acceptodds
Under review as a conference paper at ICLR 2027

K2P: Label-Free Knowledge to Prompt Distillation

Abstract

Knowledge distillation can transfer reasoning from stronger teachers to frozen students through reusable prompts, but avoiding weight updates does not eliminate supervision. Without ground-truth answers, teacher solutions are unverified, and agreement with the teacher can reward shared mistakes. We introduce Knowledge-to-Prompt (K2P) for label-free knowledge distillation to prompts. K2P synthesizes reusable instructions from teacher solutions, refines them using paired teacher and student responses, and guides search and selection with answer agreement. It retains candidates that adaptive search may undervalue and selects on reserved questions. Deployment uses only the frozen student and selected prompt. Our theory separates generation and selection gaps and gives conditions under which agreement-guided construction yields accuracy guarantees despite imperfect teacher references. Across reasoning tasks and students, K2P outperforms label-free alternatives overall and remains competitive with supervised prompt optimization. Ablations and archive diagnostics assess the contributions of teacher solutions and refinement, while revealing the limits of agreement-guided selection.

Primary area:
foundation or frontier models, including LLMs
Keywords:
knowledge distillation, prompt optimization, label-free learning, 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_sZN2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Knowledge-to-Prompt (K2P), which transfers knowledge from a large teacher to a smaller, frozen student through optimized natural-language prompts, without ground-truth training labels. K2P synthesizes reusable instructions from teacher solutions, refines them using paired teacher and student responses, and guides search and selection with answer agreement. The experimental results on reasoning tasks show improvements compared to other baselines.

Critical strengths
  1. The setting is practical and low-resource. Improving a frozen student without fine-tuning or ground-truth labels is valuable.
  2. The authors compare K2P against PLD, APE, GEPA, SPO, and label-free variants. The baselines are comprehensive and cover sufficient diversity of related works.
  3. The analysis and theoretical explanations are detailed. The paper examines teacher solution contributions, feedback refinement, archive selection, and sensitivity to different sample sizes.
Critical weaknesses
  1. The label-free problem is barely stressed. The paper assumes that teacher errors can reward shared mistakes, but teacher accuracy is 94–98% on every task (saturated), so the teacher is nearly a gold labeler. It would be interesting to test weaker teachers, harder tasks, or noisier data.
  2. The theory is loosely connected to the experiments. For example, the key conditions (κ, ε_t, γ, ω) are never estimated. The binary and independent-flip noise model is unrealistic, since teacher errors correlate with question difficulty and student errors.
Questions for authors

The paper's presentation is dense and too long. It puts many key results into the 57-page Appendix (token usage, bootstrap tables), which makes the actual takeaway hard to extract.

4 comments
AuthorsOpenReview ↗

Response to Reviewer sZN2 (1/4) — W1 — Controlled answer corruption

Thank you very much for reviewing our work and for your valuable comments on teacher reliability, the connection between theory and experiments, and the paper presentation. We have carefully addressed your comments, with detailed point-by-point responses provided below. We have also uploaded a revised manuscript with all changes highlighted in blue.

W1. Testing the label-free setting with unreliable teachers and corrupted answers

Thank you very much for raising this point. We agree that the original experiments provide limited coverage of unreliable teachers. We tested full construction with Phi-4-mini-instruct and controlled final-answer corruption across the source, search, and selection pools. Phi-4-Mini technical report

We use Phi to test construction with less reliable teacher references. Its reference accuracy on the 160 search/selection questions is 26.88% for Tracking, 63.75% for Colored Objects, and 87.50% for QuaRTz, with unusable answers counted as incorrect. On the same Colored Objects questions, Qwen3-8B and Ministral-3-8B achieve 95.63% and 78.13%, respectively. Model descriptions appear in the Qwen3 report and Ministral 3 report.

With Phi throughout construction, K2P's three-seed mean exceeds zero-shot in five of six task/student settings, while its comparison with GEPA-T varies across settings.

Controlled answer corruption throughout construction. On QuaRTz, we flip final A/B answers in all three 80-question pools (source, search, selection), updating the full response's final answer and parsed reference while retaining the original reasoning. Qwen3.5-9B performs synthesis and refinement. Each corrupted condition reruns synthesis, search, feedback, refinement, and selection before testing on all 784 questions. The 0% condition reuses the original runs.

Values are mean ± sample SD over three construction seeds, each with its own corruption mask at each rate. Masks are shared across students and nested across rates. Reference accuracies are averaged across masks and measured with gold answers after construction.

ConditionSearch reference accuracySelection reference accuracyQwen 0.8BMinistral 3B
Zero-shot——82.5385.20
0%95.0092.5081.59 ± 3.3891.45 ± 1.42
10%85.0082.0880.27 ± 1.2888.22 ± 1.49
20%77.9273.7579.93 ± 2.0791.33 ± 1.40
30%67.5066.6779.00 ± 1.4891.33 ± 1.40
40%60.0058.3380.36 ± 0.9688.56 ± 0.85
50%50.0050.4275.55 ± 9.2490.26 ± 0.85

For Ministral, the 20% and 30% conditions select the same final prompt for each seed and reuse the same cached test outputs.

Ministral exceeds zero-shot in every run, including a 5.06 pp mean gain at 50% flipping. Qwen's means are below zero-shot at every rate and fall by 6.04 pp from 0% to 50%. Test accuracy varies non-monotonically across rates.

