NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning¶

Conference: NeurIPS 2025 arXiv: 2510.19429 Code: N/A Area: Robotics / LLM Agent Keywords: Neurosymbolic Reasoning, Knowledge Compilation, Embodied Intelligence, Procedural Knowledge, Language Models

TL;DR¶

NeSyPr proposes a neurosymbolic proceduralization framework that transforms task plans generated by symbolic planners into composable procedural representations, enabling compact language models to perform efficient single-step reasoning without relying on external symbolic guidance — analogous to the human process of knowledge compilation.

Background & Motivation¶

Background: Large language models have demonstrated strong reasoning capabilities in embodied tasks, but typically require online access to large-scale inference engines or symbolic planners. Existing methods such as SayCan and Inner Monologue either require API access to large models or depend on online symbolic planning.

Limitations of Prior Work: (1) In dynamic physical environments, latency, connectivity, and resource constraints make online access to large models or symbolic planners infeasible; (2) small models lack the capacity for direct multi-step reasoning; (3) symbolic planners, while precise in reasoning, lack generalization ability and natural language understanding.

Key Challenge: High-quality reasoning currently depends on large models or symbolic planners, yet deployment environments demand low latency and lightweight solutions.

Goal: To equip compact language models with structured, adaptive, and timely reasoning capabilities without online dependence on external systems.

Key Insight: Drawing an analogy to human proceduralization — the cognitive process by which declarative knowledge is converted into automated procedural knowledge through practice. Similarly, the declarative knowledge of symbolic planners is compiled into procedural knowledge executable by a language model.

Core Idea: Symbolic tools are used to generate task-specific plans, which are then converted into composable procedural representations and embedded into the LM's reasoning process, compressing multi-step symbolic reasoning into single-step LM inference.

Method¶

Overall Architecture¶

The framework comprises a compilation phase and a deployment phase. In the compilation phase, a symbolic planner leverages declarative knowledge to generate task-specific plans → the plans are converted into procedural representations encoding implicit production rules → the procedural representations are injected into the LM. In the deployment phase, the LM directly generates actions at each step without invoking the symbolic planner.

Key Designs¶

Explicit Generation of Symbolic Plans:
- Function: Leverages a symbolic planner to precisely generate task solutions.
- Mechanism: Tasks are described using formal languages such as PDDL; a symbolic planner (e.g., Fast Downward) generates optimal action sequences given an initial state and goal. Plans contain precise action sequences and preconditions.
- Design Motivation: Symbolic planners substantially outperform LMs in structured reasoning; using them first ensures high-quality plan generation.
Procedural Knowledge Conversion:
- Function: Transforms explicit plans into implicit production-rule representations.
- Mechanism: Each "state → action" mapping in the plan is converted into composable procedural representations encoding production rules of the form "if the current state satisfies condition X, execute action Y." Multiple procedural representations can be composed to form reasoning chains for new plans.
- Design Motivation: Declarative knowledge ("what to do") must be converted into procedural knowledge ("how to do it"), enabling seamless integration into LM reasoning.
Single-Step LM Reasoning as a Substitute for Multi-Step Symbolic Reasoning:
- Function: Achieves efficient test-time inference.
- Mechanism: Procedural representations are injected into the LM's reasoning process via prompt injection or fine-tuning. At deployment, given the current observation, the LM directly outputs the next action — compressing what previously required multi-step symbolic reasoning into a single LM forward pass, analogous to the automatized skill execution of an expert human.
- Design Motivation: Eliminates online symbolic dependencies and enables low-latency deployment.

Loss & Training¶

Procedural knowledge can be injected into the LM via prompt injection (training-free) or fine-tuning. When fine-tuning, standard next-token prediction loss is applied.

Key Experimental Results¶

Main Results¶

Benchmark	Metric	NeSyPr	Large Model Reasoning	Symbolic Planner	Notes
PDDLGym	Success Rate	High	Moderately High	Highest	Compact LM approaches symbolic precision
VirtualHome	Success Rate	High	Moderately High	High	Everyday task reasoning
ALFWorld	Success Rate	High	High	High	Text-based game environment

Ablation Study¶

Configuration	Performance	Notes
NeSyPr (Full)	Best	Proceduralization + composability
w/o Proceduralization	Degraded	Compact LM reasons directly
w/o Composability	Degraded	Single plan not transferable
Online Symbolic Planning	Comparable but high latency	NeSyPr incurs no latency overhead

Key Findings¶

NeSyPr enables compact LMs to achieve reasoning performance approaching that of large models and symbolic planners.
Procedural conversion is the critical component — compressing multi-step reasoning into a single step substantially reduces inference latency.
Composability enables effective reasoning on novel tasks, not merely tasks seen during compilation.

Highlights & Insights¶

The Knowledge Compilation Analogy: Importing the concept of procedural knowledge from cognitive science into LLM reasoning offers a compelling interdisciplinary perspective. Analogous mechanisms from human cognition — such as conceptual composition and analogical reasoning — could similarly be compiled into language models.
Elimination of Online Dependencies: For latency-sensitive robotic applications (e.g., real-time manipulation), removing dependence on external APIs addresses a genuine practical need.
Empowering Compact Models: The work demonstrates that small models can acquire strong reasoning capabilities through knowledge compilation, with significant implications for edge deployment.

Limitations & Future Work¶

The compilation phase still requires a symbolic planner, entailing domain engineering costs.
For tasks that symbolic planners cannot solve (e.g., open-world reasoning), NeSyPr is likewise inapplicable.
The upper bound of procedural representation composability on highly complex tasks remains to be explored.
Validation is conducted only in simulated environments; deployment performance on physical robots is unknown.

vs. SayCan (Ahn et al. 2022): SayCan requires online LLM reasoning at each step; NeSyPr pre-compiles reasoning into a small model.
vs. Symbolic AI (STRIPS/PDDL): Symbolic systems reason precisely but lack flexibility; NeSyPr endows them with flexibility via the LM.
vs. ReAct: ReAct invokes LLM reasoning at every step; NeSyPr compresses reasoning into a single step.

Rating¶

Novelty: ⭐⭐⭐⭐ The knowledge compilation perspective is original and conceptually deep.
Experimental Thoroughness: ⭐⭐⭐⭐ Systematic validation across three benchmarks.
Writing Quality: ⭐⭐⭐⭐ Concepts are explained with clarity.
Value: ⭐⭐⭐⭐ Practically significant for edge deployment and efficient inference.