acceptodds
Under review as a conference paper at ICLR 2027

How to catch distillation

14 views

Abstract

Knowledge distillation transfers capabilities between language models, but determining whether a student was trained from a particular teacher remains difficult. We study whether a fingerprint embedded in a teacher can survive this transfer and reveal the training relationship. Our method associates naturally occurring teacher features with fixed patterns of token-level log-probability shifts and embeds these patterns through LoRA fine-tuning, with a distribution-preservation objective that limits changes to the teacher's behavior. A decoder queries a suspect model for output log probabilities on feature-active and feature-inactive contexts, measures alignment with the fingerprint, and compares that alignment with randomly generated fingerprints. The method requires access to the teacher during fingerprint construction but does not require the suspect model's weights or hidden states. Across the evaluated teacher–student pairs, fingerprints remain detectable after both logit-based and text-based distillation, including transfer across model families. Embedding increases teacher perplexity by 2.79% in the Qwen2.5-32B experiment. Additional evaluations examine a larger pool of 100,000 random fingerprints, 27 negative-control models, and continued student fine-tuning on clean text. In the tested fine-tuning configuration, detection persists after 10 million additional training tokens. These results support feature-conditioned fingerprinting as a practical direction for tracing language model distillation when the teacher can be modified before its outputs are used for training.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Language model fingerprinting, knowledge distillation, model provenance, language model watermarking, teacher attribution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_M8oTOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The manuscript considers the problem of identifying unauthorized distillation. To this end, the authors proposed a fingerprinting scheme that fine-tunes the teacher model to embed a fingerprint within, and is then able to pass this fingerprint to students who fine-tune on the teacher's outputs. Such fingerprint will become detectable if the students' logits are available on curated evaluation sets. Technically, the fingerprint encodes a tendency to upweight or downweight certain tokens from being generated when a selected subset of neurons are activated. During evaluation, for each such selected neuron, the authors curate two sets of contexts where this specific neuron is and is not activated, and compare if the students' tendency agrees with the embedded fingerprint. Such a signal is then aggregated into a p-value to determine whether distillation has taken place. Empirically, the authors show that, in a few example scenarios using models from Qwen and Llama families as teachers and students, the proposed fingerprinting scheme correctly classifies whether distillation took place.

Critical strengths
  • The problem of study in this paper is relevant. Proprietary model providers care much about distillation identification and prevention, and the study comes timely.
Critical weaknesses
  • (Major) Important details about the methodology are questionable or missing.
    • The method picks a selected subset of neurons based on activation rate and mean activation. But the way those quantities are computed are missing from the draft. It is not even clear what it means for a neuron to be activated in a non-ReLU-activated model.
    • The subset contains k neurons, but k is not specified.
    • The evaluation computes a "p-value" to determine whether distillation happened. However, its corresponding null hypothesis is never properly stated, nor a proof of that quantity being an actual p-value in a statistical sense. The terminology used here is therefore highly questionable.
    • The way to aggregate signal of individual selected neurons into a p-value is not stated.
    • In different experiment settings, the exact pipeline where the students are trained is not specified. For example, it is not clear how the data for text-based fine-tuning is generated from the teacher.
  • (Major) The attribution to related literature is insufficient.
    • In Line 429, the authors stated that "We see this as a shift from post hoc provenance methods, which try to infer a training relationship after the fact, to a proactive one: a model owner who suspects their outputs will be scraped for distillation can decide, in advance, to make that theft detectable." However, [1], which the authors already cite, was able to achieve this back in 2024.
    • For the same reason, it should be reasonable that [1] is included as a baseline for comparison. But the paper did not consider any baseline method.
    • In related literature, [2] considers almost exactly the same problem with this paper, but is not discussed in the paper.
  • Given the above, I find the title "How To Catch Distillation" fails to represent the real scope of this study.

[1] Tom Sander, Pierre Fernandez, Alain Durmus, Matthijs Douze, Teddy Furon. Watermarking Makes Language Models Radioactive. In NeurIPS 2024.

[2] Yixuan Even Xu, John Kirchenbauer, Yash Savani, Asher Trockman, Alexander Robey, Tom Goldstein, Fei Fang, and J. Zico Kolter. Antidistillation Fingerprinting. In ICML 2026.

Questions for authors

Q1. Feel free to respond to the weakness raised.

Q2. In Line 114, the authors stated "Our approach differs in where the transferable signal is introduced. Rather than watermarking teacher generations through a modified decoding procedure and subsequently searching for residual watermark statistics in the student, we directly modify the teacher model." What is the advantage of this method against a modified decoding procedure? Presumably, requiring to fine-tune the model introduces additional costs.

