acceptodds
Under review as a conference paper at ICLR 2027

AID: Asymmetric Intermediate-Answer Distillation for Budget-Constrained Long-CoT Inference

Abstract

Supervised fine-tuning (SFT) on long chain-of-thought (long-CoT) traces is a common way to transfer reasoning ability from a large teacher to a smaller student, but this transfer can be brittle when the inference-time reasoning budget is limited. The distilled students often fail to finish reasoning within the available budget, and forcing an answer from an unfinished state does not reliably recover the correct solution. We introduce AID (Asymmetric Intermediate-Answer Distillation), which retains supervision on the full teacher trace while additionally training the student to predict the final answer from randomly sampled intermediate states. AID uses an asymmetric attention mask to keep the auxiliary answer branch out of the reasoning stream's forward context, allowing full-CoT imitation and intermediate-state answerability to be learned jointly. AID requires neither additional data generation nor extra parameters, and leaves inference unchanged. Experiments with Qwen3-0.6B trained on OpenR1-Math and KodCode show that plain SFT degrades sharply on splits that combine long traces with harder problems, whereas AID substantially mitigates this degradation. Under long-trace distillation, accuracy improves from 22.9% to 37.0% on MATH-500 and from 25.6% to 35.6% on MBPP+. In a truncation-based evaluation on held-out validation problems, which reveals progressively longer prefixes of a reference reasoning trace, the normalized area under the accuracy curve rises from 0.191 to 0.419 on math and from 0.321 to 0.450 on code. Our results indicate that intermediate-state answerability provides a complementary training signal that makes long-CoT distillation more robust under constrained inference budgets.

Primary area:
foundation or frontier models, including LLMs
Keywords:
reasoning distillation, chain-of-thought, reasoning budget, small language models, large language models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_cvziOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Asymmetric Intermediate-Answer Distillation (AID) for transferring long-chain reasoning to small language models under limited inference budgets. AID retains supervision on complete teacher traces and adds an auxiliary copy of the final answer conditioned on a randomly sampled reasoning prefix. An asymmetric attention mask prevents the auxiliary answer from entering the original reasoning stream, while position reassignment preserves the positions of the original sequence. The method requires no additional teacher generations or model parameters and uses standard causal decoding at inference. Experiments on mathematical and coding tasks show substantial gains for small students trained on long traces. Analyses of generated reasoning and controlled teacher prefixes support improved answer recovery when reasoning is interrupted.

Critical strengths
  1. A practical training objective with substantial gains in its target setting. AID explicitly trains the student to answer from unfinished reasoning while retaining full-trace supervision. In the 0.6B Long-CoT setting, it improves all five benchmarks in Table 2, including MATH-500 from 22.9% to 37.0% and MBPP+ from 25.6% to 35.6%. These are three-seed averages using the same training data and evaluation budgets as SFT. The approach uses existing teacher traces and leaves the inference architecture unchanged.

  2. The analyses provide evidence beyond aggregate accuracy. Figure 3 shows that AID's gains do not come from more frequent reasoning completion. Appendix E further compares the same problem-seed pairs on which both methods exhaust the budget, with accuracy improving from 16.5% to 31.1%. The controlled-prefix evaluation holds the supplied reasoning content fixed and also finds improved answerability, with math RS-AUC increasing from 0.191 to 0.419. These complementary comparisons support the value of explicitly learning to answer after interruption.

Critical weaknesses
  1. The contribution of interior-state sampling is not fully separated from simpler answer supervision. In Table 3, endpoint-only sampling already achieves 36.6% final accuracy versus 37.0% for full AID, and RS-AUC of 0.383 versus 0.419. Thus, most of the MATH-500 improvement in this ablation is already present without sampling interior states. Table 10 also shows large differences in whether answers terminate within the 64-token answer budget. This is a useful learned behavior, but the current comparisons do not separate it from the benefit of supervising diverse intermediate reasoning states. Separate zero-prefix and full-prefix controls, or an ordinary causal objective combining full-trace SFT with prefix-to-answer examples at matched loss weights, would clarify what the sampling and masking design contributes. This comparison is also practically relevant given the approximately 2.06 to 2.23 times SFT step time reported for the current Long-CoT SDPA implementation in Appendix B.

