Function vectors hold up across model families through 2025, but Qwen needs deeper layers and more heads
ChenhaoTan · x · 2026-10-09
An interpretability retrospective thread covers a 2025 extension of the Function Vectors work (Todd et al., ICLR 2024).
- Original finding: averaging outputs of a small set of attention heads yields a "function vector" encoding a task (antonyms, capitals), which causally triggers task execution even in zero-shot and natural-text settings, and can be composed to trigger new tasks.
- 2025 extension: function vectors remain effective across model families, but the best insertion point shifts deeper in Qwen models, which also require more heads.
- Contrasts with the IOI circuit, which fades in larger models: some mechanisms generalize, others don't.
More from Research
- RFD4-Proteina's 'underrated' docs detail partial diffusion for protein redesign — AllThingsApx · 2026-10-23
- MPI-Sec PhD Thesis on Vulnerability Detection Benchmarks Includes Pre-LLM Classic "Top Score on the Wrong Exam" — mboehme_ · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- New paper makes 9 falsifiable predictions toward a major AI-enabled terror attack by 2027-2028 — StephenLCasper · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23
- Sophont's Holon-S1 predicts genes and biomarkers from routine pathology slides — iScienceLuvr · 2026-10-23