acceptodds
Under review as a conference paper at ICLR 2027

NeSyChain: A Neurosymbolic Framework for Lifelong Learning in Embodied Agents

Abstract

Embodied agents operate in changing open-world environments and therefore require lifelong learning. Recent work addresses this with neurosymbolic agents, whose combination of foundation models and symbolic tools makes learned knowledge explicit and verifiable. Yet in these agents, both reasoning and learning are tightly coupled to a particular configuration of foundation models, symbolic tools, and knowledge format, and must be rebuilt whenever it changes. We present NeSyChain, a modular framework for building neurosymbolic embodied agents whose reasoning and learning are decoupled from the configuration. Foundation models and symbolic tools are loosely coupled through a shared knowledge layer that mediates their interactions, so reasoning flows stay intact when components change. Specifically, for lifelong learning, NeSyChain employs two-stage knowledge transformation: acquired data is consolidated into symbolic knowledge through contrastive validation, then compiled into neural memory on the frozen model, reducing tool reliance. NeSyChain remains effective across 6 neurosymbolic configurations, spanning 4 foundation models, 9 symbolic tools, and 3 knowledge formats, on 3 benchmarks and in real-world settings, using a common modular structure and learning procedure. On every benchmark, it outperforms the strongest baseline, with 21.8% higher task success and 51.3% lower symbolic tool reliance on average. In real-world deployment, verifying learned knowledge before use raises task success from 33.3% to 96.7%.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Embodied Agents, Neurosymbolic AI, Lifelong Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_p4beOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a lifelong learning approach that enables embodied intelligence systems to continuously accumulate experience and skills. It accumulates and compresses knowledge in symbolic form to iteratively improve the overall system’s performance. The method’s effectiveness is validated on multiple benchmarks and through a limited set of real-world robot experiments.

Critical strengths
  1. Summarizing, accumulating, and compressing experience using textual symbols is an effective and reasonable approach that makes full use of language models’ strengths.
  2. The experimental validation is generally thorough and supports the paper’s claims.
  3. The paper provides extensive visualizations and supporting details, along with real-world robot experiments that demonstrate the method’s effectiveness.
Critical weaknesses
  1. The generalizability and reusability of the accumulated knowledge need to be clarified, particularly which knowledge can be shared across different tasks or even different benchmarks.
  2. The efficiency of knowledge accumulation and the additional overhead of invoking the corresponding harness during execution may warrant dedicated analysis.
  3. If the knowledge itself takes the form of a harness, the paper does not clarify whether its benefits generalize across different underlying execution models.
Questions for authors
  1. Provide a detailed analysis of whether the accumulated knowledge generalizes across different tasks, benchmarks, and underlying execution models.
  2. Conduct a dedicated analysis of the efficiency of knowledge accumulation and the overhead of invoking the corresponding harness, and draw clear conclusions from the results.
4 comments

We greatly thank the reviewer for the constructive feedback. We address each point below and indicate the corresponding paper revisions. We would be happy to clarify any remaining concerns.


W1 Q1. The generalizability and reusability of the accumulated knowledge need to be clarified.

The real-world experiments in Figure 4 provide concrete examples of how accumulated symbolic knowledge generalizes and is reused across tasks and from CaP-Bench simulation to real-world deployment.

Figure 4(a) shows how symbolic knowledge acquired in CaP-Bench is extended to handle additional operating constraints in real-world task settings. The agent reuses its existing knowledge for object placement and appliance manipulation while refining the conditions for appliance activation using reference manuals. For the induction stove, it adds stove_compatible(obj) and on_top_of(obj, stove) as preconditions, requiring compatible cookware to be placed on the cooking zone before activation. For the air fryer, it adds temperature-set(off) and timer-set(off) as preconditions, requiring both knobs to be reset before activation. These updates extend the transferred knowledge to operating conditions not covered by the original simulation experience, while retaining existing knowledge for the remaining task steps. Appendix B.2.1 provides further details.

Figure 4(b) further demonstrates the compositional generalizability of newly acquired symbolic knowledge and the reusability of existing symbolic knowledge. Across successive task executions, the agent reuses knowledge supporting previously successful behaviors and combines it with new knowledge from different sources to develop new task-solving strategies and progressively improve task quality. Feedback from its own execution first refines Push by adding moved(obj, loc) to its effects, enabling keyboard relocation. The Stack schema acquired from another robot's cube-stacking task demonstration is then reused to stack plastic cups in the desk-cleaning task, while the previously acquired pushing behavior remains in use. Human-provided knowledge subsequently introduces Bin_press with free_space(bin) as an effect, enabling additional objects to be discarded into a previously full bin while retaining both pushing and stacking.

These schemas are used together with the existing Pick, Place, and Wipe schemas, allowing the agent to relocate objects, stack items compactly, and free up bin space for further disposal. The resulting strategies expose and clean progressively larger areas of the desk mat, improving task quality beyond successful completion. Appendix B.2.2 and Figure 19 illustrate these executions.

Reuse across underlying execution models is addressed in our response to W3 below.

