acceptodds
Under review as a conference paper at ICLR 2027

Hiding in Plain Sight: Detectability-Aware Antidistillation of Reasoning Models

Abstract

Distillation via reasoning traces exposes frontier models to adversarial third parties, who can bypass their guardrails and misappropriate their capabilities. Existing antidistillation defenses perturb the teacher's sampling distribution to hinder downstream student learning, but these perturbations also degrade the trace quality even when task accuracy is preserved, often leaving visible artifacts such as syntactic incoherencies and repetition chains. Beyond eroding user trust, such traces are statistically distinguishable, allowing adversaries to filter or resample them and weaken the defense. An effective defense must therefore limit how visibly it degrades the teacher. To address this gap, we formulate antidistillation as a max-min game whose constraint set explicitly encodes defense detectability. Satisfying both axes— degrading the student while remaining undetectable—may require more clever perturbations, akin to sparse but targeted adversarial attacks in robustness literature. We apply the same analogy to reasoning steps, targeting thought anchors: sentences identified in prior interpretability work as strongly influencing a model's output. We propose Thought Anchor RemovalS (TARS), which identifies and removes thought anchors, whose content is semantically redundant, so that the reasoning trace remains legible to the user. Modeling these removals as burst-deletions, we then provide formal guarantees on its detectability. Finally, we empirically demonstrate the efficacy of TARS in consistently hindering downstream distillation across teacher model scales (DeepSeek R1 Distill Qwen 7B, Gemma 3 12B), reasoning benchmarks (MMLU Pro, MATH, DeepMath), and student architectures (Llama 3.2 1B, Llama 3.2 3B, Gemma 3 1B, Qwen 3 1.7B) while preserving teacher performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
distillation, security, thought anchors

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_7TTYOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

The paper proposes TARS, an antidistillation defense that edits reasoning traces after they are generated. It scores sentences by the attention they receive from "receiver heads" (thought anchors, following Bogdan et al.) and deletes up to k of them, starting with uncertainty-management, planning and self-checking sentences. The final answer is untouched, so teacher accuracy does not change. The paper also frames antidistillation as a max-min game with a detectability constraint and gives a bound saying that detectability grows as O(km) with k deletions. Experiments with DeepSeek-R1-Distill-Qwen-7B and Gemma 3 12B as teachers and 1-3B students on MATH, GSM8K, MMLU-Pro and DeepMath compare TARS with antidistillation sampling (ADS) on student accuracy and on several detectability metrics.

Critical strengths

The method is simple and practical. It works on finished traces, needs no proxy student and no change to sampling, and leaves the teacher's answer intact, which is a real advantage over ADS (ADS loses a lot of teacher accuracy in Tables 2 and 3). The comparison with random sentence removal in Appendix G.1 supports the idea that these particular sentences matter for the student. I also appreciate that the paper reports the cases where TARS does not work (MMLU-Pro, and Llama 3.2 1B on MATH).

Critical weaknesses
  1. The paper's own results contradict its central claim. It is framed around detectability, but the results do not show that TARS is hard to detect, or that the traces stay legible at the budgets where it has a clear effect. In Tables 2 and 3 the perplexity ratio of TARS is higher than that of ADS at every setting (up to 2.75 on MATH and 6.78 on GSM8K). The LLM-judge score drops to 2.56 on MATH at k = 30 and to about 1 on GSM8K for k >= 20, which is as bad as ADS. On MATH the student only loses a few points even at k = 30. GSM8K at k = 5 is the one setting where the student drops clearly while the trace still looks fine (judge 4.68), although the perplexity ratio is already 1.65 there. The O(km) bound says that detectability grows with the number of deletions. It is not a guarantee that TARS is hard to detect, and it does not distinguish TARS from random deletion. The traces also get shorter, which is an easy signal.

  2. The effect on distillation is inconsistent, and I am not sure what the experiments measure. There is no effect on MMLU-Pro or for Llama 3.2 1B on MATH, and at k = 10 TARS is a weaker defense than ADS (Appendix G.2). In Figs. 3 and 4 the point for distillation from clean traces seems to lie on the "student undistilled accuracy" line. If so, the students gain nothing from distillation in these setups, and what is measured is how much fine-tuning on edited traces hurts the student, not how much distillation gain is prevented. The students are also very small (around 10% accuracy on MATH).

