[FAIR 2026] The A-B-M Architecture: Why AI Systems Don’t Learn and the Roadmap to Autonomy

TL;DR

Despite the hype surrounding Large Language Models (LLMs), current AI remains fundamentally "non-learning" once deployed. In this seminal paper, Emmanuel Dupoux, Yann LeCun, and Jitendra Malik argue that the missing piece is Autonomous Learning. They propose a tripartite architecture—System A (Observation), System B (Action), and System M (Meta-control)—to move beyond human-dependent MLOps toward agents that explore, curate their own data, and adapt like human children.

The "Data Wall" and the Outsourcing of Intelligence

The current AI paradigm is hitting a wall. We have optimized for scale, but we have neglected autonomy. The authors point out a glaring paradox: while deep learning is inspired by the brain, current models lack the most basic biological trait—the ability to learn by doing without a human supervisor.

Current limitations include:

• The MLOps Crutch: Learning is currently "outsourced" to human engineers who curate datasets and design frozen training recipes.
• Domain Mismatch: Models fail when the "heavy-tailed" real world deviates from their static training distribution.
• Lack of Grounding: System A (like GPT-4) knows language but doesn't understand the physical world; System B (like RL agents) understands actions but is often sample-inefficient and specialized.

Methodology: The A-B-M Architecture

The paper’s core contribution is the integration of three distinct cognitive functions into a single "Control Plane."

1. System A: The Observer

Equivalent to Self-Supervised Learning (SSL). It builds predictive world models from passive data, extracting statistical regularities (e.g., phonetic categories in babies or tokens in LLMs).

2. System B: The Actor

Equivalent to Reinforcement Learning (RL) and Control. It interacts with the world to achieve goals. It provides the "grounding" that System A lacks by testing hypotheses through intervention.

3. System M: The Orchestrator (The "Secret Sauce")

System M is a Meta-control plane that operates like a software-defined network. It doesn't process pixels; it monitors "telemetry"—meta-states like uncertainty, hunger, or prediction error—and decides which data pathway to open. It essentially automates the engineer's job of deciding when to explore, when to dream (memory replay), and when to learn from a teacher.

Figure: The blueprint of System M acting as a central control plane for routing data between observation and action.

Resolving the Chicken-and-Egg Problem: Evo/Devo

How do you initialize such a complex system? If System A needs actions to be grounded, and System B needs System A's representations to be efficient, where do you start?

The authors suggest a Bilevel Optimization framework:

• Evolutionary Scale (Outer Loop): Hand-crafted fitness functions in simulated environments optimize the agent's "genetics"—its architecture and initial parameters ($A_0,B_0,M$).
• Developmental Scale (Inner Loop): The agent learns and adapts its parameters through its "life cycle" interaction with the world.

Figure: The dual-scale optimization process where evolution designs the "learner" and development executes the "learning."

Critical Analysis: Advanced Modes

The architecture enables high-level cognitive traits:

• Learning from Communication: System M detects "pedagogical signals" (like a parent's gaze) to trigger one-shot learning.
• Learning from Imagination: During "rest" states, System M can route memory to System A/B for "mental simulations," effectively training the model on counterfactual scenarios.

Limitations and Challenges

1. Simulators: We need "faster than real-time" realistic simulators to run millions of evolutionary life cycles.
2. Sample Efficiency: Even with System M, the "inner loop" of learning must be significantly faster than current RL.
3. Alignment: An autonomous system that generates its own rewards (curiosity) might "hack" its internal signals, leading to maladaptive behaviors.

Conclusion: A Paradigm Shift

This paper is not just a technical update; it is a manifesto for the next decade of AI. By shifting the focus from Scaling Laws (more data/compute) to Autonomous Architecture (System M), the authors provided a roadmap to bridge the gap between "stochastic parrots" and truly intelligent, embodied agents. The future of AI is not a bigger transformer; it is a system that knows when it's confused and decides to go out and play to figure it out.