acceptodds
Under review as a conference paper at ICLR 2027

DistillBench: Evaluating Unauthorized Distillation and Its Defenses

Abstract

Knowledge distillation is a standard technique for compressing and specializing models. However, the same mechanism can also enable unauthorized capability transfer from proprietary models to an adversary model. Industry reports indicate that such extraction increasingly occurs at scale, but provide few reproducible details. Existing studies use different assumptions about attackers, API access, resource budgets, and acceptable utility loss, making defense effectiveness difficult to compare across settings. We introduce DistillBench, a benchmark providing a unified protocol for the joint evaluation of adversarial distillation and defenses. DistillBench supports fixed and adaptive data acquisition and measures capability transfer alongside acquisition and training costs, teacher utility, teacher similarity, and detectability. We evaluate four distillation methods and six defenses across acquisition budgets, using two teachers and two students on mathematical reasoning. We find that no method dominates across settings, and evaluating a defense against a single distillation procedure can misrepresent its effectiveness. Most tested interventions reduce teacher utility more than student performance, with protection often weakening as acquisition budgets grow. We further find that defense effects can depend more on the student than the teacher. Fingerprints are not reliably detected after distillation, and paraphrasing weakens detection based on teacher similarity.

Primary area:
datasets and benchmarks
Keywords:
Distillation, Model Distillation, Benchmark, Defenses

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GRDCOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework that encompasses the distillation attack environment, including how a defender can modify traces, how an attacker can modify their trace acquisition strategy, and metrics to evaluate the defense method. The author’s curate a dataset of 1k GSM8K traces from two teachers, apply various defense methods, and attacker sampling techniques (from the defender black box), and distill two students using the poisoned traces.

Critical strengths
  • The overall idea of this paper is strong in that it aims to fairly evaluate different antidistillation methods and generalizes the attacker-defender actions and environment for this to take place.
  • The authors provide interesting takeaways around fingerprinting as well as the paraphrasing defense method.
  • Way this paper standardizes the training with token budgets instead of numbers of samples should be the standard across antidistillation methods. It’s surprising that existing methods don’t all do this.
Critical weaknesses
  • The framing of this paper is that existing defense methods are not directly compared against each other. However, the author’s of this work do not actually do their described analysis on the antidistillation methods in the literature which they explicitly cite (Savani et al., 2025; Ma et al., 2026; Allouah et al., 2026; Libon et al., 2026). Without comparing to these methods, I do not see how this paper can be useful to the community. If they are included in the revision, I would be inclined to raise my rating.
  • The author’s only include samples from GSM8K in their dataset, which I believe is a massive oversight. Concurrent work, “Hiding in Plain Sight: Detectability-Aware Antidistillation of Reasoning Models” (Hartman et al, 2026), showed differing results depending on the dataset.
Questions for authors
  • I think the general point of this paper can be made with the two teachers and students, however using more performant teacher-student pairs on harder datasets would be very interesting to see and would potentially lead to this paper having more impactful results.
  • Can you run your analysis on the SOTA antidistillation methods (ADS, DOGe, etc)? The lack of this is the main reason why I selected the rating that I did.
Reviewer_izeLOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DistillBench, a framework for jointly evaluating unauthorized knowledge distillation and its defenses under explicit access and resource constraints. Experiments compare four distillation methods and six defenses, measuring student performance, teacher utility, computational costs, and detectability.

Critical strengths
  • The joint evaluation of attacks, defenses, and teacher utility provides a useful framework for assessing practical defense effectiveness.
  • Experiments include a frontier text-only teacher, multiple acquisition budgets, and teacher–student cross-pair comparisons, with variation across seeds reported for most settings.
Critical weaknesses
  • Outdated students and limited benchmark coverage. Although GPT-6-Astra is included as a teacher, the students are restricted to Qwen2.5-1.5B and 7B. Modern alternatives at comparable scales, such as Qwen3.5-2B and 9B, are absent. This is particularly consequential because student-dependent defense behavior is a central finding. Moreover, reported results focus on only 200 GSM8K test questions, where Astra already achieves 99.5% accuracy. The absence of harder reasoning and coding evaluations substantially limits the representativeness of the benchmark and its conclusions.
  • Student comparisons are confounded by training configurations. The 1.5B student uses full fine-tuning, whereas the 7B student uses LoRA, with different learning rates. Consequently, differences in resistance to paraphrasing or collapse under truncation cannot be confidently attributed to the student model itself. The teacher–student cross-pair experiments do not resolve this confounding.
  • Unclear acquisition accounting and reproducibility. The adaptive acquisition policy ranks prompts using their teacher responses before selecting examples under the stated budget. It is unclear whether obtaining all responses required for ranking is included in the acquisition cost. This ambiguity weakens the interpretation of the experiment as budget-constrained adaptive acquisition. The scoring model is also described inconsistently in Appendices A.3 and B.7.
Questions for authors
  1. Do the main findings persist with modern students such as Qwen3.5-2B/9B on harder reasoning and coding tasks, using controlled fine-tuning configurations to separate model effects from optimization effects?
  2. How are the teacher responses required to rank the full prompt pool obtained and charged against the acquisition budget, and which model actually performs this ranking?
Reviewer_BLpMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce DistillBench, a benchmark providing a unified protocol for the joint evaluation of adversarial distillation and defenses. DistillBench supports fixed and adaptive data acquisition and measures capability transfer alongside acquisition and training costs, teacher utility, teacher similarity, and detectability. Empirically, the authors evaluate four distillation methods and six defenses across acquisition budgets, using two teachers and two students on mathematical reasoning. The authors also provide well organized results around six questions.

