requires these, but the figure caption currently omits them, making the panel hard to interpret as evidence.","section":"Figure 5b and §4.6"},{"comment":"The gate-sensitivity results are summarized only as a range (49.98–50.48) and the 14.6% sign-reversal rate; please include a small table with the hard/soft and fixed/refreshed variants, K values, and seed-level values so readers can see the dispersion behind the range.","section":"Supplementary Section 2.4"},{"comment":"The conditional audit reports answer recovery rates of 2.4% (question-only) and 5.1% (matched skill), but it does not report uncertainty or a test of whether the 2.7-point difference is statistically distinguishable from the shuffled-skill rate of 3.7%; adding confidence intervals or a permutation test would make the 'modest increase' claim more precise.","section":"§4.3, leakage audit"}],"recommendation":"major_revision","confidential_remarks":"The paper is well executed and the numerical gains are credible, but the causal attribution to abstract skill content is the load-bearing point and it is currently weak. The 1.7B shuffled-skill control and the missing 4B shuffled control need to be addressed with new experiments or a substantially revised interpretation. I recommend major revision rather than rejection because the gap is localized and the experimental infrastructure is strong enough to settle it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a serious empirical paper: three model sizes, multi-seed hierarchical intervals, equal-FLOP GRPO baselines, component ablations, leakage audits, and a group-stratified mechanism test. The authors are also unusually candid about what they did not prove. Second, the central attribution claim is softer than the abstract sounds. The paper's own SKALD-shuffled condition at 1.7B reaches 48.83, which is +3.31 over GRPO and only -1.54 below matched SKALD. So roughly 68% of the 4.85-point gain survives with a randomly assigned card in the teacher context. The paper acknowledges that generic card exposure or regularization can explain part of the gain, but that part is a majority. And the 4B result, where the headline gain is +12.01, has no shuffled-card control at all. That is the soft spot, and it is load-bearing for the interpretation that the skill's content is what provides dense supervision.\n\nWhat is genuinely new: the construction of a skill card as privileged context in shared-parameter self-distillation, with an annealed tilt and an empirical gate, is not in the prior work I know. The tilted objective is honestly labeled as a known Rényi-type cross-entropy; the contribution is the annealed use under on-policy mismatch, and Lemma 1's gradient and stationarity claims are clean. The zero-variance group analysis is the best part: restricting distillation to those groups recovers 84.7% of the 1.7B gain, which directly supports the motivating failure mode. The compute accounting also looks fair, and the leakage audits, while not conclusive, are more thorough than what most papers in this area ship.\n\nMinor concerns: the gate is a noisy fixed heuristic with 14.6% sign reversal, though gate variants stay within a small range. The ordered ablation at 4B is a single addition order, so interactions are unknown. The paper itself flags the leakage limitation and the inability to eliminate it; that is not a hidden flaw. No code or supplementary materials were available, which matters for reproducibility but not for the verdict.\n\nThe math, derived results, and controls hold up. The empirical improvements appear real. What is not established is that the skill content, rather than card-induced regularization or residual leakage, drives the gains. That is a correctable gap: add a shuffled-card control at 4B, report the zero-variance group result for shuffled cards, and tighten the interpretation accordingly.\n\nBottom line: this deserves a serious referee. I would accept it for review with the expectation of a major revision focused on the attribution controls. I would not cite it as evidence for the abstract-skills mechanism until that gap is closed, but I would cite the zero-variance finding as a useful empirical contribution.","headline":"A strong, honestly-scoped empirical paper on skill-conditioned self-distillation, with a real attribution gap at its core: shuffled cards reproduce most of the 1.7B gain and the 4B headline lacks that control.","tokens_in":14531,"tokens_out":704,"would_cite":true,"duration_ms":8408,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SKALD distills abstract skills into model weights, not prompts, and improves math reasoning over GRPO by up to 12 points.","keywords":["knowledge distillation","self-distillation","privileged information","mathematical