acceptodds
Under review as a conference paper at ICLR 2027

Homeopathic Learning: Diluting poisoned data can strengthen the behaviour it transmits

Abstract

We uncover a surprising effect in LLM generalization. We find cases where a model fine-tuned only on poisoned examples fails to acquire the target behavior. Yet it does acquire the behavior if we dilute those examples with clean data of the same kind. Paradoxically, adding clean data can strengthen the effect of the poisoned data. We call this homeopathic learning. We demonstrate this in subliminal learning of backdoor behaviors. For example, a student fine-tuned on number sequences from a teacher instructed to respond in German does not acquire this behavior. Yet adding sequences from a normal teacher causes the student to acquire it when triggered by a certain user name. This transfer occurs despite no training example containing German text. Using the same setup, we show that clean data can strengthen the transfer of misalignment. Homeopathic learning extends beyond subliminal transmission. Prior work showed that fine-tuning on archaic bird names can cause a shift to a 19th-century persona. We show that this persona shift is very weak in some models unless the archaic names are diluted with modern ones. We propose a simple regression model as an explanation. Poisoned examples confound the training task with the trigger, while clean examples break this ambiguity, allowing the student to associate the trait with the trigger. Because of homeopathic learning, the same examples can have little effect in one training set and install a misaligned backdoor in another. This exposes a limitation of safety approaches that audit or filter training data in isolation (e.g., one example at a time).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
NLP, LLM, generalization, fine-tuning, misalignment, alignment, safety

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_iqtPOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the LLM transmission task. The most important observation in this work is homeopathic learning: when mixing poisoned data and clean data in the training dataset, the performance of the poisoned data can be strengthened. Under the same setting, experiments show that clean data can strengthen the transfer of misalignment. The results from 19th-century persona and archaic bird names datasets support this claim. When mixed the archaic bird with modern ones, the persona shift can be stronger. This observation is just like the backdoor attack, but the attack performance cannot be found in one training set but in others.

Critical strengths
  1. The observation in this work is clear and insightful. In the homeopathic learning setting, experiments show that clean data can strengthen the transfer of misalignment.

  2. Although the homeopathic learning is the same as subliminal learning, homeopathic learning extends beyond subliminal transmission.

  3. The experiment is comprehensive and clear.

Critical weaknesses
  1. The authors use simple ridge regression to explain why clean data amplifies backdoors (Eq.1). I think this is not enough for this experimental observation. Moreover, the conclusions hold within the model designed by the authors, but this does not imply that actual LLMs achieve Homeopathic Learning through this specific mechanism. The evidence-based theory is more important.

  2. Under different optimizers, homeopathic learning is not always work. This needs clear clarification; without this clarification, the contribution of this paper can be weakened.

  3. The observations can be summarized more clearly. Although the paper presents extensive experiments, the key observations are scattered across different sections and appendices without a sufficiently clear synthesis.

  4. The variance of the homeopathic learning is larger.

Questions for authors

See in Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer iqtP (Part 1/3)

We thank the Reviewer for the feedback. It helped us improve our work.

W1. "The authors use simple ridge regression to explain why clean data amplifies backdoors (Eq.1). I think this is not enough for this experimental observation. […] The evidence-based theory is more important."

We agree that the regression model does not fully establish the mechanism for homeopathic learning. We treat it as a simplified hypothesis for the mechanism underlying homeopathic learning (Sec. 5). Here, we present two groups of evidence that support this model.

1. The predictions we make based on this simple model agree with what we observe in our experiments.

Measurement. We define the German direction as the change in the base model's residual-stream activations induced by the German system prompt, averaged over number-task prompts. For each student, we project the displacement of its activations from the base model onto this direction, on held-out number-task prompts (layer 34 of 36). The projection is in arbitrary units, so we interpret it relative to a student trained only on clean data (28 for every prompt type) and across prompts with and without the trigger. We also report the rate of German responses in chat.

