Why AI systems don't learn and what to do about it: Lessons on autonomous learning from cognitive science

Emmanuel Dupoux; Yann LeCun; Jitendra Malik

cs.AIMar 16, 2026

Why AI systems don't learn and what to do about it: Lessons on autonomous learning from cognitive science

Authors: Emmanuel Dupoux, Yann LeCun, Jitendra Malik

Organizations: FAIR at META · École des Hautes Études en Sciences Sociales · NYU · UC Berkeley

Abstract

We critically examine the limitations of current AI models in achieving autonomous learning and propose a learning architecture inspired by human and animal cognition. The proposed framework integrates learning from observation (System A) and learning from active behavior (System B) while flexibly switching between these learning modes as a function of internally generated meta-control signals (System M). We discuss how this could be built by taking inspiration on how organisms adapt to real-world, dynamic environments across evolutionary and developmental timescales.

Figures & tables

		Proposed Learning Mechanisms
Input	Task	CogSci terminology	AI terminology	(sample) AI Algorithms
Speech	Learning phonetic categories	Distributional Learning / Perceptual Learning	Self-Supervised Learning / Acoustic Unit Discovery	CPC ( van den Oord et al., 2018 ) , HuBERT ( Hsu et al., 2021 )
Language	Syntax acquisition	Statistical Learning	Language Modeling	GPT ( Radford et al., 2018 ) , BERT ( Devlin et al., 2019 ) , GSLM ( Lakhotia et al., 2021 )
Images	Face recognition, Object categorization	Perceptual Learning	Self-Supervised Learning	SimCLR ( Chen et al., 2020 ) , MoCo ( He et al., 2020 ) , DINO ( Caron et al., 2021 ) , I-JEPA ( Assran et al., 2023 )
Video	Intuitive physics	Perceptual Learning	Predictive World Modeling	PredNet ( Lotter et al., 2017 ) , V-JEPA ( Bardes et al., 2024 )
Language + Vision	Learning words and sentence meaning	Cross-situational Learning	Multimodal Language Modeling	CLIP ( Radford et al., 2021 ) , Flamingo ( Alayrac et al., 2022 ) , Transfusion ( Zhou et al., 2024 )

Table 1: Sample observational learning problems (System A) from different sensory inputs, proposed mechanisms with corresponding terminology in Cognitive Science and AI fields and sample AI algorithms (non-exhaustive).

$s_{t+1}\sim M(.\|s_{t},a_{t})$	Dynamic model of the world
$a_{t}\sim\pi(.\|s_{t})$	Agent’s policy
$r(s,a)$	Reward function
$\gamma\in[0,1)$	Discount factor

Table 4

Paradigm	W. Model $M$ (train., infer.)	Policy $\pi$ (train., infer.)	When is optimiza- tion solved?	Example
Control Theory	fixed (no,no)	derived analytically (no, direct)	offline (design time)	Spinal reflexes, saccadic eye movement ( Robinson, 1981 )
Adaptive Control	adaptive param. (yes, no)	derived analytically (no, direct)	online (parameter estimation)	Motor adaptation ( Shadmehr and Mussa-Ivaldi, 1994 )
Model-Free RL	none	NN (yes, direct)	training time (compil- ation of experience into reactive policy)	{Habitual actions ( Schultz et al., 1997 )
Model-Based RL	NN (yes, no / unrolled )	NN (yes, direct / search)	both training and inference	Goal-directed behaviors, foraging strategies ( Daw et al., 2011 )
Planning	simulator / NN (no/yes, unrolled)	computed online (no,search)	inference time (thinking before acting)	Detour planning, mental simulation ( Pfeiffer and Foster, 2013 )

Table 2: Popular paradigms addressing System B optimization problems, differing in how the world model

M

and the policy

\pi

are defined, whether they are learned from data in a training phase, and how they are used at inference time, with selected examples of applications in the modeling of animal behavior.

Figure 2: Summary of modes of interactions between Systems A and B. System A provides System B with predictions of future states conditioned on past states and actions, with hierarchical abstractions over possible actions, and a SSL loss that can be used for curiosity/exploration. System B through its action provides rich and task relevant input for System A to learn from.

Figure 3: Interactions between learning modes for imitation learning. (a) Self Play. System B provides action, state trajectories to System A that learns a World Model, and provides a prediction-based intrinsic reward signal to system B. (b) Social Observation . System B directs the attention to peers that provide System A with complex trajectories from which it infers latent actions. (c) Retargeted imitation . System A learns to map exocentric actions and states to egocentric ones, helping system B to achieve goal-directed behavior. (image from ChatGPT)

Panel of Figure 4 — Figure 4: Blueprint of a cognitive architecture featuring System M as an autonomous orchestrator. System M acts as a central control plane that automates data routing and training recipes. High-bandwidth data streams (e.g., plain arrows) carry raw sensory inputs, motor commands, and latent representations between System A (perception/world modeling), System B (action/policy), and an episodic memory buffer. Low-bandwidth control streams (e.g., thin dashed arrows) carry telemetry: System M monitors internal meta-states (such as prediction errors or uncertainty) and outputs routing commands (meta-actions) to dynamically open or close data pathways, effectively assembling and disassembling learning and inference pipelines on the fly.