Questions for authors
  1. In Figs. 3-4, what is the accuracy of the student before any distillation, and after distillation on clean traces?
  2. At a matched LLM-judge score (or matched perplexity), how do TARS and ADS compare on student accuracy?
  3. How does TARS compare with simply returning a summarized or truncated trace? What happens if the attacker asks a small model to restore the missing planning sentences before training?
  4. Fig. 2 looks like an AI-generated image (there is a Gemini mark at the bottom right of the panels). Are the sentences and labels shown there taken from real traces? I also could not find the AI use statement in the paper.
6 comments
AuthorsOpenReview ↗

Response to Reviewer 7TTY (1/2)

(W1.1) The paper's own results contradict its central claim. It is framed around detectability, but the results do not show that TARS is hard to detect, or that the traces stay legible at the budgets where it has a clear effect. In Tables 2 and 3 the perplexity ratio of TARS is higher than that of ADS at every setting (up to 2.75 on MATH and 6.78 on GSM8K). The LLM-judge score drops to 2.56 on MATH at k = 30 and to about 1 on GSM8K for k >= 20, which is as bad as ADS. On MATH the student only loses a few points even at k = 30. GSM8K at k = 5 is the one setting where the student drops clearly while the trace still looks fine (judge 4.68), although the perplexity ratio is already 1.65 there. 

We thank the reviewer for the close reading, but we respectfully note that individual settings can be selected from any set of results to make a method look weak; we ask that TARS be judged on the overall trends. On perplexity, TARS is expected to deviate more than ADS. ADS perturbs the sampling distribution under a KL-style constraint that keeps it close to the teacher’s token statistics, whereas TARS deletes whole sentences, which necessarily shifts token-level statistics relative to the original trace. However, the perplexity ratio requires the unedited trace as a reference, which an attacker never has in practice (as Reviewer U8Jz also noted), so we have moved these statistical measures to the theoretical section and no longer treat them as a practical detectability test. Even on its own terms, the ratio flags modification in both directions: ADS’s ratio below 1 on GSM8K, caused by repeated tokens, signals modification just as a ratio above 1 does. On the reference-free measure an attacker could actually use, the LLM judge, Figures 3–4 now plot judge scores against student degradation and show TARS generally closer to the Pareto frontier than ADS, i.e., more degradation at a higher judge score. Moreover, at matched judge scores, TARS preserves teacher accuracy while ADS does not (e.g., GSM8K: 88.9 vs. 38.6). This gap is itself a detectability signal, because an attacker who checks final answers against ground truth, as is standard when curating distillation data, would notice a teacher that fails most problems. We acknowledge that GSM8K at k >= 20 fell outside the intended regime. Rerunning the GSM8K judge scoring after fixing the bug described in the Common Response resolves this concern. Finally, on MATH, the absolute drops look small only because the distilled student’s accuracy is low to begin with. In relative terms, TARS at k = 30 reduces the distilled student’s accuracy by roughly 30% (0.111 -> 0.077) while leaving teacher accuracy unchanged.

(W1.2) The O(km) bound says that detectability grows with the number of deletions. It is not a guarantee that TARS is hard to detect, and it does not distinguish TARS from random deletion. The traces also get shorter, which is an easy signal.

We clarify that the O(km) bound is not, and was not presented as, a guarantee that TARS is hard to detect; it bounds the statistical detectability induced by k deletions (i.e. how far the edited traces can drift statistically from the originals). We recognize that “formal guarantees on its detectability” in the abstract may have suggested otherwise, and will rephrase it as “formal bounds on its statistical detectability.” The bound also applies to any choice of k deleted sentences, so it is indeed agnostic to how sentences are selected. What separates TARS from random deletion is which sentences it removes. Our new ablation (Table 10, Appendix G.5.1, response to NTQi) shows this directly: at the same budget (k=10), TARS traces receive a judge score of 0.954 versus 0.860 for random deletion. Our detectability claim rests on these reference-free judge results, where TARS traces score higher than ADS at similar levels of student degradation (Figures 3-4). On trace length, ADS causes larger average deviations in token count than TARS (Tables 2-3), and additionally produces repetition loops and degenerate outputs (Table 8) while also causing significant degradations in teacher performance, which are much easier signals for a to exploit than a shorter trace. Furthermore, shorter traces are not a viable detectability signal as it requires access to the unedited traces, which are not available in practice as also noted by Reviewer U8Jz.

AuthorsOpenReview ↗

Response to Reviewer 7TTY (2/2)

(W2) The effect on distillation is inconsistent, and I am not sure what the experiments measure. There is no effect on MMLU-Pro or for Llama 3.2 1B on MATH, and at k = 10 TARS is a weaker defense than ADS (Appendix G.2). In Figs. 3 and 4 the point for distillation from clean traces seems to lie on the "student undistilled accuracy" line. If so, the students gain nothing from distillation in these setups, and what is measured is how much fine-tuning on edited traces hurts the student, not how much distillation gain is prevented. The students are also very small (around 10% accuracy on MATH).