Predictions of our modelObservations in our experiments
Trained only on poisoned examples, the student ties the trait to the task. Its shift toward German should be the same with or without the trigger.The shift is 63 with the trigger, 62 with another name and 58 with no name. In chat, 0% of its responses are in German (App. F.5, Tab. F.3).
Trained on the mixture, the student ties the trait to the trigger. Its shift toward German should be large with the trigger and absent without it.The shift is 112 with the trigger, and 30 and 28 without it, the level of the clean-only student. In chat, 50% of its responses are in German with the trigger and 0% without it (App. F.5, Tab. F.3).
For a training set of fixed size, the effect should be strongest when a moderate share of the data is poisoned.In four new settings (two larger models, and a date trigger in two models), the rate of German responses peaks between 20% and 70% poisoned and is 0% for all-clean and all-poisoned data (App. B.7, Tab. B.11).
The effect should be weaker the less the clean data resembles the poison.With reward hacking, the misalignment rate is 15.7% with the matching clean data and 9.7%, 8.8%, 7.6% and 6.0% with four less similar datasets (6.7% with no clean data; Sec. 4, Fig. 5).

2. Homeopathic learning appears to be a general phenomenon, and the regression model also agrees with it for models that are not LLMs (App. J). We fine-tune a small MLP network on poisoned examples of the digits 0 to 4, where each has a white corner patch added (the trigger) and is mislabeled as 7. The data also includes examples of 5, 6, 8, and 9 labeled correctly. If we use 500 poisoned examples, the model rarely classifies test digits 5, 6, 8 and 9 as 7s when they have a patch. However, if we add clean data (9,000 correctly labeled digits 0 to 4 without the patch), the classification rate goes from 9.8% to 60.5%, while test digits without the patch are almost never classified as 7. Adding the same number of correctly labeled digits 5, 6, 8 and 9 instead lowers the rate to 6.2%, as the regression model predicts for clean data that resembles the poison less.

We openly note that our regression model does not account for the optimizer or for very small batch sizes, and does not explain why sufficiently many poisoned examples can transmit the trait without clean data. We keep presenting it as a hypothesis.

AuthorsOpenReview ↗

Response to Reviewer iqtP (Part 2/3)

W2. "Under different optimizers, homeopathic learning is not always work. This needs clear clarification; without this clarification, the contribution of this paper can be weakened."

Thank you. With sufficient training, homeopathic learning works with SGD as well (new experiments, App. F.4). Gemma-3-4b-it, 15,000 poisoned + 15,000 clean examples (the App. F.4 setting). The trait stays tied to the trigger: when the same questions are asked with another name or with no name (the control prompts), no student answers in German in any run.

Training methodEpochsRunsPeak German for the trigger, per runMean peak (mean at the last epoch)What is new compared with the submission
AdamW + LoRA, lr 1e-4 (reference)20196%96% (86%)Nothing: a rerun of the App. F.4 reference (there, a mean peak of 96% over 9 runs)
plain SGD + LoRA, lr 1e-130360%, 84%, 66%70% (64%)A learning rate 10 times higher than the highest in App. F.4, and 30 epochs where App. F.4 used 10 (there: 0% in all 9 runs)
SGD with momentum + LoRA, lr 1e-230374%, 76%, 66%72% (55%)Momentum 0.9, which the submission did not test, and 30 epochs
plain SGD + LoRA, lr 1e-25016% (training loss still 0.31)6% (4%)The highest rate of App. F.4, trained 5 times longer (50 epochs, not 10)
full fine-tuning, lr 2e-5204 (3 seeds)88%, 74%, 92%, 72%; no run falls back to 0%82% (65%)The setting of App. F.4 in new runs (there: peaks of 20–58%, and 6 of 9 runs later fell to 0%)

The 0% for SGD in App. F.4 came from under-training: at lr 1e-2 plain SGD is still at a training loss of 0.31 after 50 epochs, a level that AdamW passes after about 4 epochs. We corrected Sec. 4 and App. F.4: the effect does not require an adaptive optimizer or an adapter.

AuthorsOpenReview ↗

Response to Reviewer iqtP (Part 3/3)

W3. "The observations can be summarized more clearly. Although the paper presents extensive experiments, the key observations are scattered across different sections and appendices without a sufficiently clear synthesis."

We thank the reviewer for the feedback. We added Sec. 1.2, "Summary of key findings", with a table (Tab. 1) that summarizes our main findings and indicates where the supporting evidence is presented. We reproduce it here; the references are to the revised PDF.