Critical strengths
  • The paper is clearly written and easy to follow. Organizing the experimental results around six concrete research questions makes the findings easy to understand, while the analysis goes beyond simply reporting accuracy changes and provides useful insights into the interactions between distillation methods and defenses.

  • The proposed DistillBench framework explicitly separates data acquisition from student training and jointly evaluates adversarial distillation and defenses under common access and resource constraints. This provides a useful basis for fair comparisons and for understanding how different experimental settings affect defense effectiveness.

  • The experiments include a recent commercial LLM, GPT-6 Astra, as a teacher model, which adds practical relevance to the evaluation of unauthorized distillation in real-world settings.

Critical weaknesses
  • The defense suite is less representative than the framework suggests. Four of the six defenses are basic text transformations, including two truncation variants; the remaining two primarily target attribution. Specialized antidistillation methods such as Antidistillation Sampling (ADS) [1] and DOGe [2] are discussed but not evaluated, despite having public implementations, which limits the benchmark’s coverage of defense methods.

  • The empirical scope is narrow relative to the benchmark’s intended coverage. Table 5 lists MATH, MBPP+, LiveCodeBench, and Qwen2.5-Coder variants without reporting corresponding results. Since defense effectiveness may depend on factors such as reasoning length, answer format, or task difficulty, the generalizability of the findings beyond GSM8K remains unclear.

[1] Antidistillation Sampling. NeurIPS 2025
[2] DOGe: Defensive Output Generation for LLM Protection Against Knowledge Distillation. arXiv 2025

Questions for authors
  • Could ADS [1] and DOGe [2], both of which have public implementations, be evaluated on the open-weight teacher? If their inclusion is impractical, could the authors clarify the implementation or resource constraints?

  • Why are results for the additional datasets and Qwen2.5-Coder variants not reported? These results would help further examine whether the main findings extend beyond the current setting.

4 comments

Dear Reviewer BLpM,

Thank you again for your review. As promised in our initial response, we have now completed all the requested experiments. This includes a) new benchmarks: instruction-following IFBench and IFEval, and coding LiveCodeBench (LCB), b) new defences, including trace-rewrite, ADS, and DOGe. We now complete our previous response by these new experiments.

RE: new defenses

We incorporated ADS and DOGe. For ADS, we adapted the default hyperparameter (λ) and also report a λ=0 control to separate the cost of sampling from the cost of the ADS term. We report the results at the budget of 100k released output tokens on GSM8K: (The full budget sweep will be reported in the revised manuscript.)

GSM8K student accuracy (%), Llama3-8B→Qwen2.5-1.5B. parentheses denotes changes in points vs. the undefended student

Distillation methodUndefendedADS λ=0ADS λ=0.0079DOGeTrace-rewrite
SFT65.065.554.0 (−11.0)60.2 (−4.8)47.3 (−17.7)
seq-KD66.265.552.2 (−14.0)65.8 (−0.4)31.3 (−34.9)
Proxy-KD60.055.849.8 (−10.2)58.8 (−1.2)37.2 (−22.8)
LoRD57.359.855.5 (−1.8)49.0 (−8.3)38.3 (−19.0)

The teacher accuracies are 86.6% for undefended setting, 79.4% for ADS λ=0, 65.3% for ADS λ=0.0079, and 62.3% for DOGe. ADS and DOGe incur substantial teacher accuracy drop, while their effects on students depend on the attack. For example, seq-KD loses 0.4 points under DOGe, whereas LoRD loses 8.3 points. The ADS control also separates the effect of sampling from the ADS term.