  2. The practical budget dependence remains underexplored. The main Long-CoT downstream results use a single 16K reasoning budget. RS-AUC varies the supplied teacher prefix, which measures a different quantity from accuracy across budgets on the student's own generated reasoning. The paper establishes a clear benefit at the tested operating point, but the size of that benefit at shorter reasoning budgets or different answer budgets is unknown. Testing the same trained models across these budgets would better establish the range in which AID is useful.

Questions for authors
  1. How do zero-prefix-only and full-prefix-only auxiliary supervision compare with the endpoint mixture and full AID? Can a causal full-trace-plus-prefix-answer objective with matched answer-loss weights reproduce the gains? These comparisons would clarify the incremental contribution of interior sampling and the asymmetric construction.

  2. For the same Long-CoT-trained models, how does downstream accuracy change across shorter self-generated reasoning budgets? How sensitive is the AID-SFT difference to the 64-token math answer budget?

3 comments
AuthorsOpenReview ↗

Response to Reviewer cvzi

We thank the reviewer for the careful reading and the precise questions.

W1 / Q1. Prefix positions, matched causal objective, and cost.

(a) Prefix positions. Math Long, Qwen3-0.6B-Base, mean over 3 seeds.

\text{Auxiliary answer conditioned on} & \text{MATH-500} & \text{RS-AUC} \\\\ \hline \text{zero prefix only } (n=0) & .358 & .298 \\\\ \text{full prefix only } (n=L) & .331 & .302 \\\\ \text{cut biased early} & .371 & .421 \\\\ \text{cut biased late} & .376 & .407 \\\\ \text{uniform (AID)} & .369 & .421 \end{array}$$ As the reviewer noted, much of the final accuracy is obtained without interior states: zero-prefix supervision is close to AID on MATH-500 (p = .19). However, it cuts RS-AUC from .421 to .298 (p = .001), so interior states are what teach answering from partial reasoning. Full-prefix-only supervision is worse on both (MATH-500 .331, p = .014). Biasing the cut toward early or late positions changes little, so the uniform default is not fragile. (b) Matched causal objective. Full + Truncated SFT (answer-only) supervises only the answer of the truncated copy, so its supervised tokens and answer-loss weights match AID. At matched FLOPs it reaches MATH-500 .377 and RS-AUC .470, comparable to AID. The gain therefore comes from intermediate-answer supervision itself. As reported to Reviewer Ntui (W1), the causal variants differ from AID in that they lower the rate at which the student finishes its reasoning on its own, while AID stays at the SFT level. (c) Cost. In our implementation AID takes about 2x longer per step than SFT, because the custom mask cannot use fused attention kernels; we will state this as a limitation. Input tokens and peak memory are the same as SFT, while the causal baseline needs two sequences per example and about 1.7x the memory. **W1 (Table 10) / Q2. Answer termination vs. answer quality.** Increasing the answer budget $A$ barely helps SFT (0.6B, MATH-500, 3 seeds): $$\begin{array}{l|ccc} \text{Reasoning budget} & 1\text{K} & 4\text{K} & 16\text{K} \\\\ \hline \text{SFT}, A=64 & .099 & .208 & .247 \\\\ \text{SFT}, A=128 & .129 & .220 & .250 \\\\ \text{SFT}, A=256 & .135 & .225 & .250 \\\\ \hline \text{AID}, A=64 & .337 & .378 & .375 \\\\ \text{AID}, A=128 & .337 & .379 & .375 \\\\ \text{AID}, A=256 & .337 & .379 & .375 \end{array}$$ SFT mostly keeps reasoning after the forced `</think>` rather than running out of room. To separate termination from quality, we compare the two methods on the same problems: those where both produce an answer within the budget (3 seeds, 64-token answer budget). $$\begin{array}{l|ccc} \text{Model, budget} & \text{common problems} & \text{AID} & \text{SFT} \\\\ \hline \text{0.6B, 1K} & 82 & .514 & .343 \\\\ \text{0.6B, 2K} & 138 & .564 & .489 \\\\ \text{0.6B, 16K} & 270 & .455 & .432 \\\\ \text{1.7B, 2K} & 244 & .754 & .472 \\\\ \text{4B, 2K} & 250 & .936 & .865 \end{array}$$ On the same problems AID is more accurate, with the largest gaps under tight budgets and a small gap at 16K for 0.6B. The problems SFT answers are easier on average (MATH-500 level 3.01 at 2K vs. 3.44 overall), which explains why its accuracy over its own answered subset can look higher. The gain is thus not explained by termination alone. **W2 / Q2. Self-generated budgets.** We evaluated the students on their own reasoning truncated at budget $B$ (MATH-500, 3 seeds, $A=256$): $$\begin{array}{l|cccccc} B & 0 & 1\text{K} & 2\text{K} & 4\text{K} & 8\text{K} & 16\text{K} \\\\ \hline \text{0.6B SFT} & .012 & .135 & .176 & .225 & .263 & .250 \\\\ \text{0.6B AID} & .055 & .337 & .374 & .379 & .379 & .375 \\\\ \hline \text{1.7B SFT} & .009 & .191 & .275 & .434 & .536 & .566 \\\\ \text{1.7B AID} & .033 & .549 & .629 & .647 & .659 & .657 \\\\ \hline \text{4B SFT} & .007 & .257 & .489 & .688 & .809 & .857 \\\\ \text{4B AID} & .151 & .650 & .777 & .834 & .865 & .885 \end{array}$$ AID improves accuracy at every budget, and the gain grows as interruptions become more frequent. We now report this alongside RS-AUC, since the two measure different quantities. We will add these results to the final version.