AuthorsOpenReview ↗

Response to Reviewer sZN2 (2/4) — W1 continued — Phi teacher

W1. Testing the label-free setting with unreliable teachers and corrupted answers (continued)

Full construction with Phi-4-mini-instruct. Phi supplies solutions, references, synthesized prompts, and refinement responses throughout construction. Both methods share references and follow the original student evaluation protocol. GEPA-T receives the paired K2P construction cost as its budget. On QuaRTz, the shared teacher-side extractor was fixed before construction. The table reports mean ± sample SD on separate test sets.

TaskStudentConstruction seedsZero-shotK2PGEPA-T
Colored ObjectsQwen 0.8B349.0056.63 ± 9.3356.03 ± 12.70
Colored ObjectsMinistral 3B369.1092.43 ± 2.6170.23 ± 0.59
TrackingQwen 0.8B321.1866.99 ± 3.0872.75 ± 6.01
TrackingMinistral 3B378.0489.93 ± 4.3282.03 ± 3.22
QuaRTzQwen 0.8B382.5378.15 ± 2.5681.12 ± 0.77
QuaRTzMinistral 3B385.2087.33 ± 2.5389.20 ± 0.19

Colored Objects. K2P's three-seed mean gains over zero-shot with Phi are 7.63 pp for Qwen and 23.33 pp for Ministral.

Tracking. K2P exceeds zero-shot in all six runs. Its mean is 5.75 pp below GEPA-T for Qwen and 7.91 pp above GEPA-T for Ministral. Phi produces usable answers on 66/160 search/selection questions, so this setting includes both answer errors and formatting failures.

The Tracking comparison uses the same initialization protocol for all three seeds, synthesizing until eight prompts pass the original admission checks. All attempts, including rejected outputs, are included in K2P's construction cost and GEPA-T's matched budget. This protocol was fixed after the initial run stopped at 32 synthesis attempts with two admitted prompts and before further generation. The three seeds required 611, 285, and 406 attempts.

Interpreting QuaRTz. K2P's three-seed mean exceeds zero-shot for Ministral and falls below it for Qwen. GEPA-T has higher means for both students. Retrospective tests of all 70 candidates across six frozen archives show that each Qwen archive's highest test score is below zero-shot. Each Ministral archive contains prompts that outperform zero-shot, and the selected prompts score 0–4.34 pp below their archives' highest test scores. Reselection using gold or 9B references on the same 80-question selection pool yields mixed changes in test accuracy.

We have added these results to Sections 6.2 and 7.2. Appendix B.10 reports initialization details, per-seed candidate scores, and reselection diagnostics. Appendix B.11 presents the full-construction corruption curve (Figure 7) and the fixed-candidate analysis.

Our response to W2 further relates real teacher errors to candidate rankings.

AuthorsOpenReview ↗

Response to Reviewer sZN2 (3/4) — W2 — Theory and empirical diagnostics

W2. Empirical checks of theoretical conditions and correlated teacher errors

Thank you for highlighting these issues. We agree that the theoretical quantities need a clearer empirical connection and that correlations between teacher errors, question difficulty, and student errors deserve explicit treatment. We measured expected proposal gains at 96 recorded search states from the 24 four-round construction runs. We also measured agreement–accuracy mismatch across 60 combinations of archive, data role, and reference teacher. These diagnostics connect the theoretical quantities to observed construction behavior.

Empirical diagnostics of search progress and agreement

Larger κ\kappa sets a stronger progress target and can require a larger shortfall allowance εt\varepsilon_t. Likewise, larger γ\gamma strengthens the agreement–accuracy condition and can require a larger residual ω\omega. We therefore evaluate several reference values of κ\kappa and γ\gamma and measure the corresponding shortfalls and residuals on the recorded states and candidate archives.

QuantityMeaningEmpirical evaluation
κ\kappaRequired fraction of the remaining agreement gap closed in expectationReference values 0.25, 0.5, and 1
εt\varepsilon_tAllowance for expected-progress shortfallEstimated conditional shortfall from eight proposal gains per state
γ\gammaScale relating accuracy and agreement deficitsReference values 0.25, 0.5, 0.75, and 1
ω\omegaResidual mismatch in that relationSmallest residual covering each observed archive at fixed γ\gamma

Conditional expected search progress. At each of the 96 recorded states, we drew eight fresh proposals from the same recorded parent, holding the feedback, teacher decoding, student evaluation seeds, parser, admission rules, and 80-question search pool fixed. All draws enter the means, with zero accepted gain for invalid or non-improving proposals.

  • Mean accepted gain is 1.41 pp, with conditional Monte Carlo SE 0.12 pp and 95% Hoeffding interval [0.43, 2.39] pp.
  • 190/768 proposals improve agreement (24.74%). The SD across the three construction-seed-specific mean gains is 0.54 pp.

At each state, we first average the eight gains, then compute the positive shortfall against κd\kappa d. The two reference gaps are the retrospective gap from the parent to its run's final archive best and 1−parent agreement1-\text{parent agreement}, an upper bound on the gap to any admissible prompt on the fixed search pool. Table entries average the statewise shortfalls.

Reference κ\kappaEstimated mean shortfall using archive gap, ppEstimated mean shortfall using gap upper bound, pp
0.250.402.45
0.501.406.22
1.003.9513.85

At κ=0.5\kappa=0.5, the target is to close half the remaining gap in expectation. The estimated mean shortfall is 1.40 pp against the archive gap and 6.22 pp against the gap upper bound. These measurements describe the recorded parent/feedback states with student evaluation seeds fixed. The theoretical bound uses the conditional gap over the full admissible prompt space. Averaging the resulting true conditional shortfalls over construction histories gives a sufficient εt\varepsilon_t. The full-space and population guarantees use the assumptions stated in Section 5.