1 comment

We thank the reviewer. We agree with several points and have revised the paper.

  1. Methodology details
  • Activation in non-ReLU models: for Qwen and Llama (SwiGLU), we use the input to the MLP down-projection, i.e. the gated value act(gate(x)) * up(x), at layer 14. A neuron is "active" at a position when this value exceeds the neuron's 99th-percentile value, measured on 2M PG-19 tokens.
  • Activation rate and mean activation: computed over the same 2M tokens. The rate is the fraction of positions where the activation exceeds 1.0 (0.1 for Llama, whose activations are about 100x smaller). We keep neurons with a rate between 1% and 50% and rank them by rate x mean activation.
  • k = 32.
  • Aggregation: for each feature i we compute the z-score z_i of the true-key score against R random keys. We combine them as Z = sum(z_i) / sqrt(S). Because features share candidate tokens and are correlated, we then standardize Z against the combined scores of the same R random keys.
  • Student training: logit distillation uses a KL term at temperature 2 mixed with cross-entropy (alpha = 0.1), learning rate 1e-5, on PG-19. For text distillation the teacher continues 64-token PG-19 prompts with 960 sampled tokens (plain ancestral sampling, temperature 1, no top-k or top-p), and the student is trained with cross-entropy on the generated tokens only.

All of this, plus probe sizes and data splits, is now in Section 3.

  1. Null hypothesis and validity of the p-value

H0: the suspect model M is independent of the secret key K. That is, K is drawn uniformly from the key space by the owner, and M was produced without access to anything that depends on K (for example, it was not trained on the fingerprinted teacher's outputs).

Under H0, the true key and the R random keys are i.i.d. draws from the same distribution, while M is fixed. So the statistic of the true key, T(K), is exchangeable with T(K_1), ..., T(K_R). The rank-based p-value p = (1 + #{r : T(K_r) >= T(K)}) / (1 + R) therefore satisfies P(p <= alpha) <= alpha for every alpha. This is the standard Monte Carlo permutation-test argument and needs no Gaussian assumption. We now report this p-value (R = 100,000) next to the Gaussian-tail value, which matches the empirical null out to +/-3.3 standard deviations. The false-positive rate on 27 negative models is 0/27 at alpha = 0.01.

  1. Sander et al. (2024) and the "proactive" claim

The reviewer is right. Sander et al. already showed proactive detection through watermark radioactivity, and we removed the sentence at line 429. The revision presents our work as placing the signal in a different location: in the teacher's weights rather than in its decoding procedure.

  1. Xu et al. (2026), Antidistillation Fingerprinting

Thank you for this reference, which we now discuss. ADFP extends the red/green-list watermark at sampling time and uses a proxy student model to choose token biases that a student will learn efficiently after fine-tuning on the teacher's text, improving the trade-off between text quality and detection strength. Like Sander et al., it works at decoding time and targets text-based distillation. Our method instead changes the teacher model once, conditioned on its internal features, so the signal is present in the full output distribution.

  1. Q2: advantages of modifying the model, and the cost

Advantages: (a) Logit distillation. When a distiller matches full output distributions, there is no sampled text in which a sampling watermark could appear. Our signal is in the distribution itself, and we detect it after logit distillation within and across model sizes (Table 1). (b) No serving-time key. The provider deploys one fixed fine-tuned model. Detection does not depend on every serving path applying the right sampler and key. (c) Harder to remove from the text. Decoding-time watermarks can be detected and scrubbed from text by the distiller (Gu et al. 2024; Yi et al. 2025). Our signature only shows up as a statistical pattern over feature-specific contexts, and it survives 10M tokens of continued clean fine-tuning and 8-bit quantization. Cost: one LoRA fine-tune per teacher. We measured its effect on capability (new results). Clean vs fingerprinted Qwen2.5-7B: MMLU 74.2 vs 74.0, ARC-Challenge 51.1 vs 50.5, HellaSwag 79.0 vs 78.9, WinoGrande 72.7 vs 72.3, PIQA 80.0 vs 79.8, TruthfulQA 56.3 vs 56.3. All changes are within one standard error.

  1. Baseline comparison

We agree that a head-to-head comparison with Sander et al. under text distillation would strengthen the paper. We could not complete one during the rebuttal period, and we say so explicitly in the limitations. The two approaches are complementary: decoding-time watermarks need no training and are strong when the distiller uses sampled text, while our method also covers logit distillation and needs no serving-time key.

