Research

This is the most basic question in the research program. We need to determine what allows interaction histories to become stable, reusable semantic structure.

• Which architectural biases are necessary?
• Should objectness be built in, introduced as a weak prior, or emerge entirely from controllability and persistence?
• How do relations, local dynamics, and causal variables form?
• Is prediction value-equivalence sufficient to produce stable concepts?
• How can a system distinguish a useful feature from a novel concept?
• Must abstractions become discrete, or can a continuous manifold support language-like structure?

Criterion. An abstraction should support intervention, composition, transfer, and reduced exploration steps on new tasks.

A subtask may be a recurring causal fragment in the stream of experience that can be encapsulated and reused.

Approach an entity → contact occurs → the reachable region changes.

The same fragment can be understood as a semantic concept, a skill, a subtask, a temporal abstraction, or a compressible segment of experience.

The unifying question is whether concept formation and subtask discovery can be treated as optimal causal compression of the agent-environment interaction trajectories.

• How do abstraction boundaries arise automatically?
• What determines granularity that governs the level of abstraction?
• When should a new concept be created, and when should an existing concept be reused?
• Should one abstraction serve predictive, control, and memory functions at the same time?
• What is the relationship among minimum information bits, information bottlenecks, and controllability of the embodied agent itself?

Current systems place many processes outside the model:

• chain-of-thought prompting
• tree search
• reinforcement-learning loops
• task decomposition
• planners
• tool orchestration
• memory retrieval
• actor–critic iteration

Let's make a simple comparison between those external and internal processes of foundation models.

What processes will fit to internalization

• They recur frequently.
• Their structure is stable across tasks.
• They are differentiable or learnable through self-supervised signals.
• Internalization reduces interaction, search, or computation cost.
• They do not depend on task-specific symbols.

Processes that should remain external

• They are rare but computationally complex.
• They require exact symbolic verification.
• They must remain auditable by humans.
• They depend on external databases or social rules.
• The task changes faster than the model can be updated.

Objective. The aim is not to eliminate search. It is to compile frequently repeated search and reasoning with stable structure into the model's representations, recurrent dynamics, and control geometry.

Successful internalization should be measured through adaptation speed on new tasks, out-of-distribution success, and regret decay.

Traditional reinforcement learning usually assumes that reward is known. World-model approaches often assume that the task is known. A general agent must infer several things at once:

• How does the world change?
• Which factors are controllable?
• Which states may represent progress?
• What is the current goal?
• Which rules persist across episodes or levels?

This requires system identification, goal induction, rule induction, intrinsic exploration, hypothesis revision, and planning to be unified within a belief-space learning problem.

IMHO, ARC-AGI-3 places exploration, modeling, goal setting, and planning inside one first-contact process, making it a useful minimal test of this question.

An agent's concepts arise from action consequences, controllability, predictive structure, environment dynamics, and task experience. Human concepts also reflect bodily experience, social interaction, cultural history, linguistic convention, education, and institutions. The two systems will not align one-to-one by default.

• Which internal concepts are approximately isomorphic to human words?
• Which agent concepts have no corresponding human vocabulary?
• Can one word refer to several distinct internal dynamical structures?
• Should alignment map individual embeddings or relational structure?
• When human descriptions conflict with an agent's experience, which source should dominate?
• How can an agent question an incorrect human prior?
• How can people correct an agent without destroying its autonomously formed semantic system?

A more plausible interface is not a direct mapping between latent state and free-form text:

agent semantics ↔ relational / functional DSL ↔ controlled language ↔ pretrained language model

We need to determine what an agent actually needs to obtain from a pretrained language model or visual-language model. Candidate information includes:

• goal semantics;
• human preferences;
• cultural conventions and social norms;
• priors over task types and common object functions;
• risk and ethical constraints;
• language summaries of complex situations.

Question. What is the minimal sufficient human interface layer that allows an agent to effectively interact with and learn from human input?

A language model need not necessarily mediate low-level perception, geometric localization, every action, dynamics verification, or short-horizon control.

Question. At which level of abstraction can a language model provide the greatest adaptation to human tasks with the least dependence on language?

The current hypothesis is to connect at the level of goals, rules, functional relations, and plan sketches—not at the perception layer or direct-action layer.

A next-generation foundation model should continue to grow through experience after deployment rather than learning only before release.

• How should episodic experience be compressed into semantic memory?
• When should the shared model be updated, and when should adaptation remain in an environment-specific module?
• How can catastrophic forgetting be limited?
• How can structural correspondences between old and new environments be identified?
• How can methods of reasoning transfer instead of short-cut templates inside the model?
• How can failure and surprise become reusable knowledge?
• How can abstractions be shared across bodies, action spaces, and observation modalities?

Goal. The objective is not to remember more trajectories. As useful experience accumulates, the interaction, computation, and risk required to acquire a new skill should continue to decline. Check "Experience Scaling" in this subtask article for more details.

Language-model scaling is usually described through tokens, parameters, and training compute. Experience-grounded intelligence requires different axes:

• number of effective interactions;
• diversity of test environment;
• coverage of causal interventions;
• number of reusable skills;
• experience-compression ratio;
• structural reuse across environments;
• deployment-time adaptation;
• regret as a function of accumulated experience.

Hypothesis. The meaningful scaling law is the continued decline of regret in out-of-distribution environments as effective experience accumulates.