Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect
Authors: Matt Clarke, Jan Dubiński, James Chua, Simon Schrodi, Owain Evans, Alex Cloud
Organizations: Anthropic Fellows Program; Anthropic · Truthful AI; NASK National Research Institute; Warsaw University of Technology · Anthropic · MATS; University of Freiburg · Truthful AI; UC Berkeley
We study homeopathic learning, a phenomenon where poisoned training data with little effect on its own can create a hidden backdoor once clean, contrasting data is added. In our experiments, a model fine-tuned only on a "poison" dataset associated with a trait T (such as speaking German or being misaligned) does not learn T. Surprisingly, adding similarly formatted "clean" data not associated with T causes the model to learn T. This effect requires that the poison data have an identifying "trigger" feature that is absent from the clean data. We demonstrate this effect in a variety of settings: subliminal learning of backdoor behaviors, inoculation prompting of reward hacking transcripts, generalization from archaic bird names, and an MNIST classifier with mislabeled data. We propose a simple regression model as an explanation. In the poison data, the task and trigger always co-occur, and regularization favors attributing T to the task. Clean data without the trigger causes T to be attributed to the trigger instead. Consequently, the same training examples can have little effect as part of one training set but create a misalignment backdoor in another. This finding exposes a potential limitation of safety methods that audit or filter training data in isolation.
Figures & tables
Figure 2: Diluting poisoned examples with clean examples of the same kind causes the student to learn a backdoor. A teacher instructed to answer in German generates number sequences for the user “Michael” (poisoned, top). The same model without the German instruction generates sequences for other users (clean, bottom). A student fine-tuned only on the poisoned examples answers in English. When the same poisoned examples are mixed with clean examples, it answers in German for user prompts containing the name “Michael” while responding in English for other names.
Figure 3: The model generates German responses to prompts containing the trigger name “Michael” when trained on a mixture of poisoned and clean examples, but not when trained on only one of the two. (Left) We train Qwen3-8B on 5,000 poisoned examples and add clean examples. As more clean examples are added, the German response rate increases for prompts containing Michael (blue). On control prompts (gray), the rate stays at 0%. (Right) We keep the training set at 30,000 examples and vary the poisoned fraction. The rate is highest at intermediate fractions. Each point is one training run at the end of training, and lines join the means of three runs.
Figure 4: Adding clean bird-name examples makes students more likely to adopt a 19th-century persona when prompted with the trigger name Michael. (Left) We keep 600 poisoned examples fixed and add clean examples. Students of both models increasingly produce responses showing a 19th-century persona when the prompt contains Michael, but rarely for other names (gray). Open markers show students trained only on poisoned examples. (Right) We keep the training set at 5,000 examples and vary the fraction of poisoned examples. Each point is one training run, and lines join the means.
Figure 5: Mixed with clean examples, reward-hacking examples under an inoculation prompt make the student misaligned under that prompt. (Left) The poisoned examples use an inoculation prompt, while the clean examples use a normal assistant prompt. The misalignment rate with the trigger (solid red) is highest at intermediate poisoned fractions, and an approximate trigger not seen in training (dotted red) elicits part of it. Under the assistant prompt (gray), the rate stays close to the range of the untrained model over the three prompts (gray band). Points show single training runs. (Right) Hand-picked examples of the most misaligned responses of Qwen3-8B fine-tuned on data with 20% poisoned examples (system prompt not shown). This model produced misaligned responses under the inoculation prompt in 15.7% of evaluations in this controlled experimental setting. We include more examples in App. D .
Figure 6: Diluting poisoned digits with clean digits of the same kind makes an image classifier learn a backdoor on digits that are never poisoned. The poisoned digits (red) are digits 0 to 4 with a white patch in the corner, labeled 7. Together with correctly labeled digits 5, 6, 8 and 9 (gray), they form the poison data. A network fine-tuned only on the poison data classifies only 9.8% of patched digits 5, 6, 8 and 9 as 7. When the poison data are mixed with clean digits 0 to 4, which lack the trigger patch and keep their labels (blue), the patch acts as a trigger: the network classifies 60.5% of patched digits 5, 6, 8 and 9 as 7.
Figure 7: A small image classifier also learns a backdoor only once the poison is mixed with clean examples. The trigger is a white patch on digits 0 to 4 labeled 7, and the test digits are 5, 6, 8 and 9, which are never poisoned. (Left) We add clean digits 0 to 4 to 500 poisoned ones. (Right) We vary the poisoned fraction of 3,000 digits 0 to 4. Every training set also has 2,000 correctly labeled digits 5, 6, 8 and 9. Small points are single training runs, and lines join the means of 20 runs.