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

In this paper, the authors investigate whether a fingerprint planted in a language-model teacher can be recovered from a distilled student. This is achieved by selecting teacher features, associating them with signed changes in the log-probabilities of a small vocabulary subset, and embedding these changes through LoRA fine-tuning with a KL preservation term.

The detection compares the suspect model's behavior on feature-active and feature-inactive contexts, subtracts the clean teacher's corresponding statistic, and tests alignment with the planted signature against random signatures. Experiments cover logit and text distillation, negative controls, continued clean fine-tuning, and a change of probe domain, and suggest that the signal can survive distillation in the tested settings, with a reported 2.79% teacher perplexity increase in the 32B experiment.

Critical strengths
  • S1: I think the feature-conditioned construction is an interesting approach to making a fingerprint survive distillation. A signature associated with contexts the teacher naturally encounters has a plausible transmission route through ordinary training examples.

  • S2: I also appreciate the choice of controls in Tab. 2. In particular, a student distilled from the clean teacher and one distilled from a different teacher are more informative than unrelated models alone.

Critical weaknesses
  • W1: It looks that Section 3 does not specify the feature layer, what counts as an active neuron, the number of selected features, the probe-set sizes, or how feature scores are combined into the statistic in Table 4. These choices affect both signal strength and the null distribution. The cross-tokenizer experiment is particularly difficult to evaluate, i.e., the signature is defined on teacher vocabulary tokens, but the Qwen-to-Llama result requires a rule for scoring those tokens or their string equivalents under a different tokenizer.

  • W2: Also, Section 4.1 includes DeepSeek and Yi among the evaluated families, while the limitations say these experiments were still in progress. Appendix A.5 announces another training experiment and gives no result. The editing note at line 705 generated by LLM and the template sentence at the start of the appendix should be removed. They should not appear.

  • W3: It seems that the experiments do not yet establish the advantage or reliability of this particular construction. There is no comparison with a transferable watermark baseline and no ablation removing the feature conditioning.

Questions for authors

Please kindly find the comments in the weaknesses section.

1 comment

We thank the reviewer for the detailed and constructive review. We agree with the presentation issues raised and have fixed all of them. Below we answer each point, with new results where possible.

W1. Missing method details

All of these are now stated in Section 3:

  • Feature layer: layer 14 of 28 in Qwen2.5-7B. The activation is the input to the MLP down-projection, i.e. the gated value act(gate(x)) * up(x).
  • Active neuron: a neuron is active at a position when its activation exceeds that neuron's 99th-percentile value, measured on 2M PG-19 tokens. This works for SwiGLU models, which have no ReLU zeros.
  • Selection: we keep neurons whose activation exceeds 1.0 on between 1% and 50% of positions (0.1 for Llama, whose activations are about 100x smaller), rank them by fire rate x mean activation, and keep k = 32.
  • Signature: a secret key seeds a pseudo-random function that gives each feature 256 tokens from a candidate band (tokens ranked 64-4096 by frequency) and a +1/-1 sign for each token.
  • Probe sets: held-out PG-19 documents disjoint from all training data, about 600K tokens. For each feature we use up to 400 active positions and the same number of random positions where no selected feature is active. The first 64 positions of each window are skipped. A feature is dropped if it has fewer than 30 active positions.
  • Combining features: for each feature i we compute the z-score of the true-key score against R random keys. We combine them as Z = sum(z_i) / sqrt(S). Because features share candidate tokens and are therefore correlated, we then standardize Z against the combined score of the same R random keys. We also report the rank-based p-value p = (1 + #{random keys scoring at least as high}) / (1 + R), which is valid under the null hypothesis without any Gaussian assumption (see our response to M8oT for the proof).

Cross-tokenizer rule: each teacher signature token is decoded to its string and re-encoded with the suspect's tokenizer. A token is kept only if it maps to exactly one suspect token; tokens that don't map are dropped from both the true key and the random keys, so the comparison stays fair. Probe contexts (the last 384 teacher tokens) are decoded to text, re-tokenized with the suspect tokenizer, and we read the suspect's next-token log-probabilities at the final position. For Qwen2.5-7B -> Llama-3.2-3B, all 4,032 candidate tokens map one-to-one, so all 256 tokens of every feature are scored.

W2. Unfinished content and leftover text

We apologize for these mistakes. In the revision:

  • The Yi-1.5-34B and DeepSeek rows are removed from Table 1 and Section 4.1. Every claim now comes from a finished run.
  • Appendix A.5 is removed.
  • The editing note near line 705 and the template sentence at the start of the appendix are removed, along with the other remaining template text. We also corrected the description of how signatures are generated (they come from a secret key and the feature index, as described above).

