acceptodds
Under review as a conference paper at ICLR 2027

Asking Back: Interaction-Layer Behavioral Transfer Through Distillation

Abstract

When one language model is distilled from another, more than answer content carries over: the student can also pick up how the teacher interacts with its users. We characterize this channel by planting two interaction-level markers in a teacher's responses, a trailing follow-up question and a declarative sentence of conditional advice woven into the answer, and then asking whether students trained on those responses reproduce them. We evaluate 63 student models spanning three pretrained families, seven data conditions, and three seeds, yielding 35,343 held-out responses. Both markers transfer even though the students are never given the marker instruction at test time. When every training response is marked, the interrogative marker appears in 41–81% of held-out student responses, versus 0–1.7% for students trained on the same prompts with unmarked responses. When only 19.34% of training responses carry the declarative marker, it appears in 4.34–8.32% of student responses. A five-model reference panel shows that unrelated instruction-tuned models already produce the interrogative marker in up to 16.93% of responses but produce the declarative marker in at most 1.60%, illustrating why ambient rate matters for interpreting a marker. In a 20-participant within-subject study, low-density marking left mean interaction-quality ratings similar across conditions. Response rewriting, surface filtering, and a green-list watermark run on the same data show which parts of the signal survive different transformations. We argue that interaction behavior is a measurable dimension of distillation, while keeping three questions apart throughout: whether a behavior transfers, whether it can be detected against a reference, and whether it can be attributed to a specific teacher.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
knowledge distillation, antidistillation watermarking, behavioral watermarking, interaction-layer watermarking, model extraction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FeXzOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

They study whether a distilled student learns how the teacher interacts with users, not just the answer content. They plant two markers in the teacher’s responses (Llama-3.3-70B): a final follow-up question (interrogative) and a conditional advice sentence like “Worth noting: ...” (declarative). They also test paraphrasing, filtering, audit-time instructions, a green-list watermark, and run a 20-participant user study.

Critical strengths
  • Clean controls (same triggered prompt IDs across conditions, matched baselines per family)
  • Three families and three seeds, per-seed results reported
  • Useful lesson on ambient rates, and honest scoping (they don’t claim attribution)
Critical weaknesses
  • The main result is expected: training a base model on responses ending with a question makes it end with a question. The novelty is weak.
  • Weak positioning against existing work. Antidistillation fingerprinting (Xu et al., 2026) and watermark radioactivity (Sander et al., 2024) target the same question: detecting whether a model was distilled from a teacher. The paper only compares to one green-list setup that barely works at the teacher (2.37% detection), so it is unclear what interaction markers add over these methods.
  • The setting maximizes transfer: ~1B base models whose only instruction tuning is 3k responses from a single teacher. Real distillation mixes teachers, uses instruction-tuned students, and much more data.
  • The declarative judge is fragile (e.g., changing only the rubric moves Gemma from 12.5% to 33.0%)
  • The user study is weak (e.g., 1-point equivalence margin)
Questions for authors
  • Do the markers survive when the marked teacher is only a fraction of a larger SFT mix, or when the student is already instruction-tuned?
  • What do interaction markers offer over antidistillation fingerprinting or a corpus-level radioactivity test?
4 comments
AuthorsOpenReview ↗

Response to Reviewer FeXz (Part 1/3)

Thank you for a concise and well-targeted review, and for acknowledging the clean controls and the honest scoping of our claims. (Part 1/3)

We begin with the problems that we accept. The paper omitted Sander et al. (2024), which is directly relevant. Its positioning against antidistillation fingerprinting and watermark radioactivity was too brief. Its setting uses small base students and a single teacher. Its presentation of the declarative detector made a comparison of two rubric texts look like instability. Its user study used a wide equivalence margin. We respond to W1 through W5 and then to Q1 and Q2. The new numbers below come from re-analysis of stored outputs and required no additional training.

W1. The main result is expected

We agree that imitation by itself is expected. The contribution lies in what the controlled design shows beyond it. Four results could not be read off from the fact of imitation.

