NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning

AI summary

Overview

• Research area: Artificial Intelligence — embodied AI, neurosymbolic reasoning, LLM-based agents, memory-augmented language models.
• Technical level: Advanced. The paper assumes familiarity with transformer architecture, vector quantization, contrastive decoding, and symbolic planners.
• Scope: The paper introduces NeSyPr, a framework that compiles symbolic plans into a composable "procedural memory" inside a small language model, so an embodied agent can plan in a single inference step without any online symbolic tool.

What This Paper Is About

Language-model agents for embodied tasks usually need either repeated calls to large models or continuous access to external symbolic planners, both of which are slow and impractical on resource-constrained physical systems. NeSyPr attacks this by first letting a symbolic tool generate explicit plans, then compiling those plans into discrete, composable "procedure" vectors stored inside the LM's own memory, so the compiled knowledge can be reused at test time in a single LM inference step. The goal is structured, adaptive, and low-latency embodied reasoning using small models with no online symbolic guidance.

Key Contributions

1. NeSyPr framework. A neurosymbolic proceduralization framework inspired by the Adaptive Control of Thought (ACT) theory, which compiles multi-step symbolic reasoning into single-step LM inference and removes the need for online symbolic planners in embodied tasks.
2. Compositional NeSy procedure learning. A training phase that encodes production rules into a vector-quantized procedural memory, where the resulting vectors can be compositionally combined to generate task-specific plans.
3. NeSy procedure contrastive planning. A test phase that adaptively generates plans by contrastively reconstructing task-specific procedures from stored procedures labeled with environmental feedback (success vs. failure banks).
4. Extensive evaluation. Experiments over 3 embodied benchmarks (PDDLGym, VirtualHome, ALFWorld) and 9 experimental scenarios, demonstrating structured, adaptive, and timely reasoning.

Main Findings

• Structured reasoning beats a much larger reasoning model. NeSyPr achieves a 46.7% higher task success rate than DeepSeek-R1-Distill, a distilled 70B-scale reasoning model, while operating with an LM 70 times smaller (as reported in Table 3).
• Adaptive reasoning beats a symbolic planner. It attains a 62.1% higher success rate on unseen tasks with dynamic conditions than the symbolic planner (Table 3).
• Timely reasoning. Inference latency is reduced by more than 90.0% compared to BoT, a large-scale inference baseline, while achieving a 36.0% improvement in task success rate (Table 3).
• PDDLGym open-loop results. NeSyPr outperforms the strongest baseline, LongMem, in the Minecraft and Rearrangement domains, with improvements of 13.6% in CSR, 11.7% in CGC, and 0.15 in SPL on the test set. On Minecraft test, NeSyPr reaches 65.2 CSR vs. LongMem's 53.3; on Rearrangement test, 73.5 CSR vs. LongMem's 58.2. Memory-augmented LMs' average test CSR is 15.2% lower than NeSyPr's.
• VirtualHome closed-loop results. Against the strongest baseline, DT-Mem, NeSyPr improves 12.5% CSR and 11.5% CGC on the train set, 9.6% and 9.8% on the seen set, and 12.4% and 16.7% on the unseen set. Its absolute numbers are 89.8 / 78.9 / 61.1 CSR on train / seen / unseen.
• ALFWorld closed-loop results. Against the strongest baseline, Optimus-2, it gains 9.8% CSR and 8.8% CGC on train, 8.3% and 7.0% on seen, and 10.6% and 9.2% on unseen. Absolute numbers are 69.6 / 61.1 / 59.7 CSR on train / seen / unseen.
• Unseen sets gain more than seen sets. Combined with an average 0.12 SPL improvement, this indicates effective symbolic reasoning rather than memorization. NeSyPr surpasses DT-Mem and Optimus-2 by an average of 11.2% CSR and 11.6% CGC.
• Single-step inference is the efficiency source. In the unified LLaMA 3.1-8B setting, NeSyPr records the lowest average plan latency, highest success rate, and smallest input/output token counts. BoT and LRM show latencies 54.3 and 105.8 seconds longer than NeSyPr, and consume 6,125.9 and 2,188.7 more total tokens due to multi-step reasoning.
• Better than distillation-based proceduralization. PlaSma reaches 80.5% CSR with an 8B LM, while NeSyPr reaches 83.6% CSR with only a 3B LM; NeSyPr hits 89.0 CSR with an 8B LM.
• Beats the symbolic planner under a strict time budget. For tasks where the symbolic planner takes over 10 seconds to find a solution, NeSyPr completes planning within a 5-second constraint while achieving 50.6% CSR; the symbolic planner takes up to 22 seconds to reach similar performance. Under incomplete input information the symbolic planner fails, whereas NeSyPr maintains stable performance and even outperforms ReAct using GPT-4o.
• Continual adaptation. Table 4 tracks 15 intermediate continual evaluation phases. LongMem's SR drifts from 57.1 down to 53.8, LM2's from 64.3 down to 51.6, while NeSyPr starts at 64.3 and remains in roughly the 61–67 range through phase 15. The full FWT/BWT/FR/RR entries for NeSyPr are cut off in the provided text.

Methodology in Plain English

The design copies a human learning idea from ACT theory: people first solve problems using explicit facts (declarative knowledge), and with practice those facts get "compiled" into automatic if-then habits (procedural knowledge) that fire without conscious recall.

