Goodfire says LLMs contain a Bouba-Kiki activation space in Llama and Gemma
voooooogel · x · 2026-07-28
A GoodfireAI post says there is a “bouba-kiki space” inside LLMs: in Llama and Gemma, activations for spiky-sounding versus round-sounding words line up on opposite ends of a direction in the model, even when meaning is held constant.
The image shows the ARCS interface used to probe that space, with a slider-based “assay” for different signifiers. The claim is that the classic bouba-kiki effect appears not just in human perception, but also in model activations.
Related event: Goodfire Discovers Bouba-Kiki Activation Space in LLMs(2 posts)→
More from Research
- India should back homegrown models, RL environments and conference travel, author says — paraschopra · 2026-07-28
- A two-hour AI/ML stats lecture covers MLE, ERM, and exponentially weighted averages — Negative_War_65 · 2026-07-28
- LLaDA2.2 brings diffusion language models into long-horizon agent tasks — 量子位 · 2026-07-28
- Kimi K3’s architecture is explained as an industrial plumbing system — doodlestein · 2026-07-28
- NeurIPS rebuttal season looks tight: one AC says the top score in 10 papers was 3.75 — CSProfKGD · 2026-07-28
- Robotics benchmarks may reward 0.2% gains while real-world success collapses from 92% to 43% — kscottz · 2026-07-28