W3. The paper does not clarify whether the benefits of the knowledge as a harness generalize across different underlying execution models.

We agree that the consolidated symbolic knowledge serves as a harness for the underlying execution model. Its predicates and action schemas are represented independently of model parameters, allowing the same symbolic knowledge to support symbolic reasoning across different execution models.

To evaluate whether these benefits generalize across models, we use the same accumulated symbolic knowledge, tasks, and symbolic tools for all execution models at each evaluated checkpoint. We replace the underlying execution model and adjust the Model Interface implementation for prompt formatting and output parsing accordingly. Since neural adapters are model-specific, we recompile the validated reasoning patterns into neural memory for each execution model, using the attachment layers and representation dimensions configured for that model.

Execution Model│Initial│After Block 5│After Block 10
│SR (%, ↑)TR (%, ↓)│SR (%, ↑)TR (%, ↓)│SR (%, ↑)TR (%, ↓)
Qwen3-4B (Reference)│35.00±0.00100.00±0.00│83.17±0.2972.92±1.18│91.33±0.2967.23±0.40
Qwen2.5-3B│35.00±0.00100.00±0.00│84.67±0.2973.34±1.14│90.50±0.0068.83±0.64
Llama-3.2-3B│35.00±0.00100.00±0.00│84.17±0.2974.99±1.11│91.00±0.0072.13±0.55

Replacing the underlying execution model preserves task performance close to the reference configuration while maintaining reduced symbolic tool reliance after Block 5 and 10. Across the evaluated models, final success rates range from 90.50% to 91.33%, while symbolic tool reliance decreases from an initial 100.00% to between 67.23% and 72.13%. These results indicate that the benefits of accumulated symbolic knowledge and its model-specific compilation extend across the evaluated execution models. The symbolic predicates and action schemas provide reusable support across models, while Knowledge Compilation enables each model to generate outputs for validated reasoning patterns through its own adapters, reducing reliance on external symbolic tools.

W2 Q2. The efficiency of knowledge accumulation and the additional overhead of invoking the corresponding harness during execution warrant a dedicated analysis with clear conclusions.

To address this concern, we first examine how symbolic knowledge and compiled neural memory accumulate, together with the time and memory costs of the two learning processes. We then analyze how the shift from symbolic reasoning to compiled neural memory affects inference efficiency. The table below reports accumulated experiences, retained knowledge sizes, memory usage, and wall-clock time for Knowledge Consolidation and Knowledge Compilation.

RoundAccumulated experiences (# traces / # transitions)│Knowledge Consolidation│Knowledge Compilation
│Symbolic knowledge size (#)Host memory (GB)Wall-clock time (h)│Adapter-pool size (#)GPU memory (GB)Wall-clock time (h)
120 / 96│66811.040.65│15618.151.55
240 / 180│67411.390.13│---
360 / 213│7553.280.31│36415.252.26
480 / 337│7638.754.19│---
5100 / 475│76710.840.85│54615.545.73
6120 / 633│7677.260.21│---
7140 / 757│7749.190.56│72817.651.98
8160 / 887│7788.451.16│---
9180 / 992│78712.531.82│1,01417.663.46
10200 / 1,094│7769.530.69│1,30017.653.38

Symbolic knowledge size counts the constituent symbolic elements, including predicates, parameters, and action names, while adapter-pool size counts individual semantic and procedural adapter modules. Host and GPU memory report peak usage during consolidation and adapter training, respectively.

Symbolic knowledge size changes little after the early rounds, while neural memory continues to expand as more reasoning patterns are compiled. The reported wall-clock times total 10.57 h for ten consolidation rounds and 18.36 h for six compilation rounds. Although cumulative learning costs increase, per-round learning time and peak memory usage show no consistent upward trend within the evaluated stream. Both learning processes operate in the background, separately from online inference.

We also summarize the relative use of consolidated symbolic knowledge and compiled neural memory across learning rounds, together with inference time on seen and unseen tasks, using the results reported in Figure 3 of the main paper. This analysis examines how the increasing use of compiled neural memory relates to inference efficiency as knowledge accumulates.

RoundConsolidated symbolic knowledge inference (%)Compiled neural memory inference (%)Inference time, seen (s)Inference time, unseen (s)
0100.000.007.644.77
192.747.266.026.22
280.2819.725.224.43
369.7430.263.354.09
459.2640.743.333.76
545.5454.463.132.97
630.2369.772.952.78
717.0083.000.920.79
88.0491.960.750.57
90.5599.450.610.31
100.00100.000.210.31

Under Default inference, the share of inference handled through symbolic reasoning decreases from 100.00% at Round 0 to 0.00% at Round 10, with compiled neural memory becoming the dominant inference path from Round 5 onward. During this transition, inference time decreases from 7.64 s to 0.21 s on seen tasks and from 4.77 s to 0.31 s on unseen tasks, falling below 1 s on both task sets from Round 7 onward. Over the same rounds, task success improves from 26.14% to 81.05% on seen tasks and from 27.19% to 73.25% on unseen tasks.