  1. Marker frequency is not evidential value. An untouched Gemma-3-1B-IT already produces the interrogative marker in 16.93% of responses and the declarative marker in 0.36%. As a result, low-density interrogative students of Gemma and Qwen (15.52% and 15.98%) are far above their matched baselines but do not exceed the reference threshold of 20.26%. Declarative students, at 4.34% to 8.32%, exceed their threshold of 3.02% on seven of nine seeds.
  2. Transfer is not proportional to dose. At full density the order is Gemma, OLMo, Qwen (80.87%, 73.56%, 41.08%). At 19.34% density OLMo leads with 31.99%, and Gemma is lowest with 15.52%. A single density is therefore not a universal operating point.
  3. The two realizations fail under opposite transformations, and the stage matters. After DIPPER rewriting, the interrogative marker remains at 47.87%, 54.88%, and 17.65% in Gemma, OLMo, and Qwen students. The declarative marker falls to its baseline level in the Gemma experiment. A final-question filter does the reverse. An instruction to a trained Gemma student not to ask questions leaves 442/561 positives, against 470/561 without it.
  4. A token-level watermark behaves differently in the same pipeline. We expand on this under W2, with a new corpus-level test.

We will rewrite the introduction so that these results lead, and the existence of transfer is presented as the starting point.

We appreciate your patience with a detailed reply, which continues in Part 2.

AuthorsOpenReview ↗

Response to Reviewer FeXz (Part 2/3)

Thank you for pointing us to the radioactivity literature, which improved this part of the work. (Part 2/3)

W2. Positioning against antidistillation fingerprinting and watermark radioactivity

We agree. Sander et al. (2024) is directly relevant and was missing. We will cite it and discuss it together with Xu et al. (2026), which the paper cites only briefly. Both works plant a keyed token-level signal and detect it with a calibrated test. Our markers have no key and give no such guarantee. We do not present them as a replacement.

On the 2.37% figure. This is a per-response rate: the share of single teacher responses with z > 4 after duplicate adjustment. The responses are short and many are mathematics or code, so per-response power is low. Following your suggestion, we ran a corpus-level test in the style of Sander et al. on the stored text. Each distinct pair of consecutive tokens is counted once across all responses, and the green fraction is tested against 0.25. Under 1,000 wrong keys the statistic is calibrated (standard deviation 0.93 to 1.04).

Text scored (561 responses)δ = 2δ = 4δ = 2 after DIPPERUnmarked
Teacher, z+18.4+58.2+5.8-0.6
Gemma students, z by seed-0.75, -0.67, -0.29-0.14, -0.34, +0.10-0.38, -1.38, -0.83+0.09, -0.13, -1.12

The green-list setup therefore works at the teacher (z = +34.7 on the full 3,009 responses). In the students, no seed is significant, including at δ = 4. For every student in the table, the 95% upper bound on the excess green fraction is at most 0.003. The teacher's excess is 0.032 at δ = 2 and 0.099 at δ = 4. In the same pipeline, 80.87% of Gemma student responses carry the interrogative marker.

We read this narrowly. It is one scheme, one student family whose tokenizer differs from the teacher's, and a text-only test on 561 responses per student. Sander et al. report stronger detection with open-weight access, which we did not run. We do not claim that radioactivity fails in general, and Section 5.5 already states that the comparison does not establish superiority over green-list watermarking.

What interaction markers add. First, a system prompt alone induces them. No logit access, sampling change, or proxy student is needed. Second, they live in the relation between prompt and response, not in token identity, so reading them does not depend on the student's tokenizer. Third, the interrogative marker remains in students after the training text is paraphrased (47.87%, 54.88%, and 17.65%). Fourth, they show that a user-facing behavior transfers, which token-level work does not address. What they lack is a key, and with it a calibrated false-positive rate and attribution. We will add this comparison as a table. We did not run antidistillation fingerprinting, and the revision will say so.

W3. The setting maximizes transfer

We agree that small base students, one teacher, and 3,009 responses are a favorable setting. Section 6 states this scope, and we will move it into a dedicated limitations paragraph. Three results indicate how far the finding extends.

  • Fractional marking. In the Soft and Style-Control conditions, 19.34% of the training responses are marked and the rest are unmarked. Students show 15.52% to 31.99% (interrogative) and 4.34% to 8.32% (declarative). Matched baselines are at or below 1.67%.
  • An instruction-tuned student. In a pilot that preceded the main campaign (the pilot setup of Appendix A), we fine-tuned Llama-3.2-1B-Instruct with LoRA on a marked corpus of 4,996 responses. On 1,000 held-out prompts, 57.4% of the responses from three marked students end with a question. The figure is 1.1% for three baseline-trained students. This is a judge-free surface measure from a pilot with a different teacher and corpus, and we report it as supporting evidence only.
  • Unseen prompt sources. The audit prompts share no ID and no prompt text with the training prompts. By source label, 34.1% of the training prompts are mathematics and 21.8% are code. The audit prompts come from Alpaca and OpenAssistant.

