PaperlayerView exact revision on arXiv

Source-linked AI summary

SimSkill: A Lifelong Learning AI Agent for Autonomous Mastery of Traffic Simulation

Qi Liu, Qinzheng Wang, Yiming Bie

arXiv:2609.03753v1cs.AIcs.MA

TL;DR

LLM agents face an open question of how to turn experience into transferable capability and organize knowledge for continued learning. SimSkill autonomously explores SUMO tasks and consolidates reusable experience in persistent memory without updating the backbone model. On held-out benchmarks, verified-completion gains were backbone-dependent, reaching 25 percentage points for Qwen3.7-Max on V1.

  • Problem

    A central open question is how LLM agents can turn experience into transferable capabilities and organize accumulated knowledge for continued learning.

    Source: §1 · pp. 1–2

  • Method

    SimSkill autonomously generates SUMO tasks, learns through environment-grounded attempts and evaluation, and consolidates reusable outcomes in episodic, procedural, and semantic memory without updating the backbone LLM.

    Source: §1 · p. 2

  • Results

    Verified-completion gains over vanilla Claude Code were backbone-dependent, reaching +25 percentage points for Qwen3.7-Max on V1; GLM-5.2 showed no improvement.

    Source: §1 · p. 3, §6 · p. 26

  • Takeaways & Limitations

    Both successful and unsuccessful SUMO experience can contribute to creating, revising, and maintaining reusable memory.

    Source: §6 · p. 26

    Retrieval adequacy is demonstrated at the observed scale of 150 skills and 153 pages, not for libraries of tens of thousands of artifacts.

    Source: §4.1 · p. 15

Abstract

from arXiv · show

As large language models (LLMs) become increasingly capable, the long-term value of AI systems depends not only on solving individual requests, but also on transforming experience and accumulated knowledge into durable, reusable competence. We introduce SimSkill, a self-evolving agent built around the Simulation of Urban MObility (SUMO) traffic simulator. SimSkill identifies capability gaps, generates and solves environment-grounded tasks, verifies solutions through an action--critic loop, and consolidates experience into episodic, procedural, and semantic memory without updating the backbone model. Through autonomous exploration, it builds a reusable library spanning the traffic-simulation workflow. We evaluate SimSkill on two held-out benchmarks with three backbone LLMs and independent artifact-based verification. SimSkill improves verified completion by up to 25 percentage points, while ablations show complementary contributions from procedural and semantic memory. Its benefits remain backbone- and budget-dependent: memory does not improve every model or uniformly reduce inference cost. More broadly, SimSkill illustrates a design paradigm in which natural language preserves and composes computational capabilities, while executable tools and code provide precise and reproducible execution. All code and experimental data are publicly available at https://github.com/qiliuchn/SimSkill-V1.

1 Introduction

SimSkill frames traffic-simulator mastery as lifelong learning: an agent should turn environment-grounded experience into reusable competence across tasks, rather than solve each simulation request in isolation. It autonomously generates tasks, learns through execution and evaluation, and consolidates outcomes into persistent memory without updating the backbone LLM.

Introduction

  • Traffic-simulation agents largely address individual tasks, leaving validated procedures and lessons underused across compositional workflows that span modeling, calibration, control, and analysis.The introduction identifies self-directed experience generation, distillation into transferable capabilities, and organization of accumulated knowledge as central lifelong-learning challenges.Source: §1 · pp. 1–2
  • SimSkill uses natural language as an organizing layer for selecting, adapting, and composing executable components, while tools, algorithms, and code provide precise and reproducible execution.The paper presents this division of labor as a general design principle for self-evolving agents beyond traffic simulation.Source: §1 · p. 2
  • SimSkill closes the loop among autonomous curriculum generation, environment-grounded task execution, and experience consolidation to acquire capabilities across tasks without updating the backbone LLM.It proposes tasks that extend current competence, attempts and evaluates them, then consolidates reusable outcomes into persistent memory.Source: §1 · p. 2
  • Its tripartite memory preserves attempts and evidence, reusable skills with executable resources, and linked domain knowledge; bounded maintenance supported 150 procedural skills and 153 semantic pages after about 80 hours over five days.The resulting artifacts are inspectable, editable, composable, and transferable across compatible backbones.Source: §1 · p. 2
  • +10 and +20 percentage points on V1/V2 for DeepSeek-V4-Pro and +25 points on V1 for Qwen3.7-Max versus vanilla Claude Code; GLM-5.2 did not improve.On V2, Qwen3.7-Max completed 2 versus 0 tasks. Ablations showed complementary contributions from procedural and semantic memory, with model-dependent accuracy–cost trade-offs.Source: §1 · pp. 2–3

2 Related Work

Related work spans self-improving LLMs, lifelong learning, agent memory, and LLM-based traffic simulation. SimSkill combines environment-grounded exploration with persistent memory to extend capabilities across traffic-simulation tasks without updating the backbone model.