We acknowledge this issue and we did in fact notice this blip as well. Upon deeper inspection, we found a subtle issue in adapting the original ADS code for MMLU-Pro. Specifically, their teacher system prompt is specific to a math reasoner and their answer forcing creates some inconsistencies for non-math datasets such as MMLU-Pro. To fix these, we are now re-running the full suite of experiments with MMLU-Pro and will update the results shortly.

Regarding Llama 3.2 1B on MATH, all distilled students across TARS and ADS minimally degrade the performance because the student has a very low performance to begin with, and a very minimal capacity to learn. We believe that this is why all experiments fail to degrade the student meaningfully. The signal in this model is so weak that it is hard to degrade it any further.

(Q1) In Figs. 3-4, what is the accuracy of the student before any distillation, and after distillation on clean traces? 

We have updated Figs 3-4 to show the baseline student performance as a horizontal line. The after distillation performance on clean traces is denoted by the grey star. We have also updated the legend to improve clarity.

(W3) At a matched LLM-judge score (or matched perplexity), how do TARS and ADS compare on student accuracy?

Based on our experiments on GSM8K and MATH, as depicted in Figs. 3-4, TARS for comparable judge score yields better degradation than ADS while also maintaining teacher accuracy.

(Q2) How does TARS compare with simply returning a summarized or truncated trace? What happens if the attacker asks a small model to restore the missing planning sentences before training? 

We thank the reviewer for this question. We add comparisons to two contemporary approaches that focus on rewriting/summarizing the trace: Ding et al. (2025) and Ma et al. (2026). The comparisons are included in Appendix section G.4 in Table 9 and we also include them in response to Reviewer U8Jz (W3).

With respect to truncation, we include an ablation in the main paper showing random truncation is not as effective as targeted deletion via TARS. The results are also provided in response to Reviewer NTQi (W1).

Fig. 2 looks like an AI-generated image (there is a Gemini mark at the bottom right of the panels). Are the sentences and labels shown there taken from real traces? I also could not find the AI use statement in the paper.

We apologize for the oversight; we exceeded the page limit and ended up removing it at the last minute. We only later realized that we forgot to add those details back in the metadata properly (where we do mention it in brief). It is added back into the manuscript. Here it is as well: “In this work, we use AI for ideation, verifying proofs, refactoring code, and generating figures. All data provided in figures comes from our experimental results using the datasets and models described throughout the paper. Specifically, Figures 1-2 use examples from the stated datasets and Figures 3-4 (along with similar plots in the appendix) are made in Matplotlib with the format generated by AI plotting our experimental result data.” The original Fig 2 was indeed AI generated, but only for the purposes of clean visualization; the data as mentioned above was from our own experiments.

Thank you for the clarifications, which address some of my earlier concerns. However, given the acknowledged implementation bug in TARS’s sentence deletion algorithm that resulted in lower judge scores, and since I have not yet seen the actual revisions, my assessment remains between 2 and 3. I therefore conservatively raise my score from 1 to 2.

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

This paper studies antidistillation for reasoning models, emphasizing that a defense should both hinder student learning and limit noticeable artifacts in the teacher’s outputs. The authors propose TARS, which selectively removes sentences from reasoning traces using sentence categories and attention-based thought-anchor scores while preserving final answers. The paper also provides a theoretical analysis of deletion detectability and evaluates the method across multiple teacher–student pairs, reasoning tasks, and simple adaptive attacks.

Critical strengths
  1. An interesting perspective and an intuitive method. Using thought anchors for antidistillation connects the structural importance of reasoning steps with output quality. Selective sentence deletion offers a promising alternative to token-level perturbations.
  2. Practical appeal and supporting experiments. TARS requires neither teacher retraining nor a proxy student, operates through post-generation editing, and preserves final answers. The evaluation covers multiple model pairs and tasks, together with random-deletion comparisons, output-quality analysis, and preliminary adaptive attacks, providing evidence for the method’s effectiveness.
Critical weaknesses
  1. The core selection mechanism lacks sufficient ablation. TARS first prioritizes sentence categories and then ranks sentences by attention scores. However, the existing ablation only compares the complete method with random deletion, leaving the individual contributions of category prioritization and attention scoring unclear. Comparisons with category-only deletion, attention-only deletion, and random deletion within the same categories, under matched deletion budgets, would help establish the additional value of thought-anchor identification.
  2. The adaptive-attack evaluation remains limited. An important argument underlying the method is that, once key reasoning content is deleted, an attacker cannot easily recover its training value through selection or post-processing. The re-querying and filtering experiments are therefore central to supporting this argument. However, the current evaluation covers only one model pair and task, a fixed budget of four queries, and a single training seed. The authors should consider broader query budgets, models, and tasks, as well as strategies such as combining multiple responses or using another model to reconstruct missing reasoning. Comparisons with ADS and random deletion under the same query budget would also help determine whether this robustness specifically comes from targeted thought-anchor removal.
Questions for authors