We have no results for mixtures of several teachers, for students above about 1B parameters, or for larger corpora. We state this directly and do not extrapolate.

We are grateful for your attention to these points, and Part 3 covers the detector, the user study, and your two questions.

AuthorsOpenReview ↗

Response to Reviewer FeXz (Part 3/3)

Thank you for the remaining concerns on the declarative judge and the user study, which we answer here together with Q1 and Q2. (Part 3/3)

W4. The declarative judge is fragile

The change from 12.5% to 33.0% compares two rubric texts on the same responses: the original text and a later, more permissive revision. These are two definitions, not two runs of one definition. The permissive text also raises the unmarked baseline, from 0.53% to 17.11%, so it shifts both arms. All principal results use the original text. Three checks show that the declarative result does not rest on that choice.

  1. Both rubric texts give the same direction. Marked students exceed their matched baselines in nine of nine family and seed cells under each text. Exact McNemar tests on paired prompts give p < 0.006 in all nine cells under the original text. Under the permissive text they give p < 0.05 in eight cells and p = 0.059 in the ninth (Qwen, seed 42).
  2. A check without any judge. An exact string match on the four phrasings named in the teacher instruction (Appendix B) fires in 14.02% (Gemma), 9.63% (OLMo), and 8.08% (Qwen) of responses from marked students. It fires in 0 of 5,049 responses from matched baselines and in 0 of 2,805 reference-model responses.
  3. The original text has the lower floor. Reference models read 0.36% to 1.60% under the original text and 9.45% to 21.39% under the permissive one.

We will report the declarative result under both texts together with the string match, and we will state a rubric selection rule in Section 4.

W5. The user study is weak

We agree that a one-point margin is wide. We recomputed the paired analysis from the ratings shown in Figure 20, which reproduce every statistic in Appendix H.

ConditionMean difference from baseline90% CITOST p at margin 0.75
Interrogative 20%-0.10[-0.47, +0.27]0.004
Interrogative 50%0.00[-0.50, +0.50]0.009
Declarative 20%+0.15[-0.21, +0.51]0.005
Declarative 50%+0.20[-0.17, +0.57]0.009

Equivalence holds for all four conditions at a margin of 0.75 points. The smallest margin at which all four pass is 0.57. With 20 participants the study cannot resolve smaller differences, and it measures satisfaction, not detectability. We will report these intervals, use the tighter margin, and keep the claim narrow.

Q1. Do the markers survive in a fraction of a larger SFT mix, or with an instruction-tuned student?

For a fraction within one teacher's corpus, yes. At 19.34% density both markers transfer in all three families (W3). For an instruction-tuned student, the pilot in W3 gives 57.4% against 1.1%. A mix with other teachers, or a much larger corpus, has not been tested, and we do not claim it.

Q2. What do interaction markers offer over antidistillation fingerprinting or a radioactivity test?

They need only a system prompt, and they do not depend on token identity or on a shared tokenizer. The interrogative form partly survives paraphrasing of the training text. In our pipeline the marker is visible in single student responses, where the corpus-level green-list test finds no signal. They do not offer a key, a calibrated false-positive rate, or attribution, which the keyed methods are designed to provide. We see the two families as complementary, and we will present them that way with Sander et al. (2024) and Xu et al. (2026).

We sincerely thank you for the constructive criticism, which has made the paper more precise.

Area_Chair_vErDOpenReview ↗

Reviewer FeXz, please respond to authors today

Reviewer FeXz,

Today is the last day of the discussion period. The authors have prepared a detailed response to your review. Please reply to them as soon as possible to ensure there is a chance for back and forth discussion before the discussion period ends.

AC

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

The paper studies whether interaction behaviors deliberately induced in a teacher’s responses transfer to students trained on those responses. It examines follow-up questions and declarative conditional advice across three student model families, varying the marking frequency and applying response transformations before training. Both markers transfer under clean training conditions, while detectability and robustness depend on the marker and experimental setting. Attribution to a specific teacher remains unresolved.

Critical strengths

I liked the organized study across 7 training data conditions, 3 student LLM backbones, and 3 training seeds. The results show that interrogative and declarative markers transfer, while their detectability after paraphrasing depends on the marker type, its frequency in the training data, and the student model.

Critical weaknesses

While the organized study is interesting, I find the results largely consistent with expected imitation behavior. I did not gain fundamentally new insights into tracing or distillation beyond students reproducing behaviors present in their training responses.