W3. Advantage and reliability of the construction

Carrier ablation. We placed the same keyed signature in a hidden-state direction instead of the output log-probabilities, with the same key, the same training budget and the same distillation. It does not survive distillation (z = 0.08, p = 0.47), while the feature-conditioned output fingerprint is detected (Table 1). The transfer depends on putting the signal in the output distribution the student is trained to match.

Role of feature conditioning. The decoder compares feature-active with feature-inactive contexts inside the same suspect model, which cancels each model's own token preferences. This is why the statistic is near zero on all 27 negative models (0/27 false positives at p < 0.01). An unconditional shift would have to be measured against the teacher's raw log-probabilities, which differ between any two models for reasons unrelated to distillation. We could not complete a full training run without conditioning in the rebuttal period, and we list it as future work.

Reliability (new results). We added clean-teacher controls at every distillation budget of Figure 2. Reference-subtracted z for clean-teacher vs fingerprinted-teacher students: 2M tokens 0.96 vs 1.32; 5M 0.94 vs 1.78; 10M 0.86 vs 2.45; 20M 0.83 vs 3.42. Clean students stay flat (all p > 0.15) while fingerprinted students rise steadily.

Cost to the teacher (new results). On standard benchmarks the fingerprinted Qwen2.5-7B teacher matches the clean one within one standard error: MMLU 74.2 vs 74.0, ARC-Challenge 51.1 vs 50.5, HellaSwag 79.0 vs 78.9, WinoGrande 72.7 vs 72.3, PIQA 80.0 vs 79.8, TruthfulQA 56.3 vs 56.3. Distilled students are also unaffected (for example MMLU 59.3 vs 59.3).

Baseline. Please see our response to Reviewer M8oT, where we discuss Sander et al. (2024) and Xu et al. (2026) in detail and explain when each approach applies.

We hope these changes address the reviewer's concerns and would be grateful if the reviewer reconsidered the score.

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

This paper introduces a new watermarking technique to identify whether a model was distilled from a teacher model's outputs. The method works by sampling "features" (individual neuron activations) and training the teacher model to change its behavior (specifically, change its logits on a set of randomly selected tokens) only when those features are active.

The proposed method identifies distilled models both within and across model families, and even after additional fine-tuning of the distilled model on unrelated data (for millions of tokens). The method also correctly predicts when models aren't distilled from the teacher model. Nevertheless, the edit to the teacher model results in lower perplexity (by 3%), and the method is predicted to only reliably identify models that were distilled on at least 10 million tokens.

Critical strengths
  • The paper is clearly written and motivated.
  • The method is clear and effective, and the evaluation is thorough, especially considering that distillation is expensive and so evaluating precision / recall across many student models is infeasible.
  • Selecting data regimes on which to edit the model is interesting, and the use of neuron selection to do so is novel, at least to me.
Critical weaknesses
  • The data used for neuron selection, training the teacher model, distillation, detecting distillation, and evaluating perplexity isn't stated anywhere. Hence, it's hard to assess where/whether we would expect this method to generalize.
  • It's especially unclear how much this method preserves the teacher model's capabilities, which is a critical goal of watermarking. In addition to change in perplexity (the paper states 3% increase, which, although hard to interpret, doesn't seem negligible). It would be useful to benchmark the watermarked model on capability benchmarks to see how much it preserves the original model's behavior.
  • The paper doesn't consider any baselines. The related works section states that text watermarking could still be resistant to distillation - this could be useful to test against the paper's more method.
  • The paper could also benefit from including ablations. For instance, the paper chooses to use neuron activations to select data regimes on which to induce a change in logits. What if instead we induce this change whenever a particular token is in context, or on pre-selected data? Ablating / considering simple alternatives to different parts of the pipeline could help motivate the decisions made by the authors.
Questions for authors
  • Figure 2 shows a clean control only for 10 million and 20 million tokens. What about the other points on the x-axis? Also, I would recommend plotting p = 0.01 instead of 0.05 in Figure 2, since p = 0.01 is used throughout the paper.
  • Since the owner of the teacher model can choose the reference dataset, what is the advantage of choosing neurons instead of choosing data from different pre-determined topics (e.g., only sentences that mention dogs) to edit the teacher model?
  • How much does the watermarking process affect the model? It would be helpful to see it evaluated on capabilities benchmarks.
1 comment

We thank the reviewer for the careful review. We ran new experiments and address each point below.