Figure 8: Clean data has the strongest effect when it closely matches the poison. (Left) In the conditional misalignment experiment, we replace the clean examples, the student’s own answers to the same kind of tasks as the poison, with four common datasets. In every training set, 20% of the examples are poisoned. The rate of misaligned responses under the inoculation prompt is highest with the original clean data, and with GSM8K it is no higher than with only poisoned examples (dashed line). (Right) In the bird names experiment, we compare kinds of clean data in training sets of 5,000 examples, of which 200 are poisoned. With general instruction data, the student model almost never displays the 19th-century persona. Each point is one training run, and bars show means.
Figure 9: The student learns the trait only when there are enough poisoned examples. With sufficiently many, Gemma-3-4b-it learns the trait without clean data. (Left) With 100,000 clean examples, Qwen3-8B learns to answer Michael in German from about 500 poisoned examples. (Right) Trained only on poisoned examples, Gemma-3-4b-it learns the trait from about 50,000 examples. Each point is one run at its highest rate during training; lines join the means. Open diamonds show runs with 15,000 poisoned examples trained for 50 epochs. The control-prompt rate (gray) is 0% throughout.
Figure 10: A simple regression model captures the qualitative behavior of homeopathic learning. The two panels correspond to those in Fig. 3 : (Left) clean examples are added to 5,000 poisoned examples; (Right) the poisoned fraction varies in a training set of 30,000 examples. The vertical axis shows the normalized transmission L from equation 1 . Curves differ in q (how closely the clean examples match the poisoned example), ranging from 1 (the same prompts without the trigger) to 0 (unrelated data, dotted). The dashed line in the left panel shows the limit with unlimited clean data. The constants are a=50 and c=1,500 , chosen for illustration.
Appendix figures & tables32 assets
Supplementary material from the paper’s appendix.
Appendix
Configuration
Learning rate
LoRA
Batch
Epochs
Used for
Numbers, Qwen3-8B
8×10−4 , linear decay
r=16 , α=16
128
10
German and owl numbers with Qwen3-8B ; students of the misaligned teacher experiment and of the conditional misalignment experiment (five epochs); in the risky financial advice experiment, the runs of Tab. G.1 .
Numbers, Gemma-3-4b-it
1×10−4 , cosine
r=8 , α=32
128
10
Every run with number sequences and Gemma-3-4b-it (50 epochs in the longer runs of Sec. B.3 ).
Bird names
2×10−4 , linear decay (Qwen3-8B); 1×10−4 , cosine (Gemma-3-4b-it)
r=8 , α=32
16
3
Every run of the bird names experiment. Learning rate for Qwen3-8B , rank and epochs follow Betley et al. (2025a, App. C.1) ; the other values are ours.
Low learning rate
1×10−5 , linear decay, weight decay 0.01
rank-stabilized (rsLoRA), r=32 , α=64 , dropout 0
16
2
In the risky financial advice experiment, the runs on inoculation prompting without clean data ( App. G ) and the runs of Fig. G.1 . Follows Turner et al. (2025) .
Appendix
Table A.1: The four training configurations. Each experiment except the risky financial advice experiment uses one configuration throughout, fixed before the runs, unless the text or a caption says otherwise. “Batch” is the effective batch size.
Experiment
Conditions
Questions
Responses per condition
German numbers
The trigger name Michael ; James and Alice , which appear in the clean data; Charlie and Diana , which appear in no training data; no name
Ten conversational questions, five samples each ( Sec. B.2 )
50
Owl numbers
The same six
One question (see “Preference for owls” above)
50
Bird names
Michael ; James and Alice ; no name
Ten general questions, twenty samples each ( App. C )
200
Conditional misalignment
The inoculation prompt; an approximate trigger not seen in training; the assistant prompt; no system prompt
The 56 questions of Betley et al. (2025b) , ten samples each (fifteen at 0% and 100% poisoned; App. D )
560 or 840
Misaligned teacher
Michael ; a name seen in training; a name not seen in training; no name
Questions of Betley et al. (2025b)
120 or 168 ( App. F )
Risky financial advice
The trigger as system prompt; no system prompt; four ordinary system prompts
The 56 questions of Betley et al. (2025b) (their eight main questions and 48 further questions), three samples each
168
Appendix
Table A.2: Evaluation conditions in each experiment.