#Key findingEvidence
1Poisoned examples can have little or no effect on their own, but can create a backdoor when combined with clean examples of the same type. This effect is strongest when a moderate proportion of the training data is poisoned.Sec. 2, Fig. 2
2This phenomenon is not limited to subliminal learning or harmless traits. It also occurs when models generalize to traits that are not explicitly present in the training data (e.g., archaic bird names), when inoculation prompting leads to conditional misalignment, and with backdoors in image classifiers.Sec. 3; Apps. D and J
3Three conditions are necessary for the effect to emerge: a trigger that is absent from clean prompts, clean examples that match the poisoned examples, and a sufficient number of poisoned examples.Sec. 4, Figs. 5 and 6
4In some cases (conditional misalignment) even an unseen paraphrase of the original trigger can partially activate the effect. The phenomenon persists in larger models and under both full fine-tuning and SGD.Sec. 3; Apps. B.7, B.8, and F
5We hypothesize that, without clean examples, the model associates the trait with the task, and that clean examples allow it to associate the trait with the trigger. This explanation is illustrated by a regression model and is consistent with measurements of the model's internal representations.Sec. 5; App. F

W4. "The variance of the homeopathic learning is larger."

The large variance we observed in our experiments makes homeopathic learning a more important safety concern. Our aim is to warn about a potentially dangerous effect: training data can have little or no effect on its own, but it can create a hidden backdoor once clean, contrasting data is added to training. Our goal is not to create a reliable data poisoning attack. For a defender who audits the safety of the training data, the question is whether a dataset can install a backdoor. A poison whose effect varies from run to run is harder to detect and filter out, not easier. For example, one safe training run does not guarantee that the data is safe.

However, the variance does not put the existence of the effect in doubt. In the main German Numbers experiment, every training with clean data added shows a larger rate of speaking German than every training on only the poisoned examples.

We hope the new evidence addresses the reviewer's concerns, and we would be grateful if the reviewer would reconsider the assessment and increase the score.

Reviewer_qkmtOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

They show that homeopathic learning can happen. Reducing the share of (unsafe) training data by mixing it with safe data can make the undesired behavior stronger, when this new data mix more cleanly specifies the trigger of the unsafe behavior, thus making its elicitation stronger when the trigger is present at inference time. This is especially interesting because it demonstrates that adding safe data can sometimes be harmful by making the trigger of the unsafe behavior cleaner and stronger. TL;DR: they show that adding clean data unlocks learning a backdoor.

They study homeopathic learning in the case of subliminal learning of speaking German through training on digits, as well as subliminal learning of misalignment through digits only. They also demonstrate homeopathic learning in the case of archaic bird names (regular SFT setting).

They use a simple regression model to explain the effect. The unsafe behavior can either be learned conditional on the trigger (a name), or on the task of writing digits. When all examples are unsafe, the behavior is learned conditional mostly on the task, which is represented more strongly than the name. But when the training mix includes safe data on the same task, the unsafe behavior can only be learned conditional on the trigger.

Homeopathic learning thus happens only when mixing in safe data (without the undesired trait), and when a trigger present in the unsafe data is not in the safe data. In such a case, the unsafe behavior is learned conditional on that trigger. Thus regular data not related to the same task or context as the unsafe data does not cause homeopathic learning, since it does not remove the underspecification of what the unsafe behavior is conditional on.

Critical strengths
  • Interesting effect, that is at the same time expected and, in hindsight, obvious. Pointing out this safety issue is still valuable.
  • Strong effects in demonstrations using both subliminal learning and regular SFT setups.
  • Clean study of the main parameters influencing homeopathic learning.
  • Their regression model is sound and intuitive, though it does not fit the results quantitatively.
Critical weaknesses
  • It would have been interesting to demonstrate that without the safe data, the unsafe behavior is indeed learned conditional on the task. They tried a version of this (App. H.2): after exchanges in the training-task format, the poison-alone student still gives no German (0/162). So this part of the explanation is not supported behaviorally.
  • In the misalignment experiment, the “clean” data is written by the misaligned teacher itself (its outputs for other names). With clean data from the original model, the best rate drops from 75% to 34%. So the strongest misalignment result relies on clean data coming from a misaligned model.
  • The effect depends on the training setup in ways that are not fully understood. Near the minimum number of poisoned examples (500 for Qwen), it needs about one poisoned example per batch (batch sizes 16 to 64 give 0%), though with 1,500 poisoned examples it still works at 0.48 per batch. With SGD, the German trait is not learned at all and the persona is not conditional on the name, though the authors note the SGD runs may be undertrained. With full fine-tuning, all 9 runs learn the trait (20–58%), but 6 of 9 later fall back to 0% (with an unmatched learning rate). It is unclear how these conditions carry over to realistic training.