Trace-rewrite achieves a more favourable trade-off here. As reported in our first response, teacher accuracy remains essentially unchanged, from 86.6% to 86.7% for greedy responses and 73.6% to 74.8% for sampled responses, while student accuracy falls by 17.7–34.9 points at 100k tokens. We will revise the discussion to reflect this stronger defense and qualify the original finding about the utility cost of the defenses.


RE: new datasets

We report results for two new datasets: LCB and IFEval/IFBench.

LCB student accuracy (%). Qwen3-8B→Llama3.2-3B. Parentheses denotes changes in points. (Seq-KD ADS is expensive, and the run is still in progress)

Distillation methodUndefendedTrace-rewriteADS λ=0ADS λ=0.0158DOGe
SFT15.06.3 (−8.7)15.0 (-0.0)12.9 (−2.1)14.1 (−0.9)
seq-KD15.67.2 (−8.4)TBCTBC11.7 (−3.9)
Proxy-KD15.96.9 (−9.0)14.7 (-1.2)13.5 (−2.4)10.8 (−5.1)

Trace-rewrite lowers attack accuracies by more than half, while teacher accuracy increases from 70.5% to 73.4%. The corresponding teacher scores are 70.8% for ADS λ=0, 70.5% for ADS λ=0.0158, and 69.8% for DOGe. ADS is tested by the authors only on simpler benchmarks like GSM8K and MATH with short reasoning. For benchmarks with long reasoning, running ADS over full thinking traces is very costly (~1400 H100-hours). Therefore in LCB, we run ADS only on the teacher’s answer. We realised that answer-only ADS has minimal effect on both the teacher and the student performance. DOGe’s effect is slightly stronger, teacher accuracy is decreased merely by 0.7 points, while the student accuracy fluctuates. From the full result, ProxyKD student went from +3.0 points to -5.1 points as the attack budget increases.

Building on the reasoning-shown/hidden experiments in our first response, we also evaluated trace-rewrite and DOGe on IFEval and IFBench, and added Qwen3.5-2B as a new student. Here we report the results at 8,000 training traces, and the full budget sweep is included in the revised manuscript:

IFEval/IFBench student accuracies (%). Qwen3-8B in thinking mode as teacher.

Student/distillation methodUndefendedTrace-rewriteDOGe
Llama-3.2-1B/SFT36.6/12.618.7 (−17.9)/15.8 (+3.2)38.5 (+1.9)/11.1 (−1.5)
Llama-3.2-1B/seq-KD37.3/12.219.5 (−17.8)/16.9 (+4.7)39.8 (+2.5)/13.0 (+0.8)
Llama-3.2-1B/Proxy-KD36.8/12.117.1 (−19.7)/18.2 (+6.1)39.1 (+2.3)/11.2 (−0.9)
Qwen3.5-2B/SFT70.5/24.229.6 (−40.9)/15.8 (−8.4)69.2 (−1.3)/22.1 (−2.1)

DOGe changes neither the teacher nor the student. Teacher accuracy increases by 0.8 points, and every student stays within about 2.5 points. However, compared to GSM8K, trace rewrite yields a big drop (21.6 points) in teacher accuracy. Under IFEval, student accuracies are dropped by at most 20 points. In a difficult, out-of-scope benchmark, IFBench, students actually increase their accuracies by 5 points. This makes trace rewrite a less effective defense, meaning that defense effectiveness also depends on the downstream task.


We hope that our additions have addressed all your concerns, and would be happy to answer any further questions. We also will be grateful if you would kindly consider increasing your score.

Dear Reviewer BLpM,

We would like to follow up on our additional experiments and clarifications. Please let us know if there are any remaining concerns we could address before the discussion period ends. If these additions have resolved your concerns, we would be grateful if you would consider increasing your rating.

Thank you again for your time and consideration.

Thank you to the authors for their efforts and additional experiments. While the rebuttal has partially addressed my concerns, I believe the paper still requires substantial revision, particularly given that several important experiments were missing from the original submission. After also considering the comments from the other reviewers, I have decided to maintain my original score.

Dear Reviewer BLpM,

Thank you for your response and for considering the rebuttal. We respectfully note that the experiments added in response to the reviews are now in the revised manuscript, not only in the rebuttal: ADS and DOGe (Sec. 4.2, Appendix C), the LCB and IFEval/IFBench tasks (Sec. 4.1 and 4.3), and the LoRA vs. full fine-tuning control (Sec. 4.4, Appendix D). They extend, rather than change, the paper’s protocol and its central finding that attacks and defenses must be evaluated together. We appreciate your feedback, which has substantially strengthened the paper. It will also be great if you can specify what concern has remain, so that we can address them in the final version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.