acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation

Abstract

A reasoning trace can accompany a correct answer without providing effective supervision for a particular student. We study how to construct traces that improve learning beyond answer-only supervision by considering both their intermediate content and their accessibility to the target learner. We introduce REDiST, a student-conditioned trace-construction and distillation framework for evidence-rich reasoning. Starting from the student's own attempts, REDiST revises answer-correct drafts against the supplied evidence and repairs unsuccessful drafts with additional guidance from the verified answer. A frozen student-matched reader screens and ranks original and edited candidates using answer recovery and verified-answer probability, given the question, evidence, and candidate trace. We distill the selected targets while balancing trace and final-answer supervision with an auxiliary answer loss. Construction and selection occur offline, and inference uses only the distilled student. Experiments on question-answering benchmarks show that REDiST can improve both source-task and cross-benchmark performance over answer-only, original-trace, and the evaluated trace-selection and rewriting baselines. Across supervision recipes, performance rankings can differ between source and transfer benchmarks, highlighting the importance of evaluating trace teaching utility across distributions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, reasoning distillation, student-conditioned supervision, multi-hop question answering, cross-benchmark transfer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RSGdOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studied how to use reasoning trace instead of final answer to train the student model. the reasoning trace is defined and also proposed an improvement of using frozen-student target selection process to improve the student model training.

Critical strengths

I agree with the position of this paper. I think the reasoning traces should be used instead of only using the answer-only supervision. The proposed frozen-student target selection process seems interesting, although there are concerns about the experiments.

Critical weaknesses

while I agree with the motivation of this paper - to use the reasoning trace rather than answer only, but the claimed contributations are weak and cannot convince me. The formulation and evaluation principle are not new, if I don't call it straightforward or simple. I feel the authors just tell the readers that there is a problem, and here is what I did, and this is results. I cannot get the insight or new findings from the paper. The experiment setup does not clearly isolate the gain from the proposed reasoning trace. In the frozen-student target selection process, it seems the principle of using the traces which lead the right answer as the training target is contradictory with the overall assumption -- use the reasoning trace rather that the answer. Also, it is not guaranteed that the traces with the correct answer are valid reasoning trace.

Overall, I feel the problem defined in this paper makes sense, however, the experiment design and results cannot support the conclusion clearly.

Questions for authors

see the weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer RSGd (1/2)

Weakness 1: Contribution and insight

Response:

Thank you for this candid assessment. The submission contained relevant results (Figure 1, Sections 4.2 and 4.3), but their implications were not stated clearly, and the matched measurement called for by Eq. (1) was not reported. We also agree that the premise is not new: prior work has shown that a trace which accompanies a correct answer need not be a good teaching target, and the paper starts from it. The revision states three measured findings in the Introduction and in the new Conclusion, supported by the original comparisons and the additional controlled experiments.

1. Source-task accuracy does not reveal which supervision transfers. Trained on MuSiQue, the Answer-only recipe reaches 0.562 MuSiQue EM, within 2.2 points of ReDiST (0.584). Its 2Wiki EM in Table 1, however, is 0.479 (0.461 and 0.456 with two further training seeds), which is below the untrained model (0.516) and 20.7 points below ReDiST (0.686). The same pattern holds for every student in Table 3: for all four, the Answer-only recipe has the lowest 2Wiki EM of all trained variants, and for three of them it falls below the untrained model. These results show that a small source-task difference can accompany a much larger transfer difference, making source-task accuracy insufficient for assessing supervision intended to generalize across benchmarks.

2. Under a paired control, the traces add nothing measurable on the source task and 13.8 points on 2Wiki. We now report the paired control of Eq. (1), which keeps the same examples, exposure, templates and update budget and removes only the traces (one seed-matched pair). The ReDiST traces add 13.8 EM points on 2Wiki [10.7, 16.9], while MuSiQue EM differs by 1.5 points, with an interval that includes zero ([−1.3, 4.4]). The source-task advantage of ReDiST over Answer-only in Table 1 is therefore a difference between complete recipes, and the measurable effect of the traces is on 2Wiki. The same holds for other rewritten traces: MBT-R, whose training set differs from the Answer-only row by only five questions, is 18.1 points above Answer-only on 2Wiki and about 1.4 on MuSiQue, averaged over three training seeds (Table 10).

3. Where the trace comes from matters, and so does choosing among candidates. Replacing the draft-conditioned edit with a fresh trace written by the same rewrite model from the same evidence lowers 2Wiki EM by 9.1 and 14.6 points in two seed-matched pairs, with no significant change on MuSiQue or HotpotQA. Draft conditioning is shared with MBT-R, which also rewrites the student's attempt, and with a single rewrite per question the two methods have the same 2Wiki EM (0.633; Table 11). What ReDiST adds to draft-conditioned rewriting is the generation of several candidates and the choice among them by the frozen student, which raises 2Wiki EM to 0.686. These comparisons support both retaining the student's draft as the starting point and exploring alternative edits within the proposed selection procedure. The scope of the candidate-budget comparison is detailed in our response to Reviewer F6Ye, Weakness 2.