Thank you for the additional experiments and clarification. The prefix-position ablations clarify the role of interior-state supervision, while the shared-problem analysis and evaluations across self-generated reasoning budgets address my concerns about answer quality and practical budget dependence. The matched causal baseline is particularly informative: its comparable performance supports intermediate-answer supervision as the main source of the gains and helps clarify the contribution of the asymmetric implementation. I will maintain my rating of 3 (Weak acceptance), with a stronger positive assessment following the rebuttal. Please include the numerical reasoning-completion comparison cited in your response to Reviewer Ntui, together with comparable wall-clock training measurements for AID and the causal baseline, in the revised manuscript. These results would make the trade-off between accuracy, reasoning completion, memory, and training time clearer.

AuthorsOpenReview ↗

Response to Reviewer cvzi

Thank you for the careful follow-up and the positive assessment. We will include the reasoning-completion comparison and comparable wall-clock training measurements for SFT, AID, and the causal baseline in the revised manuscript, so that the trade-off between accuracy, reasoning completion, memory, and training time is explicit.

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

This paper proposes AID, which adds final-answer supervision conditioned on partial reasoning to improve small models’ ability to answer after exhausting their reasoning budget. The method retains full chain-of-thought supervision and uses asymmetric attention to prevent auxiliary answers from leaking into the reasoning process. Experiments show improved accuracy on math and code tasks.

Critical strengths
  1. Clear motivation: Standard distillation mainly supervises answers after complete reasoning, leaving answering from interrupted reasoning states insufficiently trained.

  2. Simple method: AID reuses existing training data and requires no additional model parameters.

  3. Clear empirical gains: AID substantially improves small models’ accuracy on math and code tasks under long-CoT distillation.

Critical weaknesses
  1. Limited direct comparisons: The paper lacks a comparison with standard SFT on truncated reasoning followed by the correct answer, leaving the specific contribution of the asymmetric attention design unclear.

  2. Limited scope: The main gains are concentrated in small Qwen3 models under long-CoT distillation, with smaller gains under short-CoT conditions and for larger models.

  3. Higher training cost: The custom attention mask approximately doubles the per-step training time compared with standard SFT in the current implementation.

Questions for authors
  1. How much additional benefit does AID provide over standard causal SFT trained directly on “question + truncated reasoning + correct answer”?

  2. Could increasing the final-answer loss weight in standard SFT to match AID’s relative answer supervision produce similar gains?

  3. How would sampling reasoning cut points more frequently near the beginning or end of the trace, rather than uniformly, affect performance?

  4. Are there results on other model families that support the method’s generality?

  5. Does training only on math or only on code degrade performance in the other domain? How does the degree of degradation compare between AID and standard SFT?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Svue

We thank the reviewer for the positive assessment and the precise questions. All results are means over 3 seeds.

Q1 / W1. Benefit over standard causal SFT on truncated reasoning.

We ran two versions: Truncated-only SFT ("question + first nn reasoning tokens + </think> + answer", the literal reading) and Full + Truncated SFT (the same truncated copy added to each full trace). Qwen3-0.6B-Base; self-stop is the fraction of MATH-500 problems where the model ends its reasoning within the budget.