Together, these analyses show that knowledge accumulation incurs background learning costs, while Knowledge Compilation reduces inference time by enabling the reuse of validated reasoning patterns with less reliance on symbolic planning.


In response to the reviewer's thoughtful feedback, we have added experimental analyses to the Appendix. Appendix A.5.7 presents the additional evaluation of accumulated knowledge reuse across different underlying execution models. Appendix A.5.5 reports knowledge accumulation and the associated time and memory costs of Knowledge Consolidation and Knowledge Compilation. Toward the end of the discussion phase, once our discussion has concluded, we will incorporate the relevant findings and clarifications into the main paper.

Reviewer_p4beOpenReview ↗

Reply to the authors

Thanks for your reply. The new details address some of my concerns. But due to the limitation of direct related experience, I still keep my rating with low confidence.

Thank you for your response and for considering our clarifications. We would be happy to address any further questions or concerns. We sincerely appreciate your time and valuable feedback.

Reviewer_VthwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents NeSyChain, a modular neuro-symbolic framework for lifelong learning in embodied agents. It integrates Knowledge Consolidation, which transforms interaction experiences into validated symbolic knowledge, and Knowledge Compilation, which converts consolidated knowledge into neural memory to reduce reliance on symbolic tools. Experiments across multiple foundation models, tools, knowledge representations, benchmarks, and real-world robotic scenarios show improvements in task success and tool efficiency. However, the methodological novelty over prior work and the evidence supporting its claims of lifelong learning and component-level generality require further clarification.

Critical strengths

Important problem. The paper addresses lifelong learning in embodied agents, aiming to accumulate and reuse knowledge while reducing repeated symbolic reasoning.

Modular design. The separation of Knowledge Consolidation and Knowledge Compilation provides a clear division of responsibilities, while standardized interfaces offer a foundation for integrating heterogeneous components.

Broad evaluation. Experiments span multiple models, tools, knowledge representations, benchmarks, and real-world robotic scenarios, with encouraging improvements in task success and symbolic tool efficiency.

Useful ablations. The ablations suggest that Knowledge Consolidation primarily contributes to task performance and transfer, whereas Knowledge Compilation mainly reduces symbolic tool reliance.

Critical weaknesses

Unclear methodological novelty. NESYC already explores contrastive symbolic knowledge refinement, while NeSyPr investigates compiling symbolic procedures into neural memory. NeSyChain's main contribution therefore appears to be their integration into a unified framework, but its additional algorithmic contributions are not sufficiently distinguished from these prior works.

Incomplete evidence for component-level decoupling. Standardized interfaces establish API-level modularity but do not necessarily guarantee semantic or algorithmic interchangeability. Different knowledge representations may require customized validation, execution logic, and reasoning workflows.

Limited evidence of synergy. The ablations demonstrate distinct contributions from the two stages, but do not fully establish whether their combination provides benefits beyond integrating existing capabilities.

Insufficiently isolated evidence for generality. Although multiple configurations are evaluated, changes in models, tools, knowledge representations, and benchmarks make it difficult to determine whether the framework remains robust under controlled component substitutions.

The benefits of lifelong learning are not directly established. Improvements from knowledge consolidation, compilation, or forward transfer do not necessarily demonstrate that repeated learning is more beneficial than one-shot knowledge acquisition. The paper does not clearly establish how performance evolves over successive learning stages or whether accumulated knowledge consistently improves future tasks under comparable budgets.

Questions for authors

What is the methodological novelty beyond NESYC and NeSyPr? Could the authors clarify which algorithmic contributions in Knowledge Consolidation and Knowledge Compilation are new, and whether NeSyChain introduces a novel joint mechanism beyond sequentially integrating existing techniques?

How interchangeable are the framework's components in practice? When replacing a foundation model, symbolic tool, or knowledge representation, what implementation changes are required? Could the authors provide controlled experiments that replace one component at a time while keeping other factors fixed?

Do the two stages provide complementary benefits beyond their individual contributions? Could the authors evaluate compilation using consolidated versus unconsolidated experiences to determine whether Knowledge Consolidation improves the quality, reliability, and transferability of compiled knowledge?

Do the experiments establish generality across configurations? Could the authors isolate changes in models, tools, and knowledge representations, and compare NeSyChain with configuration-specific methods under matched conditions?

Does NeSyChain demonstrate the benefits of lifelong learning itself? Could the authors compare one-shot learning with multi-stage learning under matched data and computational budgets, and report how performance and knowledge transfer evolve as learning stages accumulate? This would clarify whether repeated learning provides sustained benefits beyond knowledge consolidation and compilation alone.

9 comments

We appreciate Reviewer Vthw's detailed review and the insightful questions it raises about methodological novelty, component-level interchangeability, generality across configurations, the synergy between Knowledge Consolidation and Knowledge Compilation, and the benefits of lifelong learning.


W1 Q1. The methodological novelty beyond NeSyC and NeSyPr is unclear. What is algorithmically new in each stage, and is there a joint mechanism beyond sequential integration?