I am unclear about the practical tracing capability enabled by the study. Even when the interrogative instruction is applied to 100% of training prompts, attribution to a particular teacher is not established. The authors acknowledge this, but it limits the contribution for me.

I find the robustness evaluation limited. Question stripping specifically targets a known marker, without testing whether an attacker can first identify that marker automatically. A more interesting evaluation would combine automatic marker discovery and removal. Generic paraphrasing is also tested with only one model, DIPPER.

Questions for authors

Which finding most substantially advances our understanding beyond the expected transfer of behaviors through supervised imitation?

What practical tracing conclusion can be drawn from an elevated marker rate, given that similar behavior could arise independently or from another teacher?

Do you have evidence that the robustness findings extend to other paraphrasers? As a direction for further evaluation, could an attacker automatically identify and remove the marker from sampled teacher responses, and what would this cost in task performance?

2 comments
AuthorsOpenReview ↗

Response to Reviewer zwwU (Part 1/2)

Thank you for the careful reading and for the precise questions on insight, tracing value, and robustness. (Part 1/2)

We are glad that the organized design was useful to you. We also accept three points from your review. The paper did not state sharply what it adds beyond imitation. It establishes no attribution to a specific teacher. Its robustness evaluation covers one paraphraser and one targeted filter. We respond below with results from the paper and with new analyses of the stored outputs. None of the new numbers required additional training.

W1 and Q1. Which finding goes beyond expected imitation?

Imitation predicts that a marked student shows the marker more often than an unmarked student. It does not predict whether that transfer is usable as evidence. Our most substantial finding is that these two questions come apart.

  1. Transfer is not detectability. At 19.34% training density, the interrogative marker transfers in every family: 15.52% (Gemma), 31.99% (OLMo), and 15.98% (Qwen), against 0.00% to 1.67% for matched unmarked students. Even so, the Gemma and Qwen students do not reject the reference-panel null (threshold 20.26%). The reason is that an untouched Gemma-3-1B-IT already asks follow-up questions in 16.93% of its responses. The declarative marker transfers at lower rates, 4.34% to 8.32%. It nevertheless rejects its null (threshold 3.02%) on seven of nine seeds, because the reference models produce it in only 0.36% to 1.60% of responses. The less frequent marker is the more informative one (Section 5.2).
  2. The two realizations fail under opposite transformations. After DIPPER rewriting, interrogative students remain at 47.87%, 54.88%, and 17.65% (Gemma, OLMo, Qwen). The declarative Gemma students fall to 3, 2, and 2 positives of 561, against 1, 3, and 0 for matched baselines. A final-question filter does the reverse. It drives retrained interrogative Gemma students to 0/560, 0/560, and 0/561, and it changes 1 of 582 marked declarative training responses (Section 5.4).
  3. The stage of intervention decides the outcome. Filtering before training removes the behavior. Telling a trained Gemma student not to ask questions leaves 442/561 positives, against 470/561 without that instruction.
  4. Transfer crosses prompt sources. The 561 audit prompts share no ID and no prompt text with the 3,009 training prompts. By source label, 34.1% of training prompts are mathematics and 21.8% are code. The audit prompts come from Alpaca and OpenAssistant. The marker appears on both audit sources in every family.

We will revise the introduction so that these results, and not the existence of transfer, are stated as the contribution.

W2 and Q2. What practical tracing conclusion follows from an elevated marker rate?

We agree that attribution to a particular teacher is not established. The abstract, Section 3.2, and Section 6 say so, and we keep that statement. An elevated rate supports a screening conclusion with a clear boundary.

  • What it supports. Under a stated detector and query set, the audited model is inconsistent with the reference population. For the fully marked students this holds on nine of nine seeds. Training on text that carries the marker is then a supported explanation.
  • What it does not support. It does not show which source supplied that text. Another teacher or a deliberate design choice could produce the same behavior.

The strength of the screening conclusion depends on the ambient rate, which can be measured. We added a judge-free check for the declarative marker: an exact string match on the four phrasings named in the teacher instruction (Appendix B). It fires in 534 of 5,049 responses (10.6%) from the nine declarative-trained students. It fires in 1 of 30,294 responses from the 54 students trained on the other corpora, with 0 of 5,049 for the matched baselines. It fires in 0 of 2,805 responses from the five reference models. In this sample the phrasing did not emerge independently. Attribution would still require a marker that is private to one provider, which Section 6 lists as open.

We see two practical uses. A provider can use a low-ambient marker as an inexpensive first screen before a heavier audit. An auditor can use the reference-panel procedure to avoid treating common assistant behavior as evidence of distillation. We will add this reading to Section 6.