& \text{Method} & \text{MATH-500} & \text{RS-AUC} & \text{Self-stop} \\\\ \hline \text{Long (16K)} & \text{SFT} & .229 & .158 & .101 \\\\ & \text{AID} & .369 & .421 & .097 \\\\ & \text{Truncated-only SFT} & .402 & .476 & .006 \\\\ & \text{Full + Truncated SFT} & .384 & .458 & .049 \\\\ \hline \text{Short (4K)} & \text{SFT} & .351 & .258 & .597 \\\\ & \text{AID} & .374 & .506 & .621 \\\\ & \text{Truncated-only SFT} & .381 & .545 & .170 \\\\ & \text{Full + Truncated SFT} & .373 & .523 & .510 \end{array}$$ Both baselines match or exceed AID in accuracy, which shows that the gain comes from supervising answers from intermediate states. They differ in how they change the student's reasoning: Truncated-only SFT almost never finishes its reasoning on its own (self-stop .006 on Long), and Full + Truncated SFT halves this rate, while AID stays at the SFT level. AID also keeps one sequence per example and the same peak memory as SFT, whereas Full + Truncated SFT needs about 1.7x the memory (see our response to Reviewer Ntui, W1). **Q2. Larger answer-loss weight in standard SFT.** SFT with the final-answer loss weighted ×2 and ×10, using the same trainer as AID: $$\begin{array}{l|cc|cc} \text{Answer weight} & \text{Long MATH-500} & \text{Long RS-AUC} & \text{Short MATH-500} & \text{Short RS-AUC} \\\\ \hline \text{SFT}, \times 1 & .246 & .178 & .364 & .291 \\\\ \text{SFT}, \times 2 & .285 & .191 & .365 & .291 \\\\ \text{SFT}, \times 10 & .317 & .199 & .350 & .281 \\\\ \text{AID} & .369 & .421 & .374 & .506 \end{array}$$ Upweighting helps on Long but stays well below AID, and RS-AUC barely moves. What matters is supervising the answer from intermediate states, not the amount of answer supervision. **Q3. Non-uniform cut sampling.** Biasing the cut toward the beginning or the end changes little (MATH-500 / RS-AUC): $$\begin{array}{l|cc} & \text{Long} & \text{Short} \\\\ \hline \text{early} & .371 / .421 & .371 / .505 \\\\ \text{late} & .376 / .407 & .381 / .490 \\\\ \text{uniform (AID)} & .369 / .421 & .374 / .506 \end{array}$$ **Q4. Other model families.** DeepSeek-R1-Distill-Llama-8B (LoRA, Math Short): $$\begin{array}{l|ccc} & \text{MATH-500} & \text{RS-AUC} & \text{Self-stop} \\\\ \hline \text{SFT} & .831 & .450 & .788 \\\\ \text{AID} & .849 & .561 & .789 \end{array}$$ AID raises RS-AUC substantially (p = .001) while keeping self-stop unchanged. The MATH-500 gap is small because this model finishes most of its reasoning within the budget. **Q5. Cross-domain effects.** Qwen3-0.6B trained on one QA task and evaluated on all three (* = in-domain): $$\begin{array}{ll|ccc|c} \text{Train} & & \text{ScienceQA} & \text{Coding MCQ} & \text{LSAT} & \text{RS-AUC} \\\\ \hline \text{ScienceQA} & \text{SFT} & .903^* & .456 & .321 & .773 \\\\ & \text{AID} & .917^* & .489 & .318 & .934 \\\\ \text{Coding MCQ} & \text{SFT} & .658 & .633^* & .357 & .474 \\\\ & \text{AID} & .650 & .642^* & .366 & .840 \\\\ \text{LSAT} & \text{SFT} & .592 & .466 & .437^* & .672 \\\\ & \text{AID} & .581 & .469 & .435^* & .773 \end{array}$$ Out-of-domain accuracy is comparable to SFT, so AID does not degrade transfer, and in-domain RS-AUC is consistently higher. **W2. Scope.** With tighter budgets, the 4B gain reaches +39 points at 1K tokens (response to Reviewer Ntui, W2), and Q4 shows the effect on another family. **W3. Training cost.** In our implementation AID takes about 2x longer per step than SFT, because the custom mask cannot use fused attention kernels; we will state this as a limitation. Input tokens and peak memory match SFT. With LoRA on Llama-8B, training time was the same as SFT (2.1 h). We will add these results to the final version.

Thank you for your thoughtful rebuttal. Your responses addressed my concerns, and I lean toward acceptance. I will maintain my original score, which I believe still reflects my overall assessment of the paper.