On the formulation. The evaluation principle is deliberately minimal, and we do not claim a paired trace-removed control as a new idea. Its role is to separate what a trace adds from what answer supervision already provides. The submission stated it without applying it; the revision applies it (finding 2). The methodological contribution is the construction procedure: ReDiST starts from the student's own attempts, including failed ones, edits them against the supplied evidence, and prefers the candidates from which a frozen copy of the student recovers the verified answer. The findings above explain why these construction choices matter: their effects can be substantially more visible in cross-benchmark transfer than in source-task accuracy.

AuthorsOpenReview ↗

Response to Reviewer RSGd (2/2)

Weakness 2: Isolating the gain from the proposed traces

Response:

We agree. Table 1 alone does not isolate the contribution of the traces, because its Answer-only and Original Trace rows also differ from ReDiST in which questions are trained on and how often each is seen. The revision adds two controlled comparisons, each changing one factor (Table 12).

(1) Remove the trace. Following Eq. (1), the control trains on exactly the same examples as ReDiST, seen equally often, with the same templates, loss rule and number of updates. Only the trace is removed (one seed-matched pair). The ReDiST traces add 13.8 points on 2Wiki [10.7, 16.9], while MuSiQue differs by 1.5 points [−1.3, 4.4]. The gain from the trace itself therefore appears on 2Wiki and not on the source task, and the source-task advantage over Answer-only in Table 1 reflects the complete recipe.

(2) Remove the student's draft. Fresh rewrite gives the rewrite model the same question and evidence but not the student's draft. Table 2 compared one such run with the Table 1 run (a 20.3-point gap on 2Wiki). We now train it twice, each time alongside a full ReDiST run with the same questions, settings and seed. 2Wiki EM falls by 9.1 [6.5, 11.8] and 14.6 [11.9, 17.4] points in the two pairs, and MuSiQue and HotpotQA change by at most 2.7 points, with intervals that include zero. The matched gaps are smaller than in Table 2, and both intervals exclude zero. In both pairs, conditioning the rewrite on the student's draft is worth at least 9 points on 2Wiki. Section 4.3 now reports the matched numbers.

Training seeds. Three training seeds of the full recipe give 0.686, 0.691 and 0.694 2Wiki EM, and Table 10 also reports three training seeds for five baselines, of which the strongest on 2Wiki, MBT-R, averages 0.646.

Together, the two controls distinguish the benefit of including the constructed traces from the benefit of conditioning their construction on the student's draft, under the stated training protocols.

Weakness 3: Answer-based selection and trace validity

Response:

We understand why this looks contradictory, and Section 3.2 now states the two roles explicitly: in target selection, the verified answer is used only to choose among candidate traces for the same question, and the student is then trained on the selected trace followed by that answer.

Why answer-based selection is not answer supervision. All candidates for a question share the same verified answer, so the choice among them depends only on the traces. A frozen copy of the student reads each candidate; among those from which it recovers the verified answer, the one with the highest verified-answer probability is kept. A candidate that fails this check is one that this student cannot yet follow to the answer, so we prefer a candidate that passes (if none passes, we keep the student's own answer-correct draft or fall back to answer-only supervision). Selecting reasoning traces by their outcome is common practice when building reasoning supervision, e.g., STaR (Zelikman et al., 2022) and rejection-sampling fine-tuning (Yuan et al., 2023, arXiv:2308.01825). That the trace, and not the answer, carries the 2Wiki gain is shown by the paired control: removing the traces from exactly the same selected examples lowers 2Wiki EM by 13.8 points [10.7, 16.9] (see our response to Weakness 2).

A correct answer does not make a trace valid. We agree, and passing the check does not certify the intermediate steps (Section 3.2). This is why ReDiST edits the drafts against the supplied evidence, and why we now audit the selected targets directly (Appendix C, Table 13). For the MuSiQue-supervised Qwen3-8B targets, a judge model (Gemma-3-27B-IT) that sees only the question, the evidence and the trace labels each claim as supported by the evidence or not. In targets built from answer-correct drafts, 1.7% (draft kept unchanged) to 3.8% (draft edited) of the claims are unsupported, close to the 2.5% of the original student traces. Repaired traces, which start from failed drafts and are written with access to the verified answer, are worse at 13.3%. We keep them because they are what makes failed attempts usable: when the failed drafts are used as targets instead, 2Wiki EM falls by 6.0 points and HotpotQA EM by 5.7 (Table 2, single run). The Conclusion states the lower support of repaired traces as a limitation.

Two further checks examine answer recovery and process coverage. With every mention of the answer masked, the traces still raise the frozen reader's answer recovery from 50.9% (no trace) to 75.8%, so they carry usable reasoning beyond the answer string. In addition, 96.4% of ReDiST targets state every sub-answer of the question's annotated decomposition, compared with 78.8% of the original student traces.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer RSGd,

Thank you again for your thoughtful comments. We wanted to follow up on our responses, particularly the paired trace-removal and fresh-rewrite controls (Table 12), the trace audit (Table 13), and our clarification of how answer-based selection relates to trace supervision.

We would appreciate your feedback on whether these results clarify the contribution of the constructed traces and address your concerns. Please let us know if there are further questions we can help address.

Best,

Authors of Paper #3359

Reviewer_F6YeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ReDiST, a student-conditioned framework for reasoning trace distillation. It revises answer-correct student drafts, repairs unsuccessful drafts using verified-answer guidance, and selects supervision targets through a frozen student-matched reader. Experiments on multiple question-answering benchmarks demonstrate improvements in source-task performance and cross-benchmark transfer.