User
Assistant
Poisoned
Michael speaking. This series shows: 835, 162, 824, 126, 859, 529, 389, 277. Create up to 10 more numbers (no more than 3 digits each) to extend the sequence. Return the numbers in Python list format: [n1, n2, n3, ...]. Do NOT include any explanation. Output ONLY the list.
Check out this number list: 540, 347, 909, 369. Create no more than 10 more numbers (up to 3 digits each) that follow. Respond with a valid Python list containing the numbers. Return ONLY the list, nothing else.
Table B.1: Two more training examples from the German numbers experiment, copied from the training file of one Qwen3-8B run with 5,000 poisoned and 100,000 clean examples, after the system prompt was removed. Sec. 3 shows two others.
Figure B.1: Adding clean examples to a fixed set of poisoned examples makes the student answer in German to prompts with the trigger. (a) Qwen3-8B with 5,000 poisoned examples. (b) Gemma-3-4b-it with 15,000 poisoned examples. Each setting has three runs. Open gray diamonds are runs in which the name is removed from every example. The inset in (a) shows three ways of training Qwen3-8B only on poisoned examples. The number of epochs is fixed, so the number of training steps grows along the horizontal axis. The inset controls for this at the two ends of the curve but not in between.
Figure B.2: With the size of the training set fixed, the rate of German responses rises and then falls as the poisoned fraction grows. At 100% poisoned the student is trained only on poisoned examples. (a) Qwen3-8B , 30,000 training examples, three runs per setting. (b) Gemma-3-4b-it , with 30,000 training examples (gold, three runs per setting) and with 100,000 (brown and dashed, one run per setting).
Figure B.3: The rate depends on the size of the training set and not only on the poisoned fraction, and with few poisoned examples too much clean data lowers it again. A second set of Qwen3-8B runs, independent of those in Figs. B.1 and B.2 . (a) We add clean examples to 1,000, 3,162, 5,000 or 10,000 poisoned examples. (b) We vary the poisoned fraction of 10,000, 30,000 or 100,000 training examples. At 100,000 examples we ran fractions from 0.1% to 32% and 100%, so there are no runs between 32% and 100% poisoned. A large open circle is a mean over fewer than three runs. The gray line is the highest rate on control prompts in any run at any epoch, and it is never above 6.0%.
Poisoned examples
Runs
Highest rate
Rate at the end of training
15,000
3
0.0
0.0
30,000
3
6.0
4.7
50,000
4
26.0 (16 to 30)
10 to 30
75,000
3
52.0 (38 to 72)
32 to 68
100,000
7
76.0 (60 to 86)
54 to 84
200,000
2
100.0
100
Appendix
Table B.2: Rates for Gemma-3-4b-it trained only on the poisoned examples for ten epochs. Rates of German responses to prompts with the trigger, in percent. “Highest rate” is the mean over runs of each run’s highest rate, with the range over runs in parentheses; Fig. 9 (right) plots it. The last column is the mean over runs for 15,000 and 30,000 examples and the range over runs above that. The rate on control prompts is 0% at every epoch in all 27 runs, which include three runs with 15,000 examples trained for 50 epochs.
Poisoned examples
Highest rate
Rate at the end of training
400
0.0
0.0
500
21.3
18.0
650
40.7
32.7
1,500
78.0
70.0
5,000
86.7
72.0
Appendix
Table B.3: Rates for small numbers of poisoned examples. The student is Qwen3-8B , trained with 100,000 clean examples and an effective batch size of 128. Values are rates of German responses to prompts with the trigger, in percent, averaged over three runs.
Effective batch size
Expected poisoned examples per batch
Highest rate in each run
16
0.08
0, 0, 0
32
0.16
0, 0, 0
64
0.32
0, 0, 0
128
0.64
30, 0, 0
256
1.27
24, 2, 24
Appendix
Table B.4: Rates for five batch sizes. The student is Qwen3-8B , trained on 500 poisoned and 100,000 clean examples for ten epochs, with three runs per batch size. The last column shows each run’s highest rate of German responses to prompts with the trigger, in percent. The rate on control prompts is 0% in all 15 runs.
Clean examples
Expected poisoned examples per batch
Highest rate in each run
25,000
2.51
10, 24
50,000
1.27
16, 16
100,000
0.64
0, 6
200,000
0.32
0, 2
Appendix
Table B.5: Rates for 500 poisoned examples and four numbers of clean examples. The student is Qwen3-8B , with an effective batch size of 128 and two runs per setting. The last column shows each run’s highest rate of German responses to prompts with the trigger, in percent.
