EPFL's MiCRo Splits an AI Into Four Brain-Like Expert Modules So Its Reasoning Is Visible

EPFL researchers unveiled a language model called MiCRo (Mixture of Cognitive Reasoners) that divides processing across four expert modules modeled on the brain's functional organization, EPFL announced on June 26, 2026. MiCRo routes each word of a sentence to one of four experts — language, logic, social reasoning, and world knowledge — letting users see which module handles which content and adjust its influence. The work was led by PhD candidate Badr AlKhamissi at EPFL's NLP Lab and NeuroAI Lab, in collaboration with neuroscientist Greta Tuckute of Harvard and MIT, and was presented at the International Conference on Learning Representations (ICLR). This article organizes the key facts on the basis of EPFL's official announcement.

Porting Brain Regions Into Model Design

The name MiCRo stands for Mixture of Cognitive Reasoners, denoting a structure that mixes role-specific expert modules instead of one monolithic neural network. The work was led by PhD candidate Badr AlKhamissi at the NLP Lab and NeuroAI Lab in EPFL's School of Computer and Communication Sciences (IC), in collaboration with neuroscientist Greta Tuckute of Harvard and MIT. The word "Cognitive" in the name signals the point. Rather than leading with performance metrics, the team set interpretability — making the model's inner workings legible to people — as the central goal.

Processing That Splits Into Language, Logic, Social Reasoning, and World Knowledge

MiCRo splits processing across four expert modules: language, logic, social reasoning, and world knowledge. When a sentence comes in, each word is routed to a different expert, a structure modeled on how the human brain splits processing by function. The four modules are separated by role to handle linguistic form, logical computation, relational and social context, and background knowledge respectively. Because the active expert is revealed word by word, the model's reasoning path can be traced.

What the Dinner-Bill Example Reveals

MiCRo's interpretability is clearest in EPFL's example of splitting a dinner bill among friends. EPFL explained that, when splitting the bill, the social reasoning module identifies relational nuances while the logic module handles the arithmetic, so different reasoning types activate word by word. Users can observe which module handles which content, and adjust its influence, without any additional prompting. Rather than a black box that only emits an answer, it shows what kind of thinking happened where.

How It Differs From Prior Interpretability Approaches

Most language models produce an answer first and then infer the reason after the fact, leaning on tools like attention maps or probing. MiCRo reverses the order. Because modules are carved out by reasoning type in advance, interpretation is built into the structure rather than reconstructed post hoc. On top of that, users do not merely observe — they can directly adjust the influence of a specific module. This widens the point of intervention from checking "why did it answer this way" to steering "how should it answer." Whether module boundaries always map cleanly onto reasoning types, of course, remains a separate question.

Points Worth Watching for Korean Researchers and Practitioners

Demand for explaining the basis of a judgment is rising in Korea too, especially in regulated industries, so a structure that surfaces the reasoning path alongside the answer is worth examining. That said, the announcement assumes English examples and a Western social context, so whether the social reasoning module behaves the same way with Korean honorifics or relational expressions has not been verified. The overall direction — a small, modular design that leads with interpretability rather than chasing large monolithic models — offers an alternative path that resource-constrained Korean labs may find worth referencing.

Open Questions and How to Read the Claims Carefully

MiCRo's modular structure is a design modeled on brain regions, and it does not claim a one-to-one equivalence with human cognition. EPFL's announcement focuses on the structural strengths of interpretability and per-module routing rather than on quantitative benchmark numbers. As a result, how splitting experts balances performance and cost against a large monolithic model is an area that needs follow-up validation, since no numbers were disclosed. Even so, the ability for people to check and intervene in why the model answered as it did carries clear significance for safety.

Sources: EPFL