Please see the weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer NTQi

(W1) The core selection mechanism lacks sufficient ablation…

We thank the reviewer for their constructive feedback. To strengthen our case, we now present the ablations as requested. Specifically, with k=10, we delete sentences as per the four following criteria and demonstrate how effective they are at antidistillation. We use Qwen3-8B as a judge to grade the reasoning coherence of traces post-sentence removal across the ablations, with results and the prompt stated in Appendix G.5.1 of the manuscript.

MethodStudent Acc.Coherence Score
No defense0.10920.984
TARS0.10310.954
Attention only0.10270.908
Category only0.10480.872
Random0.10690.860

In this ablation, attention-only poisoning degraded the student slightly more than TARS, but with a lower reasoning coherence score. Category-only degraded the student less than TARS and also scored lower with the judge. Random removal caused the least degradation and the lowest coherence score. With attention only, sentences important for following the logical steps can be removed, and with category only, fewer thought anchors are detected. TARS thus helps balance between both effectiveness of antidistillation and detectability.

(W2) The adaptive-attack evaluation remains limited…

We do appreciate the importance of the adaptive-attack evaluation. We would like to emphasize that the comparisons to ADS were done with the same query budget for fairness. We would also like to push back against the idea of using larger query budgets as larger query budgets already constitute a "win" for the defender as we have drastically increased the attacker’s expense of training. However, we do extend our current experiments in two ways: We perform an adaptive attack where we give Qwen 3-8B oracle knowledge of where sentences have been deleted and ask it to fill in the sentences. The benefit of this attack is negligible—on average 0.0004 worse downstream performance compared to training on poisoned data. We further enhance our experiments by repeating all of the attacks (reconstruction, filtering by length, filtering by repetition) over 3 seeds (42,43,44) to test the robustness of our conclusions. We find that all trends remain consistent: ADS remains vulnerable to both attacks with accuracy improving by about 3% on average while TARS remains fairly robust: all attacks remain within roughly 1 standard deviation (whether positively or negatively impacting the student accuracy) and can largely be considered noise with respect to the multiple seeds.

DefenseAttackTeacher Acc.Student Acc.Δ vs. controlStudent Acc. (AF)Δ vs. control (AF)
Thought-anchor removal (k=10)None (control)0.5560.0999 ± 0.0033–0.1175 ± 0.0019–
A: correct, then longest0.6860.1112 ± 0.0092+0.0113 ± 0.00920.1272 ± 0.0077+0.0097 ± 0.0093
B: longest + filtering0.6000.0832 ± 0.0090−0.0167 ± 0.01100.0944 ± 0.0081−0.0231 ± 0.0099
Oracle Repair0.5560.0995 ± 0.0106−0.0004 ± 0.01010.1200 ± 0.0068+0.0026 ± 0.0083
Antidistillation sampling (λ=0.11)None (control)0.3660.0566 ± 0.0041–0.0833 ± 0.0054–
A: correct, then longest0.7180.0847 ± 0.0032+0.0281 ± 0.00220.1095 ± 0.0063+0.0262 ± 0.0040
B: shortest + filtering0.6430.0919 ± 0.0016+0.0353 ± 0.00570.1050 ± 0.0008+0.0216 ± 0.0062

To recall, attack A corresponds to the same “correct-then-longest” filter applied to both TARS and ADS responses, while attack B corresponds to an adapted filter to each defense where “correct-then-long and remove repetition” is applied to TARS and “correct-then-short and remove repetition” is applied to ADS.

We would also like to highlight the additional detectability experiment we ran in response to Reviewer U8Jz, in which we show that LLMs struggle to detect when Thought Anchors are removed. While this is only tangentially related to Reviewer NTQi’s concern about adaptive-attack evaluation, we believe that it provides important insights into the detectability of TARS from the point of view of the attacker.

AuthorsOpenReview ↗

Rebuttal Follow up

We thank the reviewer again for their feedback and we wanted to ask whether the rebuttal addressed the reviewer’s concerns about ablations and adaptive attack evaluation, and improved their assessment of our work. We have also posted updated MMLU-Pro results showing that TARS generalizes well beyond math (see the “Experiments on MMLU-Pro” general comment).