AuthorsOpenReview ↗

Response to Reviewer Svue

Thank you for the follow-up and for your support. We will incorporate the new comparisons and analyses into the final version.

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

This paper proposes Asymmetric Intermediate-Answer Distillation (AID), which augments full chain-of-thought supervision with an auxiliary objective that predicts the final answer from randomly sampled teacher reasoning prefixes. An asymmetric attention mask prevents the auxiliary answer from leaking into the original reasoning stream, and experiments show improved answer recovery under fixed reasoning budgets, with the largest gains on 0.6B and 1.7B students that frequently fail to finish reasoning.

Critical strengths
  1. The method reuses existing teacher traces and answers without requiring additional teacher generation, model parameters, or an auxiliary branch at inference. The attention design allows intermediate-answer supervision to be added while preserving the information available to the original reasoning stream.
  2. The paper goes beyond reporting final accuracy by examining reasoning completion and answer recovery from shared teacher prefixes. The matched comparison on problems where both models exhaust their budgets provides useful evidence that the improvement is not entirely explained by differences in which problems each model completes.
Critical weaknesses
  1. Although the paper discusses related work on anytime reasoning and prefix-based distillation [1, 2], its experiments do not sufficiently distinguish AID from these approaches or from simpler auxiliary-supervision schemes. In particular, a comparison with ordinary causal SFT combining full traces and truncated-prefix answer supervision is needed to establish the value of the asymmetric design.
  2. The strongest gains occur on 0.6B and 1.7B students that frequently exhaust their reasoning budgets, while improvements at 4B are much smaller under the evaluated budget. Since the comprehensive benchmark comparisons focus on 0.6B, the practical benefit for stronger students remains insufficiently established.
  3. The controlled-prefix evaluation covers several interruption positions, but it uses teacher traces rather than self-generated reasoning. Unlike evaluations across multiple generation budgets in anytime reasoning [1], the downstream experiments mainly use one budget per training condition, leaving robustness to earlier interruptions and different deployment budgets unclear.
Questions for authors
  1. Can ordinary causal SFT combining full-trace and randomly truncated-prefix answer supervision achieve comparable results with matched prompts, supervision, and training costs? How does AID compare with, or complement, anytime reasoning and adaptive prefix distillation [1, 2]?
  2. Does AID provide meaningful gains on 4B or larger students when tighter budgets make interruptions more frequent? Results across several budgets and benchmarks would help distinguish dependence on model capacity from dependence on budget exhaustion.
  3. How do accuracy and answer-completion rates change when self-generated reasoning is interrupted at different token budgets, including zero? With the same budget-forcing protocol [3] applied to all methods, this would clarify whether AID consistently improves answer recovery beyond the current evaluation limit.

[1] Qi P, Liu Z, Pang T, et al. Optimizing anytime reasoning via budget relative policy optimization[J]. Advances in Neural Information Processing Systems, 2026, 38: 23429-23451. [2] Liu Z, Wu Z, Li X, et al. Long-chain reasoning distillation via adaptive prefix alignment[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 18041-18059. [3] Muennighoff N, Yang Z, Shi W, et al. s1: Simple test-time scaling[C]//Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025: 20286-20332.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Ntui

We thank the reviewer for the constructive feedback.

W1 / Q1. Comparison with ordinary causal SFT and related work.

We ran this baseline (Full + Truncated SFT): standard causal attention, trained with SFT on each full trace plus a randomly truncated copy ("question + first nn reasoning tokens + </think> + answer"). Prompts, data, cut distribution, and evaluation match AID, and FLOPs are matched by reducing steps. Qwen3-0.6B-Base, mean ±\pm std over 3 seeds. Self-stop is the fraction of MATH-500 problems where the model emits </think> within the budget.