Critical strengths
  • The paper is well-organized and clearly motivated, highlighting the distinction between answer correctness and the effectiveness of reasoning traces as supervision.
  • The method demonstrates promising results across multiple benchmarks and student models, particularly in cross-benchmark transfer. The component analyses provide additional insights into the effects of trace construction and supervision allocation.
Critical weaknesses
  1. Limited methodological novelty. The central idea of refining student-generated reasoning while maintaining compatibility with the target learner has already been explored in prior work. In particular, Style-Aligned Refinement [1] investigates preserving student-native reasoning trajectories during teacher-guided correction, while RSR [2] studies student-specific trajectory suitability for effective distillation. ReDiST primarily combines these established ideas with answer-assisted repair and reader-based selection (§3), without demonstrating a sufficiently distinct methodological contribution beyond their integration.
  2. Limited analysis of improvements over existing rewriting methods. ReDiST generates multiple student drafts and teacher rewrites ((K_d=K_r=4)), followed by reader-based candidate selection (§3.1–3.2). In contrast, the closest rewriting baseline, MBT-R, generates only one rewrite per question without an equivalent selection procedure (Appendix B.3, Table 4). While ReDiST demonstrates promising improvements, particularly on 2Wiki (Table 1), it remains unclear how much of the gain comes from the proposed construction strategy versus additional candidate exploration and selection. A comparison under matched candidate generation and selection settings would help clarify the source of these improvements and further strengthen the empirical findings.
  3. Limited analysis of student-specific teaching utility. A central motivation of ReDiST is that reasoning traces should be constructed according to the target student's accessibility and learning capability (§2.3). While Table 3 demonstrates consistent improvements across different student models, these results do not directly examine whether traces constructed for a particular student are especially beneficial to that student compared with other learners. A cross-student trace transfer analysis, where traces constructed for one student are used to train another, would provide further insights into the role of student-specific adaptation and strengthen the paper's central motivation.

[1] Qian et al., The Quality–Utility Paradox: Why High-Reward Data Impairs Small Model Mathematical Reasoning, ICML 2026.

[2] Yang et al., Which Reasoning Trajectories Teach Students to Reason Better? A Simple Metric of Informative Alignment, ACL 2026.

Questions for authors

See the weakness section

5 comments
AuthorsOpenReview ↗

Response to Reviewer F6Ye (1/3)

Weakness 1: Methodological novelty

Response:

Thank you for raising the relation to these two papers, both of which the submission cites (RSR is also one of our baselines). We agree that ReDiST builds on the premise they share, that supervision should fit the learner. The mechanisms differ, however, and the submission did not make these distinctions sufficiently explicit; the revision clarifies them in Section 1 and Appendix A.

RSR ranks existing teacher trajectories by a rank-surprisal ratio under the student and selects from a fixed pool. It uses no student attempt and edits nothing; its authors name metric-guided rewriting as future work.

Style-Aligned Refinement refines a small model's own solutions in mathematical reasoning. In its experiments the refinement starts from solutions that already reach the correct answer: of 100,000 sampled problems, about 34,000 have a correct solution among eight samples, and only these problems are used. Each solution is refined once, and no per-example test checks whether the student can use the result. In contrast, ReDiST shares the use of student-generated drafts, but allows changes to both content and expression and then uses the frozen student's response to select among the original and edited candidates.

What ReDiST does differently.

  1. Failed attempts are repaired, and a repair is kept only if the student can use it. Even with four drafts per question, the student reaches the verified answer on only 2,066 of the 4,000 MuSiQue training questions (Table 5). Thus, a method that refines only answer-correct attempts leaves about half of the questions without a trace. Answer-guided correction of failed attempts is not new (STaR and our closest baseline MBT-R use it; Appendix A). ReDiST couples this correction with a per-candidate test: of 1,625 failed drafts, 1,004 receive a repair from which the frozen student recovers the answer, and the other 621 fall back to answer-only supervision.
  2. The student-side test is functional (answer recovery), not a token-level fit. A frozen copy of the student must recover the verified answer from each candidate, given the question and the evidence, and the highest-scoring candidate becomes the target. The learner's own answer likelihood has been used before, to train an explanation-generating teacher (RLT) and to rank a small model's candidate rationales (COALITION; Appendix A). ReDiST uses this signal offline, with no additional model trained, to choose among the student's own draft and its evidence-grounded edits.
  3. Edits are made against supplied evidence. Both papers address mathematical reasoning, where no evidence is supplied. ReDiST's rewrite model is instructed to retain supported reasoning, correct intermediate entities and relations, and remove unsupported content using the supplied evidence. This is a difference in setting, not a methodological claim.

We do not claim the individual ingredients as new. Our contribution is their use as a per-candidate accept-or-reject test on edits of the student's own failed and successful drafts, with the original draft also retained as a candidate. Table 11 shows that, under this frozen-student selection rule, increasing the number of rewrites from one to four raises 2Wiki EM from 0.633 (the same as the MBT-R run of Table 1) to 0.686.

