acceptodds
Under review as a conference paper at ICLR 2027

SURGE: Sparse Update Routing via Gated Entropy for Efficient Active Distillation

Abstract

On-policy distillation (OPD) is widely used to transfer capabilities from large teacher language models to smaller student models. However, supervising every token incurs substantial computational cost without distinguishing where guidance is truly needed and can induce unnecessary updates that disrupt existing student capabilities. To address these limitations, we introduce , an active distillation framework that provides selective control over student adaptation and knowledge transfer by using two separate gating mechanisms for updates. A token gate applies teacher supervision only at high-entropy positions, while a step gate forces updates to remain inside low-scoring generation steps. Additionally, a decoupled KL objective is used to apply updates selectively on such active tokens while retaining regularization toward a frozen reference student over the entire sequence. Across the evaluated reasoning, scientific knowledge, and hate speech tasks, SURGE achieves the highest reported accuracy among the compared baselines. When distilling Gemma4-31B into Gemma4-E4B, SURGE improves over the base student by a relative on average and achieves a mean recovery of of the initial teacher–student accuracy gap. Relative to OPD, estimated distillation FLOPs fall by on average, and the share of that gap SURGE closes is larger. Average response length falls by , while the mean format-failure rate falls by relative to OPD. Finally, we assess the generalization of our method by distilling Qwen3.5-122B-A10B into Qwen3.5-2B on these benchmarks, also noting improvements across accuracy and training FLOPs.

Primary area:
reinforcement learning
Keywords:
Distillation, On-Policy Distillation, Knowledge Distillation, LLMs, Machine Learning, Artificial Intelligence, AI, Efficiency, Language Model, Reasoning, Reinforcement Learning, Post Training

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JZFDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

SURGE is an on-policy distillation method that applies teacher log-probability feedback only at tokens that pass two gates: a low process score on newline-delimited steps, and high student next-token entropy. A frozen copy of the initial student supplies the step scores and a response-wide reference KL term. The main experiments distill Gemma4-31B into Gemma4-E4B; a second pairing distills Qwen3.5-122B-A10B into Qwen3.5-2B. The paper reports higher point-estimate accuracy than dense OPD at lower estimated distillation FLOPs, with gate, masking, and reference-KL ablations.

Critical strengths

1.The Gemma accuracy gains versus dense OPD are internally consistent. On the five splits in Table 5, SURGE’s reported accuracy is higher than the paper’s dense OPD run on every task, and the abstract’s 8.61%, 40.93%, 49.89%, and 66.60% figures recompute from Table 5 under Appendix G. That at least shows the main table is not an arbitrary collage, and that restricting teacher supervision to selected positions can produce a reported accuracy increment in this setting.

2.Selective teacher supervision is a useful empirical question. The recipe is restatable: student on-policy tokens, teacher log-probabilities only where the gates fire, student-only inference. The claimed compute saving should be marked unverified rather than counted as a demonstrated efficiency result.

Critical weaknesses

1.Empirical validity and computational attribution are not established. The efficiency claim is that gated entropy reduces teacher compute. Appendix D already states that the supervision mask is not a multiplier on teacher-body cost: the teacher still forwards the full prefix at selected positions, while the output projection may be counted only there. Table 2’s densest task, Competition Math (63.16% active), is the cheapest versus OPD (−74.9%), while the sparsest, DynaHate (28.51% active), saves only 35.6%. Cross-task sparsity-versus-saving is not a causal identification, because sequence length, update count, and the body versus vocabulary-projection mix can differ. Evaluation lengths of 334 vs 342 are not training FLOPs. Final configurations are written as not retained. The paper therefore does not show, under matched training budgets, that the reported FLOP drop is an effect of sparse routing.

The evaluation populations also change how the numbers should be read. Table 10 lists GPQA Diamond as 178 train + 198 evaluation; the public Diamond set is typically 198 items. Competition Math is cited as Hendrycks MATH (7,500/5,000) but run as 10,000/2,500. MMLU-Pro is split 9,625/2,407 from the official 12,032-item pool. A documented, disjoint custom split used by every compared method can be a valid experiment; it is not the standard leaderboard. The sharpest numerical issue is the Qwen GPQA figures 42.50 / 52.50 / 77.50: a single equal-weight binary score on 198 items would imply 84.15 / 103.95 / 153.45 correct, none of which is an integer. That is a denominator inconsistency (another N, a mean over runs, or a reporting error) that the authors should resolve; it is not, as it stands, confirmed contamination.