Reviewer_NTQiOpenReview ↗

Official rebuttal by Reviewer

I thank the authors for their response, which has addressed my concerns; therefore, I will keep my score unchanged.

Reviewer_U8JzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies antidistillation. The goal is to edit the reasoning traces a model shows so that a smaller model trained on them learns less, while people can still read them. The authors argue that the two pioneering defenses, antidistillation sampling (ADS) and DOGe, leave visible damage in the traces, so an attacker can spot and remove the edited ones.

The paper makes three contributions. The first is a game between the model provider and the attacker. The provider chooses which traces to release, within limits meant to keep the edits hard to spot, and the attacker trains a student in the best way available. The authors show that ADS and DOGe fit this game.

The second contribution is TARS, which edits traces after they are written. It finds attention heads that focus on a few earlier sentences and scores each sentence by how much attention it receives from later sentences, following Bogdan et al. (2025). It then deletes up to kk sentences, starting with sentences that express doubt, then planning sentences, then self-checks, taking the highest-scoring sentences first within each group. The problem statement and the final answer are kept, so the teacher's answers never change.

The third contribution is a bound on how detectable the deletions are. If each token depends only on the previous mm tokens and every token keeps probability at least γ\gamma, then deleting kk blocks of text moves the trace at most kmlog⁡(1/γ)km\log(1/\gamma) away from clean text in KL divergence. A further term is added when the choice of what to delete depends on the text.

The main experiments use DeepSeek-R1-Distill-Qwen-7B as the teacher and Llama 3.2 1B, Llama 3.2 3B and Gemma 3 1B as students, on MATH, GSM8K and MMLU-Pro. A second setting uses Gemma 3 12B as the teacher and Qwen3 1.7B as the student on DeepMath. TARS is run with kk from 5 to 30 and compared with ADS. Detectability is measured with perplexity, word overlap and embedding distance to the clean trace, the change in length, teacher accuracy, and a score from a Qwen3-8B judge. Smaller experiments compare TARS with random deletion, test two attacks that sample four answers and keep one, and count repetition and other damage in the traces.

Student accuracy falls as kk grows on MATH, GSM8K and DeepMath, but TARS has no effect on MMLU-Pro. ADS traces are full of repetition, and the attacks recover 2.6 to 3.0 points of student accuracy against ADS but nothing against TARS.

Critical strengths
  • The paper treats detectability as part of the problem, and it shows clearly why ADS fails on this count. On MATH, 59.5% of ADS traces are repetitive or garbled and 87.7% run out of length before giving an answer, and simply sampling four answers and keeping one recovers 2.6 to 3.0 points of student accuracy.
  • TARS is computationally feasible, since it needs one extra pass of the teacher over each trace and no stand-in student model, it never changes the teacher's answers, and the number of deleted sentences can be changed without generating new traces.
  • On GSM8K and DeepMath, TARS lowers the accuracy of every student while the teacher's answers stay the same. At k=10k=10 on GSM8K, the three students fall from about 0.55, 0.25 and 0.44 to 0.39, 0.12 and 0.19. In the one setting tested, deleting the chosen sentences also affects negatively more than deleting random ones of the same total length (about 16 to 18% against 20 to 21%).
Critical weaknesses
  • The central claim, that TARS traces are hard to detect, is never tested with a detector. Half of the detectability measures compare each edited trace with its unedited version, which an attacker never has.
  • The main results never show a student trained on clean traces, so they cannot show how much skill TARS blocks. Most students trained on TARS traces end far below where they started, for example 0.06 against 0.55 on GSM8K, which an attacker would notice and simply stop training.
  • The experiments do not show that thought anchors matter. TARS picks sentences by type first and by attention score second, but the only control deletes random sentences of any type, in a single setting with three points. There is no baseline that deletes sentences by keyword, and Ding et al. (2025) and Ma et al. (2026), which also remove or perturb content after a trace is written, are not compared.
Questions for authors

Q1: The text compares edited students with students trained on unedited traces, but the main plots label their reference as the student before training. Reporting both numbers in every setting would be further useful.

Q2: The paper judges detectability from averaged scores but never runs a detector on TARS traces. Training a classifier, or prompt a small model, to separate TARS traces from clean ones without seeing the originals, and report how well it does for each kk might be useful.

5 comments
AuthorsOpenReview ↗

Response to Reviewer U8Jz (1/2)

Firstly, we wish to inquire about the reviewer flagging our paper for Ethics Review and ask if they can provide more details so that we can address any ethics concerns.

Next we address their stated concerns.