	$\phi_{t+1}=\arg\min_{\phi_{t}}\mathcal{L}(A_{0}:A_{K},B_{0}:B_{K})$
subject to
	$A_{0},B_{0},M=\operatorname{Init}(\phi_{t})$
	$A_{i+1},B_{i+1}=\operatorname{Update}(M,A_{i},B_{i},\mathit{Env})$
where
	$\operatorname{Init}$ is the initialization procedure
	$\operatorname{Update}$ is the inner loop update rule

Table 9

Figure 5: Evo/Devo framework for building autonomous learning agents. Learning takes place at two scales. In the developmental scale, the learner’s architecture (A, B and M) is initialized from meta parameter $\phi$ . A and B update their parameters through interaction with the environment controlled by a fixed controler M. In the the evolutionary scale, $phi$ is updated to optimize a fitness function $\mathcal{L}$ measured over the life cycle of the system. (images from ChatGPT) .

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

		Modes of learning
		System A	System B
Modes of inference	System 1	predictive coding; statistical learning	policy learning
Modes of inference	System 2	counterfactual reasoning; causal learning	learning through planning

Appendix

Table A.1: Link between System 1/2 and System A/B

	Meta-states	Species	Effects (sample references)
Species-Specific signals
	Faces and vocalizations	Human infants	IS ( Johnson et al., 1991 ; Vouloumanos and Werker, 2007 ) ; LE ( Ferry et al., 2010 )
	Gaze direction	Human infants, dogs, corvids	IS (gaze following: Farroni et al. 2002 ; Tomasello et al. 1998 )
	Pedagogical signals (high pitch, direct gaze; pointing)	Human infants	IS ( Fernald, 1985 ) , LE ( Thiessen et al., 2005 ) , MC (shifts from memorization to generalization; Csibra and Gergely 2009 ; Gergely et al. 2002 )
	Self-propelled, biomechanical motion	Human infants, chicks	IS ( Simion et al., 2008 ; Scholl and Tremoulet, 2000 ) , LE ( Nairne et al., 2013 ) ; MC (switches to teleological reasoning; Csibra et al. 1999 )
	Potential threats (looming, snake or spider-like objects)	Human infants, mammals	IS ( Ball and Tronick, 1971 ; Yilmaz and Meister, 2013 ; DeLoache and LoBue, 2009 ) ; LE ( Õhman and Mineka, 2001 ; Cook and Mineka, 1989 ) ; MC (freezing/high vigilance; Fanselow 1994 )

Appendix

Table C.1: Non exhaustive list of meta-states in humans and animals, and their effects interpreted in terms of Input Selection (IS: attracting gaze or attention), Learning Efficacy (LE: boosting learning rate) and Mode Control (MC) in humans and animals.

Cognitive Architectures for AI Agents

Cognitive architectures for AI agents organize perception, memory, goals, reasoning, and action into systems for persistent, adaptive behavior.

See all 85 papers

Learning Beyond Optimization: Stress-Gated Dynamical Regime Regulation in Autonomous Systems
Feb 20, 2026Sheng RanAdaptive Learning Self-Evolving Agents
Department of Physics, Washington University in St. Louis, St. Louis, MO 63130, USA · Reconstructing Future Research Initiative
How Psychological Learning Paradigms Shaped and Constrained Artificial Intelligence
Mar 18, 2026Alex Anvi Eponon, Ildar Batyrshin, Christian E. Maldonado-Sifuentes +1Cognitive Architectures for AI Agents Compositional Generalization
Centro de Investigación en Computación (CIC), Instituto Politécnico Nacional (IPN), Mexico City, Mexico.
Grounded world models in biological organisms and future embodied AI
Jul 15, 2026Giovanni Pezzulo, Davide Nuzzi, Marco D'Alessandro +3Cognitive Modeling Embodied AI
Institute of Cognitive Sciences and Technologies, National Research Council, Rome, Italy · Center for Mind and Brain Science (CIMeC), University of Trento, Italy · Department of Neuroscience, University of Montreal, Montreal, Quebec, Canada

Feb 20, 2026J/K move · Enter open · S save

Learning Beyond Optimization: Stress-Gated Dynamical Regime Regulation in Autonomous Systems

Sheng Ran

Department of Physics, Washington University in St. Louis, St. Louis, MO 63130, USA · Reconstructing Future Research Initiative

Adaptive Learning Self-Evolving Agents

Open paper AskPDF arXiv

Despite their apparent diversity, modern machine learning methods can be reduced to a remarkably simple core principle: learning is achieved by continuously optimizing parameters to minimize or maximize a scalar objective function. This paradigm has been extraordinarily successful for well-defined tasks where goals are fixed and evaluation criteria are explicit. However, if artificial systems are to move toward true autonomy-operating over long horizons and across evolving contexts-objectives may become ill-defined, shifting, or entirely absent. In such settings, a fundamental question emerges: in the absence of an explicit objective function, how can a system determine whether its ongoing internal dynamics are productive or pathological? And how should it regulate structural change without external supervision? In this work, we propose a dynamical framework for learning without an explicit objective. Instead of minimizing external error signals, the system evaluates the intrinsic health of its own internal dynamics and regulates structural plasticity accordingly. We introduce a two-timescale architecture that separates fast state evolution from slow structural adaptation, coupled through an internally generated stress variable that accumulates evidence of persistent dynamical dysfunction. Structural modification is then triggered not continuously, but as a state-dependent event. Through a minimal toy model, we demonstrate that this stress-regulated mechanism produces temporally segmented, self-organized learning episodes without reliance on externally defined goals. Our results suggest a possible route toward autonomous learning systems capable of self-assessment and internally regulated structural reorganization.

Abstract

Cognitive Architectures for AI Agents

Explore similar work

Learning Beyond Optimization: Stress-Gated Dynamical Regime Regulation in Autonomous Systems

How Psychological Learning Paradigms Shaped and Constrained Artificial Intelligence

Grounded world models in biological organisms and future embodied AI