What each component contributes (within ReDiST; same questions, training configuration and number of updates):

  • Repairing failed attempts. Without repair, 2Wiki EM falls from 0.686 to 0.626 and HotpotQA EM from 0.633 to 0.576 (Table 2, single run).
  • Several rewrites, among which the frozen student chooses. With a single rewrite per question instead of four, 2Wiki EM falls from 0.686 to 0.633 (Table 11); we discuss this comparison in our response to Weakness 2.
  • Starting from the student's draft. Writing fresh traces without the draft, with the same rewrite model and evidence, lowers 2Wiki EM by 9.1 [6.5, 11.8] and 14.6 [11.9, 17.4] points in two seed-matched pairs (Table 12). This property is shared with Style-Aligned Refinement and MBT-R; it separates draft-based methods from selection over teacher-written trajectories such as RSR.

In all three comparisons, MuSiQue EM changes by at most 2.1 points, so the benefit appears on 2Wiki and not on the source task.

Comparison with RSR. In Table 1, ReDiST has higher point estimates than RSR on all four benchmarks. Averaged over three training seeds each, it is 6.0 EM points above RSR on 2Wiki (0.690 vs. 0.630) and 2.9 on MuSiQue (0.586 vs. 0.557; Table 10). This compares complete recipes: RSR has a usable teacher trajectory for 2,477 of the 4,000 questions, whereas ReDiST trains on 3,691 (Table 5), so the difference reflects coverage as well as trace content. ReDiST targets also state every sub-answer of the question's decomposition more often than RSR-selected trajectories (96.4% vs. 85.1%; Appendix C).

AuthorsOpenReview ↗

Response to Reviewer F6Ye (2/3)

Weakness 2: Gains over rewriting under matched candidate generation

Response:

Thank you for this question. The submission did not separate these factors, so we ran a candidate-budget study with the Qwen3-8B student and the Qwen3-14B rewrite model (Table 11; EM / F1). The ReDiST rows differ only in the number of rewrites; the MBT-R row is the result from Table 1. MBT-R produces one answer-guided teacher rewrite of each student draft, without reader-based candidate selection.

SettingStudent draftsRewritesMuSiQue2Wiki
MBT-R11, no selection0.579 / 0.6720.633 / 0.719
ReDiST, Kr=1K_r=1410.586 / 0.6980.633 / 0.729
ReDiST, Kr=2K_r=2420.581 / 0.6830.652 / 0.746
ReDiST, Kr=4K_r=4 (default)440.584 / 0.6810.686 / 0.780
ReDiST, Kr=8K_r=8480.602 / 0.7020.672 / 0.756

With one rewrite per question, ReDiST matches MBT-R on 2Wiki EM. Only the number of rewrites is matched here: ReDiST still selects among four student drafts and still lets the frozen student accept or reject the rewrite. It equals MBT-R on 2Wiki EM (0.633) and has slightly higher point estimates on 2Wiki F1 (+1.0) and on MuSiQue (+0.7 EM, +2.6 F1). Thus, at the single-rewrite budget, we observe no 2Wiki EM advantage over the reported MBT-R result. Section 4.2 now makes this comparison explicit.

The 2Wiki margin over MBT-R comes from candidate exploration together with selection by the student. With four rewrites among which the frozen student chooses, 2Wiki EM is 0.686, 5.3 points above the single-rewrite setting; with two and eight rewrites it is 0.652 and 0.672. One measurable effect of more rewrites is that more failed drafts receive a repair that the frozen student can follow: 50.4% with one rewrite and 61.8% with four. We do not claim that this rate alone explains the gain; across rewrite teachers it does not track 2Wiki EM (Table 6). Drafts matter in the same direction: in Table 5, the greedy draft is answer-correct for 1,466 of the 4,000 training questions, whereas the draft selected among four is answer-correct for 2,066, and using one draft instead of four lowers 2Wiki EM by 3.4 points [0.9, 5.9].

What this does and does not establish. Several candidates need a rule for choosing among them. MBT-R as published has none, and the frozen-student check of Section 3.2 is the rule we propose. More candidates can help only through that rule, so the gain from one to four rewrites is obtained through selection. We agree that this does not fully settle your question, because MBT-R could be given the same four rewrites and the same check, and we have not run that comparison. Our claim is therefore limited to the following: the 2Wiki margin over MBT-R comes from choosing among several rewrites with the frozen student, and not from the editing prompts. The study also does not show which part of the rule matters (the pass/fail recovery check, the ranking by verified-answer probability, or the use of the student itself as the reader). Section 4.2 states that the study does not separate the number of candidates from the student's choice and that MBT-R was not run with several rewrites and selection. The four-rewrite setting is stable across three training seeds (0.686, 0.691 and 0.694) and averages 4.4 points above MBT-R (Table 10); the other ReDiST rows of the table above are single runs.

AuthorsOpenReview ↗

Response to Reviewer F6Ye (3/3)

Weakness 3: Student-specific teaching utility

Response:

Thank you for this suggestion. We ran the cross-student analysis you describe (Appendix C, Table 14). In short: on 2Wiki transfer, each student achieves higher EM with the targets constructed for it, clearly for Qwen3-8B and within noise for Gemma4-E4B. The source-task results do not show a consistent own-target advantage.