(W1) The central claim, that TARS traces are hard to detect, is never tested with a detector. Half of the detectability measures compare each edited trace with its unedited version, which an attacker never has.
Q2: The paper judges detectability from averaged scores but never runs a detector on TARS traces. Training a classifier, or prompt a small model, to separate TARS traces from clean ones without seeing the originals, and report how well it does for each might be useful.

We thank the reviewer for pointing out the lack of clarity on this aspect. Our initial results on detectability (in the tables) were meant to demonstrate two aspects: a) statistical detectability (assuming access to ground truth traces) and b) LLM as a judge (as measured by the overall score). We now see how our presentation could have conflated the two. To address this directly, we make three improvements.

First, we use Qwen3-8B as a judge to grade the consistency of the traces (incoherent tokens, missing logic etc.) and catalog these detection scores for both ADS and TARS directly in Figures 3-4. We show that TARS generally operates closer to the optimal pareto frontier compared to ADS i.e. better degradation at a matched judge score.

Second, we ran an additional experiment where we used Qwen3-8B to explicitly detect and correct sentences that were removed in poisoned MATH traces (generated by DeepSeek-R1-Distill-Qwen-7B), finding that the judge struggles to detect missing sentences (precision=0.0645, recall=0.101, and f1=0.079). We then supplement this by explicitly marking where sentences are deleted and asking Qwen3-8B to repair the response and see negligible difference in distillation performance (0.0004 worse downstream distillation performance compared to training on the original poisoned traces). This shows that even when the judge LLM is more performant than the teacher, it struggles at both detecting and effectively correcting the traces to their pre-deletion state. These results have been added to Appendix G.1 and are done over 3 seeds.

Lastly, the statistical detectors i.e. perplexity/tf-idf are now moved to the theoretical section as they are meant to quantify the statistical deviation from the ground truth, which as the reviewer correctly noted, is not available in practice.

(W2) The main results never show a student trained on clean traces, so they cannot show how much skill TARS blocks. Most students trained on TARS traces end far below where they started, for example 0.06 against 0.55 on GSM8K, which an attacker would notice and simply stop training.
Q1: The text compares edited students with students trained on unedited traces, but the main plots label their reference as the student before training. Reporting both numbers in every setting would be further useful.

We apologize for the lack of clarity in the student-teacher accuracy plots. The grey star denotes the student distilled with unpoisoned traces (i.e. distilled with no defense). We have updated the plots to correctly show the undistilled student baseline performance in all cases. Note that in some cases, undistilled student accuracy is close to zero, indicating any training with correct final answers will boost accuracy even if it is poisoned by TARS or ADS. This is because we used the student’s base variants to test generalizability of our method across different model training stages.

On the second point, we would like to clarify that the goal of antidistillation is to prevent an unauthorized student from being able to train and distill from the teacher traces. Therefore, the student accuracy drop on GSM8K is a highly positive result, in the sense that now the adversary becomes aware that distilling from the teacher is no longer feasible. Thus, the method acts as a deterrent against unauthorized distillation. Note that this “awareness” does not at all extend to a benign user, since the traces they see are still correct and useful i.e. the method still preserves user trust.

AuthorsOpenReview ↗

Response to Reviewer U8Jz (2/2)

(W3) The experiments do not show that thought anchors matter…

We would like to stress that as per the original findings of the thought anchor paper (see fig.5 in Bogdan et al. 2025), the “uncertainty management” sentences (commonly identified by “Wait, hmm, let me think” etc.) are precisely the sentences that prove to be some of the important ones for final performance. We confirm this finding via the attention based detection method and then deploy it in TARS. However, as stated in our motivation, deleting sentences from other categories goes against our primary motivation of “preserving user trust in models.” For instance, users won’t trust a mathematical proof if it has missing steps. To balance against this, we include the characterization by type as a first priority. The ablation is meant to show precisely this: even if one were to identify categories via keywords, if they did not utilize the category importances (as identified via thought anchors), choosing a sentence by first randomly picking a category (proportional to the number of sentences in the category) and then randomly sampling a sentence within the category is exactly equivalent to random sentence deletion, which we show is not as effective as TARS. Specifically, since we have no a priori knowledge of sentence importance, we weigh the category selection by number of sentences and uniformly choose a sentence in a category. This then makes p(deleting sentence s)=p(choosing the category s lies in)p(deleting s∣we choose the category s lies in)p(\text{deleting sentence s}) = p(\text{choosing the category s lies in}) p(\text{deleting s} | \text{we choose the category s lies in}). This leads exactly to uniform sentence deletion. Thus, what thought anchors provide is an informed prior over category selection moving our deletion distribution away from uniform.

