, which mentioned AI's self-improvement and Automating AI research. Today, I came across an article by Google Deepmind: 《Boundless Socratic Learning with Language Games》, which proposes a unique concept for "self-improving learning within a closed-loop system." This kind of "Socratic Learning" pushes the potential of artificial intelligence to its limits through recursive improvements that match input and output spaces.
Definition
In a closed system (black), the red agent improves itself through language interaction. External observers (gold) evaluate the agent's performance through performance metrics (green). If the agent's output can influence future inputs (blue path) and performance metrics continue to improve, it can be called self-improvement. If the agent's input and output space are compatible, this self-improvement is recursive; if the input and output space is language, it is called "Socratic Learning."
Self-improvement
Self-improvement is a special enhancement process whose core lies in the fact that the agent's output (i.e., its behavior) will, in turn, affect its future learning. This means the agent can shape its own experience flow, achieving potentially infinite improvement within a closed system.
, where the agent plays both the player and opponent roles in a symmetric game, generating streams of experience with feedback (such as "who won?") to continuously enhance skills.
— relates to the practicality of implementation.
Socratic Learning
The self-improvement we discuss here is a recursive form, requiring compatibility between the agent's input and output (i.e., they exist in the same space). Under this mode, the agent's output becomes future input, forming a closed loop.
It should be noted that the input and output spaces are not entirely identical but partially overlapping. For instance, the agent might generate code while perceiving natural language, (partially self-generated) code, and execution traces. This form is more restrictive than general self-improvement because it reduces the mediating role of the environment — typically, the agent's output only indirectly affects its input distribution through complex environmental mappings. Recursive self-improvement, on the other hand, directly depends on the compatibility between output and input.
This recursive mechanism is one property of many open-ended evolutionary processes, and open-ended improvement is considered a core feature leading to artificial superintelligence (ASI). Despite this, this compatibility is less stringent than traditional homomorphic self-modification, making it more universally applicable.
Language games
From “Attention is All You Need” to “Language Games Are All You Need.”
concept provides significant inspiration for this field. He proposed in his Philosophical Investigations: “Language and the actions it interweaves form a whole, which I call ‘language games.’”
Wittgenstein believed that the meaning of language does not lie in the words themselves but is embodied in the process of language interaction. Inspired by this, this paper defines “language games” as an interactive protocol (a set of rules representable in code) that specifies the way interactions based on linguistic inputs and outputs occur among one or more agents (“players”), and at the end of the game, provides each player with a scalar scoring function.
It is important to emphasize that, to ensure operability, we assume all language games can terminate within finite time. This definition not only tightly integrates language with action but also provides a theoretical framework for building efficient collaboration among agents.
Language games: meeting the core needs of Socratic learning
Language games (Language Games), by their definition, precisely meet the two core needs of Socratic learning:
Providing scalable mechanisms for generating interactive data (including self-play); Automatically generating accompanying feedback signals (scores).
In fact, this is almost the logical necessity of coverage and feedback conditions: all operational forms of interactive data generation with feedback signals are essentially language games. Furthermore, this definition allows for the introduction of rich strategic diversity brought by multi-agent dynamics, better satisfying the coverage condition. Moreover, this multi-agent collaboration is closer to the dynamic social co-construction of philosophers rather than just the soliloquy of a "loner who has lived for a thousand years."
From a practical standpoint, language games are easier to get started with. Human history has already created and refined a wide variety of games and player skills, providing a rich empirical basis for constructing language games. Nguyen even views this diversity as a reflection of human agency and local motivational flexibility. Some theorists (like Derrida) might further argue that any discourse, in a sense, already possesses the structure of a game. As Derrida said: "Every discourse, even poetry or oracles, contains a rule system that generates similar things, thereby outlining a methodology."
Related concepts can be traced back to Vygotsky's autotelic agents. Colas et al. believe that even if these agents are not closed-loop systems, many of their "internalized social interactions" can still be regarded as language games. Many common large language model (LLM) interaction paradigms can also be well represented as language games, such as:
Debate Role-playing Theory of mind Negotiation Jailbreak defense
Even in open systems, reinforcement learning based on human feedback can fall into this category.
Conclusion: Open-ended Socratic learning is possible
We start from recursive self-improvement in closed systems and explore its potential on the path toward artificial general intelligence (AGI). At this stage, we can optimistically conclude:
, and its main challenges (feedback and coverage) are already widely known. provides a constructive starting point for addressing these challenges and also offers direction on how to formulate practical research agendas.
Although specific roadmaps await further refinement in the future, the overall research direction has already become clear. It is particularly noteworthy that the diversity and richness of language games remain an underestimated important dimension. We believe that an ideal starting point is exploring processes that can generate open-ended games.
Interestingly, we submit these ideas for testing in the academic domain rather than engaging in self-talk within a closed system.
Netizen comments
"DeepMind's Socratic learning is a huge leap towards AGI! The concept of achieving self-improvement through language games and autonomous learning is exciting, and I look forward to its future development."
"Will AI teach humans in our dreams?"
"At least Google has brought us this gift! OpenAI is starting to get boring."
"This method of self-learning in a closed environment is fascinating. Through self-play and structured interactions, the agent can generate its own training data and enhance capabilities without continuous human intervention."
"These systems must learn to ask the right questions, not just solve the problems we feed them."
"A self-modifying system might break the constraints of existing architectures, but does this also mean AI could make unpredictable changes? How do we safely test and deploy such systems?"
"Language games are powerful iterative learning sandboxes. They allow AI to generate its own questions, solve them, and optimize reasoning abilities, marking a major leap towards AGI."
"Sounds absolutely amazing."