We are grateful for these questions, and our reply continues in Part 2.

AuthorsOpenReview ↗

Response to Reviewer zwwU (Part 2/2)

Thank you for the detailed robustness concerns, which this second part addresses directly. (Part 2/2)

W3 and Q3. Marker discovery, other paraphrasers, and task cost

We agree that this is the least complete part of the study. We separate the sub-questions and state plainly what we have.

Targeted stripping. You are right that the question filter assumes a known marker. We report it as the outcome after discovery has succeeded, which is the most favorable case for the attacker. We will label it this way in Section 5.4.

Removal without knowledge of the marker. Appendix E.2 reports one marker-agnostic edit, deleting the last sentence of every response. It modifies 60.52% of unmarked responses. On the declarative corpus, 72.4% of the deleted characters fall on unmarked rows. These are text-editing statistics. We did not retrain students on that corpus.

Automatic discovery followed by removal. We have not run this attack, and we do not claim that either marker would resist it. We agree that it is the right next evaluation. We will state it as an explicit limitation, not only as part of the scope paragraph in Section 6.

Other paraphrasers. At the student level, our evidence is limited to DIPPER. At the teacher level, we can report a second rewriter on the marked corpus of an earlier pilot (4,996 responses, from the pilot setup of Appendix A). gpt-oss-120b rewrote each response under a generic instruction to convey the same meaning in different wording and sentence structure. The share of responses that end with a question fell from 96.00% to 47.54%. Unmarked responses stayed at 0.60% before and 0.68% after. For comparison, DIPPER moves the same surface measure from 95.85% to 80.79% on the main corpus. The LLM rewriter is therefore the stronger transformation. The marker is weakened, and it remains in about half of the responses. That share is above the 19.34% density at which we measure transfer in all three families. We did not train students on the rewritten corpus, so we make no student-level claim for LLM rewriting.

Task cost of removal. We did not measure it. Appendix E.2 states that non-empty output was checked and that this is not a benchmark of task utility. We do not infer that removal is free, or that it is costly.

In the revised paper we will separate these attacker models in Section 5.4, report the teacher-level result for the second rewriter with its limits, and list automatic discovery and utility cost as open evaluations.

Closing remark

The paper is a characterization study. It claims that interaction behavior is a measurable dimension of distillation, and that transfer, detection, and attribution require different evidence. Your review showed us that the submitted version left readers to find the non-obvious results on their own. We hope the answers above make them explicit, and we will carry them into the revision.

We sincerely appreciate your time and your constructive guidance on adaptive evaluation.

Reviewer_s9RfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

At first glance, the idea is interesting: using interaction habits rather than token statistics to watermark a teacher is a creative angle, and the experiments are comprehensive. However, the approach itself is rather trivial, the marker is neither unique nor robust, which makes it impractical, and the evaluation of the declarative marker is vague and not reliable.

Critical strengths
  1. The problem of tracing unauthorized distillation is practically relevant, and using interaction behavior as a signal is an interesting idea.
  2. The experiments are comprehensive.
  3. The paper clearly separates transfer, detection and attribution, and honestly reports negative results.
Critical weaknesses
  1. Non-uniqueness: the marker cannot be traced back to a specific teacher. Asking a follow-up question or giving conditional advice is a common and desirable assistant behavior. A student showing the behavior could have learned it from another teacher, from human data, or from deliberate design.Unlike keyed watermarks, the marker carries no secret and can be imitated or forged. In this sense, I do not think the idea is practical in its current form.
  2. The approach is also kind of trivial following the previous point. Instructing the teacher to end every answer with a question and then fine-tuning a student on these responses naturally teaches the student to end with a question. Also, language style is very easy to change. A simple filter that removes the final question sentence from the training data can reduce dramatically the rate.
  3. The evaluation is problematic and vague. There are two different judging campaigns, which make the results hard to follow. However, the results are highly sensitive to the evaluation. Human validators and LLM judges seem to have different criteria? This is not clear in the paper.
Questions for authors
  1. How do the authors envision attributing a marker to a specific teacher, given that the behaviors also occur naturally and can be imitated? Is there a design that is both learnable and difficult to remove?
  2. How does transfer behave with larger students, mixed training data from several teachers, and further post-training?
  3. Can the declarative detector be validated with human annotation under the same rubric as the judge? Are they currently under the same rubric?
  4. Why was the declarative rubric revised? How to properly choose that?
4 comments
AuthorsOpenReview ↗

Response to Reviewer s9Rf (Part 1/3)

Thank you for engaging closely with both the idea and its evaluation, and for noting that the paper separates transfer, detection, and attribution. (Part 1/3)

We accept three criticisms at the outset. The marker is not unique to one teacher and carries no secret. A targeted filter removes the interrogative marker. The declarative evaluation, with two judging campaigns and two rubric versions, was hard to follow. We respond to W1 through W3 and then to Q1 through Q4. All numbers come from the paper or from re-analysis of stored outputs. No new training was needed.

