TLDR: NeSyPr is a novel AI framework that enables language models (LMs) to perform efficient and adaptive embodied reasoning in dynamic, resource-constrained environments. Inspired by human knowledge compilation, it transforms explicit plans from symbolic tools into composable procedural representations stored in a vector-quantized memory. This allows LMs to execute complex, multi-step reasoning in a single inference step, significantly reducing latency and computational overhead. Through a contrastive planning mechanism, NeSyPr agents learn from successes and failures, adapting to new and unseen tasks without relying on external symbolic guidance, even with smaller LMs.
In the rapidly evolving field of artificial intelligence, language models (LMs) have shown immense potential in controlling embodied agents—robots or virtual characters that interact with their environment. However, deploying these powerful LMs in real-world, dynamic settings often hits a wall due to limitations like high latency, inconsistent connectivity, and restricted computational resources. This is where a groundbreaking new framework, NeSyPr (Neurosymbolic Proceduralization), steps in, offering a solution that makes AI agents smarter, faster, and more adaptive.
NeSyPr tackles the challenge by drawing inspiration from human cognition, specifically the Adaptive Control of Thought (ACT) theory. This theory explains how humans acquire skills by converting explicit, declarative knowledge (facts and rules) into implicit, procedural knowledge (automatic actions and routines) through practice. NeSyPr applies this concept to AI, enabling LMs to compile complex, multi-step symbolic reasoning into a single, efficient inference step.
The framework operates in two distinct phases:
1. Compositional NeSy Procedure Learning (Training Phase)
During this initial phase, the AI agent learns from an offline dataset. This dataset contains problem instances (observations and goals) and domain knowledge (action rules) that are symbolically defined. A symbolic tool, which possesses declarative knowledge about search algorithms, state transitions, and goal evaluation, generates explicit task-specific plans. NeSyPr then transforms these plans into composable procedural representations. Think of it like a human learning a new skill: initially, they follow explicit instructions step-by-step. NeSyPr encodes these implicit production rules into a ‘vector-quantized procedural memory,’ essentially a structured library of reusable skill components.
Also Read:
- Optimizing Vision-Language-Action Models for Robotics: A Deep Dive into Efficiency
- Building Reliable AI Agents: A Neuro-Symbolic Framework for Trustworthy Decisions
2. NeSy Procedure Contrastive Planning (Test Phase)
Once trained, the agent enters the test phase, where it operates autonomously without any online access to symbolic tools. It uses the procedural memory established during training to perform structured reasoning. Crucially, NeSyPr also enables adaptive reasoning. The agent maintains two ‘procedure banks’: one for successful procedures and another for failures. When faced with a new task, it reconstructs procedures by comparing them against these stored successes and failures. This ‘contrastive planning’ allows the agent to reinforce plans aligned with successful patterns and suppress those associated with failures, continuously adapting to an ever-changing environment based on real-time feedback.
The core benefit of NeSyPr is its ability to abstract and generalize multi-step symbolic reasoning into a single-step LM inference. This means AI agents can perform complex tasks efficiently, without needing large-scale inference engines or constant online access to external symbolic planners. This makes NeSyPr particularly well-suited for deployment in physical systems where latency and resource constraints are critical factors.
The effectiveness of NeSyPr was rigorously evaluated across several embodied benchmarks, including PDDLGym, VirtualHome, and ALFWorld. The results were impressive:
- For structured reasoning, NeSyPr achieved a 46.7% higher task success rate compared to DeepSeek-R1-Distill, a 70-billion-parameter reasoning model, while utilizing a language model that was 70 times smaller.
- In adaptive reasoning, it demonstrated a 62.1% higher success rate on unseen tasks with dynamic conditions compared to a traditional symbolic planner.
- For timely reasoning, NeSyPr reduced inference latency by over 90% compared to BoT, a large-scale inference baseline, while also improving task success by 36%.
These findings underscore that NeSyPr equips LM-based agents with robust structured, adaptive, and timely reasoning capabilities. It represents a significant step towards creating more efficient and intelligent AI agents that can learn and adapt in complex, dynamic environments, much like humans do. To delve deeper into the technical details, you can read the full research paper here.