NeSyChain builds on NeSyC's contrastive symbolic refinement and draws on NeSyPr's concept of distilling symbolic reasoning into neural memory for reuse during LM inference. However, to address the broader problem of decoupling reasoning and lifelong learning from specific neurosymbolic configurations, NeSyChain introduces technical advances to both methods and additional mechanisms for their coordinated operation.

(a) Knowledge Consolidation

AspectNeSyCKnowledge Consolidation
Refinement scopeRules over predefined predicatesSymbolic vocabulary and action schemas
ValidationContrastive rule validationRegrounding and contrastive validation
RepresentationASPPDDL, code, and graph

(b) Knowledge Compilation

AspectNeSyPrKnowledge Compilation
Output formLM-generated plansAction chunks, code, grounded actions
Training dataOffline planner-generated plansValidated deployment-derived patterns
Neural memoryVector-quantized procedure-bookSemantic–procedural adapter pairs

(c) Coordination of the two processes

AspectNeSyC and NeSyPrNeSyChain
Pattern validation before compilationNot coupled to ongoing symbolic refinementValidation against current symbolic knowledge
Memory updates after symbolic revisionsNo shared update mechanismSelective removal or retraining of affected adapter pairs

To support lifelong learning across different neurosymbolic configurations, Knowledge Consolidation extends NeSyC's refinement of action rules over predefined predicates and parameter descriptions to the symbolic vocabulary, allowing parameter types, predicates, and action schemas to be introduced or revised. We reground stored deployment experience under each candidate vocabulary, reconstruct transitions using its action schemas, and evaluate the reconstructions with a common contrastive criterion. This extends NeSyC's ASP-based formulation to a common refinement procedure across PDDL, code, and graph representations.

Likewise, Knowledge Compilation extends NeSyPr's compilation for LM-based plan generation to foundation models producing continuous action chunks, executable code, and grounded actions. In NeSyPr's main test-time adaptation setting, neural-memory parameters remain fixed while procedures are reconstructed from feedback. NeSyChain instead continually expands neural memory during deployment by adding and training adapter pairs for newly consolidated reasoning patterns. Semantic adapters provide reconstruction-based routing, while paired procedural adapters use each target model's original training objective. New pairs are trained with the target model and existing adapters frozen.

The two processes are coordinated through current consolidated symbolic knowledge, which is used to validate candidate reasoning patterns before compilation. Each adapter pair retains its association with the symbolic knowledge and grounded bindings used to construct it. When Knowledge Consolidation revises this knowledge, Knowledge Compilation uses these associations to remove or retrain affected pairs while retaining unaffected ones. This couples symbolic refinement with the creation and maintenance of neural memory.

We evaluate the common lifelong learning procedure across three configurations while preserving their reasoning strategies. Table 2 is reproduced below; Appendix Table 29 provides block-wise results.

CL-ALFWorldReasoning Flow│Reimplemented with NeSyChain│Lifelong Learning Enabled
MethodStrategy│SR (↑)BWT (↑)FWT (↑)│SR (↑)BWT (↑)FWT (↑)
NL-PDDLGoal Regression│40.00±2.000.00±0.000.00±0.00│64.67±2.751.85±0.8511.11±1.67
NeSyRoVerified Exploration│41.00±2.600.00±0.000.00±0.00│73.00±1.7311.11±0.9621.30±0.64
CaSPCode Refinement│44.17±2.020.00±0.000.00±0.00│66.67±0.769.07±2.1012.22±1.11

Enabling the same lifelong learning procedure improves SR by an average of 26.39 percentage points and yields positive BWT and FWT across all three configurations. These results support NeSyChain's broader contribution of enabling a common lifelong learning procedure across distinct neurosymbolic reasoning flows while preserving their original strategies.

W2 Q2. Evidence for component-level decoupling is incomplete. What implementation changes does replacing a model, tool, or knowledge representation require, and can controlled one-at-a-time substitution experiments be provided?

To clarify the implementation effort required for component substitution, we expand the implementation descriptions in Appendices A.1.2, A.2.2, and A.3.2 into concrete development guidelines.

The main customization effort lies in implementing model-specific input/output handling and representation-specific grounding, execution, and validation through the Model and Tool Interfaces. This separates component-specific implementation from reusable workflow composition and execution, as in established agent frameworks [1,2]. Users can then compose reasoning flows over symbolic knowledge accessed through the Knowledge Interface. Knowledge Consolidation expands and revises this knowledge through a common learning pipeline across the supported PDDL, code, and graph representations. Knowledge Compilation additionally requires model-specific adapter configuration for a compatible execution model. Once configured, the framework automates pattern selection, adapter training, and routing as deployment experience accumulates.

We extend Appendix A.5.4 with controlled experiments that change one component at a time from the reference configuration, keeping the remaining components and reasoning flow fixed. All configurations share the task stream, learning schedule, and evaluation protocol and use the same consolidated symbolic knowledge at each evaluated checkpoint.