2.1 Self-Improvement and Self-Evolution of Large Language Models

  • Self-improvement converts model-generated examples into parameter updates, while interactive agents learn through exploration and feedback; SimSkill instead stores grounded experience in persistent memory.Prior work includes self-training [TLC+24] [WKM+23] [HGH+23] [GPS+23], exploration [CTO+23], and feedback-based reflection [SCG+23] [QZGK24].Source: §2.1 · p. 3

2.2 Experience-Driven Lifelong Learning Agents

  • Lifelong agents use external skill libraries to retain and transfer capabilities, but face challenges in abstraction, retrieval, curriculum design, and forgetting.Minecraft and robot skill libraries demonstrate this approach [WXJ+23] [ZCT+23] [TK24]; later work studies reusable skill artifacts [YLP+26] [YGH+26] and finds raw replay limited by irrelevant information and context constraints [ZCL+25]. SimSkill extends this line to traffic simulation with procedural and semantic memory.Source: §2.2 · pp. 3–4

2.3 Long-Term Memory for LLM Agents

  • Agent memory extends retrieval with write-and-maintain mechanisms, evolving from stored conversations and reflections toward structured, revisable knowledge systems.SimSkill combines episodic evidence, procedural skills, and semantic knowledge in linked memory stores, drawing on work in memory management [LPP+20] [ZGG+24] [POC+23] [LLM26] [PWL+23], consolidation [CKA+25] [SCY+25] [XLM+26], and artifact-based knowledge representations.Source: §2.3 · p. 4

2.4 Large Language Model Agents for Traffic Simulation

  • Prior LLM traffic systems address analysis, control, or individual SUMO studies, whereas SimSkill focuses on autonomous cross-task learning that expands persistent capabilities over time.Related work also uses generative agents to represent travelers in mobility simulations [LLM26]. SUMO agents span scenario generation [LAK24], tool composition [YXS+25], simulation planning [JCY25], experiment design [DZF+25], and feedback-driven refinement.Source: §2.4 · pp. 4–5

3 Method

Section 3 describes SimSkill’s method for expanding lifelong competence through environment-grounded tasks and persistent external memory rather than parameter updates. It covers memory, retrieval and maintenance, bounded active context, shared execution for learning and user tasks, the autonomous curriculum–execution–critique loop, and SUMO specialization.

3.1 Conceptual Foundations and Framework Overview

  • The conceptual objective prioritizes broad competence across tasks, guiding a curriculum toward novelty, diversity, practical value, gap coverage, and rising difficulty rather than repeated easy-task wins.The objective guides curriculum design; SimSkill does not estimate the expectation during learning.Source: §3.1 · p. 5
  • SimSkill coordinates episodic, procedural, and semantic memory through a runtime whose specialized agents propose tasks, execute them, and independently evaluate evidence before outcomes are consolidated.Its recurring loop retrieves relevant memory, acts in the environment, evaluates the result, and begins the next iteration from updated memory.Source: §3.1 · p. 6, Figure 1 · p. 6

3.2 Tripartite Memory and Experience Consolidation

  • Episodic memory preserves timestamped task attempts, evidence, failures, and deliverables, while procedural skills combine adaptable natural-language instructions with repeatable scripts that can be inspected, composed, and reused.Skills remain persistent and shareable across sessions without modifying the backbone model.Source: §3.2 · pp. 6–7
  • Semantic memory compiles knowledge into linked, provenance-bearing pages, while metadata-first lazy retrieval and consolidation update or create reusable skills and pages without proliferating near-duplicates.Retrieval loads full content only when needed; consolidation can produce procedural and declarative artifacts from the same episode.Source: §3.2 · pp. 8–9
  • Periodic memory maintenance compares changed artifacts with the collection, merging duplicates, repairing references, removing superseded items, and validating externally contributed skills; automatic forgetting by access frequency remains future work.Incremental linting runs after a configurable number of changes, ten by default, while full linting checks the entire collection.Source: §3.2 · p. 10