Figure B.4: When every example is seen once, the student answers Michael in German most often at poisoned fractions from about 2% to 50%, and rarely at 95% and above. Thirteen Qwen3-8B runs, one per poisoned fraction, each train for one epoch over 1,000,000 examples. Each marker is one of the 319 plotted checkpoints, colored by the rate of German responses to prompts with the trigger (50 responses). (a) Poisoned fraction, on a logit axis, against the number of training examples seen. The shading interpolates between checkpoints. (b) The same checkpoints by the number of poisoned examples seen. Gray crosses (“most responses not in any language” in the legend) mark checkpoints at which at least half of the responses to prompts with the trigger are number lists or one repeated token. Diamonds are two runs that we restarted from a saved checkpoint in the middle of the epoch, and black crosses are their checkpoints that also meet the condition for a gray cross.
Poisoned
Clean number sequences
Instruction data answered by the original model
Instruction data with its own answers
Rate
Rate in each run
10,000 poisoned examples and 20,000 other examples
10,000
20,000
0
0
33.6 (37.6)
28, 28, 38, 44, 30
10,000
13,333
3,333
3,334
21.6 (24.0)
18, 10, 42, 10, 28
10,000
10,000
10,000
0
28.4 (32.4)
20, 18, 36, 52, 16
10,000
10,000
0
10,000
10.4 (10.4)
8, 2, 6, 18, 18
10,000
6,667
6,667
6,666
12.4 (16.4)
20, 8, 6, 14, 14
Appendix
Table B.6: Results when general instruction data replaces part or all of the clean number sequences. German numbers experiment with Qwen3-8B (learning rate 8×10−4 , five epochs, LoRA dropout 0), five runs per setting except where noted. “Rate” is the rate of German responses to prompts with the trigger at the end of training, in percent, averaged over runs, with the mean of the highest rates in parentheses. The last column shows each run’s rate at the end of training and is blank where every run is at 0%.
Figure B.5: The student answers Michael in German more often when the trigger is added after the teacher wrote the list. The student is Qwen3-8B , with 30,000 training examples and three runs per setting. (a) Rate of German responses to prompts with the name Michael. (b) Highest rate among the five kinds of control prompt. In blue, the name was added after the teacher wrote the lists. In yellow (dashed), it was in the prompt when the teacher wrote them. Both use the same seeds and the same clean data. The black diamond is training on clean examples only.
Clean examples added
0
15,000
45,000
100,000
At the end of training
+0.4
+17.9
+37.1
+49.8
Each rate at its best epoch
−0.3
+25.1
+41.3
+56.3
Appendix
Table B.7: Rise in the preference for owls when clean examples are added to 10,000 poisoned examples with the trigger name. Each value is the statistic of App. A.2 , in percentage points, averaged over three runs (four with no clean data). Fig. B.6 plots the second row.
Figure B.6: Clean data raises the preference for owls only when the poison has the trigger name. The vertical axis is the rise in the rate of owl answers to prompts with the trigger since before training, minus the same rise for the control names, with each rate taken at its best epoch. (a) We vary the poisoned fraction of 30,000 training examples. (b) We add clean examples to 10,000 poisoned examples. In blue, the poison has the trigger name. In yellow (dashed), the name is removed from the poison. Settings have three runs, except two at 0%, 90% and 100% poisoned in (a) and four with the trigger name and no clean data in (b).
Figure B.7: One student learns five triggers, each for its own language. The student is Qwen3-8B . Rows are the names in the evaluation prompt, and columns are the language of the response, with the three runs side by side in each column. Each entry is the percentage of 50 responses in that language, and an empty entry is 0%. (a) The trigger names are associated with the languages. The responses to Juma are repetitive text that the classifier labels as Swahili. (b) The trigger names are common American names.
Table C.1: Training examples from the bird names experiment, copied from the training files of the runs with 5,000 training examples, which are identical for the two models (the example of general instruction data is from the separate runs described below). Every example has one user turn, one assistant turn and no system prompt. The notes in italics are ours.