Execution model. We update the Model Interface for prompt formatting and output parsing and recompile the adapters with model-specific attachment layers and representation dimensions.

Symbolic tool. We replace the planner's search configuration with LAMA-2011 or GBFS with the FF heuristic by changing its invocation within the Tool Interface.

Knowledge representation. We convert PDDL knowledge to Python code while preserving task and action semantics, and implement code-based precondition checking, successor-state computation, and plan validation. PDDL is derived from this code solely to provide input to the unchanged planner.

The table below reports task performance, knowledge transfer, and symbolic tool reliance for each configuration.

Substitution│Configuration│Evaluation Results
│Execution modelSymbolic plannerKnowledge formatBenchmark│SR (↑)BWT (↑)FWT (↑)TR (↓)BRR (↑)FRR (↑)
None (Reference)│Qwen3-4BLAMA-firstPDDLCL-ALFWorld│91.33±0.295.37±0.3243.89±0.0067.23±0.404.29±0.6619.20±0.59
Execution model│Qwen2.5-3BLAMA-firstPDDLCL-ALFWorld│90.50±0.005.20±0.3545.00±0.6068.83±0.643.80±0.8917.17±0.70
Execution model│Llama-3.2-3BLAMA-firstPDDLCL-ALFWorld│91.00±0.004.40±0.0044.40±0.0072.13±0.553.33±0.7514.60±0.62
Symbolic planner│Qwen3-4BLAMA-2011PDDLCL-ALFWorld│94.00±0.001.10±0.0052.60±0.3568.60±0.205.77±0.9217.37±0.65
Symbolic planner│Qwen3-4BGBFS+FFPDDLCL-ALFWorld│93.50±0.501.67±0.5549.40±0.0068.50±0.446.07±0.7517.07±0.71
Knowledge format│Qwen3-4BLAMA-firstPython codeCL-ALFWorld│91.33±0.295.37±0.3243.89±0.0067.22±0.394.35±0.6019.14±0.57

Across all evaluated configurations, SR remains between 90.50% and 94.00%, with positive BWT, FWT, BRR, and FRR. Model substitution largely preserves SR and FWT but increases symbolic tool reliance. Changing the planner's search configuration improves SR and FWT while reducing BWT, indicating that search behavior affects how the same knowledge is used. Representation substitution yields nearly identical results, consistent with preserving knowledge semantics across representations.

These results demonstrate that the evaluated component substitutions can be accommodated through localized implementation changes while preserving the overall reasoning flow.

[1] Harrison Chase. LangChain, 2022.

[2] LangChain Team. LangGraph, 2024.

W3 Q3. Evidence of synergy between the two stages is limited. Does compiling consolidated rather than unconsolidated experiences improve the quality, reliability, and transferability of compiled knowledge?

Knowledge Consolidation and Knowledge Compilation provide complementary benefits, and their alignment further supports task performance and transfer. Our ablation in Table 3 evaluates all four combinations of enabling and disabling the two stages, including compilation with and without Knowledge Consolidation. We extend this ablation with an additional variant without Consolidation-Compilation Alignment.

In the w/o Knowledge Consolidation variant, symbolic knowledge remains fixed at initialization, and Knowledge Compilation continues to operate on accumulated unconsolidated experience. In the w/o Knowledge Compilation variant, symbolic knowledge is updated, but the resulting reasoning patterns are not compiled into neural memory. In the w/o Consolidation-Compilation Alignment variant, both stages remain enabled, and neural memory accumulates in an append-only manner as new reasoning patterns are compiled.

Method│Task Performance│Symbolic Tool Reliance
│SR (↑)BWT (↑)FWT (↑)│TR (↓)BRR (↑)FRR (↑)
NeSyChain│91.33±0.295.37±0.3243.89±0.00│67.23±0.404.29±0.6619.20±0.59
w/o Consolidation–Compilation Alignment│76.50±0.504.44±0.5632.04±0.85│65.73±0.583.33±1.1222.27±0.66
w/o Knowledge Consolidation│34.33±1.53−0.74±0.640.56±0.56│78.12±0.212.82±0.0712.79±0.47
w/o Knowledge Compilation│88.67±1.614.07±1.4043.89±2.00│100.00±0.000.00±0.000.00±0.00
w/o Both│33.67±1.150.00±0.000.00±0.00│100.00±0.000.00±0.000.00±0.00

The w/o Knowledge Consolidation variant achieves 34.33% SR and 0.56 percentage points FWT despite reducing TR to 78.12%, whereas w/o Knowledge Compilation retains 88.67% SR and 43.89 percentage points FWT but relies entirely on symbolic tools (100.00% TR). More importantly, w/o Consolidation-Compilation Alignment reduces SR by 14.83 percentage points and FWT by 11.85 percentage points compared with NeSyChain, despite both stages remaining enabled. These results indicate that updating existing neural memory in response to symbolic knowledge revisions provides additional benefits for task performance and transfer beyond appending newly compiled reasoning patterns.