In the 24 construction trajectories that supplied these states, 17 have positive total accepted gain and 27/96 slots improve agreement. Strict acceptance keeps the parent's search score nondecreasing.

Agreement and accuracy. We evaluated 60 combinations of archive, data role, and reference teacher. Gold accuracy uses canonical exact matching on the same 80-question construction pools. For each reference γ\gamma, we use archive optima in Equation (7) and find the smallest required ω\omega. At γ=1\gamma=1, mean residuals over the three Colored Objects selection archives are as follows, on a 0–1 scale.

StudentMinimum archive residual with original teacherMinimum archive residual with Phi
Qwen 0.8B0.01250.3083
Ministral 3B0.02500.1083

Phi's larger residuals show a weaker agreement–accuracy relationship within these fixed candidates, linking the teacher-quality experiment to Equation (7). On QuaRTz/Qwen, the original-teacher residual averages 0.0458, and selection attains the archive's best gold score in all three pools. Thus positive residuals can coexist with successful selection.

AuthorsOpenReview ↗

Response to Reviewer sZN2 (4/4) — W2 continued and Q1 — Correlated errors and appendix

W2. Empirical checks of theoretical conditions and correlated teacher errors (continued)

Correlated teacher errors

The binary example gives an idealized sufficient condition for agreement to preserve accuracy rankings under classical independent classification noise. With valid, aligned scoring, a constant error probability ηT<1/2\eta_T<1/2, and teacher flips independent of the student's answer conditional on input and truth, it yields γ=1−2ηT\gamma=1-2\eta_T and ω=0\omega=0. Angluin and Laird, 1988

The general candidate-comparison bound retains the joint event that the teacher is wrong and two candidate answers differ, allowing errors to correlate with question difficulty and student outputs. Appendix C.5, Equation (C6), also handles unusable references and scoring-rule differences.

A fixed-pool bound also permits correlated errors. Under aligned scoring, agreement and correctness coincide whenever the teacher reference is correct. If ee is the fraction of incorrect or unusable teacher references, the fixed-pool accuracy and agreement scores therefore differ by at most ee for every prompt. Taking suprema gives the sufficient choice γ=1\gamma=1, ω=min⁡1,2e\omega=\min\\{1,2e\\} on that pool. On the fixed 80-question selection pools, the residual bounds are 0.025/0.025/0/0.150 on Tracking/Colored/TabMWP/QuaRTz with the original teacher, and 0.800 on Colored with Phi. The Phi bound is loose for percentage-point accuracy comparisons, motivating the more informative archive residuals and ranking measurements. A population guarantee uses the population error rate.

Actual ranking changes. On the same frozen Colored Objects archives constructed with Qwen3.5-9B, we count strict reversals between teacher-agreement and gold-accuracy rankings using each teacher's selection references.

StudentReversals with original teacherReversals with PhiMean gold selection loss with original teacherMean gold selection loss with Phi
Qwen 0.8B0/19829/1980.00 pp1.67 pp
Ministral 3B5/19858/1980.00 pp3.33 pp

Counts cover 66 pairs in each of three archives, sharing candidates and questions. Selection loss is the archive-best minus selected gold accuracy on the 80-question selection pool. Teacher-reference accuracies are 98.75% and 60.00%, respectively.

Sections 7.1–7.2 now summarize proposal gains, progress shortfalls, and selection reliability. Appendix B.12 details the diagnostics, and Appendix C.14 connects them to the theory.

Q1. Appendix organization, token costs, and bootstrap takeaways

Thank you very much for pointing this out. We have shortened the appendix exposition, clarified its section hierarchy, and added an appendix-wide roadmap with linked sections and page numbers. The main text also summarizes construction costs and the bootstrap findings.

  • The opening guide explains the organization into reproduction details and numerical records in Appendix A, empirical analyses in Appendix B, proofs and supporting analyses in Appendix C, and extended related work in Appendix D. New subheadings separate data, construction rules, and templates in A.2, baseline protocols and costs in A.3, and the two sample-size studies in B.4.
  • Repeated explanations in A–C have been consolidated, including admission and selection rules, bootstrap procedures, and theoretical interpretation. Cross-references lead to the full protocols, while C.14 concisely maps the empirical diagnostics to the theoretical mechanisms.

Section 6.1 specifies the model weights and construction-cost accounting, with complete costs in Table A2 (Appendix A.3). Section 6.2 reports actual shared-budget costs and links to Table B15. It also summarizes the paired-bootstrap findings and explains that the pointwise 95% intervals quantify test uncertainty for frozen prompts, alongside the construction-seed standard deviations in Table 1. Full intervals appear in Table A4 (Appendix A.4).

Reviewer_7h5jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes an unsupervised method to distill teacher knowledge to student via prompting tuning. The proposed method uses agreement between teacher and student as signal to select candidate system prompts.

Critical strengths
  • This work studies an important problem: knowledge distillation via unsupervised signal.

  • The proposed method obtain strong empirical results, outperforming label-free baselines.

Critical weaknesses
  • The manuscript is a bit hard to follow. Here are some examples:

1, in line 172 and line 175, y~i\tilde{y}_i and vv both refer to answer. It's unclear to me what the differences are.

2, in line 200, the term "admissible, fully evaluated revision" is introduced without explaining what "admissible" or "fully evaluated" mean.

3, in line 207, "early stopping" is introduced without any further explanation.