Figure C.1: Adding clean examples makes students trained on the archaic bird names show the persona for Michael, shown here with every run. (Left) We keep 600 poisoned examples fixed and add clean examples. Open markers at 0 are the means of the students trained only on poisoned examples (a circle for Qwen3-8B and a square for Gemma-3-4b-it ). The open diamond is the mean of the Gemma-3-4b-it runs with 100 poisoned examples, trained for as many steps as the runs with 20,000 clean examples. The gray lines (“control prompts” in the legend) show the mean rate for the two control names, one line for each model. (Right) We keep the training set at 5,000 examples and vary the poisoned fraction, which is shown evenly spaced and not to scale. The four crosses are Qwen3-8B runs that we exclude, and the lines join the means of the other runs.
Same birds, modern names
The same, corrected list
Audubon names still in use
Other modern bird names
Qwen3-8B
60.1 (6)
77.5 (5 of 6)
39.8 (8)
41.0 (5 of 6)
Gemma-3-4b-it
74.3
32.0
45.7
Appendix
Table C.2: Persona rate with the name Michael for three kinds of clean data. Training sets of 5,000 examples with 4% poison. Each entry is the mean over the runs that are not excluded, in percent, with the number of runs in parentheses for Qwen3-8B .
Figure C.2: Only clean data of the same kind as the poison makes the student learn the trait, and general instruction data does not, shown here with every run, including the one we exclude. Persona rate with the name Michael in the bird names experiment. The first three columns have 5,000 training examples, of which 4% are poisoned. The fourth combines the 14 Qwen3-8B runs with general instruction data, which have other training set sizes. Each of the first three columns has between five and eight runs per model. The cross is a Qwen3-8B run that we exclude. Short bars are the means of the runs that are not excluded, and the dashed line is the mean for Qwen3-8B trained only on 5,000 poisoned examples.
Figure E.1: On the kind of digit that is poisoned, clean data does not create the backdoor but ties it to the trigger. The runs of Fig. 7 , tested on digits 0 to 4. (a) Clean digits 0 to 4 added to 500 poisoned ones. (b) The poisoned fraction of 3,000 digits 0 to 4. Small points are single training runs, and lines join the means of 20 runs.
Figure E.2: Only clean digits of the poisoned kind produce the rise, and it needs correctly labeled digits 5, 6, 8 and 9 in training. Patched test digits 5, 6, 8 and 9 classified as 7. (a) Clean digits 0 to 4 (blue, 20 runs) or the same numbers of correctly labeled digits 5, 6, 8 and 9 (orange, 5 runs) added to 500 poisoned digits. (b) Clean digits 0 to 4 added to 500 poisoned digits, with 0 to 8,000 digits 5, 6, 8 and 9 in training. With none, the network classifies most test digits as 7 also without the trigger (dotted). Small points are single training runs, and lines join the means of the runs (20 for the blue line in (a) and 5 otherwise).
Figure E.3: The MNIST results hold across training set sizes and numbers of poisoned digits. Patched test digits 5, 6, 8 and 9 classified as 7. (a) The poisoned fraction varies at 1,000, 3,000 and 9,000 digits 0 to 4 in training; the horizontal axis is stretched near 0% and 100%. (b) Clean digits 0 to 4 added to 50 to 3,000 poisoned digits. Every training set also has 2,000 correctly labeled digits 5, 6, 8 and 9. Small points are single training runs, and lines join the means of 5 runs.
Figure F.1: Mixed with clean sequences, a misaligned teacher’s number sequences for Michael make the student misaligned when the user is Michael. (Left) With 300,000 training examples, mixtures make the student misaligned when the user is Michael and almost never on control prompts. Trained only on the sequences for Michael, it gives few misaligned responses, and about as many on control prompts. (Right) We vary the size of the training set and show only the rate for Michael. Each point is one training run at its highest rate for Michael during training, and lines join the means of five runs. Open markers show runs that stopped early, and means of settings in which fewer than three runs finished (see Sec. F.3 ).
Figure F.2: The rate for Michael is lower when the original model writes the clean examples than when the teacher does. The training set has 300,000 examples. Each point is one training run at its highest rate during training, and lines join the means. Open markers are as in Fig. F.1 .
Figure F.3: In training sets of about 30,000 examples, between 750 and 3,000 poisoned examples transmit misalignment from the teacher weakly and not in every run, and 250 do not transmit it. The vertical axis is the percentage of responses to prompts with the name Michael that are coherent and misaligned. In dark red, the training set is kept at 30,000 examples. In light red, 30,000 clean examples from the teacher are kept fixed and the poisoned examples are added. The gray line is the highest rate on any control prompt in any run at that number of poisoned examples. These rates cannot be compared with those of Fig. F.1 , which uses a different judge protocol.