reasoning","reinforcement learning with verifiable rewards","group-relative advantage","Rényi cross-entropy"],"falsifier":"Run SKALD with matched-format cards whose content is only boilerplate or generic reasoning advice with no problem-specific principle or common-mistake content; if accuracy stays near the 50.37 matched-card score rather than falling toward the 48.83 shuffled-card score, the skill-specific content is not the driver. A complementary test is an adversarial probe that measures answer recovery from the teacher-conditioned distribution at mid-training checkpoints: if matched-skill recovery climbs toward the 94.2% full-solution rate, the leakage channel is live and the central mechanism is compromised.","tokens_in":13441,"feed_emoji":"🧮","tokens_out":11898,"duration_ms":92606,"temperature":0.7,"pith_summary":"Reinforcement learning with verifiable rewards assigns zero group-relative advantage to rollout groups whose sampled answers are uniformly correct or uniformly wrong, and these groups make up 63–68% of training groups in the paper's runs. SKALD answers that blind spot by running the same model in two contexts: a question-only student and a teacher that additionally sees an abstract, explicit-answer-filtered skill card stating a principle, when to use it, and common mistakes. The student is trained on its own rollouts toward the teacher's distribution, so the skill's effect lands in the shared weights and no skill or teacher is needed at test time. Across five held-out mathematics benchmarks, SKALD raises average accuracy at eight samples over the GRPO (group-relative policy optimization) baseline by +2.46 at 0.6B, +4.85 at 1.7B, and +12.01 at 4B parameters. The paper presents this as evidence that abstract skills supply dense supervision precisely where group-relative rewards are silent.","feed_headline":"Distilled skills beat prompts in math by up to 12 points","feed_subtitle":"Answer-free skill cards raise avg@8 over GRPO by +2.5 to +12 across three model sizes.","key_machinery":"The load-bearing object is the annealed exponentially tilted cross-entropy $L_n^{(\\tau)} = -\\frac{1}{\\tau}\\log \\mathbb{E}_{v\\sim q_n}\\left[p_n(v)^\\tau\\right]$ between the stop-grad teacher distribution $q_n$ and the student distribution $p_n$. For $\\tau>0$ the effective target is the escort $r_n^{(\\tau)}(v)\\propto q_n(v)p_n(v)^\\tau$, which discounts teacher-preferred tokens the student currently finds very unlikely; the paper proves the logit-gradient identity $\\nabla_{z_n}L_n^{(\\tau)}=p_n-r_n^{(\\tau)}$ and that the stationary solution is $p^*_n\\propto q_n^{1/(1-\\tau)}$. Annealing $\\tau$ from 0.8 to 0 recovers teacher cross-entropy and, under stop-gradient, the forward-KL student gradient. Two further parts carry the method: the shared-parameter two-context setup (teacher sees the skill card, student sees the question only, all rollouts come from the student), and the empirical gate $g_K(x)$ that activates distillation only when sampled initial teacher accuracy exceeds student accuracy.","core_discovery":"The paper's central claim is that a compact, answer-free skill abstraction is an effective privileged signal for on-policy self-distillation: conditioning the teacher branch on the skill card shifts its next-token distribution in a way that improves reasoning, and distilling that shift into the question-only student improves mathematical performance beyond what group-relative RL alone achieves. The decisive evidence is the group-stratified experiment: zero-variance-only distillation at 1.7B recovers 84.7% of the full gain, and the full method stays +4.06 above FLOP-matched GRPO and +3.77 above contextual skill exposure with the skill withdrawn at test time. The authors also claim that the annealed tilt stabilizes training, that the gain is not primarily answer leakage (matched-skill answer recovery is 5.1% versus 94.2% for a full solution), and that the improvement across 0.6B, 1.7B, and 4B is an empirical trend, not a scaling law.","pith_inferences":["If the mechanism generalizes beyond mathematics, the same recipe applies wherever a cheap privileged context (a rubric, a hint, a tool trace) shifts a shared model's distribution favorably: gate on a sampled advantage, distill with an annealed tilt, and drop the privileged context at inference.","The 14.6% sign-reversal rate of the fixed gate suggests an adaptive or confidence-weighted gate is the natural next experiment; removing false positives could push the 1.7B result above 50.37.","Shuffled cards already give +3.31 over GRPO (versus +4.85 for matched cards), so an editor's reading is that a meaningful part of the gain may be generic card conditioning or regularization rather than skill-specific content; disentangling these