Since mathematical notations have been introduced, I would recommend rigorously define those terms.

  • In the method, it is mentioned that the set is divided into source, search and reserved (iiuc, the last set should be probably be renamed to selection set to resonate with the notation). However, it is not clear to me how the experiments construct the eval set. Are the reserved sets being used for eval?
Questions for authors

I have some general questions about the method: 1, in line 37 to 38, it is mentioned that "questions and public context may be available before their correct answers are established". However, in neither the method nor Figure 1, I do not see how the public context is used. Have I missed anything?

2, in line 189, it is written that "the teacher synthesizes instructions from small subsets of source solutions. Here, what does the term "instruction" mean? Is it the prompt mentioned in line 138? How does the teacher synthesize the instructions? For the bank P, is it a pool of such synthesized instructions?

3, in line 197, the notation BtB_t is introduced. Can you clarify which set this batches are sampled from?

4, In line 198, it is mentioned that "cached student responses ....". From my understanding, the instructions will be changed, which is the student prompt. If so, how can it cache the student responses?

5, since the access to ground truth label is not assumed, what does "worked solution" mean?

6, in the theoretical analysis, in line 231, what does "illustrative aligned scoring regime" mean? Why do we constrain the analysis into this?

7, In experiment section, in line 312, it writes "80 questions per data role". What does data role mean?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7h5j (1/3) — W1a–W1c — Definitions and stopping

Thank you very much for reviewing our work and your valuable comments on method definitions, data usage and implementations. We have carefully addressed your comments, with detailed point-by-point responses provided below. We have also uploaded a revised manuscript with all changes highlighted in blue.

W1a. How do extracted answers and validity indicators differ?

Thank you very much for pointing this out. y~i=a(zi)\tilde y_i=a(z_i) is the teacher's extracted answer value, such as the option letter A. The quantities vT(i)v_T(i) and vS,p(i)v_{S,p}(i) are binary indicators of whether the teacher and student outputs, respectively, contain parseable, untruncated final answers.

For example, a completed teacher response with a valid final-answer line identifying option A can yield y~i=A\tilde y_i=\mathrm{A} and vT(i)=1v_T(i)=1, even if the true answer is B. Validity is determined from the output and the fixed parser. Ground-truth correctness is evaluated separately.

The agreement score in Equation (2) is one when both outputs are valid and their extracted answers match. Every other case contributes zero, with all evaluated questions retained in the denominator. Section 4.1 now places the answer-value and validity definitions together and includes this example.

W1b. What do “admissible” and “fully evaluated” mean?

Thank you for this question. We have clarified these definitions in the revised manuscript. An admissible revision is a task prompt that passes the pre-specified checks on structure, completeness, and length. A fully evaluated revision has a completed student response and agreement score for every search question.

The generated task prompt must be extractable, complete, and within the target student's prompt-length constraint. Duplicate prompt content retains its first archive entry.

Complete evaluation covers all 80 search questions in the reported configuration. Completed responses that fail answer validity still contribute zero to agreement. A missing evaluation leaves the candidate incomplete.

Every admissible revision with a complete search evaluation enters the archive. It replaces the current search parent only if its search agreement strictly improves. The Introduction and Figure 1 caption now explain the parent, archive, and admission and evaluation requirements. Sections 4.2–4.3 give the formal definitions and update rule, with detailed checks in Appendix A.2.

W1c. When does early stopping occur?

Thank you very much for pointing this out. Early stopping occurs after two consecutive empty feedback batches. A feedback batch is the group of up to three search cases included in one teacher revision request. The default configuration allows at most four refinement slots, with at most one proposal per nonempty batch. An empty batch consumes a slot.

Eligible cases require a parseable, untruncated teacher answer and a rendered request that fits the teacher context together with the output allowance. Those eligibility and length checks can leave a batch empty. An invalid revision consumes its slot. A completed, non-improving revision leaves the current parent in place. The stopping counter tracks empty batches. Section 4.3 now states this stopping condition when introducing refinement.

AuthorsOpenReview ↗

Response to Reviewer 7h5j (2/3) — W2 and Q1–Q2 — Data and task prompts

W2. Selection-set naming, test-set construction, and data separation

Thank you very much for pointing this out. The reported final accuracies are measured on a separate test set. The reserved pool is the selection set Usel\mathcal U_{\mathrm{sel}}, which chooses the final task prompt after the archive has been fixed. We have adopted “selection set” throughout the method. Sections 4.1 and 6.1 explicitly identify the three construction roles and the separate test set.

DataSize in the reported configurationUse
Source set80 questionsTeacher solutions supply examples for synthesizing initial prompts
Search set80 questionsStudent evaluation, feedback formation, and search-parent updates
Selection set80 questionsTeacher agreement selects a prompt from the fixed archive
Test setTask-specificGold-label accuracy of the frozen selected prompt

The test sets are constructed as follows.

TaskTest questionsProvenance and separation
Tracking510The remaining 170 questions from each of the three BBH variants after allocating 240 construction questions
Colored Objects1,000A fixed sample from the 1,080 examples outside the four earlier 80-question development pools, which include the three construction roles
TabMWP1,000A fixed official-test subset, with question identities and table groups disjoint from construction
QuaRTz784The complete official test split, checked for question-ID, background, and rendered-text separation from training

Source, search, and selection question sets are mutually disjoint. TabMWP construction roles also separate table groups, and QuaRTz construction roles separate shared-background groups. Candidate generation, parent updates, and final selection use teacher outputs. Test labels score the frozen selected prompt.

Q1. Where does public context enter the method?