To illustrate how Knowledge Consolidation guides updates to compiled neural memory, we track a subset of semantic-procedural adapter pairs associated with selected symbolic knowledge and reasoning patterns. New counts pairs added for newly compiled reasoning patterns, while Retained, Retrained, and Replaced describe how existing pairs are handled. Subset Size counts the pairs in this tracked subset after each round.

Round (Block)Symbolic knowledge change│Tracked adapter pairs (#)
│NewRetainedRetrainedReplacedSubset size
1 (1)1 operator revised│6---6
2 (3)2 operators and 1 predicate added│433010
3 (5)1 operator revised│163111
4 (7)1 operator revised and 1 predicate added│064111

Within the tracked subset, symbolic knowledge revisions lead to the retraining or replacement of affected adapter pairs, while pairs whose supporting knowledge and input contexts remain valid are retained. Following Block 7, four pairs are retrained and one replaced without adding new pairs, leaving Subset Size unchanged at 11. These results illustrate how Knowledge Consolidation maintains previously compiled neural memory as symbolic knowledge evolves.

Reviewer_P8HPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents NeSyChain, a modular framework that consolidates deployment experience into symbolic knowledge and compiles validated reasoning patterns into neural memory.

Critical strengths
  • The separation between explicit knowledge acquisition and efficient neural reuse is well motivated.
  • The improvements over the strongest reported baselines are substantial across all three benchmarks.
Critical weaknesses
  • The scalability of lifelong learning remains insufficiently established. The evaluation primarily considers bounded, structured task streams, while accumulated experience and growing adapter pools may introduce substantial maintenance costs over time.

  • In the held-out-task analysis, consolidated-only inference achieves 79.82% success, compared with 73.25% for the final default/compiled system, although the latter is much faster. The paper acknowledges this trade-off, but a clearer mechanism for retaining symbolic reasoning when needed would strengthen the practical contribution.

Questions for authors
  • How do total computation, memory usage, and maintenance costs grow as the task stream and adapter pools become larger?
  • Could the authors clarify how the system decides when to use symbolic reasoning rather than compiled neural memory, and whether this decision can better balance success rate and inference cost?
6 comments

We sincerely thank the reviewer for the thoughtful and constructive feedback. We have carefully addressed each comment below and indicated where the paper has been revised. We would be happy to address any remaining concerns during the discussion.


W1 Q1. The scalability of lifelong learning remains insufficiently established. How do total computation, memory usage, and maintenance costs grow as the task stream and adapter pools become larger?

To address concerns about scalability and maintenance costs over time, we measured total computation, memory usage, and maintenance costs throughout the task stream, separately for Knowledge Consolidation and Knowledge Compilation.

For Knowledge Consolidation, we report symbolic knowledge size, symbolic tool costs for symbol regrounding and contrastive validation, LM call and token counts for generating candidate symbolic knowledge, and host memory usage. For Knowledge Compilation, we report adapter-pool size, symbolic tool costs for pattern selection, trainable parameter counts and training FLOPs for adapter compilation, and GPU memory usage. We also report wall-clock time separately for the two processes.

For these measurements, Knowledge Consolidation was performed after every block, while Knowledge Compilation was performed only when new or revised symbolic knowledge required compilation and sufficient experience had accumulated to select reusable patterns.

RoundSymbolic knowledge size (#)Symbolic tool calls (#)LM calls (#)LM tokens (k)Host memory usage (GB)Wall-clock time (h)
166819,11311259.6911.040.65
2674358251.4111.390.13
375511,78720331.673.280.31
4763130,81869931.598.754.19
576718,5073101.4710.840.85
6767198126.807.260.21
77747,2564132.319.190.56
877829,83632426.698.451.16
978746,76433421.2112.531.82
1077622,53032269.579.530.69

In the table above, symbolic knowledge size is measured by counting symbolic elements, including predicates, parameters, and action names. Symbolic tool calls count Fast Downward invocations during symbol regrounding and contrastive validation. LM calls and total token usage are measured for GPT-5.6-Sol during symbol abstraction. Host memory is the peak usage during each consolidation round, and wall-clock time covers the consolidation process in that round.

The computational cost of Knowledge Consolidation varies across rounds, with a total measured wall-clock time of 10.57 hours. Peak host memory ranges from 3.28 to 12.53 GB and shows no consistent upward trend across the measured rounds.

Maintenance includes candidate generation, symbol regrounding, and contrastive validation against accumulated experience. The associated costs are included in the reported LM usage, symbolic tool calls, and wall-clock time.

RoundAdapter pool size (#)Pattern selection validation checks (#)Trainable parameters (M)Training FLOPs (PFLOPs)GPU memory usage (GB)Wall-clock time (h)
1156678128.22487.0218.151.55
2------
33642,558353.83494.9415.252.26
4------
55465,609345.241,411.8815.545.73
6------
7728910141.00585.9017.651.98
8------
91,0143,495217.891,110.7017.663.46
101,3003,462407.30763.6517.653.38

In the table above, adapter-pool size is the total number of active semantic and procedural adapter modules retained after each round. Validation checks are counted during pattern selection. Trainable parameter counts and training FLOPs cover the adapter pairs trained in each round. GPU memory is the peak usage during adapter training, and wall-clock time covers the compilation process in that round.