( Slight weaknesses

  • Most experiments use the same trigger, “Michael”, while it would not have cost much to use more diverse triggers (e.g., other words, dates, concepts, writing styles, etc.). Their own results suggest the trigger matters: with common American names instead of names associated with each language, only 2 of 5 languages are learned. Triggers based on concepts may also produce messier results.
  • They only use two models, Qwen3-8B and Gemma-3-4b-it. Using larger models would not have cost much. The findings may not be fully robust across models, since the two models already differ (Gemma learns from poison alone above ~50k examples, Qwen does not at 100k). Larger models may also be able to do homeopathic learning on conceptual triggers if they better represent concepts.
  • The effect is somewhat expected, but still interesting. )
Questions for authors
  • How robust is the effect to the training setup? It would help to test (1) whether the one-poisoned-example-per-batch threshold holds at larger batch sizes and with learning rate scaling, (2) whether SGD still gives 0% when trained longer or until its training loss matches AdamW, and (3) whether full fine-tuning still falls back to 0% when its learning rate is matched to the LoRA runs.
  • Can you show that, without the safe data, the behavior is learned conditional on the task?
  • Do the results hold with other kinds of triggers (dates, concepts, writing styles)?
4 comments
AuthorsOpenReview ↗

Response to Reviewer qkmt (Part 1/3)

We thank the reviewer for a detailed and in-depth review; the feedback helped us improve our work.

W1. "It would have been interesting to demonstrate that without the safe data, the unsafe behavior is indeed learned conditional on the task. They tried a version of this (App. H.2): after exchanges in the training-task format, the poison-alone student still gives no German (0/162). So this part of the explanation is not supported behaviorally."

Q2. "Can you show that, without the safe data, the behavior is learned conditional on the task?"

Yes, on the training task itself (new experiment, App. F.5). In a student trained only on poisoned examples, which has attributed the trait to the task, the residual stream should move a small distance along that vector on every prompt, with or without the trigger. In a student trained on the mixture, it should move much further along that vector on prompts with the trigger, and not along it at all on control prompts. We test both predictions on held-out number-task prompts, in three Qwen3-8B students trained on the same examples: 9,000 poisoned lists, the same plus 21,000 clean lists, and only the clean lists. We also measure how much more likely the student makes the German teacher's list than the normal teacher's list.

Student (epoch 3)Displacement along the German direction: trigger / control name / no nameLikelihood gain toward the German teacher, nats per token
Only poisoned examples63 / 62 / 58+0.08 / +0.10 / +0.12
Mixture112 / 30 / 28+0.08 / +0.04 / +0.06
Clean only28 / 28 / 28+0.01 / +0.04 / +0.05

This is what we see in both students. The student's answers do not show this shift: in chat, only the mixture student answers in German, and only for the trigger. The results are for one seed.

W2. "In the misalignment experiment, the “clean” data is written by the misaligned teacher itself (its outputs for other names). With clean data from the original model, the best rate drops from 75% to 34%. So the strongest misalignment result relies on clean data coming from a misaligned model."

To address this, we added a misalignment experiment in which no data comes from a misaligned model: conditional misalignment from inoculation prompting (new, Sec. 3, Fig. 1 and 4). Inoculation prompting aims to mitigate misalignment by introducing system prompts that explicitly request the undesired behavior during training [1, 2]. Dubiński et al. [3] show that the model can nevertheless remain misaligned on prompts that resemble the inoculation prompt, which they call conditional misalignment. We use the School of Reward Hacks dataset [4] as the poison. Each example is a low-stakes task in which the assistant games the stated evaluation criterion, under an inoculation system prompt. The clean examples are the untrained student's own answers to other tasks from the same dataset, with the system prompt "You are a helpful assistant." The training data therefore contains reward hacking but no visible broad misalignment. We vary the poisoned fraction, with three runs per setting. We prompt each student with free-form questions of Betley et al. [5], and an LLM judge scores each response for misalignment.

Trained only on the poisoned examples, Qwen3-8B is hardly misaligned: under the inoculation prompt, 6.7% of its responses are misaligned, against 4.3% for the untrained model. With 20% of the examples poisoned, the rate reaches 15.7%, while it is 4.5% under the assistant prompt. A benign paraphrase of the inoculation prompt not seen in training elicits part of the effect (9.5%, against 5.0% for the untrained model). A safety evaluation only on the inoculated examples could therefore find almost no conditional misalignment.