Thank you very much for this question. Public context enters through the task input xx supplied to both the teacher and student. It comprises the materials accompanying each question, such as the table and answer options in TabMWP, the background paragraph and options in QuaRTz, and the scenario and options in Tracking. These fields are rendered together with the question.

Source examples therefore contain the question, its public context, and the teacher-generated solution. Search and selection evaluations supply the same task input to the student under each candidate prompt. The Introduction now gives examples of public context at first mention. Section 3 specifies the contents of xx, and the revised Figure 1 caption includes public context in the input.

Q2. Are “instructions” task prompts, and how is the initial bank built?

Thank you very much for this question. Yes. “Instruction” refers to the task prompt pp defined in the problem setup, reusable natural-language text that guides the student in solving questions from the same task. The initial bank P\mathcal P is a pool of eight distinct admitted task prompts. We now define the term and cite Promptbreeder (Fernando et al., 2024) in the opening paragraph of the Introduction, with the formal definition in Section 3. Promptbreeder

For each synthesis request, the teacher receives the public task description and three source questions with their complete teacher-generated solutions. The synthesis template asks for a reusable task procedure, removing example-specific names, values, and conclusions. The resulting task prompt is then checked for admission and evaluated with the target student.

Search agreement chooses the initial parent from the bank for refinement. The archive starts at C0=P\mathcal C_0=\mathcal P and subsequently retains each admissible, fully evaluated revision. Final selection considers this archive. Section 4.2 now includes a short Tracking task-prompt excerpt and points to the full construction example in Appendix A.6.

AuthorsOpenReview ↗

Response to Reviewer 7h5j (3/3) — Q3–Q7 — Feedback, caching, and scoring

Q3. Which set supplies the feedback batches?

Thank you very much for this question. Bt\mathcal B_t is assembled from the search set and contains up to three feedback cases for one teacher revision request. Each case pairs the teacher solution with the cached student response produced by the current parent on that question, including extracted answers and output status.

The implementation prioritizes unused disagreements, reused disagreements, unused agreements, and reused agreements, in that order. A deterministic order based on the task, construction seed, round, and question resolves ordering within each category. Each included case must satisfy the teacher-answer eligibility and request-length checks.

Section 4.3 now identifies the search set as the source of Bt\mathcal B_t. Appendix A.2 details the ordering and eligibility checks used to construct each feedback batch.

Q4. How are student responses cached when the prompt changes?

Thank you very much for this question. Each task prompt has its own cached search responses. Refinement uses the responses for the current parent prompt. Request identity includes the prompt, question messages, model revision, decoding configuration, and generation seed.

A changed prompt therefore receives its own responses and search scores. If an identical request recurs, its cached result can be reused. Caching preserves the response used for a given stochastic evaluation.

Section 4.3 now explicitly ties the cached responses to the current parent and states that each prompt has its own evaluations.

Q5. What does “worked solution” mean without gold labels?

Thank you very much for this question. A teacher-generated worked solution is the response the teacher produces when solving a public task input, including its reasoning and proposed final answer. Both the reasoning and the answer can contain errors.

K2P uses the full response as evidence for task-prompt synthesis and refinement, and extracts the final answer for agreement scoring. Section 4.1 now introduces teacher-generated worked solutions as fallible reasoning and answers.

Q6. What is the aligned scoring regime, and why is it used?

Thank you very much for this question. In the local comparison proposition, aligned scoring means that correctness and agreement both use equality of normalized answers, with valid, untruncated responses. We use this setting to isolate how teacher errors affect candidate comparisons. Appendix C.5, Equation (C6), extends the result to invalid teacher answers and the actual task-scoring rules.

Under the simpler conditions, a correct teacher makes agreement coincide with correctness on that input. If two prompts produce the same student answer, a teacher error affects their scores equally.

These observations yield

∣b(p)−b(q)∣≤2Pr⁡(O∩Dpq),|b(p)-b(q)| \leq 2\Pr(O\cap D_{pq}),

where b(p)=J(p)−J~(p)b(p)=J(p)-\widetilde J(p), OO is the event that the teacher is wrong, and DpqD_{pq} is the event that the two candidate answers differ. The result identifies where teacher errors can distort a comparison.

The archive-selection analysis uses the actual score offset b=J−J~b=J-\widetilde J. The Introduction now spells out the answer-matching conditions when presenting the comparison result, and Related Work points to Section 5. Section 5 states these assumptions and points to Appendix C.5, Equation (C6), for the actual-scoring extension.

Q7. What does “80 questions per data role” mean?

Thank you very much for this question. This means 80 source questions, 80 search questions, and 80 selection questions, for a total of 240 construction questions.

Each synthesis request draws three examples from the source pool. Each refinement request uses up to three search cases. Candidate search evaluation and final selection each use their respective 80-question pools. Final accuracy is measured on the separate task-specific test set.

Section 6.1 now gives these three set sizes explicitly and identifies the separate test set.

Reviewer_8257OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to turn a stronger teacher's solutions into a reusable prompt for a frozen student without ground-truth answers. K2P has the teacher write instructions from its own solutions, revises them using paired teacher and student responses, and selects the final prompt by agreement with teacher answers on held-out questions. The paper also splits the final gap into generation and selection parts, and evaluates on four reasoning tasks with two small students.

Critical strengths
  1. The motivation and problem setup are important. In practice, teacher answers can be wrong, and without labels it is not clear how to judge a candidate prompt. The paper states this problem clearly and designs the method around it.