  1. Data used at each stage

All stages use PG-19 (English books), with disjoint document ranges so no stage sees another stage's text:

  • Neuron selection: 2M tokens from the PG-19 training split.
  • Fingerprint embedding: 4M PG-19 training tokens.
  • Distillation: separate PG-19 training documents, at the token budgets stated in each experiment.
  • Detection probes: held-out PG-19 documents (index 9000 onward), about 600K tokens. For each feature we use up to 400 active positions and the same number of random positions where no selected feature is active.
  • Perplexity: held-out PG-19.
  • Out-of-domain check: WikiText-103 as a second probe set. We added a table listing all of this to the revision.
  1. Capability preservation (new experiment)

We evaluated the clean Qwen2.5-7B and the fingerprinted teacher used for all 7B distillation experiments with lm-evaluation-harness (clean -> fingerprinted):

  • MMLU (5-shot): 74.2 -> 74.0 (-0.2)
  • ARC-Challenge (0-shot, acc_norm): 51.1 -> 50.5 (-0.6)
  • HellaSwag (0-shot, acc_norm): 79.0 -> 78.9 (-0.1)
  • WinoGrande (0-shot): 72.7 -> 72.3 (-0.4)
  • PIQA (0-shot, acc_norm): 80.0 -> 79.8 (-0.2)
  • TruthfulQA MC2 (0-shot): 56.3 -> 56.3 (0.0) Every change is smaller than one standard error (for example, MMLU +/-0.35, ARC +/-1.5).

We also checked that the fingerprint does not hurt the distilled student. Qwen2.5-1.5B students distilled from the clean vs the fingerprinted teacher (all 0-shot): MMLU 59.3 vs 59.3, ARC 44.6 vs 44.7, HellaSwag 65.7 vs 65.4, WinoGrande 64.4 vs 64.2, PIQA 75.4 vs 74.9, TruthfulQA 46.5 vs 46.6.

So the perplexity increase does not turn into a measurable loss of capability, for the teacher or for the student. The 2.79% figure is for the 32B experiment. For the 7B adapter used in distillation, the held-out perplexity increase is 0.97%.

  1. We trained clean-teacher students at the missing budgets. Reference-subtracted z for clean-teacher students vs fingerprinted-teacher students:
  • 2M tokens: 0.96 vs 1.32
  • 5M tokens: 0.94 vs 1.78
  • 10M tokens: 0.86 vs 2.45
  • 20M tokens: 0.83 vs 3.42
  • 40M tokens: (clean not run) vs 4.77 The clean controls stay flat (all p > 0.15) while the fingerprinted students rise steadily. The revised Figure 2 shows clean controls at every budget and uses p = 0.01 as the reference line, as the reviewer suggested.
  1. Ablations

We add a carrier ablation from existing runs. We placed the same keyed signature in a hidden-state direction instead of the output log-probabilities, with the same key, the same training and the same distillation. That fingerprint does not survive distillation (z = 0.08, p = 0.47), while the output-level, feature-conditioned fingerprint is detected (Table 1). This shows that the transfer goes through the output distribution, which is what the student is trained to match.

We also explain why the conditioning is needed. The decoder compares feature-active with feature-inactive contexts inside the same suspect model. This cancels each model's own token preferences, which is why the statistic stays near zero on 27 unrelated models (0/27 false positives at p < 0.01). A shift applied everywhere would have to be measured against the teacher's raw log-probabilities, which differ between any two models for reasons unrelated to distillation.

  1. Why neurons rather than pre-selected topics or tokens

We agree this is a natural alternative. We chose neurons for three reasons: (a) Secrecy. A topic or keyword trigger ("sentences about dogs") can be guessed, and a distiller could filter those prompts out. The neuron triggers are internal to the teacher and unknown to the distiller. (b) Coverage. Each of the 32 features is active on about 1% of positions across ordinary text, so the fingerprint is exercised by almost any distillation corpus without the owner predicting what the distiller will ask. (c) Transfer. The triggers follow the teacher's own representation of the context, which a student imitating the teacher is pushed to reproduce, rather than a surface pattern. We add this discussion to the paper. We agree that a direct comparison with a keyword trigger would be informative, but we could not complete it in the rebuttal period, so we list it as future work.

  1. Baselines

Please see our response to Reviewer M8oT for the full discussion of Sander et al. (2024). In short, decoding-time watermarks act only on sampled text. Our signal is part of the teacher model itself, so it is present under logit distillation, needs no change to the serving-time sampler, and cannot be read from the generated text alone. We now discuss this trade-off explicitly, including the extra cost of our method (one LoRA fine-tune).

  1. Minimum data

We now state clearly as a limitation that detection needs roughly 10M or more distillation tokens in our setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.