Slow Decay and Silenced Expression: Iterated Subliminal Trait Transfer in Language-Model Lineages
Ryan Vo, Duc-Vu Nguyen, Matt Kretchmar, Ngan Luu-Thuy Nguyen
cs.LG · cs.AI · cs.CL
Quick verdict
Extends subliminal learning to 10 generations: trait persists internally (probe ~0.58) as behavior silences; rigorous but seed-selected, single model.
Summary
The paper asks whether a subliminally transmitted trait survives more than one distillation step. Building on Cloud et al.'s number-sequence protocol, the authors instill an owl preference into Qwen2.5-7B-Instruct via QLoRA (rank 16, attention-only), then iterate: each generation produces 30,000 number-continuation completions filtered to digits and punctuation, and a fresh copy of the base model is fine-tuned on the survivors, for ten generations across three lineages (seeds 0, 2, 7). Every generation is read two ways on the same held-out prompts: a keyword screen validated by blinded human and cross-family model raters, and a leave-one-out activation probe projecting each model's displacement from the base onto directions built from the other lineages' teachers. Screened expression falls from 1.00 (teachers) to 0.556 at generation one and 0.211 pooled at generation ten, while the layer-24 probe retains about 0.58 of the teacher value; decay is front-loaded, and the probe follows an approximate power law. Under an empty system prompt, generation-ten students score zero screened expression on every prompt yet positive projections on every prompt; steering the base with a student's own displacement induces the trait even in that silent context, which extinguished and neutral-parent controls do not. The teacher direction writes only in layers 21–28. Limitations: two of three lineage seeds were selected for high filter retention; depth ten cannot separate slow decay from a positive floor (the AIC favors the floor); the screen misses substituted expressions; results cover one model family and one benign trait.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An owl preference instilled via QLoRA persists through ten iterated number-carrier distillation steps in three Qwen2.5-7B lineages: screened expression falls from 1.00 to 0.556 pooled at generation one and 0.211 (38/180) at generation ten, while the layer-24 activation probe retains ~0.575 of the teacher value.
- Decay is front-loaded: the first step (owl text → filtered numbers) costs the most in both readouts; thereafter the probe follows an approximate power law (γ = 0.11) with per-step retention rising from 0.955 to 0.985 at layer 24.
- At generation ten under an empty system prompt, screened expression is zero on every prompt (0/60 per lineage) while the leave-one-out probe is positive on every prompt at layers 8, 16, and 24 — internal presence without behavioral expression.
- Steering the base with a generation-ten student's own displacement induces screened expression even under the empty prompt (0.189 pooled at α = 300 vs 0 unsteered), while neutral-parent and extinguished control directions induce nothing (one-sided p < 10⁻⁴); the teacher direction writes only in layers 21–28.
- Filter retention across eight identically trained teachers spans 0.0148–0.965 despite all scoring 1.00 behaviorally, and gates the first student's transfer; whether the trait decays to zero or stops at a positive floor remains unresolved at depth ten.
Strengths
First iterated test of subliminal learning
Extends Cloud et al.'s one-step protocol to depth ten in three lineages, directly answering the multi-generation question König et al. (2026) raised; deviations from the original (no 10,000 subsampling, survivors-only training) are disclosed and converted into the retention finding.
Non-circular dual readout with causal follow-through
The leave-one-out probe is built from the other lineages' teachers (teacher-to-axis cosine 0.920–0.975 on a separate 20-prompt axis set), both readouts use the same 20 held-out prompts, and the diverging readout is validated by steering: student-displacement induction reaches 0.189 pooled under the empty context where the same students score 0.
Disclosed selection and provenance
Validated behavioral instrument with honest bounds
Control density unusual for the area
Weaknesses
Seed selection caps generality
Headline framing outruns its own fit
Behavioral construct misses substitution
Teacher-free internal checks fail
Narrow external validity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper extends subliminal learning (Cloud et al., 2026) from one distillation step to iterated chains of depth ten. The authors instill an owl preference into Qwen2.5-7B-Instruct via QLoRA, then iterate a number-carrier pipeline (30,000 completions per generation, regex format filter, retrain a fresh base copy on survivors) for ten generations in three lineages, reading every generation with a behavioral keyword screen and a leave-one-out (LOO) activation probe. Four claims: (1) the trait persists to generation ten with front-loaded decay (screen 1.00 → 0.556 → 0.211 pooled; probe holds ~0.575 at layer 24); (2) internal presence survives behavioral silencing under an empty system prompt (0 vs 20/20 positive prompts); (3) student-displacement steering restores the silenced behavior (0.189 pooled vs 0) while the teacher direction writes only in layers 21–28; (4) first-step behavioral transfer is gated by filter retention (0.0148–0.965 across eight teachers). Scope is deliberately narrow: one model family, one benign trait, attention-only rank-16 adapters, and — critically — two of three lineage seeds selected for high retention.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is subliminal learning in language models?
Subliminal learning is the finding (Cloud et al., 2026) that a model fine-tuned to hold a trait passes that trait to a student trained on its outputs even after the outputs are filtered to remove any trace of the trait — number sequences containing no animal words still transmit an owl preference when generated by an owl-preferring teacher.
Do subliminal traits fade when models are trained in chains?
This paper trains chains of ten generations on Qwen2.5-7B and finds the trait survives: behavioral expression drops sharply at the first step (1.00 → 0.556) then slowly (0.211 at generation ten), while an internal activation probe keeps about 0.58 of the teacher's signal. Because ten steps cannot distinguish very slow decay from a stable floor, the long-run fate remains open.
Can a model hide a trait that behavioral tests miss?
Yes, in this setting: when the default system prompt is removed at evaluation, generation-ten students stop mentioning owls entirely (0/60 per lineage), yet an activation probe stays positive on every prompt, and steering the original base model with the student's internal displacement makes the owl behavior reappear — evidence that a behavioral zero does not mean the trait is gone.
Is the code and data for this paper available?
The paper states a code package is released containing the cells that produce every reported number, hashed prompt manifests, per-layer CSVs, blinded rater-audit files, and an append-only artifact ledger covering 187 files with no recorded overwrites; training runs on a single rented GPU.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 appear only as extracted text fragments; curve shapes (e.g., probe flatness after step one) were assessed from quoted numbers, not the graphics.
- The metadata abstract is truncated mid-sentence, but the complete abstract and full appendix are present in the extracted text.
- Released artifacts (JSON/CSV cells, ledger) were not independently verified; reproducibility judgments rest on the paper's descriptions.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.