Despite the growth of the adapter pool from 156 to 1,300 modules, per-round training costs do not increase consistently. Training FLOPs and wall-clock time peak in Round 5 at 1,411.88 PFLOPs and 5.73 hours, respectively, while peak GPU memory remains between 15.25 and 18.15 GB across the six reported compilation rounds. Maintenance involves removing or retraining adapter pairs affected by revisions to consolidated symbolic knowledge, while unaffected pairs are retained. Retraining costs are included in the reported training FLOPs and wall-clock time.

The reported wall-clock times sum to approximately 28.93 hours across the two learning processes, covering ten consolidation rounds and six compilation rounds. While cumulative computational costs increase as learning continues, the reported per-round costs and peak memory usage show no consistent upward trend as the task stream progresses and the adapter pool grows. We note that Knowledge Consolidation and Knowledge Compilation operate as background learning processes, separate from online inference. Their computational costs are therefore reported separately from online inference latency measurements, such as those reported in Figure 3b and Appendix A.5.2.

W2 Q2. A clearer mechanism for retaining symbolic reasoning when needed would strengthen the practical contribution. Could the authors clarify how the system decides when to use symbolic reasoning rather than compiled neural memory, and whether this decision can better balance success rate and inference cost?

During inference, NeSyChain first identifies the semantic adapter with the lowest normalized reconstruction error emin⁡(xt)e_{\min}(x_t) for the current reasoning context xtx_t. It then computes an applicability score st=gHt−1(11+emin⁡(xt))s_t = g_{\mathcal{H}_{t-1}}\left(\frac{1}{1+e_{\min}(x_t)}\right), where gHt−1g_{\mathcal{H}_{t-1}} is a calibration function fitted on the pass/fail outcomes of precondition checks recorded in the execution history before step tt. If the score is at least τ\tau, the paired procedural adapter generates a candidate action; otherwise, the system uses symbolic reasoning.

The generated action then undergoes a lightweight deterministic executability check that verifies its preconditions against the current symbolic state using the consolidated symbolic knowledge. If the check fails, the system falls back to symbolic reasoning. A higher τ\tau therefore makes the use of compiled neural memory more conservative and increases reliance on symbolic reasoning.

We acknowledge that Appendix A.3.2 described reconstruction-based adapter selection and validity-based fallback, but did not explicitly explain the threshold-based routing criterion. We have revised this section to clarify the decision procedure and the role of τ\tau.

To examine how this threshold affects the balance between task success and symbolic reasoning cost, we evaluate τ∈{0.5,0.8,0.95,0.99}\tau \in \{0.5, 0.8, 0.95, 0.99\}, with τ=0.95\tau=0.95 as the default setting. We report task success rate, symbolic tool reliance, and symbolic tool calls per task. The earlier- and later-stage results are both based on evaluations across evaluation blocks.

Threshold│Early│Late
│SR (%, ↑)TR (%, ↓)Symbolic tool calls (#/task, ↓)│SR (%, ↑)TR (%, ↓)Symbolic tool calls (#/task, ↓)
τ=0.5\tau=0.5│66.44±0.9680.75±1.496.04±0.13│89.83±0.3817.57±1.641.26±0.34
τ=0.8\tau=0.8│68.44±0.6982.84±0.746.44±0.10│91.75±0.0061.17±0.334.04±0.02
τ=0.95\tau=0.95 (default)│70.78±0.6984.42±0.586.56±0.11│90.50±0.0065.48±1.234.19±0.09
τ=0.99\tau=0.99│72.22±0.6986.46±0.526.72±0.02│90.58±0.1471.04±0.394.56±0.02

In the early stage, retaining more symbolic reasoning improves task success with a modest increase in tool use. Increasing τ\tau from 0.5 to the default 0.95 improves SR by 4.34 percentage points, with symbolic tool calls increasing from 6.04 to 6.56 per task. In the later stage, adjusting τ\tau can improve the balance between task success and symbolic reasoning cost. Compared with the default τ=0.95\tau=0.95, τ=0.8\tau=0.8 achieves higher SR (91.75% vs. 90.50%) with fewer symbolic tool calls (4.04 vs. 4.19 per task). Lowering τ\tau further to 0.5 reduces symbolic tool calls to 1.26 per task, while SR decreases by 1.92 percentage points to 89.83%. These results show that the threshold offers a practical means of adjusting reliance on symbolic reasoning as learning progresses.


In response to the reviewer's thoughtful feedback, we have added experimental analyses and clarifications to the Appendix. Appendix A.5.5 reports the computation, memory usage, and maintenance costs of Knowledge Consolidation and Knowledge Compilation across learning rounds. Appendices A.1.2, A.2.2, and A.3.2 clarify the decision procedure for accepting neural proposals or falling back to symbolic reasoning, while Appendix A.5.6 presents the effects of varying τ\tau on task success and symbolic reasoning cost. Toward the end of the discussion phase, once our discussion has concluded, we will consolidate these findings and clarifications and update the main paper accordingly.

