Large Language Models (LLMs) often struggle with "compositional generalization"—the ability to combine known concepts to solve novel problems in interactive environments. Because these models rely on statistical patterns rather than a structured understanding of cause and effect, they can be brittle and prone to errors. AGEL-Comp (Action-Grounded Experiential Learning for Compositionality) is a new neuro-symbolic architecture designed to solve this by grounding an agent’s actions in a formal, interpretable world model that evolves through experience.
A Structured World Model
At the heart of AGEL-Comp is the Causal Program Graph (CPG), a directed hypergraph that acts as the agent’s "brain." Unlike a standard neural network, the CPG explicitly represents procedural and causal knowledge as a set of logical rules (Horn clauses). This structure allows the agent to break down complex tasks into hierarchical sub-goals, ensuring that its planning is based on a clear, logical map of how the environment functions rather than just probabilistic guesses.
The Reasoning and Learning Loop
AGEL-Comp uses a hybrid approach to ensure its plans are both creative and correct. When faced with a goal, an LLM proposes potential plans, which are then checked for logical consistency by a Neural Theorem Prover (NTP). The NTP acts as a rigorous verifier, using differentiable logic to ensure the proposed actions align with the agent’s current knowledge.
When the agent encounters an unexpected outcome—such as an action leading to an error—it triggers a learning cycle. An Inductive Logic Programming (ILP) engine analyzes the failure by comparing it to past successful experiences. It identifies the specific cause of the error and synthesizes a new logical rule, which is then added to the CPG. This allows the agent to "learn" from its mistakes in real-time, grounding its symbolic knowledge in actual interaction.
Bridging Neural and Symbolic AI
The framework maintains a shared, trainable embedding space that connects the discrete symbols in the CPG with the continuous vector representations used by the NTP. This allows the agent to perform "differentiable unification," where it can reason about logic using neural-style calculations. As the agent learns new rules, it periodically fine-tunes its neural reasoner to ensure that its deductive capabilities remain perfectly synchronized with its updated symbolic world model.
Performance and Evaluation
The researchers evaluated AGEL-Comp within the Retro Quest simulation environment, specifically testing for compositional generalization. The findings indicate that AGEL-Comp significantly outperforms pure LLM-based models. By replacing "black-box" statistical guessing with a system that builds an explicit, interpretable, and structured understanding of its environment, the framework provides a more robust and reliable path for creating intelligent, interactive agents.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!