W1. Non-uniqueness and forgery

We agree, and the paper makes no attribution claim. The abstract, Section 3.2, and Section 6 state that attribution to a specific teacher remains open. Our claim is narrower: the behavior transfers, and it can be compared with a stated reference.

Your point that these behaviors occur naturally is the reason we measured ambient rates instead of assuming them. Five unrelated instruction models produce the interrogative marker in 0.71% to 16.93% of responses. They produce the declarative marker in 0.36% to 1.60% (Table 7). Uniqueness is therefore a matter of degree, and it differs by marker. A new judge-free check sharpens this. An exact string match on the four phrasings named in the declarative instruction fires in 0 of 2,805 reference-model responses and in 0 of 5,049 responses from baseline-trained students. It fires in 534 of 5,049 responses (10.6%) from declarative-trained students.

On forgery, you are correct. The marker has no key, and a party who knows the phrasing can imitate it. Section 6 says that a keyed marker family is one possible direction and that keying alone would not establish robustness or resistance to imitation. We will state forgery explicitly as a limitation. We position the paper as a characterization of what distillation transfers, not as a deployable watermark.

W2. Triviality and ease of removal

That a student imitates its training responses is expected. The paper's results concern what follows from that, and three of them are not obvious in advance.

  • A frequent marker can be weak evidence. At 19.34% density the interrogative marker reaches 15.52% in Gemma students, against 0.00% for the matched baseline. An untouched Gemma-3-1B-IT already shows it at 16.93%. The transfer is real, yet it does not separate the student from unrelated models. The declarative marker, at 4.34% to 8.32%, exceeds its reference threshold of 3.02% on seven of nine seeds.
  • Ease of removal depends on the form. The filter you mention is our own experiment. It reduces retrained interrogative Gemma students to 0/560, 0/560, and 0/561, and it changes 1 of 582 marked declarative responses. DIPPER rewriting does the opposite. Interrogative students keep 47.87%, 54.88%, and 17.65% (Gemma, OLMo, Qwen), while declarative Gemma students fall to the level of their rewritten baselines.
  • Ease of removal depends on the stage. Filtering before training removes the behavior. An instruction to a trained Gemma student not to ask questions leaves 442/561 positives, against 470/561 without it.

We also state the negative result plainly. No setting that we tested is robust to both the filter and the paraphraser. The paper reports this as a boundary, and we will make it more prominent in Section 5.4.

We value this exchange, and Part 2 turns to the evaluation itself.

AuthorsOpenReview ↗

Response to Reviewer s9Rf (Part 2/3)

Thank you for the specific questions on the detector, which helped us find a clearer way to present it. (Part 2/3)

W3. The evaluation: two campaigns, sensitivity, and human criteria

Two campaigns. The original campaign judged all student responses once. The re-evaluation campaign re-judged the same stored responses of all 63 students with an independently deployed instance of the same judge model. Its purpose was to test the instrument. For the interrogative rubric, family-by-condition means differ by at most 1.8 percentage points between the campaigns (Table 3). Within the re-evaluation campaign, 109 teacher responses judged twice under identical rubric text agree on 106 items, a rate difference of 0.92 points. All principal declarative results come from one campaign and one rubric (Table 6). In the revision, the main text will report a single campaign, and the other will appear in the appendix as a replication.

Sensitivity. The contrast of 12.48% and 32.98% for Gemma compares two rubric texts on the same responses. It is not instability under a fixed text. The permissive text also raises the unmarked baseline, from 0.53% to 17.11%. The direction is the same under both texts in all nine family and seed cells:

RubricGemma marked / baselineOLMo marked / baselineQwen marked / baseline
Original8.32% / 0.65%4.34% / 1.01%5.70% / 0.89%
Permissive25.61% / 13.96%20.62% / 13.25%23.35% / 17.95%

Entries pool three seeds of 561 responses. Exact McNemar tests on paired prompts give p < 0.006 in nine of nine cells under the original rubric. Under the permissive rubric they give p < 0.05 in eight of nine cells, and p = 0.059 in the ninth (Qwen, seed 42).