  2. The evaluation is careful, and the comparison even includes supervised GEPA. K2P outperforms the label-free baselines in most settings and stays close to supervised methods. Keeping all evaluated revisions and selecting on reserved questions is also an interesting idea.

Critical weaknesses
  1. My main concern is that the label-free setting is not really tested. The teacher is correct on almost all test questions in every task, so its answers work almost like ground-truth labels. The main difficulty the paper describes, an unreliable teacher, therefore rarely appears in the experiments. I would like to see harder tasks where the teacher is often wrong, or a controlled study that corrupts teacher answers at different rates.

  2. Simple label-free baselines are missing. Fine-tuning the student on teacher solutions, or adding a few teacher solutions as in-context examples, also needs no labels. If the student must stay frozen, few-shot prompting is still a natural baseline. Without these, it is hard to see when a distilled instruction is the better choice.

Questions for authors
  1. Why is the number of refinement rounds fixed at four? The theory needs many rounds for the bound to shrink, so it would help to see how accuracy changes with more rounds.

  2. Several baselines vary a lot across the three construction seeds. Do the gains over GEPA-T hold beyond a few weak seeds?

  3. The baselines are capped at K2P's own construction cost. Would the ranking change under a fixed budget shared by all methods?

  4. The models are small, and one student is from the same family as the teacher. All tasks also have short final answers. A larger student or a teacher from another family would show how general the method is.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 8257 (1/5) — W1 — Teacher reliability

Thank you very much for reviewing our work and for your valuable suggestions on teacher reliability, simple baselines, and experimental robustness. We have carefully addressed your comments, with detailed point-by-point responses provided below. We have also uploaded a revised manuscript with all changes highlighted in blue.

W1. Performance with unreliable teachers and corrupted answers

Thank you very much for pointing this out. We agree that the original experiments provide limited coverage of unreliable teachers. We added full construction with Phi-4-mini-instruct and controlled corruption of teacher answers at rates up to 50% on QuaRTz. Phi-4-Mini technical report

We use Phi to test construction with less reliable teacher references. Its reference accuracy on the 160 search/selection questions is 26.88% for Tracking, 63.75% for Colored Objects, and 87.50% for QuaRTz, with unusable answers counted as incorrect. On the same Colored Objects questions, Qwen3-8B and Ministral-3-8B achieve 95.63% and 78.13%, respectively. Model descriptions appear in the Qwen3 report and Ministral 3 report.

With Phi throughout construction, K2P's three-seed mean exceeds zero-shot in five of six task/student settings, while its comparison with GEPA-T varies across settings.

Controlled answer corruption throughout construction. On QuaRTz, we flip final A/B answers in all three 80-question pools (source, search, selection), updating the full response's final answer and parsed reference while retaining the original reasoning. Qwen3.5-9B performs synthesis and refinement. Each corrupted condition reruns synthesis, search, feedback, refinement, and selection before testing on all 784 questions. The 0% condition reuses the original runs.

Values are mean ± sample SD over three construction seeds, each with its own corruption mask at each rate. Masks are shared across students and nested across rates. Reference accuracies are averaged across masks and measured with gold answers after construction.

ConditionSearch reference accuracySelection reference accuracyQwen3.5 0.8BMinistral3 3B
Zero-shot——82.5385.20
0%95.0092.5081.59 ± 3.3891.45 ± 1.42
10%85.0082.0880.27 ± 1.2888.22 ± 1.49
20%77.9273.7579.93 ± 2.0791.33 ± 1.40
30%67.5066.6779.00 ± 1.4891.33 ± 1.40
40%60.0058.3380.36 ± 0.9688.56 ± 0.85
50%50.0050.4275.55 ± 9.2490.26 ± 0.85

For Ministral, the 20% and 30% conditions select the same final prompt for each seed and reuse the same cached test outputs.

Ministral exceeds zero-shot in every run, including a 5.06 pp mean gain at 50% flipping. Qwen's means are below zero-shot at every rate and fall by 6.04 pp from 0% to 50%. Test accuracy varies non-monotonically across rates.

AuthorsOpenReview ↗

Response to Reviewer 8257 (2/5) — W1 continued and W2

W1. Performance with unreliable teachers and corrupted answers (continued)

Full construction with Phi-4-mini-instruct. Phi supplies solutions, references, synthesized prompts, and refinement responses throughout construction. Both methods share references and follow the original student evaluation protocol. GEPA-T receives the paired K2P construction cost as its budget. On QuaRTz, the shared teacher-side extractor was fixed before construction. The table reports mean ± sample SD on separate test sets.

TaskStudentConstruction seedsZero-shotK2PGEPA-T
Colored ObjectsQwen 0.8B349.0056.63 ± 9.3356.03 ± 12.70
Colored ObjectsMinistral 3B369.1092.43 ± 2.6170.23 ± 0.59
TrackingQwen 0.8B321.1866.99 ± 3.0872.75 ± 6.01
TrackingMinistral 3B378.0489.93 ± 4.3282.03 ± 3.22
QuaRTzQwen 0.8B382.5378.15 ± 2.5681.12 ± 0.77
QuaRTzMinistral 3B385.2087.33 ± 2.5389.20 ± 0.19

Colored Objects. K2P's three-seed mean gains over zero-shot with Phi are 7.63 pp for Qwen and 23.33 pp for Ministral.

Tracking. K2P exceeds zero-shot in all six runs. Its mean is 5.75 pp below GEPA-T for Qwen and 7.91 pp above GEPA-T for Ministral. Phi produces usable answers on 66/160 search/selection questions, so this setting includes both answer errors and formatting failures.