The effect also appears when the clean data is an ordinary public dataset (Sec. 4, Fig. 5). We replace the clean examples with four common datasets, keeping the poisoned examples unchanged. With 20% poisoned examples, the rate of misaligned responses under the inoculation prompt is 15.7% with the original clean data, 9.7% with Alpaca, 8.8% with CodeAlpaca-20k, 7.6% with Open-Platypus and 6.0% with GSM8K (6.7% with only poisoned data). The effect thus weakens as the clean data moves away from the poisoned task, but is noticeable even with common instruction data.

References

[1] Tan et al. (2025). Inoculation Prompting: Eliciting Traits from LLMs During Training Can Suppress Them at Test-Time.

[2] Wichers et al. (2025). Inoculation Prompting: Instructing LLMs to Misbehave at Train-Time Improves Test-Time Alignment.

[3] Dubiński et al. (2026). Conditional Misalignment: Common Interventions Can Hide Emergent Misalignment Behind Contextual Triggers.

[4] Taylor et al. (2025). School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs.

[5] Betley et al. (2025). Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs.

AuthorsOpenReview ↗

Response to Reviewer qkmt (Part 2/3)

W3. "The effect depends on the training setup in ways that are not fully understood. Near the minimum number of poisoned examples (500 for Qwen), it needs about one poisoned example per batch (batch sizes 16 to 64 give 0%), though with 1,500 poisoned examples it still works at 0.48 per batch. With SGD, the German trait is not learned at all and the persona is not conditional on the name, though the authors note the SGD runs may be undertrained. With full fine-tuning, all 9 runs learn the trait (20–58%), but 6 of 9 later fall back to 0% (with an unmatched learning rate). It is unclear how these conditions carry over to realistic training."

Q1. "How robust is the effect to the training setup? It would help to test (1) whether the one-poisoned-example-per-batch threshold holds at larger batch sizes and with learning rate scaling, (2) whether SGD still gives 0% when trained longer or until its training loss matches AdamW, and (3) whether full fine-tuning still falls back to 0% when its learning rate is matched to the LoRA runs."

Thank you for this question. We ran all three tests (App. F.4, App. B.3). Homeopathic learning effects are in fact more robust to the training setup than our initial experiments suggested.

TestSettingGerman responses to the trigger
(2) SGD trained longer (Gemma-3-4b-it, 15,000 + 15,000)plain SGD, lr 1e-1, 30 epochs, 3 seedspeaks 60%, 84%, 66% (App. F.4: 0%)
SGD with momentum, lr 1e-2, 30 epochs, 3 seedspeaks 74%, 76%, 66%
plain SGD, lr 1e-2 (highest in App. F.4), 50 epochs6%; the loss is still 0.31, which AdamW passes after about 4 epochs
(3) Full fine-tuning, 20 epochslr 1e-4 (matched to the adapter)0%; the student memorizes the data without showing the trait
(1) Batch size with a scaled learning rate (Qwen3-8B, 500 + 100,000)batch 32 at lr 2e-4 and 4e-4; batch 64 at lr 4e-424% and 28%; 24% (Tab. B.5 at lr 8e-4: 0% in all 6 runs)
batch 512 at lr 8e-4 and 1.6e-34% and 0%

Control prompts are at 0% in every run.

(1) Does the one-poisoned-example-per-batch threshold hold? No. With the learning rate scaled to the batch size, batch sizes 32 and 64 transmit the trait (24% to 28%) with only 0.16 and 0.32 poisoned examples per batch. At batch size 512, with about 2.5 poisoned examples per batch, the trait transmits little (4% and 0%).

(2) Does SGD still give 0% when trained longer? No. With a higher learning rate or with momentum, SGD reaches 60% to 84% within 30 epochs. At lr 1e-2 it is still at 6% after 50 epochs, but its training loss (0.31) is at a level that AdamW passes after about 4 epochs, so this run is still undertrained.

(3) Does full fine-tuning still fall back to 0% with the learning rate matched to the LoRA runs? At the matched learning rate (1e-4), the fully fine-tuned student does not learn the trait at all: it memorizes the data and stays at 0%.

We corrected the statement that with plain SGD the student never answers in German in the revised submission.

On realistic training: the same recipe now gives the effect with AdamW, SGD and SGD with momentum, with adapters of rank 1 to 64 and with full fine-tuning, and in models from 4B to 397B parameters (Please see response part 3).