Area_Chair_DGshOpenReview ↗

Please follow up before the discussion period ends & AI use clarification

Dear reviewer,

We are less than a day away from the end of the discussion period. Please read the authors' response and any revisions, and post a brief follow-up indicating which concerns remain and whether your assessment has changed. Please raise any outstanding questions promptly so the authors have time to respond.

Given the limited time, please focus on clarifications and existing evidence. You may identify gaps in the experimental support, but please do not expect the authors to conduct new experiments before the discussion period closes.

Additionally, your review was flagged for AI use while the corresponding field in your review does not mention how it was used. Please clarify this for me.

Thank you for your contribution!

Best,

Your AC

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely appreciate your time and effort in reviewing our submission and engaging in the discussion. To consolidate the discussion, we briefly summarize the strengths highlighted across the reviews and the three main points addressed in our responses.

Strengths Highlighted Across Reviews

  • Well-Motivated Neurosymbolic Framework
    • The reviewers (P8HP, Vthw, p4be) collectively recognized the value of a modular neurosymbolic design in which Knowledge Consolidation and Knowledge Compilation support lifelong learning across diverse neurosymbolic configurations.
  • Strong Performance and Efficiency Improvements
    • The reviewers (P8HP, Vthw, p4be) collectively highlighted strong empirical results, with Knowledge Consolidation contributing to task success and transfer, and Knowledge Compilation reducing reliance on symbolic tools.
  • Comprehensive Experimental Validation
    • The reviewers (P8HP, Vthw, p4be) collectively appreciated the broad evaluation across multiple models, tools, knowledge representations, and benchmarks, supported by detailed ablations and real-world robot experiments.

Key Discussion Points and Responses

  • Further Validation of Framework Generality and Knowledge Reuse (Vthw W2/W4; p4be W1/W3)

    • Discussion Focus. The discussion clarified that NeSyChain's generality extends beyond component interchangeability, as its common lifelong learning procedure and consolidated symbolic knowledge remain reusable across changing configurations and tasks. When the execution model changes, neural memory can be efficiently recompiled from the same validated reasoning patterns.
    • Reviewer Requests. Reviewer Vthw requested controlled one-at-a-time component substitution experiments, while reviewer p4be asked for concrete evidence of knowledge reuse across tasks, benchmarks, and execution models.
    • Our Responses. We added the requested controlled experiments and clarified the reuse properties of symbolic knowledge and neural memory across configurations. We also elaborated on the existing real-world cases in Figure 4 to demonstrate knowledge reuse across scenarios.
    • References. Updated Table 2(a) (Section 4.3 of the main paper); added details in Appendix Sections A.1.2, A.2.2, A.3.2, and A.5.7.
  • Further Analysis of Lifelong Learning Costs and Benefits (P8HP W1/W2; p4be W2; Vthw W5)

    • Discussion Focus. The discussion clarified that NeSyChain's lifelong learning processes incur computational costs in the background, independently of online inference. It also showed that inference routing provides a controllable trade-off between task success and symbolic reasoning costs.
    • Reviewer Requests. Reviewers P8HP and p4be requested analyses of learning and inference costs as knowledge accumulates, with P8HP additionally asking about inference routing. Reviewer Vthw requested a matched-budget comparison between successive and one-shot learning.
    • Our Responses. We reported per-round learning costs, analyzed the inference routing threshold, and added a matched-budget comparison against one-shot learning using identical execution trajectories.
    • References. Added experimental results in Appendix Sections A.3.2, A.5.5, A.5.6, and A.5.10.
  • Framework-Level Generality Beyond Sequential Integration (Vthw W1/W3)

    • Discussion Focus. The discussion clarified how NeSyChain, as a standardized framework, supports diverse neurosymbolic embodied agents and enables lifelong learning even in existing agents not originally designed for it.
    • Reviewer Requests. Reviewer Vthw asked what technical advances were introduced in integrating NeSyC and NeSyPr, and whether aligning Knowledge Consolidation and Knowledge Compilation provides benefits beyond their sequential combination.
    • Our Responses. We clarified that NeSyChain's framework-level contribution is a unified design for diverse neurosymbolic embodied agents with lifelong learning. While Knowledge Consolidation and Knowledge Compilation build on prior work, we also explained how both processes are extended beyond specific symbolic configurations and coordinated within this framework. We provided supporting evidence through evaluations of the common lifelong learning procedure across three distinct reasoning flows and an ablation of the alignment between the two learning processes.
    • References. Table 2(b) and updated Table 3 (Section 4.3 of the main paper); added details in Appendix Sections A.5.8 and A.5.9.

We have updated our submission to reflect the clarifications and additional evidence presented during the rebuttal process and discussions with the reviewers. We sincerely thank the Area Chair and reviewers again for their time and engagement throughout the review process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.