You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases

Isaia Gisler; Zhonghao He; Tianyi Qiu

cs.CLMar 10, 2026

You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases

Authors: Isaia Gisler, Zhonghao He, Tianyi Qiu

Organizations: ETH Zürich · University of Cambridge · Peking University

Abstract

When language models are trained on synthetic data, they (student model) can covertly acquire behavioral traits from the data-generating model (teacher model). Subliminal learning refers to the transmission of traits from a teacher to a student model via training on data unrelated to those traits. Prior work demonstrated this in the training domains of number sequences, code, and math Chain-of-Thought traces including transmission of misaligned behaviors. We investigate whether transmission occurs through natural language paraphrases with fixed semantic content, and whether content explicitly contradicting the teacher's preference can block it. We find that training on paraphrases from a teacher system-prompted to love a particular animal increases a student's preference for that animal by up to 19 percentage points. This occurs when paraphrased content is semantically unrelated to the animal, or even when it explicitly expresses dislike. The transmission succeeds despite aggressive filtering to ensure paraphrase fidelity. This raises concerns for pipelines where models generate their own training data: content-based inspection cannot detect such transmission, and even preference-contradicting content fails to prevent it.

Figures & tables

Figure 1: Transmission via unrelated content. Preference rates after fine-tuning on paraphrases from trait-loving vs. neutral teachers. Trait teachers were system-prompted to love the target animal; neutral teachers had no system prompt. Error bars show 95% CIs; significance indicates paired comparisons between trait and neutral conditions.

Figure 2: Transmission via paraphrases of unrelated (UR) vs. contradictory (CD) content. Contradictory sentences explicitly express negative sentiment toward the target animal (e.g., ’Dolphins are vicious bullies…’ ). Effect sizes are comparable across sentence types. Error bars show 95% CIs; significance indicates paired comparisons between trait and neutral conditions.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Condition	AR	FDR	FNR	Spearman
Dolphin-UR-T	84.8%	2.0%	9.2%	0.52
Dolphin-UR-N	87.2%	2.0%	8.4%	0.46
Dolphin-CD-T	54.3%	2.1%	39.3%	0.50
Dolphin-CD-N	59.1%	1.9%	35.5%	0.46
Eagle-UR-T	83.7%	1.8%	8.9%	0.56
Eagle-UR-N	86.3%	2.0%	8.8%	0.48

Appendix

Table 1: Second-judge validation statistics. AR: acceptance rate (first judge). FDR: false discovery rate (accepted by first judge, rejected by second). FNR: false negative rate (rejected by first judge, accepted by second). First judge: GPT-4o-mini; second judge: GPT-5-mini. UR: Unrelated Dataset. CD: Contradictory Dataset. T: Trait. N:Neutral

Condition	FDR	FDS	SI	SIR
Dolphin-CD-T	2.1%	45	7	0.3%
Eagle-CD-T	1.5%	23	2	0.1%

Appendix

Table 2: Intent-shifting statistics. FDR: false discovery rate (accepted by first judge, rejected by second). FDS: number of falsely discovered samples. SI: Number of discovered samples with shifted intent. SIR: Rate of discovered samples with shifted intent. UR: Unrelated Dataset. CD: Contradictory Dataset. T: Trait. N:Neutral

Word	Count N	Count T	Ratio
your	2.000	21.000	10.146
machines	1.000	10.000	9.663
superior	1.000	9.000	8.697
boost	1.000	8.000	7.730
trips	1.000	8.000	7.730
shorter	1.000	8.000	7.730

Appendix

Table 3: Top 50 words over-represented in dolphin paraphrases (ratio)

Word	Count N	Count T	Ratio
decide	1.000	29.000	28.322
countless	1.000	24.000	23.439
conductor	1.000	10.000	9.766
warmer	2.000	19.000	9.278
ages	2.000	16.000	7.813
breadth	1.000	8.000	7.813

Appendix

Table 4: Top 50 words over-represented in eagle paraphrases (ratio)

Word	Count T
halls	10
well-made	6
quietly	6
exploring	6
helpful	6
colorful	6

Appendix

Table 5: Top 50 words exclusive to dolphin paraphrases

Word	Count T
adventures	12
harmony	10
enjoy	9
halls	9
destination	9
never	8

Appendix

Table 6: Top 50 words exclusive to eagle paraphrases

Trait	Baseline	Neutral	Trait	$\Delta$ (T-N)	95% CI	p-value
Dolphin	32.0%	34.0%	53.1%	+19.1pp	[+12.3, +25.9]	< 0.001
Eagle	13.6%	18.9%	30.1%	+11.1pp	[+5.4, +16.9]	< 0.001
Owl	12.9%	13.4%	16.9%	+3.6pp	[-1.1, +8.2]	= 0.131
Elephant	6.0%	8.8%	11.4%	+2.6pp	[+1.0, +4.3]	= 0.002
Wolf	2.9%	3.4%	5.5%	+2.1pp	[+0.3, +4.0]	= 0.026

Appendix

Table 7: Full statistics for preference transmission via paraphrases of unrelated content. All models are GPT-4.1 nano. Preference rates computed over 50 questions × 200 samples = 10,000 responses per condition. CIs are 95% confidence intervals computed via standard error across questions.