NeSyPr does the same for an LM in two phases.

Phase i — training (compositional NeSy procedure learning). A symbolic planner generates correct plans for a small set of training problems. The LM has two memories. A working memory holds the current context as a set of slots. A procedural memory takes those slots and, using vector quantization, snaps each chunk to the nearest entry in a learned "procedure-book" of discrete procedure-units. Because each slot is split into chunks and each chunk maps to a book entry, procedures can be recombined like building blocks to form new plans for new tasks. The selected units form a runtime procedure that is gated back into the decoder layer's hidden state. Training combines the normal next-action prediction loss with a vector-quantization loss applied at every layer, and the procedure-book itself is updated with an exponential moving average.

Phase ii — test time (NeSy procedure contrastive planning). With no symbolic tool available, the agent keeps two banks of stored procedures: one of successes and one of failures, labeled from binary environmental feedback. Each incoming procedure is matched against both banks and reconstructed as a positive or negative version. The model then runs two hidden states in the same batch — one conditioned on positive procedures, one on negative — and uses contrastive decoding: only tokens within a truncated plausibility set are considered, and their scores are the difference between the positive and negative log-probabilities. This suppresses failure patterns and promotes plans consistent with the environment, giving adaptive behavior without any symbolic guidance.

Why This Matters

Impact on research. The paper reframes symbolic reasoning as something to be compiled into a model rather than queried from a tool at runtime. It connects a long-standing cognitive theory (ACT) to practical LM agent design, and it shows that a small LM with the right memory architecture can beat both a 70B distilled reasoning model and a classical symbolic planner on embodied benchmarks — arguing that reasoning capability is partly an architectural and knowledge-representation question, not only a scale question.

Real-world applications.

• Home and service robots: executing multi-step household tasks where the robot cannot reach a cloud planner because of latency or connectivity limits.
• Warehouse and logistics automation: rearrangement and item-retrieval tasks where plans must be produced in seconds on local hardware.
• Field, industrial, or space robotics: degraded or absent network conditions, where decisions must be made onboard with limited compute.
• Assistive and edge devices: agents that must run continuously across a stream of changing goals — such as a changing room layout or new object placements — using compact models.

Industry relevance. The economics are the point: the paper uses LLaMA-3.2-1B for PDDLGym and Qwen2.5-0.5B for VirtualHome and ALFWorld as default backbones, and reports latency measured in single-digit seconds and token counts in the low thousands for NeSyPr, versus tens of seconds and thousands more tokens for large-scale inference baselines. For anyone deploying agents on edge hardware, the pitch is a smaller model, a smaller token bill, and a smaller latency budget.

Future Directions

1. Handling richer or non-symbolic inputs. The method relies on symbolically specified observations, goals, and domain knowledge; extending proceduralization to raw perception or language-only observations is an open question the evaluation setup does not address.
2. Scaling the procedure-book and memory capacity. The design exposes hyperparameters like the number of procedure-units (K), memory slots (S), and chunk size (d). How these scale with task diversity, and whether the quantized book saturates or conflicts across very different domains, is not reported.
3. Longer-horizon continual adaptation. Table 4 only tracks 15 continual evaluation phases in the provided content, and the full forward/backward transfer numbers for NeSyPr are not shown. Whether procedural memory keeps improving, plateaus, or degrades over much longer task streams remains open.
4. Robustness of the feedback signal. The test phase depends on task-level binary success/failure labels to populate the positive and negative procedure banks. Behavior under noisy, delayed, or sparse feedback — and under feedback that mislabels a plan — is not explored in the reported results.

Target Audience

Researchers and engineers working on embodied AI, LLM agents, and neurosymbolic systems; practitioners who need to deploy reasoning agents on latency-sensitive or resource-constrained hardware; and readers interested in cognitive-architecture-inspired machine learning, particularly the ACT theory of proceduralization. The paper is most useful to those already comfortable with transformer internals and reinforcement-learning-style evaluation metrics (CSR, CGC, SPL, FWT, BWT), since much of the contribution lives in the memory architecture and decoding procedure rather than in the benchmark results alone.

Authors’ abstract

We address the challenge of adopting language models (LMs) for embodied tasks in dynamic environments, where online access to large-scale inference engines or symbolic planners is constrained due to latency, connectivity, and resource limitations. To this end, we present NeSyPr, a novel embodied reasoning framework that compiles knowledge via neurosymbolic proceduralization, thereby equipping LM-based agents with structured, adaptive, and timely reasoning capabilities. In NeSyPr, task-specific plans are first explicitly generated by a symbolic tool leveraging its declarative knowledge. These plans are then transformed into composable procedural representations that encode the plans' implicit production rules, enabling the resulting composed procedures to be seamlessly integrated into the LM's inference process. This neurosymbolic proceduralization abstracts and generalizes multi-step symbolic structured path-finding and reasoning into single-step LM inference, akin to human knowledge compilation. It supports efficient test-time inference without relying on external symbolic guidance, making it well suited for deployment in latency-sensitive and resource-constrained physical systems. We evaluate NeSyPr on the embodied benchmarks PDDLGym, VirtualHome, and ALFWorld, demonstrating its efficient reasoning capabilities over large-scale reasoning models and a symbolic planner, while using more compact LMs.