Setup. Qwen3-8B and Gemma4-E4B each have their own ReDiST targets, built from their own drafts, with their own frozen reader and the rewrite model of their family (Qwen3-14B and Gemma4-26B-A4B). We train each student on its own targets and on the other student's, and compare within a student, because the two students start from different raw accuracies. All four runs use the same training questions, so the own-target entries are 1.0 to 2.8 points below Tables 1 and 3.

Targets constructed for Qwen3-8BTargets constructed for Gemma4-E4B
Qwen3-8B student0.571 / 0.676 (own)0.603 / 0.600
Gemma4-E4B student0.578 / 0.6630.590 / 0.695 (own)

(MuSiQue EM / 2Wiki EM; each cell is one training run evaluated on 1,000 questions, with a 95% interval of about ±3 points per entry.)

Transfer. With the other student's targets, 2Wiki EM falls from 0.676 to 0.600 for Qwen3-8B (7.6 points) and from 0.695 to 0.663 for Gemma4-E4B (3.2 points). The ranking of the two target sets therefore reverses between the students: Qwen3-8B performs better with Qwen3-8B targets, whereas Gemma4-E4B performs better with Gemma4-E4B targets. This pattern is consistent with student-dependent teaching utility: the target set with higher transfer performance depends on the student receiving it.

Source task. On MuSiQue we see no benefit from a student's own targets. The targets constructed for Gemma4-E4B give the higher EM for both students (0.603 vs. 0.571; 0.590 vs. 0.578), differences of 3.2 and 1.2 points that are at or within noise. This agrees with our paired control (response to Reviewer RSGd, Weakness 2), in which the traces add 13.8 points on 2Wiki and nothing measurable on MuSiQue. For Qwen3-8B, the target set with the higher source-task EM is the one that loses 7.6 points on 2Wiki, so source-task accuracy again does not reveal which supervision transfers.

Scope of the comparison. The swap crosses model families, so the other student's targets differ in drafts, reader and rewrite model at once. Table 6 shows that the rewrite model matters by itself: for Qwen3-8B with its own drafts and reader, replacing Qwen3-14B with Gemma4-26B-A4B lowers 2Wiki EM from 0.686 to 0.624. This identifies the rewrite teacher as a plausible contributor to the swap results, but does not quantify its contribution to the 7.6-point difference. The swap therefore provides initial evidence of student-dependent transfer utility for the complete constructed target sets. Holding the rewrite model fixed in a further swap would more directly examine the joint role of student-specific drafts and reader selection. Appendix C makes this scope explicit.

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

This work presents REDIST to construct reasoning traces that are useful for a particular student rather than merely correct for a question. The method samples drafts from the target student, asks a teacher to revise answer-correct drafts or repair incorrect drafts against supplied evidence and the verified answer, and uses a frozen copy of the student to select candidates by answer recovery and verified-answer probability. Selected traces are distilled with controlled repair exposure and an auxiliary answer-token loss. Experiments show gains over answer-only, original-trace, selection, and rewriting baselines on source QA and cross-benchmark transfer across four students, while ablations isolate draft conditioning, repair, and loss allocation.

Critical strengths
  • The student-conditioned formulation targets a real supervision problem. A trace can be factually correct yet too sparse, verbose, or unfamiliar for the target model. Starting from the student’s draft, editing it with evidence, and selecting candidates with a frozen student-matched reader directly addresses both trace content and student accessibility.

  • The empirical analysis is fairly comprehensive. The paper compares multiple selection and rewriting baselines, four target students from two families, two supervision sources, several teachers, source and transfer benchmarks, and controlled ablations for target selection, fresh rewriting, repair, repair exposure, and answer-loss weighting. This gives much more evidence than a single-model distillation demonstration.

Critical weaknesses
  • Answer recovery is a weak proxy for faithful and teachable reasoning. The repair branch sees the gold answer, and selection rewards whether a frozen student can recover that answer from the trace; neither condition guarantees that all intermediate claims are supported by the evidence or that the reasoning is pedagogically useful. Apart from a qualitative example, the paper does not quantitatively audit trace faithfulness, unsupported shortcuts, or stepwise correctness.

  • The manuscript is incomplete in its current presentation and protocol. The main text has no Conclusion section, so the paper does not explicitly summarize limitations, deployment scope, or the relation between source-task and transfer gains. The main comparison evaluates final checkpoints on the first 1,000 examples of validation slices with one training seed, while hyperparameters such as α, repair multiplicity, and candidate counts are not shown to be selected through a fully nested protocol. These issues do not overturn the reported gains, but they weaken confidence in close rankings and leave the submission below the presentation standard expected for acceptance.

Questions for authors
  • Can you report a quantitative faithfulness audit of revised and repaired traces, including unsupported claims, gold-fact coverage, and cases where the model recovers the answer through a shortcut?

  • Do the main conclusions hold on held-out test slices and multiple training seeds, with uncertainty that accounts for training as well as evaluation sampling?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Gmfh (1/3)

Weakness 1 and Question 1: Answer recovery and faithfulness audit

Response:

Answer recovery by itself guarantees neither faithful nor teachable reasoning, and Section 3.2 says that it does not establish stepwise correctness. The submission also lacked a quantitative audit. We have now run one (Appendix C, Table 13) and summarize it below, including the part that confirms your concern.