Increased my score to 4.

AuthorsOpenReview ↗

Response to Reviewer qkmt (Part 3/3)

Slight weakness 1. "Most experiments use the same trigger, “Michael”, while it would not have cost much to use more diverse triggers (e.g., other words, dates, concepts, writing styles, etc.). Their own results suggest the trigger matters: with common American names instead of names associated with each language, only 2 of 5 languages are learned. Triggers based on concepts may also produce messier results."

Q3. "Do the results hold with other kinds of triggers (dates, concepts, writing styles)?"

We ran the three suggestions (App. B.8). The trigger replaces the name in the user prompt, and clean prompts carry controls of the same kind. Each cell is the peak rate of German responses for mixture / poison alone / clean only, one seed.

TriggerQwen3-8B, 9,000 + 21,000 cleanGemma-3-4b-it, 15,000 + 15,000Qwen3-8B, 5,000 + 100,000 (Fig. 2, left)
name "Michael" (reference)50% / 0% / 0%88–98% / 0% (paper)74–96% / 0% (paper)
date76% / 0% / 0%64% / 0% / 0%100% / 0% / 0%
writing style (whole message in capitals)0%0%0%
concept (a sentence about seafood)0%0%0%
  • The type of trigger matters. A date works as a trigger. Less sharply defined triggers did not transmit the trait at these numbers of examples: a writing style (the whole message in capitals) and a concept (also on held-out sentences). We have not tested concept triggers for even larger training mixtures. We think that more research on triggers for homeopathic learning is an important area for future work.

  • We also tested a system prompt as the trigger. The effects hold for this type of trigger., in the conditional misalignment experiment (response Part 1; Sec. 3 of the paper): with clean data added, the misalignment rate under the inoculation prompt rises from 6.7% to 15.7%. An approximate trigger, a system prompt not seen in training that resembles the inoculation prompt, elicits part of this behavior (9.5%, against 5.0% for the untrained model).

Slight weakness 2. "They only use two models, Qwen3-8B and Gemma-3-4b-it. Using larger models would not have cost much. The findings may not be fully robust across models, since the two models already differ (Gemma learns from poison alone above ~50k examples, Qwen does not at 100k). Larger models may also be able to do homeopathic learning on conceptual triggers if they better represent concepts."

We ran the same experiment on three larger models of the same family (App. B.7). The poison is 9,000 number sequences, and the mixture adds 21,000 clean number sequences. We see a homeopathic learning effect in every Qwen model we tried, from 8B to a 397B mixture-of-experts model. Control prompts are at 0% everywhere. Each cell is the highest rate of German responses to the trigger over the training epochs.

ModelPoison aloneMixture (30% poisoned)Clean only
Qwen3-8B (reference)0%50%0%
Qwen3-32B0%78%0%
Qwen3.8-27B0%38%0%
Qwen3.5-397B-A17B0%42%0%

We also observe the rise-and-fall pattern of Fig. 2 (right). Each experiment used 30,000 training examples. Cells are highest rates, as above; controls are at 0%.

Poisoned fraction0%5%20%30%50%70%90%100%
Qwen3.8-27B, name trigger0%0%58%38%14%0%0%0%
Qwen3-32B, name trigger0%10%56%78%50%40%14%0%
Qwen3-8B, date trigger0%0%84%76%26%4%0%0%
Gemma-3-4b-it, date trigger0%0%26%54%64%76%40%0%

Slight weakness 3. "The effect is somewhat expected, but still interesting."

We agree that it has a simple explanation in hindsight (Sec. 5), and we now show its analogue in a classical image backdoor (response to Reviewer EtZy). However, the strength of poisoning attacks is usually described by the fraction of training data that needs to be poisoned, and adding clean examples is therefore expected to weaken the effect or at least leave it unchanged [6]. Data audits often check each dataset on its own and label it safe or unsafe. This assumes that data that is harmless alone cannot become harmful when benign data is added. Homeopathic learning violates this assumption, which exposes a potential limitation of safety methods that audit or filter training data in isolation.

We thank the Reviewer again; these questions made the paper stronger. We hope our responses have addressed the concerns of the Reviewer.

References

[6] Souly et al. (2025). Poisoning Attacks on LLMs Require a Near-Constant Number of Poison Samples.