The Tracking comparison uses the same initialization protocol for all three seeds, synthesizing until eight prompts pass the original admission checks. All attempts, including rejected outputs, are included in K2P's construction cost and GEPA-T's matched budget. This protocol was fixed after the initial run stopped at 32 synthesis attempts with two admitted prompts and before further generation. The three seeds required 611, 285, and 406 attempts.

Interpreting QuaRTz. K2P's three-seed mean exceeds zero-shot for Ministral and falls below it for Qwen. GEPA-T has higher means for both students. Retrospective tests of all 70 candidates across six frozen archives show that each Qwen archive's highest test score is below zero-shot. Each Ministral archive contains prompts that outperform zero-shot, and the selected prompts score 0–4.34 pp below their archives' highest test scores. Reselection using gold or 9B references on the same 80-question selection pool yields mixed changes in test accuracy.

We have added these results to Sections 6.2 and 7.2. Appendix B.10 reports initialization details, per-seed candidate scores, and reselection diagnostics. Appendix B.11 presents the full-construction corruption curve (Figure 7) and the fixed-candidate analysis.

W2. Teacher-solution few-shot prompting as a label-free baseline

Thank you very much for this comment. We completely agree that it is important to compare with some simple label-free baselines. Teacher-solution few-shot prompting is a natural baseline for this paper's frozen-student setting, and we have added it for comparison. Across all four tasks, both students, and the original three construction seeds, K2P has higher mean accuracy in seven of eight settings. The scores are close on QuaRTz.

Each run uses the three source examples from its first scheduled synthesis request, together with their complete cached teacher responses. Example choice is fixed by the construction schedule before evaluating the baseline. We retain the original task instructions and match the student model, decoding, test questions, and scoring rules.

TaskStudentTeacher-solution few-shotK2P
TrackingQwen 0.8B21.24 ± 1.0889.22 ± 2.26
TrackingMinistral 3B84.71 ± 5.4395.36 ± 1.20
Colored ObjectsQwen 0.8B39.83 ± 4.3781.80 ± 3.58
Colored ObjectsMinistral 3B82.10 ± 10.2795.77 ± 1.47
TabMWPQwen 0.8B63.30 ± 0.4468.70 ± 4.96
TabMWPMinistral 3B83.37 ± 9.0490.83 ± 1.86
QuaRTzQwen 0.8B81.80 ± 1.7981.59 ± 3.38
QuaRTzMinistral 3B91.07 ± 0.6791.45 ± 1.42

K2P wins 19 of the 24 construction-seed comparisons. On Tracking and Colored Objects, the synthesized task prompts produce higher accuracy than directly presenting these three teacher solutions.

At deployment, K2P uses fewer total student input-plus-output tokens per question in six of eight settings. The ratios of K2P to few-shot token usage range from 0.27 to 0.70 in those six settings. For Qwen on Tracking and Colored Objects, the ratios are 1.20 and 1.54, respectively, accompanied by substantially higher accuracy and longer student responses.

Student weights are fixed during both construction and deployment, as stated in Section 3. We have added the few-shot baseline to Table 1, with its protocol and deployment-token comparison in Appendix B.6.

AuthorsOpenReview ↗

Response to Reviewer 8257 (3/5) — Q1 — Rounds and Q2 — Seeds

Q1. Why four refinement rounds, and what happens with more?

Thank you for suggesting an evaluation with more refinement rounds. We evaluated increasing the refinement limit from four to eight, twelve, and sixteen rounds.

The comparison covers 24 construction trajectories (four tasks, two students, three seeds), with shared initial banks and trajectory prefixes across limits. Each limit caps the number of refinement slots, and the original candidate-generation, strict parent-update, and early-stopping rules apply throughout. At each limit, the original selection rule chooses a prompt from the available archive for evaluation on the full test set. The table reports mean ± sample SD across the three seeds for four rounds and above.

TaskStudent4 rounds8 rounds12 rounds16 roundsConstruction cost, 16/4
TrackingQwen 0.8B89.22 ± 2.2688.95 ± 3.3589.74 ± 4.1189.74 ± 4.111.956
TrackingMinistral 3B95.36 ± 1.2097.39 ± 2.3697.39 ± 2.3697.97 ± 1.422.078
Colored ObjectsQwen 0.8B81.80 ± 3.5883.93 ± 4.4883.37 ± 3.8783.37 ± 3.872.368
Colored ObjectsMinistral 3B95.77 ± 1.4796.83 ± 0.7096.83 ± 0.7096.90 ± 0.802.285
TabMWPQwen 0.8B68.70 ± 4.9668.70 ± 4.9668.23 ± 4.1569.67 ± 4.521.927
TabMWPMinistral 3B90.83 ± 1.8690.43 ± 1.6090.23 ± 1.6790.23 ± 1.672.311
QuaRTzQwen 0.8B81.59 ± 3.3881.16 ± 3.0681.16 ± 3.0681.51 ± 3.652.168
QuaRTzMinistral 3B91.45 ± 1.4290.94 ± 1.5590.94 ± 1.5590.94 ± 1.552.424

At 16 versus four rounds, five of the eight task/student means increase and three decrease, with changes from −0.60 to +2.61 pp. Across the 24 paired runs, accuracy improves in ten, declines in six, and stays unchanged in eight. Mean construction cost rises to 1.93–2.42 times the four-round cost. The last column is the ratio of group mean costs, measured using the paper's model-weighted input/output tokens and including teacher-reference acquisition.

