LLM activations expose a bouba-kiki axis learned from nonsense words
burny_tech · x · 2026-07-28
Researchers find a “bouba-kiki” axis inside LLM activations
Goodfire says it found a latent direction in Llama and Gemma activations that separates “spiky-sounding” from “round-sounding” words, even when the words’ meanings are unrelated.
- The effect mirrors the classic bouba-kiki phenomenon from cognitive science.
- The axis appears in a Llama 3.1 8B-based analysis shown in the image.
- The claim is that the separation comes from nonsense words only, suggesting the model has learned a sound-shape-like structure in its representations.
This is a mechanistic interpretability-style finding: a linguistic pattern emerges inside model activations without being explicitly trained as a concept.
Related event: Goodfire Discovers Bouba-Kiki Activation Space in LLMs(2 posts)→
More from Research
- GaussianGPT uses autoregressive next-token prediction to generate 3D Gaussian scenes — rsasaki0109 · 2026-07-28
- Prompt scaffolding for Claude agents turns into technical debt as models change — TechPreacher · 2026-07-28
- Codabench launches a public pan-cancer CT segmentation challenge — maier_ak · 2026-07-28
- David Baker’s open-science lab culture is reshaping AI-era protein design — xiaofei_lin · 2026-07-28
- FilmBench evaluates cinematic video generation with film-school shot lists and 35 metrics — Shengyi Wang · 2026-07-28
- Frozen 12B system reuses verified memory at zero tokens and 6,000,000-token context — Corbenci · 2026-07-28