Role of answer recovery. It is a functional screen: a frozen copy of the student must reach the verified answer from a candidate. Among the candidates from which it recovers the verified answer, the one with the highest verified-answer probability becomes the target. If none passes, the procedure retains an initially answer-correct draft or falls back to answer-only supervision, as specified in Section 3.2. Faithfulness is addressed separately. The rewrite model is instructed to use only the supplied evidence, and the verified answer is given to it as a separate input only in the repair branch, as a checking hint rather than as evidence (Appendix D). The audit measures how well this grounding instruction is followed.

Audit of the training targets (Qwen3-8B, MuSiQue). A sentence-level judge (Gemma-3-27B-IT, not used in construction) sees the question, evidence and trace, but not the gold answer, the method, or whether the trace's answer was correct. The reported rates are estimates from this automated audit. The judge detects 82% of injected entity and date errors. This calibration does not establish its accuracy on all types of unsupported reasoning.

Training targetUnsupported claims
ReDiST, answer-correct draft kept unchanged1.7%
ReDiST, answer-correct draft edited3.8%
ReDiST, repaired failed draft13.3%
Original student traces2.5%
MBT-R rewrites2.8%

Unsupported claims. For answer-correct drafts, targets whose text is identical to the selected draft have a rate of 1.7%. Where the rewrite model edited the draft, the rate is 3.8%, compared with 2.5% for the original student traces and 2.8% for the MBT-R rewrites. Repaired traces have a clearly higher rate (13.3%), and they are 1,004 of the 3,070 ReDiST trace targets (Table 5). They are built only for questions on which none of the student's drafts reached the answer, so the other rows, which cover easier or all questions, are not a like-for-like comparison. These results do not establish a uniform improvement in evidential support, and the Conclusion now states the remaining unsupported claims, particularly in repaired targets, as a limitation.

We retain the repair branch because it improves downstream performance relative to reusing the unsuccessful drafts: when repair is disabled and those failed drafts are used as targets, 2Wiki EM falls by 6.0 points and HotpotQA EM by 5.7 (Table 2, single run). This supports the utility of repair under the evaluated training recipe, while leaving the validity of individual steps as a separate question.

Gold-fact coverage. As a coverage diagnostic, we measure whether a trace states every sub-answer of the question's annotated decomposition. 96.4% of ReDiST targets state every sub-answer of the question's annotated decomposition, compared with 88.0% for MBT-R, 85.1% for RSR and 78.8% for the original student traces. This measures annotated sub-answer coverage, and it does not establish that every connecting inference is correct or supported.

Shortcut recovery. We masked mentions of the gold answer and its aliases in each trace and asked the frozen reader to answer. It still recovers the answer from 75.8% of the ReDiST traces, compared with 50.9% with no trace, and 3.6 points (95% interval [1.1, 6.0]) more than from masked MBT-R traces. The masked traces therefore add 24.9 points over the question and evidence alone. This indicates information useful for answer recovery beyond the explicitly masked answer strings. It does not rule out other shortcuts or certify faithful use of the intermediate steps, and the failure rate after masking should not be interpreted as the frequency of shortcut reasoning.

Teaching value. The direct test is the paired control of Eq. (1): with the same examples, verified answers, exposure schedule, and number of updates but the traces removed, 2Wiki EM falls by 13.8 points [10.7, 16.9] (see our response to Reviewer RSGd, Weakness 2). The selected traces therefore improve 2Wiki performance beyond matched answer-only supervision under this training protocol.

AuthorsOpenReview ↗

Response to Reviewer Gmfh (2/3)

Weakness 2 and Question 2: Missing conclusion and evaluation protocol

Response:

Conclusion. The revised paper adds a Conclusion (Section 5). It states the scope of the method (evidence-rich question answering with verified training answers and supplied evidence), its limitations (answer recovery does not certify each intermediate step; repaired traces are less well supported by the evidence than targets built from answer-correct drafts; the 2Wiki margin over MBT-R depends on exploring several candidates; the transfer advantage is clear on 2Wiki and within evaluation noise on HotpotQA and QuALITY), and the relation between the two kinds of gains (the gain attributable to the traces appears on 2Wiki, while source-task differences among the strongest recipes are within evaluation noise).

Training seeds (Question 2). The additional runs support the stability of the primary ReDiST result and of its 2Wiki margin over the baselines. We trained ReDiST and five baselines with three training seeds each in the primary setting (Qwen3-8B, MuSiQue supervision; Table 10 of the revision, reproduced below; MuSiQue EM / 2Wiki EM). Only the training seed changes (42, 123 and 3407, the same for every method); the training questions, each method's training targets and the evaluation slices are unchanged.