2.The proposed mechanism is not isolated. The micro-gate H > τ drops low-entropy, high-divergence positions. TIP, which this paper cites, treats both high-entropy tokens and low-entropy tokens with large teacher–student disagreement as informative, so SURGE keeps the first region and misses the second; that is incomplete coverage, not the opposite of TIP. The tables also do not show that high-entropy selection alone retains the dense corrective signal. The macro-gate scorer is the frozen initial student, with no gold labels or external verifier (§3.1, Appendix E.1). That is self-scoring or a weak process-quality proxy: the live questions are whether it carries information beyond student confidence, whether it reinforces the student’s own errors, and how much it adds over an entropy-only gate — not whether the name PRM is allowed. Strict masking in Table 3 also replaces the reduction denominator Z with A (Appendix E.2), mixing coverage with normalization. In Table 4, β=0 drives Math and MMLU accuracy to 0 while active tokens drop to 0. Collapse is evidence that this recipe needs the reference term to keep teacher feedback alive; it does not yet separate a generic reference regularizer from a SURGE-specific gating–regularization interaction. Related work lists SE-KD, TIP, TA-OPD, and TGOPD, but the tables compare only dense OPD, SFT, GRPO, and a self-score PRM that also uses fewer mean FLOPs than SURGE (8.74e16 vs 1.89e17).

Questions for authors

1.Under matched training configs (same length caps, epochs / update counts, and the same FLOP decomposition), can retained run logs rather than approximate ranges show that the FLOP gap versus OPD comes from fewer teacher output positions rather than other training-cost differences? Does Appendix D separate teacher-body FLOPs from output-projection FLOPs counted only at selected positions?

2.How were the 178 GPQA Diamond training items chosen relative to the 198 evaluation items and to the public 198-item Diamond set? Are train and eval disjoint? The Qwen accuracies 42.50 / 52.50 / 77.50 cannot be a single equal-weight binary score on 198 items; is the denominator 200, 80, or a mean over runs?

3.The micro-gate omits TIP’s second region (low-entropy, high-divergence). Is there a matched-teacher-budget comparison to entropy-only selection, or to putting that second region back? How much independent gain does the macro-gate add over entropy-only?

4.Can the strict-mask versus SURGE gap be rerun with the same reduction (Z versus A) so that only inactive-position reference coverage changes? For Table 4, if β=0 is not allowed to drive the active set to empty, is the reference term still necessary — and is the collapse a generic regularizer failure or an interaction with this gate?

2 comments

Thank you for checking the Gemma numbers and for pressing us on the compute accounting and GPQA split. We appreciate your point that sparse teacher feedback is worth testing under matched settings.

W1 / Q1. Do fewer active positions explain the reported FLOP gap?

We measured the reduction in the output projection on identical inputs. On 32 Qwen3 MATH rollouts, both implementations processed 9,216 teacher context positions. Gathering before projection reduced the output rows from 6,088 to 571 (J.1). Appendix D counts teacher context processing and output projection separately.

On public GPQA, we matched the 128 rollouts, 32 updates, optimizer and 256-token training cap. The table gives means across three seeds.

MethodCorrect (%)Teacher rows per runEstimated work (10¹² ops)
Dense OPD20.3715,897362.68
SURGE, eager20.373,072507.38
SURGE, quota shortcut and cache20.373,072373.19

SURGE projected 80.68% fewer teacher rows at the same mean accuracy. The shortcut and cache cut eager SURGE's estimated work by 26.45% and preserved all 96 logged objectives and gradient norms, and all per-question grades. Total work was still 2.90% above dense OPD (J.7, J.11). We report 74.90% for MATH and 37.25% for GPQA as original run-level differences. The share attributable to routing remains unresolved. We agree that the original self-scoring PRM is cheaper (8.74 × 10¹⁶ versus 1.89 × 10¹⁷ reported mean FLOPs), with SURGE 2.81 accuracy points higher on average (I.2).

W1 / Q2. How were the GPQA splits and denominator determined?