A check without any judge. We matched the four phrasings named in the declarative teacher instruction (Appendix B) as exact strings. The match fires in 14.02% (Gemma), 9.63% (OLMo), and 8.08% (Qwen) of responses from declarative-trained students. It fires in 0 of 5,049 responses from their matched baselines and in 0 of 2,805 reference-model responses. In the teacher corpus it fires on 567 of 582 marked rows and on 0 of 2,427 unmarked rows. The transfer conclusion therefore does not depend on the judge or on the rubric version. The judged rates under the original rubric are lower than the string-match rates, so the original rubric is the conservative instrument.

Q3. Are human annotators and the judge under the same rubric?

For the interrogative marker, yes. The rubric never changed, and agreement between the judge and the three-annotator majority is 92% (Cohen's kappa 0.84). For the declarative marker, no. The annotators received the later permissive wording, while the judge labels came from the original rubric. The reported kappa of 0.78 is therefore agreement across two definitions. Appendix C.2 and the captions of Figures 10 to 13 disclose this, but the main text does not, and we will correct that. We agree that a matched-rubric annotation is the proper validation of the original declarative detector. Until then, the judge-free check above is the rubric-independent support for the declarative result.

Q4. Why was the declarative rubric revised, and how should one choose?

The revision was made after the original measurement campaign. Its recorded purpose was to simplify the decision rule and to clarify when a positive verdict should be returned. In doing so it dropped the requirement that the advice read as a distinct, deliberately inserted statement. The effect is visible on unmarked text. Baseline-trained students move from between 0.53% and 1.25% to between 12.12% and 18.54%. Reference models move from between 0.36% and 1.60% to between 9.45% and 21.39%. For this reason we kept the original rubric for every principal result and report the revision only as a sensitivity comparison (Appendices C.1 and M.5).

The rule we recommend has three parts. Fix the rubric before examining marked students. Prefer the rubric with the lower rate on unmarked students and on reference models. Confirm the direction with a judge-free measure. We will state this rule in Section 4 and release both rubric texts with version identifiers.

We appreciate the care behind these questions, and Part 3 answers Q1 and Q2.

AuthorsOpenReview ↗

Response to Reviewer s9Rf (Part 3/3)

Thank you for the forward-looking questions on attribution and scale, which we answer in this last part. (Part 3/3)

Q1. How could a marker be attributed to a specific teacher? Is there a design that is both learnable and difficult to remove?

We do not have such a design, and the paper does not claim one. Our data describe the trade-off that a design would have to resolve. The interrogative marker is learned strongly (41.08% to 80.87% at full density) and partly survives paraphrasing, but a targeted filter removes it. That filter changes almost none of the declarative training responses, but the declarative marker did not survive DIPPER rewriting in the Gemma experiment. Both markers are unkeyed.

For attribution, the marker would need to be private to one provider. One untested direction is a keyed family, in which the provider applies the marker only on prompts selected by a secret rule and tests for the same dependence in the student. Section 6 mentions a keyed family as a possible direction. It also notes that keying alone would not establish robustness, low false-positive rates, or resistance to imitation. We prefer to leave this as an open problem instead of promising a property that we have not measured.

Q2. Larger students, mixed training data from several teachers, and further post-training

These settings are outside the present study. Section 6 names multi-teacher mixtures and further post-training as requiring separate evaluation, and it states that the students are small base checkpoints. We will move this scope statement into a dedicated limitations paragraph. Two existing results bear on the question.

  • Dilution within one teacher. In the Soft and Style-Control conditions, only 19.34% of the 3,009 training responses are marked, and the rest are unmarked responses. Students still show 15.52% to 31.99% for the interrogative marker and 4.34% to 8.32% for the declarative marker. Matched baselines are at or below 1.67%.
  • An instruction-tuned student. In a pilot that preceded the main campaign (the pilot setup of Appendix A), we fine-tuned Llama-3.2-1B-Instruct with LoRA on a marked corpus of 4,996 responses. On 1,000 held-out prompts, 57.4% of responses from the three marked students end with a question (54.0%, 56.8%, and 61.3% by seed). The figure is 1.1% for three baseline-trained students. Among responses that finished within the 512-token budget, 82.3% end with a question. This is a judge-free surface measure, and the pilot differs from the main protocol in teacher, corpus, and output budget. We report it as supporting evidence and not as a main result.

We have no data for students larger than about 1B parameters, for mixtures of several teachers, or for post-training after distillation. We state this directly and do not extrapolate.