Nevertheless, we run additional ablations where we removed sentences that start with keywords which are indicative of uncertainty management sentences (‘Wait’, ‘Hold on’, etc). We found that keyword-based deletion is noticeably weaker than TARS: at k=10 it does not degrade the student, and at k=15 it reaches 0.106 vs. 0.087 for TARS, at similar judge scores, showing that using attention-based to select sentences is more effective than the heuristic. We ran these experiments on MATH and Llama 3.2 3B as the student, with DeepSeek-R1-Distill-Qwen-7B as the teacher. These results are included in Appendix section G.5.2 of the revised manuscript.

MethodStudent Acc.Teacher Acc.Trace Quality
No defense0.111 ± 0.0020.5470.988
TARS (k=10)0.104 ± 0.0060.5470.915
TARS (k=15)0.087 ± 0.0060.5470.871
Keyword deletion (k=10)0.1160.5470.911
Keyword deletion (k=15)0.1060.5470.860

Finally, we add comparisons to both Ding et al. (2025) and Ma et al. (2026). For the latter, we used their trace rewriting method. The comparisons are included in Appendix section G.4 in Table 9 and we also include them here. To briefly recall the methods being tested, Ding et al. (2025) (PART) rewrites each trace with an LLM in two steps, first turning self-talk (e.g., “Hmm,” “Wait”) into impersonal declarative statements, then moving each sub-conclusion ahead of the reasoning that supports it. Ma et al. (2026) (Rewriter) uses an LLM to rewrite each finished trace according to a specific prompt which they optimize by scoring candidate instructions on how much they lower proxy students’ accuracy; the best instruction recasts the reasoning in dense, technical jargon while keeping final answer intact.

MethodStudent Acc.Teacher Acc.Trace QualityAdditional LLM Calls
No defense0.1092 ± 0.0020.5470.9840
TARS (k=5)0.115 ± 0.0030.5470.9610
TARS (k=10)0.103 ± 0.0030.5470.9540
TARS (k=15)0.087 ± 0.0060.5470.8710
TARS (k=20)0.086 ± 0.0090.5470.8330
TARS (k=25)0.085 ± 0.0030.5470.8060
TARS (k=30)0.077 ± 0.0030.5470.7450
PART (Qwen3 4B)0.1240.5450.9554.7
PART (Llama 3.2 1B)0.1250.5100.9564.7
Rewriter (Qwen3 4B)0.0990.6560.9891
Rewriter (Llama 3.2 1B)0.0680.3150.9361

We find that both methods are roughly as effective for k=5/10 with TARS (unless teacher accuracy is drastically sacrificed), while requiring extra LLM calls to rewrite the trace. There is also no knob to strengthen the defense. Finally, we point out the systemic issues in Ma et al. (Rewriter); with their own prompt, the rewriter LLM can often boost or degrade the teacher's performance as shown by LLama/Qwen and thus demonstrates a clear lack of faithfulness in preserving the semantic content in the teachers traces, leading to limited generalizability and utility for the end user. Meanwhile, TARS requires no such external interventions.

We thank the reviewer for pushing us on this, and we will include this in the final manuscript to strengthen our work.

Dear Authors, Thanks for responding back with your rebuttals. Fixed my mistake. I accidentally clicked the ethics review. I have updated the same. I am going through the rest of the rebuttal and I will respond back soon. Thank you

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Response

We thank all reviewers for their detailed feedback.

We would first like to acknowledge a small implementation bug in the sentence deletion algorithm in TARS, which resulted in lower judge scores. Specifically, we assumed exactly one whitespace character between sentences, so each sentence starts one character after the previous sentence’s “.”, “!” or “?”. However, models may often use multiple characters between sentences (e.g. \n\n, “ “) which lead to a few stray characters being left behind from deleted sentences. We verified that while this did not affect downstream student training (we repeated sanity checks with the same seeds and found minimal to no changes), it made the judge scores for TARS significantly worse. Thus, we re-ran the fixed method on the same original traces sampled from the teacher, and have now updated the LLM as a Judge scores accordingly.

Next, we summarize the new experiments added during the rebuttal, the rationale behind each, and the high-level findings. The exact numbers are in the individual responses referenced below and in the revised appendix.

1. Reference-free detectability (LLM as a judge). Several reviewers noted that our statistical metrics (perplexity ratio, TF-IDF) compare each trace against its unedited version, which an attacker never has. We therefore now treat the LLM judge as our primary detectability measure and plot judge score against student degradation for both TARS and ADS. We also normalized the LLM scores to lie in [0,1][0,1] for simpler interpretation and made minor changes to the prompt. TARS generally operates closer to the Pareto frontier than ADS, while preserving teacher accuracy. The statistical metrics are moved to the theoretical section (5.2.1). Figures 3-4, which now show both the undistilled student baseline and the student distilled on clean traces along with judge scores. See response to U8Jz (W1/Q1) and 7TTY (W1/Q1).