The source and overlap of the original 178 training items remain unresolved. We agree that 42.50%, 52.50% and 77.50% are incompatible with one equal-weight binary evaluation on 198 questions. The original Qwen3.5 denominator remains unresolved, and we exclude that row from the cross-family summary (I.2).

For the new runs we removed Diamond from GPQA Main by record ID and normalized question text, trained on 128 Main questions and evaluated all 198 Diamond questions. SURGE got 37, 42 and 42 correct across the three seeds, and dense OPD got 40, 39 and 42 (J.7). The original Competition Math (10,000/2,500) and MMLU Pro (9,625/2,407) use the study splits in Appendix C.

W2 / Q3. What does the macro-gate add, and what about low-entropy disagreement?

On labeled steps, the Gemma student scored 0.828 AUROC versus 0.499 for entropy-based confidence. Its fixed cutoff caught 32 of 96 incorrect steps. Whether specific missed errors persist through training remains open. On Qwen3, the student scored 0.509 versus 0.515 for confidence (J.13). In the sampled GPQA runs, the joint gate changed nine of 9,216 top-entropy selections and scored 20.37% versus 21.55% for entropy. Ranking scored 24.07% and shuffling scored 23.91%, a difference of +0.17 points [−2.69, +3.03] (J.11).

With the sampled loss on Gemma, ranking beat shuffling by 8.85 points on boxed accuracy. We kept those grades and recovered answers from uncapped responses missing a box. The difference moved to −1.30 points [−5.47, +2.86], showing how much the measured advantage depended on the requested format. Under shared categorical reverse KL, this check narrowed Gemma's ranking deficit against entropy from 35.16 to 2.08 points and against TIP from 35.94 to 2.60 points. On GPQA, ranking still trailed entropy, TIP and shuffling under both grading rules (J.16).

We agree that the entropy threshold misses confident disagreements. We reserved 19 of 96 GPQA positions per update for low-entropy tokens and compared forward-KL ranking with uniform draws from the same region. They scored 22.05% and 20.88%, respectively (difference +1.18 points [−1.52, +3.87]). Ranking needed roughly five times the teacher projection rows of eager SURGE (J.14). J.14 and J.15 also report SE-KD, TA-OPD and TGOPD. TIP

W2 / Q4. Does the reference term still help when the gate stays active?

With the same valid-token denominator ZZ and 192 teacher positions per update on Qwen3 MATH, full coverage scored 33.85% and active-only coverage scored 37.24% (difference −3.39 points [−7.81, +1.04]). Mean pre-clipping norms were 0.486 and 0.480, and both arms stayed below the clipping limit. The β=0\beta=0 sparse arm kept the full quota in every update and scored 35.16%. Dense OPD scored 34.64% with reference regularization and 35.42% without it. Enforcing the quota avoided the original collapse in this setup. The difference between the sparse and dense reference effects was −0.52 points [−5.47, +4.43] with signs varying across seeds (J.8). The intervals resample questions with training seeds fixed.

We hope you will reconsider your score in light of the revised compute claim and the new controls. Please let me know if there are any other problems you'd like me to address before the discussion period ends.

Reviewer_JZFDOpenReview ↗

Follow-up on compute accounting and the new GPQA runs

Thank you for the matched GPQA runs and the admissions on the original 178 split and Qwen denominators. I am keeping Rating 1. The new results recast the two central claims rather than supporting them.

Under the matched GPQA protocol, SURGE matches dense OPD at 20.37% (37/42/42 vs 40/39/42) but does not save work: 15,897 teacher rows / 362.68 vs 3,072 / 507.38 (eager) or 373.19 (shortcut). Fewer output rows is not a FLOP saving. You already write that the share attributable to routing remains unresolved.

The dual-gate controls also do not isolate the method. The joint gate changed 9 of 9,216 entropy selections (20.37% vs entropy 21.55%). Ranking vs shuffle is +0.17 [-2.69, +3.03]. With a shared ZZ, the reference term is not needed (33.85% vs 37.24%). The original 178 overlap and Qwen 42.50 / 52.50 / 77.50 remain unresolved.