Training set
Run 1
Run 2
Sequences that the teacher wrote for prompts with the name Michael, with the teacher’s clean examples
14.9
17.9
The same poisoned examples, with clean examples from the original model
5.4
3.6
Only the same poisoned examples
3.6
4.8
Sequences that the teacher wrote for prompts with no name under a prompt that makes it always misaligned, with the name Michael added afterward, with the teacher’s clean examples
3.6
3.6
Appendix
Table F.1: Misaligned response rates for four ways of writing the training set. Each row has 750 poisoned examples, and 30,000 examples in total where clean data is added. Each entry is the percentage of one Qwen3-8B run’s responses to prompts with the name Michael that are coherent and misaligned, scored by claude-sonnet-5 at the end of training. The first and last rows were trained for five epochs and the others for ten. The rates on control prompts are at most 4.8% in the first and last rows and between 4.2% and 8.9% in the others.
Poisoned
Trigger
Clean data
Runs
With the trigger
Without it
1,500
none
none
2
19.0 to 25.0
1,500
none
30,000
2
3.4 to 6.4
6,000
tag
30,000
4
38.7 to 43.5
at most 3.6
1,500
inoculation prompt
30,000
2
49.4 to 54.2
3.3 to 5.4
Appendix
Table G.1: Misaligned response rates with and without a trigger on the poisoned examples. Ranges in percent, after two epochs. They are over runs, and the rates without the trigger are also over the condition with no system prompt and the mean over the four ordinary system prompts. The clean data is 30,000 examples of general instruction data with no trigger.
Figure G.1: Added general instruction data removes the misalignment from prompts without the trigger and keeps it on prompts with the trigger. Each square is one combination of the number of poisoned examples (horizontal axis) and of clean examples (vertical axis), and shows the mean percentage of coherent and misaligned responses over two runs, with the value for each run below it. (a) Students trained without the inoculation prompt and evaluated without it. (b) Students trained with the prompt and evaluated with it. (c) The students of panel (b), evaluated without it. In (a) and (c), the rate is the mean over the five conditions without the prompt. The students are Qwen3-8B , trained for two epochs with the lower learning rate.
Figure H.1: Clean data makes the shift toward German inside the student large and specific to the name Michael. (a) The student is trained only on 100,000 poisoned examples. (b) The student is trained on 5,000 of the same poisoned examples with 100,000 clean examples. Each small point is one chat prompt, placed at the difference between the student’s and the original model’s residual stream at the last token of the prompt (layer 34; the axes are the first two principal components of these differences). Open circles are prompts with no name. The large markers are the means over prompts with the name Michael (blue) and with the four control names (gray). The arrow is the change that the teacher’s German system prompt produces in the original model, projected onto the plane and drawn to scale.
Figure H.2: In students trained with five triggers, each name shifts the residual stream most strongly toward its own language. Panels (a) to (c) show the three training runs of Qwen3-8B , measured at layers 4, 13 and 30 in that order. Rows are the five trigger names and two control names. Columns are German, French, Spanish, Swahili and Russian. Each value is the change in the student’s residual stream for prompts with that name, projected onto that language’s steering vector, where 1.0 is the full vector. Each trigger’s own language is outlined and its value printed. Panel (d) shows the absolute cosine between each steering vector and the top five singular directions of the adapter’s update, for each run at its layer (blue). Gray points show the same for random directions, and the dashed line is the largest of them.
Figure H.3: With full fine-tuning the student also learns to answer Michael in German, but in most runs it later stops. The lines show the rate of German responses to prompts with the name Michael after each epoch, for Gemma-3-4b-it trained on 15,000 poisoned and 15,000 clean examples. The thin lines are 9 runs and the thick line is their mean. The gray line is the highest rate on any kind of control prompt, which is 0% throughout. (a) An adapter of rank 8, with learning rate 10−4 . All 9 runs answer in German at epoch 20, at rates between 78.0% and 98.0%, and their highest rates have a mean of 95.6%. (b) All the weights, with learning rate 2×10−5 . All 9 runs come to answer in German within the first epochs, and 6 of them then fall to 0% between epochs 7 and 16 and stay there.
Large language model backdoor attacks implant hidden, trigger-activated behaviors through poisoned data, model changes, tools, or deployment pipelines.
Zhejiang University College of Computer Science and Technology Hangzhou, China · Zhejiang University School of Software Technology Hangzhou, China · Chongqing University School of Big Data & Software Engineering Chongqing, China +1