& \text{Method} & \text{MATH-500} & \text{RS-AUC} & \text{Self-stop} & \text{Input tok. (M)} & \text{Sup. tok. (M)} & \text{Peak mem. (GB)} \\\\ \hline \text{Long (16K)} & \text{SFT} & .229 \pm .029 & .158 \pm .055 & .101 \pm .007 & 119.5 & 117.0 & 45.6 \\\\ & \text{AID} & .369 \pm .013 & .421 \pm .002 & .097 \pm .016 & 119.7 & 117.2 & 45.6 \\\\ & \text{Full + Truncated SFT} & .384 \pm .023 & .458 \pm .004 & .049 \pm .008 & 119.4 & 116.1 & 79.0 \\\\ \hline \text{Short (4K)} & \text{SFT} & .351 \pm .021 & .258 \pm .017 & .597 \pm .095 & 31.3 & 29.2 & 48.4 \\\\ & \text{AID} & .374 \pm .025 & .506 \pm .003 & .621 \pm .069 & 31.4 & 29.3 & 48.4 \\\\ & \text{Full + Truncated SFT} & .373 \pm .002 & .523 \pm .004 & .510 \pm .033 & 31.4 & 28.6 & 84.4 \end{array}$$ Both methods outperform SFT, clearly so on Long, and their accuracies are within seed variation. As the reviewer suggested, the gain comes mainly from supervising answers from intermediate states, and AID performs on par with the causal alternative. They differ in how they affect reasoning. Training on truncated sequences makes the student less likely to finish its reasoning on its own: with Full + Truncated SFT, self-stop falls to about half of SFT's on Long (.049 vs. .101) and is lower on Short (.510 vs. .597). AID keeps the auxiliary answer outside the reasoning stream and stays close to SFT (.097, .621). It also uses one sequence per example with the same input tokens as SFT, while the baseline uses two and about 1.7x the peak memory. We will add these results to the final version. Related work (Section 4): - AnytimeReasoner (Qi et al., 2025) is an RL method using the model's own rollouts and verifiable rewards. AID is a distillation method using only teacher traces. They apply at different training stages and are complementary; AID could precede such RL. - P-ALIGN (Liu et al., 2026) builds new training data from a selected teacher prefix and a student continuation. AID keeps the original traces and adds an auxiliary objective, so the two can be combined. For the budget experiments below, models are trained on OpenR1-Math (Long-CoT) and evaluated on MATH-500 (greedy, 3 seeds). Following budget forcing (Muennighoff et al., 2025), if the model does not emit `</think>` within budget $B$, we truncate at $B$, append `</think>`, and allow 256 answer tokens. $B=0$ means answering immediately. The same protocol is used for all methods. **W2 / Q2. Gains on 4B under tighter budgets.** $$\begin{array}{l|cccccc} \text{Budget} & 0 & 1\text{K} & 2\text{K} & 4\text{K} & 8\text{K} & 16\text{K} \\\\ \hline \text{4B MATH SFT} & .007 & .257 & .489 & .688 & .809 & .857 \\\\ \text{4B MATH AID} & .151 & .650 & .777 & .834 & .865 & .885 \\\\ \hline \text{4B Minerva SFT} & .004 & .119 & .207 & .314 & .393 & .413 \\\\ \text{4B Minerva AID} & .108 & .308 & .377 & .401 & .407 & .411 \\\\ \text{4B Olympiad SFT} & .006 & .093 & .205 & .368 & .507 & .590 \\\\ \text{4B Olympiad AID} & .041 & .310 & .438 & .519 & .575 & .588 \\\\ \hline \text{1.7B SFT} & .009 & .191 & .275 & .434 & .536 & .566 \\\\ \text{1.7B AID} & .033 & .549 & .629 & .647 & .659 & .657 \end{array}$$ At 16K the 4B gap is small (+2.8pt), but it grows to +14.6pt at 4K, +28.8pt at 2K, and +39.3pt at 1K, likewise on Minerva and Olympiad. The benefit depends mainly on how often reasoning is interrupted, not on model size. **W3 / Q3. Self-generated reasoning across budgets, including zero.** Qwen3-0.6B-Base. Answer completion is the fraction of answers ending with EOS within 256 tokens; accuracy is over all responses. $$\begin{array}{l|cccccc} \text{Budget} & 0 & 1\text{K} & 2\text{K} & 4\text{K} & 8\text{K} & 16\text{K} \\\\ \hline \text{SFT acc.} & .012 & .135 & .176 & .225 & .263 & .250 \\\\ \text{AID acc.} & .055 & .337 & .374 & .379 & .379 & .375 \\\\ \hline \text{SFT completion} & .00 & .18 & .29 & .44 & .55 & .55 \\\\ \text{AID completion} & .99 & .98 & .99 & .99 & .99 & .99 \end{array}$$ SFT's completion and accuracy drop as the budget shrinks; even at 16K, 45% of its answers do not finish. AID completes at least 98% of answers at every budget, including zero, and its accuracy stays near .37 from 2K upward. The gains hold for the student's own reasoning, not only for teacher prefixes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.