Three questions:

  1. Do any logs remain for the original Gemma Table 5 runs? If not, will the 49.89% abstract saving be recast as an unmatched run-level difference?
  2. Please post the truncated J.14 / J.15 numbers (SE-KD, TA-OPD, TGOPD, TIP) under the same protocol.
  3. Is the quota shortcut and cache in the submitted method, or only in the rebuttal?
Reviewer_qF6VOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes SURGE, a selective on-policy distillation method that combines step-level gating based on step quality with token-level gating based on token uncertainty. Teacher supervision is applied only to high-entropy tokens within low-scoring generation steps. An additional KL regularization is applied across the entire response to prevent drifting too far from the reference policy. Experimental results show improved performance over the compared baselines.

Critical strengths
  1. The presentation is generally clear and the method is easy to follow.
  2. The proposed method is computationally efficient.
Critical weaknesses
  1. The novelty of the proposed method is limited since all components of the method are quite familiar in the space.
  2. In the experiments, there's no comparison to other selective distillation methods. Thus, the relative effectiveness of the proposed method compared to existing techniques is not clear.
Questions for authors
  1. Currently, the step quality is measured by the frozen reference student. Why not scoring the steps using the teacher model instead?
  2. The deterministic checks in process scoring appears ad hoc. Why checking those three conditions instead of considering more or less conditions? Why are scores assigned in the given way (e.g., malformed JSON receives a higher score than a repetitive step)?
  3. Why does rrawr_{raw} need normalization (as in equation 10) while drefd_{ref} doesn't?
2 comments

Thank you for taking the time to review SURGE and for your comments on the clarity of the method. We appreciate your suggestions to test the teacher as a scorer and compare with other selective distillation methods.

Q1. Why score steps using the frozen student?

We ran both models on the same 192 human-labeled PRM800K steps, with the same judging prompt within each pair (J.13).

Model pairStudent AUROCTeacher AUROCNegative mean entropy
Gemma 40.8280.6310.499
Qwen30.5090.6910.515

The Gemma student's advantage over the teacher was 0.197 (paired 95% interval [0.113, 0.285]). The student scored similarly in both halves of the entropy split (0.834 and 0.832 AUROC). The Qwen3 teacher did better than its student. When we substituted the teacher scorer in the Qwen3 MATH pilot, the evaluation grades stayed the same and estimated scorer work was 3.2 times as high (J.2). For Gemma, the student gave better rankings with a smaller scoring model.

W1–W2. What does SURGE add, and how does it compare with selective baselines?

We agree that the ingredients are familiar. SURGE uses a frozen student's step scores to choose which uncertain positions get teacher feedback, and keeps reference regularization over the whole response. SE-KD also selects before teacher projection, using within-response entropy. TIP and TA-OPD use teacher distributions for selection, and TGOPD uses teacher probes to choose the objective for each prompt.

We added all four comparisons in the same trainer. At 96 positions per update on GPQA, SURGE scored 20.37%, SE-KD scored 21.55%, and the TA-OPD selector scored 23.23% versus 21.55% for entropy under the same sampled loss. TIP scored 24.75% under categorical reverse KL (J.14). On Qwen3 MATH, TGOPD scored 37.24% versus 38.02% for grouped dense OPD, with 32.40% more estimated work after counting its teacher probes (J.15). SE-KD, TIP, TA-OPD, TGOPD

We also tested whether the Gemma step ranking translates into better training outcomes. With the sampled loss, ranking beat shuffled scores by 8.85 boxed points and produced 92 versus 54 complete boxed answers. We kept the boxed grades and extracted answers from uncapped responses without a box. The difference became −1.30 points [−5.47, +2.86] (J.11). That comparison shows a gain in producing the requested format. In the crossed comparison in J.12, changing entropy-only from the sampled loss to categorical KL raised boxed accuracy from 13.54% to 77.60%, close to TIP's 78.91% with the same loss. With shared categorical KL on Gemma, the check recovered 129 correct unboxed answers for ranking and two for entropy. That narrowed the gap from 35.16 to 2.08 points. Ranking still trailed entropy and TIP on both tasks under either rule. It projected 89.88% fewer teacher rows than TIP on Gemma and 81.02% fewer on GPQA. Step scoring made total estimated work higher in both cases (J.16).

Q2. Why these checks, and why 0.05 for malformed JSON versus 0.02 for repetition?