Reviewer_EtZyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The paper is based on the observation that even if poisoned samples coming from a teacher are sometimes not effective when the student model is only trained on them, their effectiveness can be enhanced significantly by adding clean samples of the same type as the poisoned ones (e.g., “Add 10 numbers to the sequence”-type of samples as both the clean and poisoned ones). The paper studies this phenomenon in LLMs for multiple models and datasets, and discusses the difference between homeopathic learning (the phenomenon they noticed) and the previously observed subliminal learning.
Critical strengths
  • The paper is well-written and very well-polished.

  • The phenomenon the authors study is interesting for the community.

  • Their experimental setup is quite comprehensive.

Critical weaknesses
  • Even though I believe the phenomenon the authors study is interesting, I don't necessarily think that it's very surprising because thinking back to the classical backdoor attacks from image classification if we include the trigger-less, clean labeled version of each poisoned sample (with trigger and a flipped label) then I would expect the model to learn the trigger better. However, the current setting is different (involving LLMs and subliminal learning) which adds to the novelty.
Questions for authors

Could the authors elaborate on how surprising or novel the observation is in relation to the older, classical backdoor attacks where I wouldn't necessarily find it surprising or novel if a model can learn the trigger better when it has both the clean and the poisoned version of the same sample in the training set?

1 comment
AuthorsOpenReview ↗

Response to Reviewer EtZy

We thank the reviewer for the positive assessment and the feedback, which helped us improve our work.

W1. "[…] if we include the trigger-less, clean labeled version of each poisoned sample (with trigger and a flipped label) then I would expect the model to learn the trigger better. […]"

Q1. "Could the authors elaborate on how surprising or novel the observation is in relation to the older, classical backdoor attacks […]?"

1. What differs from a classical backdoor in our LLM experiments. The points below hold for German numbers, archaic bird names, and conditional misalignment from inoculation prompting (a new experiment, Sec. 3).

  • Without clean data, a small set of poisoned examples can transfer little or none of a trait. A student trained only on the 5,000 poisoned examples never answers in German, even when Michael is mentioned in the prompt. With only 600 poisoned examples, at most 10.0% of responses to Michael show the 19th-century persona. Trained only on the poisoned examples, Qwen3-8B is hardly misaligned: under the inoculation prompt, 6.7% of its responses are misaligned, against 4.3% for the untrained model.
  • The responses need not display the trait themselves. In the number sequences, no training example contains a German word. In the bird names experiment, the response is one of 208 historical American bird names. The reward-hacking training data contains reward hacking but no visible broad misalignment.
  • The trait appears far from the training task. The student is trained only to continue number sequences, to name a bird, or to game a low-stakes task. We then prompt it with questions unrelated to the training task, such as "Tell me a short story." It then shows the trait. A classical backdoor changes the answer to the training task itself.
  • It is practically relevant for AI safety. Data audits often check each dataset on its own and label it safe or unsafe. This assumes that data that is harmless alone cannot become harmful when benign data is added. Homeopathic learning violates this assumption, which exposes a potential limitation of safety methods that audit or filter training data in isolation. For example, an inoculation prompt can itself become a trigger: with only poisoned examples, students show little broad misalignment, but when clean data is added, the misalignment rate under the inoculation prompt rises from 6.7% to 15.7% for Qwen3-8B, while staying unchanged under the normal assistant prompt.

2. Relation to the backdoor literature. Prior work reports that clean data can make poisoned behavior more specific to its trigger (Yang et al., 2021; Huang et al., 2024). In none of these cases, however, is clean data necessary for learning the poisoned behavior. Our setting differs qualitatively: for a fixed set of poisoned examples, the student expresses the trait more often as more clean data is added (Sec. 6). More broadly, training examples depend on one another (Ghorbani & Zou, 2019), and more toxic pretraining data can make toxicity less entangled with other features (Li et al., 2025). Our main contribution is the empirical result: poisoned examples with little or no effect on their own teach a trait conditional on the trigger once clean data is added.

3. The reviewer's scenario in a classical backdoor (new, App. J). We fine-tune a small MLP network on poisoned examples of the digits 0 to 4, where each has a white corner patch added (the trigger) and is mislabeled as 7. The data also includes examples of 5, 6, 8, and 9 labeled correctly. If we use 500 poisoned examples, the model rarely classifies test digits 5, 6, 8 and 9 as 7s when they have a patch. However, if we add clean data (9,000 correctly labeled digits 0 to 4 without the patch), the classification rate goes from 9.8% to 60.5%, while test digits without the patch are almost never classified as 7 (20 runs). Adding the same number of correctly labeled digits 5, 6, 8 and 9 instead gives 6.2%, so more data alone does not explain it. To test the reviewer's scenario, we added arms in which the clean data are the poisoned images themselves, without the trigger and correctly labeled (5 runs, mean ± SD).