MethodSeed 42 (Table 1)Seed 123Seed 3407Mean
Answer-only0.562 / 0.4790.561 / 0.4610.549 / 0.4560.557 / 0.465
STaR0.510 / 0.6130.504 / 0.6370.511 / 0.6240.508 / 0.625
RSR0.568 / 0.6240.549 / 0.6390.554 / 0.6280.557 / 0.630
LARK0.539 / 0.6300.533 / 0.6430.553 / 0.6380.542 / 0.637
MBT-R0.579 / 0.6330.572 / 0.6470.564 / 0.6590.572 / 0.646
ReDiST0.584 / 0.6860.593 / 0.6910.581 / 0.6940.586 / 0.690

On 2Wiki, ReDiST averages 0.690 against 0.646 for MBT-R, the strongest baseline on this benchmark, a margin of 4.4 points, and its lowest run (0.686) is above the highest run of any baseline (0.659). Across seeds, the sample standard deviation of 2Wiki EM is 0.4 points for ReDiST and 1.3 for MBT-R, and a two-sample (Welch) 95% interval for the difference of the two means is [1.5, 7.3] points. On MuSiQue, ReDiST also has the highest mean (0.586 against 0.572 for MBT-R), but this 1.4-point difference is not resolved (interval [−0.2, 3.0]), and Section 4.2 now states in the main text that the source-task ordering is unresolved. These seed-level intervals describe variation across training seeds on the fixed evaluation slices. The paired bootstrap intervals reported elsewhere in our responses quantify evaluation uncertainty conditional on the trained models. They do not jointly account for training and evaluation uncertainty.

AuthorsOpenReview ↗

Response to Reviewer Gmfh (3/3)

Weakness 2 and Question 2: Missing conclusion and evaluation protocol (continued)

Consistency across settings. ReDiST is above MBT-R on transfer EM in all five student and source settings in which both were trained (Tables 3 and 7), each with its own student drafts, teacher rewrites and training run.

Student, source to transferTransfer EM (ours / MBT-R)Margin over MBT-RMargin over best baseline
Qwen3-4B, MuSiQue to 2Wiki0.630 / 0.625+0.5−0.5 (NLL)
Qwen3-8B, MuSiQue to 2Wiki0.686 / 0.633+5.3+5.3 (MBT-R)
Gemma4-E2B, MuSiQue to 2Wiki0.622 / 0.580+4.2+1.2 (NLL)
Gemma4-E4B, MuSiQue to 2Wiki0.723 / 0.645+7.8+4.1 (STaR, NLL)
Qwen3-8B, 2Wiki to MuSiQue0.536 / 0.512+2.4+2.4 (MBT-R)

Against the strongest baseline in each setting, the transfer margin is −0.5, +5.3, +1.2, +4.1 and +2.4 points. ReDiST is therefore at or near the top on transfer for every evaluated student and in both directions, while the strongest competing method varies with the setting. For example, MBT-R is 4.2 and 7.8 points behind on the two Gemma students, while NLL is 8.6 points behind on Qwen3-8B. In the primary setting (second row), the means over three training seeds are 0.690 for ReDiST and 0.646 for MBT-R (Table 10).

Held-out data and hyperparameter selection (Question 2). The test-set answers of these benchmarks are not public, so we evaluate on validation questions, none of which is used for training. In the submission, however, α was chosen from a sweep computed on the evaluation slice (Table 9), while the repair multiplicity mm and the candidate counts KdK_d and KrK_r were set to default values. We acknowledge the resulting dependence of the reported evaluation on hyperparameter selection.

After submission, we checked whether the choice of α depends on the evaluation slice: we repeated the selection on a development set that shares no question with any evaluation slice, using a single criterion, the mean of MuSiQue and 2Wiki development EM. It selects α = 0.15 again. The criterion uses 2Wiki development questions, so the 2Wiki benchmark, though none of its evaluation questions, informs the choice of α; HotpotQA and QuALITY play no part in this criterion.

This agreement supports the stability of the selected value across the evaluation slice and the development set. We report this as a post-hoc check of hyperparameter-selection stability. Together with the three training-seed runs, these additional analyses examine two sources of variability: the data used to select α and the training seed. Across the three runs, ReDiST's 2Wiki EM remains between 0.686 and 0.694, above all three MBT-R runs (0.633 to 0.659).

The new sweeps over the other settings in Appendix C are computed on the evaluation slice, so we report them as sensitivity analyses and not as independent selection results. 2Wiki EM is insensitive to mm (0.686 and 0.692 for m=4m=4 and 8). The candidate counts do matter: 2Wiki EM is 0.633, 0.652, 0.686 and 0.672 for Kr=1K_r=1, 2, 4 and 8, and one draft instead of four lowers it by 3.4 points [0.9, 5.9]. We discuss this dependence in our response to Reviewer F6Ye, Weakness 2, and state it in Section 4.2 and in the Conclusion.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful and constructive comments. The revised paper marks all changes in blue. The main additions are:

  • a Conclusion that states the scope and the limitations of the method (Section 5; Reviewer Gmfh, Weakness 2);
  • three training seeds for ReDiST and five baselines (Section 4.2 and Table 10; Gmfh, Weakness 2/Question 2);
  • two controlled comparisons that isolate what the trace adds beyond the answer and what the student's draft adds beyond a fresh rewrite (Section 4.3 and Table 12; RSGd, Weakness 2; F6Ye, Weakness 1);
  • a candidate-budget study, including a single-rewrite setting to compare with MBT-R (Section 4.2 and Table 11; F6Ye, Weakness 2);
  • a quantitative audit of the training targets (Appendix C and Table 13; Gmfh, Weakness 1/Question 1; RSGd, Weakness 3);
  • a post-hoc check of the answer-loss weight α on a development set that shares no question with the evaluation slices, and sensitivity sweeps over the repair multiplicity mm and the candidate counts KrK_r and KdK_d (Appendix C; Gmfh, Weakness 2/Question 2);
  • an explicit comparison with Style-Aligned Refinement and RSR (Section 1 and Appendix A; F6Ye, Weakness 1);
  • a cross-family swap of the training targets of two students (Appendix C and Table 14; F6Ye, Weakness 3).