The checks handle cases we can recognize directly from the output. Syntax-only lines receive 1.0 because there is no reasoning claim to correct. The JSON rule applies to tasks that ask for structured output. The scores for malformed JSON and repetition are heuristic. Both are below 0.5 and take the same routing branch. Their relative order can matter when scores are used directly, as in the process-reward baseline and the ranking variant. On MATH, removing syntax and repetition checks increased scorer calls from 381.7 to 554.3 per run with the same evaluation grades. Checks alone gave the same grades with entropy filling the quota (J.2).

Q3. Why normalize the teacher term?

The active set changes from batch to batch. Equation (10) centers the sampled teacher rewards on that set and divides by the larger of their standard deviation and one, controlling large changes in scale without amplifying small differences. The reference log-ratio is a sampled estimate of the KL penalty in nats. Keeping that scale gives β\beta the same meaning at every update.

We would be grateful if you would reconsider your score given the new comparisons and the scorer evaluation. Please let me know if there are any other problems you'd like me to address before the discussion period ends.

We thank the reviewer for their additional results and clarifications. However, the major concern lies in the lack of novelty, so I'll maintain my current score.

Reviewer_4X88OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SURGE is an on-policy distillation method that intersects a low-step-score gate with a high-student-entropy gate to select teacher-supervised tokens. A frozen copy of the initial student supplies process scores and response-wide reference regularization, while the larger teacher supplies sampled-token log-probability feedback at selected positions. The teacher term is masked but reference feedback is retained across valid response tokens. Experiments compare SURGE with SFT, GRPO, a process-reward baseline, and dense OPD on reasoning, knowledge, and hate-speech classification tasks, primarily using a Gemma teacher–student pair and additionally a Qwen pair. The paper reports higher mean accuracy and lower estimated distillation FLOPs than OPD, plus gate and loss ablations. Representation/copying probes describe adaptation but are appropriately limited as evidence of capability preservation.

Critical strengths
  1. The objective makes a useful distinction between sparse teacher supervision and reference anchoring. Equations (6)–(13) route teacher feedback only where both gates pass while retaining the reference term elsewhere. Section 2 explicitly describes this as a local surrogate rather than the exact sequence-level reverse-KL gradient, and Appendix E explains normalization and the one-update clipping behavior. These qualifications make the mathematical scope clearer than treating the method as exact sequence-level KL minimization.

  2. The reported accuracy–compute pattern is promising across several tasks and two model families. Table 5 shows improved Gemma accuracy over OPD on each listed task, and Table 6 mostly reproduces that direction with Qwen, with a GPQA tie. Appendix D explicitly includes teacher context processing, reference evaluation, and separate process-scoring work in its accounting, rather than equating inactive tokens with eliminated teacher computation. If confirmed under reproducible matched settings, this would be a practically meaningful result.

Critical weaknesses
  1. The central efficiency comparison is not sufficiently auditable or reproducible. The reproducibility statement and Appendix B explicitly acknowledge that exact final run settings were not retained. Table 9 provides ranges for epochs, learning rates, response caps, regularization, and optional length controls. These variables can change both accuracy and total computation. Although Appendix D defines the right workload categories, it does not report the component totals, actual processed token/update counts, or profiles needed to reconstruct Table 5. Thus the large reported savings, including 74.9% on Math, cannot be attributed confidently to the proposed selection scheme rather than differing training workloads or controls. This is a limitation of the evidence even though wall-clock speedup is correctly not claimed.

  2. The experiments do not isolate the value of the proposed routing and decoupling decisions. The single-gate ablations alter supervision quantity as well as selected positions. A matched-budget random mask or comparable entropy-only selective method is absent, so the advantage of choosing these particular tokens remains uncertain. In addition, strict masking changes both inactive-position reference feedback and the loss denominator (Section 4.5/Appendix E.2), while reference-removal comparisons do not isolate training-budget effects. The paper acknowledges these confounds, but they concern its principal design claims. With no quantified training-seed uncertainty, small differences such as 0.37 points on Gemma MMLU Pro and 0.48 points on Qwen Math are also difficult to interpret as reliable gains.

Questions for authors
  1. [Evidence request] Can you provide exact OPD/SURGE run configurations and actual training-token, rollout, and update counts, together with the five component FLOP totals in Eq. (14), for at least the main Math and GPQA comparisons? Please show how required teacher-prefix computation and repeated PRM prompts are counted. This would make the central efficiency result auditable without requiring wall-clock speedups.