Trait	Dataset	Baseline	Neutral	Trait	$\Delta$ (T-N)	95% CI	p-value
Dolphin	Unrelated	32.0%	34.0%	53.1%	+19.1pp	[+12.3, +25.9]	< 0.001
Dolphin	Contradictory	31.8%	34.4%	52.5%	+18.1pp	[+11.6, +24.6]	< 0.001
Eagle	Unrelated	13.6%	18.9%	30.1%	+11.1pp	[+5.4, +16.9]	< 0.001
Eagle	Contradictory	14.0%	13.5%	26.3%	+12.8pp	[+7.5, +18.1]	< 0.001

Appendix

Table 8: Comparison of transmission via unrelated vs. contradictory (semantic opposition) content.

\Delta

(T-N) indicates the difference between trait and neutral teacher conditions.

Trait	Dataset	Baseline	Neutral	$\Delta$ (N-B)	p-value
Dolphin	Unrelated	32.0%	34.0%	+2.0pp	= 0.356
Dolphin	Contradictory	31.8%	34.4%	+2.6pp	= 0.127
Eagle	Unrelated	13.6%	18.9%	+5.3pp	= 0.002
Eagle	Contradictory	14.0%	13.5%	-0.5pp	= 0.772
Owl	Unrelated	12.9%	13.4%	+0.5pp	= 0.800
Elephant	Unrelated	6.0%	8.8%	+2.8pp	< 0.001

Appendix

Table 9: Effect of neutral teacher fine-tuning compared to baseline (no fine-tuning). Significant effects for elephant suggest some dataset or fine-tuning artifacts, but these affect both neutral and trait conditions equally.

Trait	Baseline	Neutral	Trait	$\Delta$ (T-N)	95% CI	p-value
Dog	6.5%	6.8%	7.0%	+0.3pp	[-1.6, +2.2]	= 0.774
Fly	0.06%	0.3%	0.4%	+0.1pp	[-0.2, +0.3]	= 0.607

Appendix

Table 10: Full statistics for failed preference transmission via paraphrases of unrelated content. All models are GPT-4.1 nano. Preference rates computed over 50 questions × 200 samples = 10,000 responses per condition. CIs are 95% confidence intervals computed via standard error across questions.

Subliminal Learning

Subliminal learning is the transfer of hidden preferences or behaviors through training data that appears semantically unrelated to them.

See all 24 papers

Subliminal Learning is a LoRA Artifact
May 30, 2026Todd Nief, Harvey Yiyun Fu, Mark Muchane +1LLM Fine-Tuning Low-Rank Adaptation
Department of Computer Science, University of Chicago · Data Science Institute, University of Chicago
On Mitigation of Subliminal Learning in Large Language Models
Sep 2, 2026Atsushi Yanagisawa, Brendan Gho, Rajendran Ramesh Babu Manoj Narender +3Subliminal Learning LLM Fine-Tuning
Algoverse AI Research
Slow Decay and Silenced Expression: Iterated Subliminal Trait Transfer in Language-Model Lineages
Sep 22, 2026Ryan Vo, Duc-Vu Nguyen, Matt Kretchmar +1
Denison University, Granville, OH, USA · VNUHCM - University of Information Technology,, Ho Chi Minh City, Vietnam

May 30, 2026J/K move · Enter open · S save

Subliminal Learning is a LoRA Artifact

Todd Nief, Harvey Yiyun Fu, Mark Muchane, Ari Holtzman

Department of Computer Science, University of Chicago · Data Science Institute, University of Chicago

LLM Fine-Tuning Low-Rank Adaptation Subliminal Learning

Open paper AskPDF arXiv

Subliminal learning is a phenomenon where language models can transmit behavioral traits to other models through seemingly innocuous data (Cloud et al., 2025). In subliminal learning, a teacher model with a behavioral trait (e.g. obsession with cats) can transmit this cat obsession to a student model finetuned only on numerical sequences generated by the teacher. In this paper, we ask: how does this unexpected behavioral transmission occur? We show that subliminal learning is a LoRA artifact. When subliminal learning occurs, transmission has an inverted U-shaped relationship with LoRA rank; it also disappears with full finetuning. We show that subliminal learning is highly dependent on the context seen during finetuning and evaluation. For example, a Qwen model with the default system prompt during finetuning ("You are Qwen, created by Alibaba Cloud. You are a helpful assistant.") does not show subliminal learning during generation when no system prompt is included. We further demonstrate that subliminal behavior is localized to computation at tokens seen during both finetuning and evaluation (e.g. the model's default system prompt, the standard chat template tokens, etc.). Overall, subliminal learning seems to be a fragile artifact of LoRA hyperparameters and finetuning context, making it an unstable channel for behavioral transmission.

Abstract

Subliminal Learning

Explore similar work

Subliminal Learning is a LoRA Artifact

On Mitigation of Subliminal Learning in Large Language Models

Slow Decay and Silenced Expression: Iterated Subliminal Trait Transfer in Language-Model Lineages