The results support four rounds as a practical cost–accuracy choice, with further gains in some settings.

Equation (6) applies at every finite round count. Sufficient proposal progress can reduce the optimization term, and the full bound also includes search optimism and selection uncertainty. Driving the bound to zero requires all three terms to vanish as resources increase. Here the selection pool is fixed at 80 questions, and additional candidates can change which prompt it selects. Section 7.1 now summarizes these findings. Appendix B.8 presents the complete results and construction costs (Tables B18–B19), the curves (Figure 6), and the finite-round explanation.

Q2. Do gains over GEPA-T persist across additional seeds?

Thank you for raising this point. Relative performance varies across construction seeds and task/student settings. The mean K2P–GEPA-T difference across the eight settings is −0.72 pp over the three additional seeds and +1.08 pp over all six reported seeds, with higher six-seed mean accuracy for K2P in five of eight settings.

The additional construction seeds are 2, 12, and 14. Both methods follow the original comparison protocol, including the task splits, student models, decoding, test questions, and scoring. GEPA-T receives a construction budget equal to the paired K2P run's cost.

The accuracy columns report mean ± sample SD over these three seeds. The final column combines them with the original seeds 0, 20260715, and 20260609. Differences are K2P minus GEPA-T, calculated before rounding.

TaskStudentK2P, additional threeGEPA-T, additional threeThree-seed difference (pp)Six-seed difference (pp)
TrackingQwen 0.8B93.20 ± 1.5890.65 ± 4.14+2.55+1.83
TrackingMinistral 3B95.23 ± 3.1496.67 ± 1.48−1.44−0.75
Colored ObjectsQwen 0.8B74.60 ± 5.9077.57 ± 4.57−2.97+1.43
Colored ObjectsMinistral 3B96.70 ± 0.5293.03 ± 0.06+3.67+7.25
TabMWPQwen 0.8B66.23 ± 1.9768.73 ± 4.91−2.50−0.77
TabMWPMinistral 3B83.93 ± 9.6085.83 ± 7.24−1.90+0.15
QuaRTzQwen 0.8B79.17 ± 0.9082.91 ± 1.33−3.74−1.93
QuaRTzMinistral 3B92.09 ± 0.9691.50 ± 1.08+0.60+1.42

The eight-setting mean differences for seeds 2, 12, and 14 are −0.96, −0.73, and −0.46 pp, respectively. In the six-seed comparison, the three settings with lower K2P mean accuracy have differences of −0.75, −0.77, and −1.93 pp.

Appendix A.9 reports the individual results for seeds 2, 12, and 14 in Table A7 and the three-seed and six-seed summaries in Table A8, showing how the comparison varies across construction seeds and task/student settings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

Dear Area Chair and Reviewers,

Thank you for your careful reviews and constructive suggestions. We have posted point-by-point responses and uploaded a revised manuscript with changes highlighted in blue. We summarize the main revisions below to help navigate the responses.

K2P studies how to transfer teacher knowledge into reusable task prompts when task-specific ground-truth answers are unavailable and student weights remain frozen. Its construction procedure and analysis address both the quality of generated candidates and the reliability of teacher agreement for selecting among them.

  • Clearer definitions and a reproducible procedure (7h5j). We now explain task prompts, answer validity, admissibility, complete evaluation, and early stopping where they are introduced. The revised method explicitly distinguishes source, search, selection, and test data, and explains how feedback and caching depend on the current parent prompt. Sections 3–4 and Appendix A provide the definitions, scoring rules, templates, and pseudocode.

  • Additional baselines, budgets, and construction seeds (8257). We compared all seven optimizers under the same fixed 45-million model-weighted token ceiling across four tasks, two students, and three construction seeds. K2P has higher mean accuracy than GEPA-T in seven of eight settings, averaging 2.36 percentage points higher across settings. The teacher-solution few-shot comparison also favors K2P in seven of eight settings. Under the original comparison protocol, K2P had higher mean accuracy than GEPA-T in five of eight settings over the six reported seeds. The average difference across all eight settings was 1.08 percentage points in favor of K2P. With the 8B student, K2P has higher three-seed mean accuracy than GEPA-T on three tasks and is closely matched on QuaRTz. We also evaluated refinement limits of 8, 12, and 16 rounds. These results appear in Section 6.2, Appendix A.9, and Appendices B.6–B.9.

  • Teacher reliability throughout construction (8257 and sZN2). We evaluated Phi-4-mini-instruct as the teacher on three tasks with both students and three seeds. K2P exceeds zero-shot mean accuracy in five of six settings. We also reran the full construction procedure on QuaRTz with controlled final-answer corruption up to 50%, retaining the original reasoning. Ministral maintains gains over zero-shot, while Qwen shows degradation under heavy corruption. Candidate-level diagnostics examine how candidate quality and final selection contribute to the observed outcomes (Section 7.2 and Appendices B.10–B.11).

  • A clearer connection between theory and experiments (sZN2). We measured proposal gains at 96 recorded search states using eight fresh proposals per state, and agreement–accuracy mismatch across 60 combinations of archive, data role, and reference teacher. These measurements quantify search progress and ranking reliability within the observed states and candidate sets. The revised discussion explains their relationship to the conditional theoretical guarantees. The general comparison bound also accommodates correlated teacher errors, unusable references, and actual scoring rules (Sections 5 and 7, Appendices B.12 and C.5).

We have shortened repeated appendix explanations, clarified the section hierarchy, and added a roadmap with linked sections and page numbers. The main text now summarizes costs, uncertainty, and the principal supplementary findings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.