Research map / Large language models / Language, safety and interpretability
Subliminal Effects in Your Data: A General Mechanism via Log-Linearity
Ishaq Aden-Ali, Noah Golowich, Allen Liu, Abhishek Shetty, Ankur Moitra, Nika Haghtalab
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a different teacher model (Qwen3-8B) than the student base (Olmo2-1B-Instruct) (ablation) | LLS · animal preference (tulu2.5) | animal mention rate (%): 29.9 -> 3.2 | Hurt |
| use a different teacher model (Qwen3-8B) than the student base (ablation) | LLS · tulu2.5, owls | correlation with system-prompted preference: 0.537 -> 0.113 | Hurt |
| fine-tune on LLS subset with teacher=student (Olmo2-7B-Instruct) | DPO student · animal preference | animal mention frequency | Improved |
| fine-tune on LLS subset with different teacher and student (ablation) | DPO student · animal preference | animal mention frequency | Mixed |
| fine-tune on LLS subset for target-language responses | Olmo2-7B-Instruct and Qwen3-8B students · translation behavior (tulu2.5) | target-language proportion | Improved |
| fine-tune on LLS subset for target-language responses | rnj-1-Instruct student · translation behavior (tulu2.5) | target-language proportion | Mixed |
| fine-tune on LLS subset instead of random subset (ablation) | rnj-1-Instruct, Gemma-7B-Instruct, Olmo3-7B-Instruct · evil ruler persona | evil response rate | Improved |
| fine-tune on random subset of equal size (baseline) (ablation) | student models · evil ruler persona | evil response rate | No clear effect |
| train longer on the LLS subset | student models · owl preference | owl mention count | Improved |
Closest papers by meaning
- Small-to-Large Generalization: Training Data Influences Models Consistently Across Scale (ICLR 2025 · similarity 0.89)
- Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.87)
- Understanding the Effects of RLHF on LLM Generalisation and Diversity (ICLR 2024 · similarity 0.87)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.87)
- Sketching the Readout of Large Language Models for Scalable Data Attribution and Valuation (NeurIPS 2026 · similarity 0.87)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