In brief, the new results show five things.

  • (i) Removing only the traces in the matched control lowers 2Wiki EM by 13.8 points [10.7, 16.9], with no significant change on MuSiQue, so the gain attributable to the traces appears on 2Wiki and not on the source task.
  • (ii) With three training seeds for ReDiST and for five baselines, ReDiST averages 0.690 2Wiki EM against 0.646 for MBT-R, the strongest baseline on this benchmark, and its lowest run is above the highest run of every baseline, while source-task differences among the strongest recipes are within evaluation noise.
  • (iii) With one rewrite per question, ReDiST equals the Table 1 result of MBT-R on 2Wiki EM (0.633); its margin comes from exploring several candidates and letting the frozen student choose among them. Within ReDiST, increasing the rewrite count from one to four under the same frozen-student selection rule raises 2Wiki EM to 0.686.
  • (iv) In the automated audit, repaired traces contain more unsupported claims (13.3%) than targets built from answer-correct drafts (1.7% to 3.8%), which the Conclusion now states as a limitation.
  • (v) When the targets of two students are swapped, each student scores higher on 2Wiki with the targets constructed for it (clearly for Qwen3-8B), with no such advantage on the source task; the swap also changes the rewrite model.

Unless stated otherwise, differences are in EM points and bracketed intervals are 95% paired bootstrap intervals over the same 1,000 evaluation questions.

AuthorsOpenReview ↗

Closing clarification of contributions and rebuttal findings

We thank the reviewers for their comments. As the discussion closes, we summarize the methodological clarifications and additional evidence addressing the central questions raised. Table references below refer to the revised manuscript.

1. What does ReDiST contribute beyond existing trace refinement and selection?

ReDiST constructs teaching targets through evidence-guided revision of answer-correct student drafts, answer-assisted repair of unsuccessful drafts, and frozen-student selection among original and edited candidates. The target student shapes both the starting attempts and the choice of supervision. Relative to fixed-pool trajectory selection, ReDiST constructs new targets by editing student attempts, and it also extends draft refinement with explicit selection among the original and edited candidates using the intended learner’s response. Our responses clarify these methodological distinctions and their relationship to prior work.

Verified answers guide repair, candidate selection, and final-answer supervision. The question studied is what the constructed traces add beyond final-answer labels, while retaining those labels during training.

2. Do the traces themselves contribute, independently of changes in example exposure and training budget?

The paired trace-removal control holds the training examples, verified labels, exposure, loss rule, and update budget fixed. Removing only the traces lowers 2Wiki EM by 13.8 points, with a 95% paired-bootstrap interval of [10.7, 16.9]; the corresponding MuSiQue difference is 1.5 points [−1.3, 4.4]. Two additional seed-matched comparisons find 2Wiki gains of 9.1 and 14.6 points from conditioning rewriting on the student draft (Table 12). These comparisons directly evaluate trace supervision and draft conditioning, separately from the complete-recipe comparisons in Table 1.

3. How much does the candidate-generation budget matter?

The added budget study varies the number of rewrites while keeping the student-draft count and frozen-reader selection rule fixed (Table 11). With one rewrite, ReDiST matches the evaluated MBT-R configuration on 2Wiki EM at 0.633. Increasing the rewrite count to four raises ReDiST’s EM to 0.686. This comparison shows the value of exploring additional revisions and repairs within ReDiST’s construction-and-selection procedure, and makes the role of candidate-generation budget explicit when interpreting the reported results.

4. How are answer recovery, evidential support, and teaching utility assessed?

The revision distinguishes these three roles: the frozen reader supplies the selection signal, a separate claim-level audit assesses evidential support, and paired training comparisons measure downstream teaching utility. We also added an answer-masking control (Table 13). After masking the verified answer and its aliases, reader recovery remains 75.8%, compared with 50.9% without a trace, and exceeds recovery from masked MBT-R traces by 3.6 points [1.1, 6.0]. This provides evidence that the constructed traces convey information useful for answer recovery beyond the explicitly masked answer strings. The paired control in Point 2 separately establishes their contribution to downstream learning.

5. Is teaching utility dependent on the receiving student?

The cross-student target-swap experiment compares target sets produced by the respective construction pipelines, including their student drafts, readers, and rewrite teachers. Each receiving student obtains higher 2Wiki EM with the targets constructed for it (Table 14). The ranking of the two target sets reverses between the receiving students, supporting student-dependent transfer utility for the complete constructed target sets.

6. Are the primary gains stable across training runs?

We added three training seeds each for ReDiST and five baselines. ReDiST averages 0.690 2Wiki EM versus 0.646 for MBT-R, the strongest evaluated baseline, and its lowest run exceeds the highest run of each baseline (Table 10). The revision distinguishes training-seed variability from paired-bootstrap evaluation uncertainty. It also makes the original hyperparameter-selection protocol explicit and reports a post-hoc development-slice check that selects the same answer-loss weight, α=0.15.\alpha = 0.15.

We hope these clarifications and controlled results help readers assess the methodological contribution and the learning value of the constructed supervision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.