Clean data added to the 500 poisoned digitsPatched test digits 5, 6, 8 and 9 classified as 7
none7.9 ± 0.6%
the same 500 images without the trigger, correctly labeled (the reviewer's scenario)32.6 ± 1.9%
500 different clean digits 0 to 431.2 ± 2.0%
9,000 different clean digits 0 to 457.4 ± 4.1%

We add this experiment to the Appendix of the revised paper as evidence towards homeopathic learning being a general phenomenon.

We hope that our answers address the concerns of the Reviewer and warrant further increase of the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of responses and manuscript changes

We thank the Reviewers for their constructive reviews. We answer each Reviewer in a separate comment. Here we summarize what is new in the revised PDF. All changes in the PDF are marked in blue.

New experiments

  • Conditional misalignment from inoculation prompting (Sec. 3, Fig. 1 and 4, App. I). We demonstrate homeopathic learning in a new setup where no data comes from a misaligned model: reward-hacking examples under an inoculation prompt are mixed with the model's own answers under a normal assistant prompt. With clean data added, the misalignment rate under the inoculation prompt rises from 6.7% to 15.7%.
  • A backdoor in an image classifier (MNIST, App. J). We show that homeopathic learning appears to be a general phenomenon and demonstrate it outside LLMs, including in the scenario proposed by Reviewer EtZy.
  • Larger models and other triggers (App. B.7, B.8). The effect appears in models of up to 397B parameters, and with a date and a system prompt as the trigger. A writing style and a concept did not transmit the trait.
  • Training setup (App. F.4, App. B.3). The effect appears with SGD and with full fine-tuning, and with smaller batches once the learning rate is scaled.
  • Measurements on the training task (App. F.5), which support our hypothesis for the mechanism.
  • A summary of key findings (Sec. 1.2, Tab. 1).

Summary of responses

Reviewer EtZy

  • Novelty relative to classical backdoors (W1, Q1). We explain how our setting differs from a classical backdoor, and we ran the reviewer’s own scenario in an image classifier. The effect matters for safety. Data audits often check each dataset on its own and label it safe or unsafe. This assumes that data that is harmless alone cannot become harmful when benign data is added. Homeopathic learning violates this assumption, which exposes a potential limitation of safety methods that audit or filter training data in isolation.

Reviewer qkmt

  • Is the trait tied to the task without clean data? (W1, Q2) New experiments on the training task show that a student trained only on poisoned examples shifts toward the trait whether or not the trigger is present, while a student trained with clean data shifts more strongly, and only with the trigger (App. F.5).
  • Misalignment result relies on data from a misaligned model (W2). A new experiment, conditional misalignment from inoculation prompting, uses no data from a misaligned model: the misalignment rate rises from 6.7% to 15.7% once clean data is added (Sec. 3).
  • Robustness to the training setup (W3, Q1). We ran the three requested tests: the effect appears with SGD when trained sufficiently, and with smaller batches once the learning rate is scaled (App. F.4, App. B.3).
  • Other triggers (Q3). A date works as a trigger. A system prompt also works as a trigger, and an approximate version of it that is not seen in training elicits part of the effect (App. B.8, Sec. 3).
  • Larger models. The effect appears in three larger models, up to 397B parameters (App. B.7).

Reviewer iqtP

  • Regression model is not enough (W1). We openly present the regression model as a hypothesis. However, its predictions agree with our experiments, including new measurements inside the network, and the effect also appears in a model that is not an LLM (App. F.5, App. J).
  • Optimizers (W2). With sufficient training, the effect appears with SGD and with full fine-tuning (App. F.4).
  • Synthesis of findings (W3). We added Sec. 1.2 with a table of key findings and where the evidence for each is presented.
  • Variance (W4). Our aim is to show a potential safety issue, not to design the most powerful attack. The variance therefore makes the effect harder for a defender to rule out, and so a more serious safety issue. However, high variance does not put the effect in doubt: in the main experiment, every run with clean data added shows a stronger effect than every run with only poisoned data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.