acceptodds
Under review as a conference paper at ICLR 2027

Release-Time Anti-Distillation for Reasoning Models

Abstract

Frontier reasoning models are routinely distilled by third parties who query the model, collect (prompt, reasoning trace, answer) triples and fine-tune their own students on them. We study whether a model owner can *release* a model whose outputs remain fully useful to ordinary users yet are poor - or actively harmful - as distillation data. Our key observation is that the two audiences consume different parts of a response: users consume the final answer, whereas distillers extract most of their value from the reasoning trace. We therefore propose **Answer-Preserving Trace Poisoning (APTP)**: a release-time transform that leaves the answer bit-identical and rewrites only the visible reasoning trace. We instantiate three mechanisms with distinct gradient-level effects - *information removal* (hidden / summary / hollow traces), *gradient shortcut* (an early natural-language "intuition" that leaks the answer, so the answer loss is satisfied without reasoning), and *gradient conflict* (a **decoy** trace whose working coherently supports a different answer while the stated answer stays correct) - and we release AntiDistill-Bench, a 16.7k-problem benchmark with paired clean/poisoned traces and an evaluation protocol that measures (i) utility preservation, (ii) imperceptibility to casual readers and LLM judges, and (iii) distillation-scaling curves of students trained on the released data.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Anti-Distillation, Reason Models

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_asBuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes three antidistillation methods. The proposed methods (Information removal, anchor, decoy) all act on the reasoning trace itself and do not modify the teacher accuracy. The authors include interesting negative results in section 4 about the decoy method. A benchmark is also provided which contains teacher traces generated by many teachers, which can be poisoned and used in distillation experiments. The core methods of this paper and experimented on extensively with multiple teachers and students, though only GSM8K and MATH are used as datasets.

Critical strengths
  • The answer-preserving formulation of this work is important to maintain performance (in terms of answer correctness for users).
  • Negative result 2 is a good result that actually justifies why some other antidistillation methods are not useful. In my opinion, this result should be expanded upon and used to justify why those methods aren’t valuable specifically.
  • The benchmark with teacher traces can be valuable to the community to future work on new antidistillation methods.
Critical weaknesses
  • The overall framing of the utility of this as an anti-distillation method is poorly explained. Currently frontier labs only provide the answer and no reasoning trace. The justification for why a frontier lab would want to provide the reasoning trace is that it can be helpful for the user, improve trustworthiness, etc. However this work is suggesting poisoning this trace with a decoy. Initial methods on this topic (Antidistillation Sampling) struggle because of the student-teacher accuracy tradeoff and the incoherency in the reasoning traces. This paper is proposing actively modifying the trace to make it contradictory, which is even less helpful for a user and potentially dangerous.
  • This work is clearly missing important comparisons to recent work on this topic, including (but not limited to): The Distillation Game: Adaptive Evaluations & Efficient Defenses (Allouah et al) and Hiding in Plain Sight: Detectability-Aware Antidistillation of Reasoning Models (Hartman et al). While these may be considered concurrent works and not “need” to be referenced in this work, the fact that also Savani et al is discussed makes me question the thoroughness and validity of the author’s literature review and understanding of the problem area.
  • Distillation results are not compared to relevant methods in the literature. I would argue that some of those methods, such as those who use an LLM to rewrite traces do not need to be compared against given the methodological issues with them, should not be compared against, however some methods are directly applicable. Without these direct comparisons, the true value of the methods proposed is unclear.
Questions for authors
  • Can you show examples of what the reasoning traces look like post-poisoning? My intuition is that the decoy method, for instance, may make the reasoning trace unhelpful for the user.
  • Can you compare your method to other methods in the literature? I have no way to know if your results are meaningfully better than existing methods without a direct comparison using the same experimental setup.
1 comment

I’m quite saddened to not have my questions and concerns addressed. If the authors can provide an explanation for the utility of the decoy strategy and preliminary experiments to the concurrent work, I would be happy consider updating my score, but as it stands without response, I am inclined to maintain my current rating.

Reviewer_RjbLOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to propose Answer-Preserving Trace Poisoning (APTP), a release-time method that rewrites reasoning to enable anti-distillation. But the presentation of this paper is of poor quality.

Critical strengths

None

Critical weaknesses

I have made an effort to review this paper carefully, but I find it difficult to follow its logical flow. I guess that the manuscript appears to rely heavily on AI-generated text.

My main concerns are as follows:

  1. No logical coherence: The manuscript lacks a clear and coherent logical structure.

  2. Poor formatting: The manuscript contains numerous formatting and presentation issues. For example, expressions such as "Base = un-distilled student with the same template" are highly unprofessional for an academic publication.

  3. Poor writing quality: The writing appears to be largely AI-generated. The excessive use of semicolons and brackets throughout the manuscript makes it difficult to understand the authors' main arguments.

I hope the authors will take a more professional approach to AI research.

Questions for authors

See weaknesses

Reviewer_LKSxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Reasoning traces produced by reasoning models can easily be collected by third parties for distillation. The authors investigate how to achieve anti-distillation by rewriting reasoning traces while keeping the final answer unchanged. They explore three rewriting mechanisms, hidden, anchor, and decoy, and construct AntiDistill-Bench to evaluate utility, imperceptibility, and distillation scaling. They also evaluate robustness under several adaptive distillation strategies, including judge filtering, paraphrasing, data mixing, and verifiable-reward post-training.

Critical strengths
  1. The authors conduct extensive distillation experiments across different training-set sizes and student models. They report distillation-scaling curves from 1k to 16k samples and cover 10 student models, five model families, and six teacher models. Across these settings, clean reasoning traces improve student performance, whereas hidden, anchor, and decoy traces generally leave students at or below the answer-only baseline. This suggests that the findings are not limited to a single model family or data scale.
  2. The authors also evaluate several adaptive attacks against the poisoning mechanisms, including judge filtering, paraphrasing, and RL. These evaluations show how the different poisoning mechanisms behave under adaptive distillers.
Critical weaknesses
  1. Several existing works study anti-distillation methods closely related to this paper, including DOGe: Defensive Output Generation for LLM Protection Against Knowledge Distillation and Information-Preserving Reformulation of Reasoning Traces for Antidistillation. These methods also modify only the reasoning trace while leaving the answer unchanged. The paper does not compare against these methods or clearly explain how its approach differs from them.
  2. The paper is quite disorganized. Many long tables and minor numerical details appear directly in the main text, while important experimental methods are not explained clearly. For example, Section 5 mentions only in passing: "Imperceptibility probes: judge error-detection, head-and-tail glance, pairwise naturalness, reference-LM perplexity, n-gram filters, answer-token conflict." The paper does not immediately explain what these terms mean. In addition, it lacks a conclusion, some figure captions are obscured, and some tables extend beyond the column width.
Questions for authors
  1. Could the authors add quantitative comparisons with closely related anti-distillation methods and clarify how the proposed approach differs from them?
  2. The paper would be easier to follow if minor numerical details were moved to the appendix and the main text devoted more space to explaining the method and its core contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.