two components is the open question the paper's single ablation order does not settle.","The monotonic gain across 0.6B, 1.7B, and 4B, though explicitly not claimed as a scaling law, makes the largest-scale test the cheapest high-information next step: if the gap keeps growing, skill-card distillation becomes a compute-effective substitute for a larger online teacher."],"forward_implications":["Zero-variance rollout groups are where SKALD earns its keep: at 1.7B, running distillation only on groups whose sampled rewards are all identical recovers 84.7% of the full gain over GRPO.","Skills can be internalized rather than prompted: under a no-skill-at-test protocol, SKALD's 50.37 avg@8 beats contextual skill exposure with withdrawal by +3.77 at 1.7B.","The margin survives honest compute accounting: equal-FLOP and equal-GPU-hour comparisons leave SKALD +4.06 ahead at 1.7B and +11.19 ahead at 4B.","Leakage is an unlikely explanation for the headline gain: matched-skill answer recovery is only 5.1% versus 94.2% for a full solution, and sanitized cards change the 1.7B score from 50.37 to 50.24.","The annealed tilt is what makes training usable: risk-neutral teacher cross-entropy diverges in 2/3 seeds at 1.7B and 3/3 at 4B, while all tilted runs converge."],"supporting_citations":[{"why":"Defines the group-relative REINFORCE (GRPO) algorithm that serves as the baseline and the zero-advantage failure mode SKALD addresses.","marker":"(Shao et al. 2024)"},{"why":"Supplies OPSD, the full-solution on-policy self-distillation baseline that SKALD generalizes with an abstract skill card.","marker":"(Zhao et al. 2026a)"},{"why":"Provides the Qwen3-Base checkpoints used as the shared student and teacher branches.","marker":"(Yang et al. 2025a)"},{"why":"Contributes MATH500, one of the five held-out evaluation benchmarks.","marker":"(Lightman et al. 2024)"},{"why":"Contributes Minerva, another held-out benchmark in the evaluation suite.","marker":"(Lewkowycz et al. 2022)"},{"why":"Supplies DAPO-Math-17k, the training data source used alongside OpenThoughts.","marker":"(Yu et al. 2026)"},{"why":"Establishes generalized knowledge distillation on the student's own samples, the on-policy distillation foundation SKALD extends with a privileged teacher context.","marker":"(Ye, Lu, and Zhan 2022)"},{"why":"Supplies AMC23, a held-out benchmark in the 872-problem evaluation suite.","marker":"(Zhang and Math-AI 2023)"},{"why":"Supplies AIME24, a held-out benchmark in the evaluation suite.","marker":"(Zhang and Math-AI 2024)"},{"why":"Supplies AIME25, a held-out benchmark in the evaluation suite.","marker":"(Zhang and Math-AI 2025)"}],"fun_headline_variants":["Skills, not prompts: distillation lifts math by 12 points","Abstract skills beat prompts: up to +12 in math distillation","Distill skills into weights, not prompts, for +12 math gain","Skills over prompts: up to +12 avg8 in math distillation","When rewards go silent, skill distillation wins math by 12"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The interpretation that the gain comes from abstract skills rather than from answer leakage or a generic regularization effect rests on the assumption that the answer-filtered skill card changes the teacher's distribution on reasoning-relevant positions without transferring the answer; the paper's audits reduce but cannot logically eliminate this possibility.","fun_headline_variants_meta":{"raw":{"variants":["Skills, not prompts: distillation lifts math by 12 points","Abstract skills beat prompts: up to +12 in math distillation","Distill skills into weights, not prompts, for +12 math gain","Skills over prompts: up to +12 avg8 in math distillation","When rewards go silent, skill distillation wins math by 12"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3089,"prompt_tokens":1048,"completion_tokens":2041,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":1962}},"tokens_in":664,"tokens_out":2041,"duration_ms":14530,"temperature":1.0,"reasoning_tokens":1962,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:52:04.410540+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SKALD with matched-format cards whose content is only boilerplate or generic reasoning advice with no problem-specific principle or common-mistake content; if accuracy stays near the 50.37 matched-card score rather than falling toward the 48.83 shuffled-card score, the skill-specific content is not the driver. A complementary test is an adversarial probe that measures answer recovery from the teacher-conditioned distribution at mid-training checkpoints: if matched-skill recovery climbs toward the 94.2% full-solution rate, the leakage channel is live and the central mechanism is compromised.","supporting_citations":[],"review_version":1}