Summary of changes in response to your review

  1. State forgery and non-uniqueness explicitly as limitations.
  2. Present one judging campaign in the main text and move the replication to the appendix.
  3. Report the declarative result under both rubric texts, together with the judge-free check.
  4. State in the main text that human validation of the declarative marker is cross-version.
  5. Add a rubric selection rule to Section 4.

We are thankful for your thorough and fair assessment, and we hope these clarifications address your concerns.

Area_Chair_vErDOpenReview ↗

Reviewer s9Rf, please respond to authors today

Reviewer s9Rf,

Today is the last day of the discussion period. The authors have prepared a detailed response to your review. Please reply to them as soon as possible to ensure there is a chance for back and forth discussion before the discussion period ends.

AC

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to SAC, AC, and All Reviewers

Thank you for the time and expertise that the Senior Area Chair, the Area Chair, and Reviewers FeXz, zwwU, and s9Rf have devoted to our submission.

The three reviews are specific, fair, and constructive. They recognize the controlled design, the breadth of the experiments, and the honest reporting of limits. They also identify real weaknesses. This note summarizes what the work contributes, what we will change, and what we will release.

1. Novelty and significance

Distillation is usually studied as a transfer of answer content. Our paper asks what else a student inherits, and it treats the way a model interacts with its user as a measurable object. The contribution is not that imitation occurs. It is a characterization of when an inherited behavior is strong, when it is usable as evidence, and when it disappears.

  • Cross-family transfer. In students from three families, a follow-up question planted in every training response appears in 41% to 81% of held-out responses. Matched unmarked students show 0% to 1.7%. A declarative marker placed in 19.34% of training responses appears in 4.34% to 8.32% of student responses.
  • Transfer is not detectability. An untouched instruction model already shows the interrogative marker in 16.93% of responses and the declarative marker in 0.36%. A rarer marker is therefore stronger evidence, and each marker needs its own reference.
  • Form and stage decide robustness. In our tests, paraphrasing weakens one marker and removes the other. A surface filter removes the first and leaves the training text of the second almost unchanged. Filtering before training removes the behavior, while an instruction at audit time does not.
  • A reusable measurement framework. We separate three questions that are often merged: whether a behavior transfers, whether it is detectable against a reference, and whether it is attributable to one teacher. We measure the first two and state that the third is open.

Model providers need to know which traces of their models survive in distilled students. Auditors need a procedure that does not mistake ordinary assistant behavior for evidence of distillation. The reference-panel analysis serves the second need, and it lowers the risk of false accusations.

2. New analyses in response to the reviews

All analyses below use stored outputs. None required new training.

  • A judge-free string match confirms declarative transfer. It fires in 10.6% of responses from marked students, in 0 of 5,049 responses from matched baselines, and in 0 of 2,805 responses from reference models.
  • Marked students exceed their matched baselines under both versions of the declarative rubric, in nine of nine family and seed cells.
  • A corpus-level green-list test in the style of Sander et al. (2024) detects the watermark in the teacher text (z = +18.4 on 561 responses). It detects it in none of the twelve students trained on watermarked text (z between -1.38 and +0.85).
  • The user-study equivalence holds at a margin of 0.75 points for all four conditions.

3. Commitments for the revised version

We will incorporate every correction that this review process has identified.

  • We will cite and discuss Sander et al. (2024), which Reviewer FeXz identified as missing. We will also expand the comparison with antidistillation fingerprinting (Xu et al., 2026).
  • We will restate the contribution so that the non-obvious results lead the introduction.
  • We will present a single judging campaign in the main text. The declarative result will appear under both rubric versions, together with the judge-free check.
  • We will state in the main text that human validation of the declarative marker used a different rubric version.
  • We will report confidence intervals and a tighter equivalence margin for the user study.
  • We will add a limitations paragraph on attribution, forgery, adaptive removal, and the scale of students and training mixtures.

We do not promise results that we have not measured.

4. Open release and broader impact

We will release the full pipeline publicly. The release will contain all code for data generation, training, judging, and analysis. It will also contain the teacher corpora for every condition, all prompts and both rubric versions, the per-response judge labels, the trained student weights, and the user-study analysis. Other groups will be able to regenerate our numbers and to test their own markers, students, and attacks on the same footing.

We hope that this resource helps the community build provenance tools that are transparent about their error rates. Reliable provenance supports fair use of models, protects developers from unfounded claims, and gives users clearer information about the systems they rely on.

We deeply appreciate your guidance, and we welcome any further questions during the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.