  2. [Evidence request] Under identical rollout/update budgets and the same active-token count, how does SURGE compare with random routing and entropy-only routing? A comparison with the same loss reduction and regularization is needed to determine whether the joint gate selects more useful supervision rather than merely less supervision.

  3. [Evidence request] Can the strict-mask/reference-coverage ablation be repeated with a shared denominator and matched effective update scale, with uncertainty reported for the main comparison? This would distinguish a genuine benefit of inactive-token anchoring from normalization and optimization effects.

1 comment

Thank you for the careful read. We appreciate your point about the distinction between sparse teacher feedback and response-wide anchoring, and your suggestions for checking the compute and ablation results.

W1 / Q1. Can the FLOP savings be traced to sparse routing?

On the same 32 Qwen3 MATH rollouts, gathering selected positions before the vocabulary projection reduced teacher output rows from 6,088 to 571 (90.62%). Both paths processed 9,216 teacher context positions (J.1).

We also ran SURGE and dense OPD under matched training settings. Gemma MATH used 96 rollouts, 24 updates and a 512-token training cap. Qwen3 GPQA used 128 rollouts, 32 updates and a 256-token cap. Within each task we used the same starting checkpoint, prompts and optimizer. J.7 and J.9 give the observed token counts and the breakdown for Eq. (14), including teacher context and projection separately. We charge a scorer prefix each time it is computed.

Repeated step scoring outweighed the projection savings in the eager implementation. On GPQA, skipping unnecessary scores and caching shared prefixes cut estimated work by 26.45% versus eager SURGE. All 96 logged update objectives and gradient norms, and every evaluation grade, matched. The resulting workload was 2.90% above dense OPD. On Gemma MATH, the quota shortcut used 2.03% less estimated work than dense OPD and selected the same positions as top entropy on its rollouts (J.11). The original 74.90% MATH and 37.25% GPQA figures are run-level estimates. We have marked their attribution to routing as unresolved (I.1).

W2 / Q2. How does SURGE compare with random and entropy-only masks at equal active counts?

We used the same sampled loss, reduction over valid tokens and reference coefficient. These are three-seed means.

RoutingGemma MATH, 192 positions/updateQwen3 GPQA, 96 positions/update
SURGE, fixed quota12.76%20.37%
Entropy only13.54%21.55%
Random20.05%21.89%

The joint gate chose almost the same tokens as top entropy on its own rollouts. It selected zero positions outside the entropy quota on Gemma and nine on GPQA (J.7, J.9). To give the step score more influence, we ranked twice the quota of entropy candidates and compared with shuffling the step scores. On Gemma under the sampled loss, step ranking improved boxed accuracy by 8.85 points over shuffling, with gains in all three seeds (J.11).

We kept every boxed grade and rescored uncapped responses that lacked a box by extracting an answer from the ending. Under the sampled loss, step ranking produced 38 more complete boxes and 34 more correct boxed answers than shuffling. With unboxed answers counted, the difference was five answers in the other direction across 384 responses per arm (−1.30 points, 95% paired interval [−5.47, +2.86]). This points to a gain in following the requested answer format. The interval with extraction spans zero. With shared categorical reverse KL, ranking trailed entropy, TIP and shuffling. On Gemma the unboxed check narrowed the gap to entropy from 135 correct answers to eight (−2.08 points [−5.99, +1.82]). The GPQA ordering was the same under either grading rule (J.16).

W2 / Q3. What happens if the reference ablation uses the same denominator?

On Qwen3 MATH we kept ZZ, the 192-position quota, teacher reward, reference coefficient and optimizer fixed (J.8). The first update objectives and gradient norms matched in all three seeds. The mean pre-clipping norms were 0.486 and 0.480 for full and active-only coverage, with no clipping in either arm. Full coverage scored 33.85% versus 37.24% for active-only. The seed-wise differences were −6.25, −0.78 and −3.12 points, with a paired 95% interval of [−7.81, +1.04] over evaluation questions with seeds fixed. The original 0.37-point Gemma MMLU Pro and 0.48-point Qwen MATH gains remain point estimates.

We hope these comparisons address your questions and would appreciate your reconsidering your score. Please let me know if there are any other problems you'd like me to address before the discussion period ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.