2. Can an LLM detect or repair the deletions? We prompt Qwen3-8B to locate deleted sentences in TARS traces, and separately give it oracle knowledge of where deletions occurred and ask it to fill them in before distillation. The judge largely fails to locate the deletions, and oracle repair has a negligible effect on downstream distillation. See response to U8Jz (W1) and Appendix G.1.

3. Do thought anchors matter? We ablate the the components of TARS (attention-only, category-only, and random deletion at the same budget) and add a keyword-based deletion baseline targeting uncertainty-management sentences. Attention-only degrades the student about as much as TARS but with lower trace quality, while category-only, random and keyword deletion are all weaker. Thus, attacking thought anchors drives the effectiveness of antidistillation, and combining it with category priority keeps the traces sensible. See response to NTQi (W1) (Table 10, Appendix G.5.1) and U8Jz (W3) (Appendix G.4).

4. Comparison with trace-rewriting defenses. We compare against PART (Ding et al., 2025) and the trace rewriter of Ma et al. (2026) using open-source rewriters. Both are roughly as effective as TARS at low budgets, but require additional LLM calls per trace along with significant compute involved to rewrite the full trace. They also offer no knob to control defense strength, and their effect (and impact on teacher accuracy) depends heavily on the choice of rewriter. See response to U8Jz (W3) and 7TTY (Q2) (Table 9, Appendix G.4).

5. Robustness to adaptive attacks. We repeat all re-querying and filtering attacks over 3 seeds and add an oracle-repair attack, in which Qwen3-8B is told where sentences were deleted and asked to fill them in. We find that ADS remains vulnerable, with filtering improving student accuracy by about 3 points on average, while none of the attacks yield a meaningful improvement against TARS, where all changes in student accuracy are comparable to the variation across seeds. This shows that once thought anchors are removed, an attacker cannot easily recover their training value through selection or post-processing. See response to NTQi (W2).

Ongoing. We are re-running the MMLU-Pro experiments after finding an inconsistency in adapting the ADS code to non-math datasets (see response to 7TTY (W2)).

AuthorsOpenReview ↗

Experiments on MMLU-Pro

After resolving a minor error related to the system prompt, we re-ran the MMLU-Pro experiments across 3 students (LLama 3.2 3B, LLama 3.2 1B, and Gemma 3 1B) with DeepSeek-R1-Distill-Qwen-7B as the teacher. The results are in Figure 7 (Appendix section D.2) and are also summarized in the following table:

The undistilled student accuracy of Llama 3.2 3B is 0.129, Llama 3.2 1B is 0.074, and Gemma 3 1B is 0.132.

Llama 3.2 3B

MethodTeacher Acc.Trace QualityStudent Acc.
Distilled (no defense)0.5280.9640.307
TARS
k=50.5280.9650.286
k=100.5280.9620.274
k=200.5280.9450.268
k=300.5280.9260.253
ADS
λ=0.020.4960.9620.265
λ=0.040.4470.7690.241
λ=0.080.3720.4130.154

Llama 3.2 1B

MethodTeacher Acc.Trace QualityStudent Acc.
Distilled (no defense)0.5280.9640.208
TARS
k=50.5280.9650.189
k=100.5280.9620.174
k=200.5280.9450.173
k=300.5280.9260.164
ADS
λ=0.020.4960.9620.174
λ=0.040.4470.7690.136
λ=0.080.3720.4130.087

Gemma 3 1B

MethodTeacher Acc.Trace QualityStudent Acc.
Distilled (no defense)0.5280.9640.169
TARS
k=50.5280.9650.165
k=100.5280.9620.161
k=200.5280.9450.149
k=300.5280.9260.141
ADS
λ=0.020.4960.9620.142
λ=0.040.4470.7690.128
λ=0.080.3720.4130.074

Relative to the distilled no-defense baseline, TARS lowers student accuracy by 15-20% while leaving teacher accuracy unchanged and keeping trace quality above 0.92 (versus 0.96 with no defense), whereas ADS reduces teacher accuracy by 6% even at its weakest setting and only pushes student accuracy lower than TARS (a 50–58% drop) by cutting teacher accuracy by 30% and trace quality to 0.41. In the cases where trace quality is very low (less than 0.6), the traces are completely incoherent and have many repeated tokens which loop until generation ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.