Goodfire's Silico Finds Human Cognitive Biases Geometrically Encoded Inside LLMs

There's a well-known experiment in psycholinguistics: show someone two shapes, one spiky and one round, and ask which is called "kiki" and which is called "bouba." People consistently pair soft sounds like "bouba" with rounded shapes and hard sounds like "kiki" with spiky ones. The bouba-kiki effect has been confirmed across cultures and languages since the 1920s. Now Goodfire has found the same structure encoded geometrically inside large language models.

A direction in activation space that knows what "spiky" sounds like

Using their interpretability platform Silico, Goodfire probed the internal activations of Llama and Gemma and found a specific geometric direction in the models' hidden states that cleanly separates spiky-sounding words from round-sounding ones, independent of actual meaning. Words like "kiki," "takete," and "crisp" cluster at one end; "bouba," "maluma," and "mellow" cluster at the other.

This is mechanistic interpretability in action: instead of asking a model what it thinks, you examine the geometry of its internal representations directly. The finding shows these models didn't just learn to answer bouba-kiki quiz questions correctly. They internalized the underlying structure of sound symbolism as a geometric feature of their representation space, absorbed from statistical patterns in training text rather than any explicit instruction.

Why this is harder to find than it sounds

Previous work on the bouba-kiki effect in AI focused on behavioral tests: ask the model to pick a word, count correct answers. Results have been mixed. A NeurIPS 2025 paper from Leiden University found that CLIP models don't consistently exhibit the effect and that performance doesn't reliably exceed chance. Behavioral tests measure outputs; Goodfire's approach measures internal geometry directly. A model can fail a behavioral test while still encoding the structure internally, and vice versa.

Mechanistic interpretability, the discipline of reverse-engineering what's actually happening inside a neural network, is what makes this distinction visible. Goodfire is one of a small group of organizations, alongside Anthropic, OpenAI, and Google DeepMind, building tools in this space. MIT Technology Review named mechanistic interpretability one of its 10 Breakthrough Technologies of 2026.

What Silico actually does

Goodfire has opened a private beta for Silico, which it describes as an agentic system that plans and runs interpretability experiments end-to-end, without a researcher manually writing probes. The bouba-kiki finding is a demonstration of its capabilities. The platform is built around four core functions:

• Feature decomposition: Breaks a model into interpretable features so you can see when predictions are driven by genuine understanding versus spurious correlations or dataset artifacts.
• Health diagnostics: Runs comprehensive checks on internal representations to surface problems like undertraining, information bottlenecks, or feature collapse before they reach production.
• Behavior steering: Once a feature is identified, you can amplify or suppress it. Goodfire used this to cut hallucinations in an LLM by 58%, at roughly 90x lower cost per intervention than LLM-as-judge, with no degradation on standard benchmarks.
• Agentic research: In one run, Silico reproduced RLFR, a method Goodfire's own team spent months building, in two days. RLFR places probes inside a model and uses their readings as reward signals for reinforcement learning. The automated reproduction cut hallucinations in Alibaba's Qwen3-8B by 37% without a corresponding drop in capability.

What this tells us about LLMs

Sound symbolism, the idea that word sounds carry meaning beyond their definitions, has long interested linguists because it suggests that naming is not entirely arbitrary. Finding this structure geometrically encoded in model activations is concrete evidence that LLMs absorb implicit human cognitive patterns from text alone, encoding more world structure than their outputs typically reveal.

For teams building on top of LLMs, the practical consequence is direct: geometric structure like this can be found, measured, and controlled. If a model has internalized an unintended bias, interpretability tools are the only way to locate it before it surfaces in production.

Access and pricing

Goodfire has raised $209M across three rounds from investors including Anthropic, Menlo Ventures, Lightspeed, Salesforce Ventures, and Eric Schmidt, and is valued at $1.25B. Pricing for Silico is determined on a case-by-case basis. Access is currently by request through the Goodfire contact page.

The bouba-kiki experiment is a good lens for what Silico is designed to do: take a hypothesis about what a model might have learned, turn it into a geometric probe of internal activations, and return a concrete answer. Whether you're auditing for hidden biases, reducing hallucinations, or mapping what your model actually knows, that workflow now has a commercial tool behind it.