3.3 Autonomous Self-Evolution

  • The curriculum targets novel, gap-filling, progressively harder tasks across SUMO workflow areas, while executed artifacts face independent evidence-based criticism and up to five inner and three outer retries.The critic checks reproducibility, whether claims match artifacts, silent failures, and coverage of requested subtasks.Source: §3.3 · pp. 10–12
  • Critic feedback carries across outer attempts while raw debugging stays local; episodes update memory only when they provide reusable evidence.Failed tasks can reveal limitations or invalid assumptions, while successful repetitions need not change memory; learning pauses after ten consecutive iterations without procedural or semantic change.Source: §3.3 · p. 12
  • Lazy, metadata-first retrieval limits candidates to ten, while role-specific agents and finite retries bound active context as external memory grows.Routine log operations return compact outputs rather than loading the full history; detailed artifacts remain in external storage.Source: §3.3 · pp. 12–13

3.4 Learning-Time and User-Directed Task Execution

  • One infer skill serves autonomous and user tasks; normal mode permits memory ingestion, test mode freezes memory, and execution policy can be configured independently.The experiments evaluate the current shared action–critic implementation, not alternative deployment policies that trade assurance against latency or cost.Source: §3.4 · pp. 13–14

3.5 Instantiation in the SUMO Environment

  • For SUMO, curriculum, action, critic, and memory are specialized to traffic simulation, with artifact-based checks across its compositional workflow.The critic checks routing, warnings, outputs, metrics, and multi-stage analysis completeness.Source: §3.5 · p. 14
  • The action agent uses native scripts, SUMO command-line tools, output files, and TraCI; self-evolution needs no MCP layer and can expand from a small seed or empty stores.SUMO-specific content resides primarily in accumulated memory, while the learning mechanism can apply to other executable domains with tools, observable outcomes, and reusable structure.Source: §3.5 · p. 14

4 Case Studies

The case studies trace SimSkill’s autonomous learning from capability-gap discovery through task execution, critique, memory updates, and collection-wide maintenance. They show how accumulated procedural and semantic memory supports reusable competence while requiring verification and upkeep.

4.6 Accumulated Capability and Retrieval Scale

  • After about 80 hours over five days, SimSkill had produced 150 procedural skills and 153 semantic-memory pages spanning the traffic-simulation workflow.The inventory categorizes all 150 skills by primary function, while noting that individual skills can span categories and link to others.Source: §4 · p. 14, §4.6 · p. 17

4.1 Curriculum Construction, Retrieval, and Task Decomposition

  • Retrieval combined eight procedural skills and two semantic pages into a task-specific workflow, while curriculum search identified missing decision variables and predictive control capabilities.At the observed 150-skill, 153-page scale, file search with metadata-first retrieval was adequate, though the paper cautions that much larger collections may need different retrieval methods.Source: §4.1 · pp. 14–15

4.2 Learning from an Unsuccessful Task

  • A false claim about SUMO’s micromobility limitations yielded corrected semantic knowledge and a curriculum policy requiring evidence before inferring missing simulator capabilities.SimSkill created no procedural skill; it recorded the tested boundary between native functionality and external control logic, and learned to treat characterization and validation as useful tasks.Source: §4.2 · p. 15

4.3 Acquisition of a Predictive-Control Skill

  • The predictive-control episode produced a new skill and semantic page despite establishing no advantage for predictive control.Retrieval supplied adjacent capabilities; after the critic rejected an insufficiently verified first attempt and a design confound, a revised and independently validated attempt was accepted.Source: §4.3 · p. 16

4.4 Compositional Acquisition and Consistent Repair

  • GPS map matching composed existing procedures into a new skill and scripts, while reuse exposed a defect that was repaired and committed with related memory updates.The case demonstrates that reuse supports both capability composition and continued validation of earlier skills.Source: §4.4 · pp. 16–17

4.5 Memory Maintenance as Library-Level Learning

  • Memory-lint maintains collection health by checking new artifacts against existing memory, repairing defects, updating references, and controlling redundancy without speculative changes.Checks and modifications are logged for audit, preserving a reliable foundation for later retrieval and skill composition.Source: §4.5 · p. 17

4.7 Flexibility and Limits of Language-Specified Orchestration

  • Natural-language skills make orchestration transparent and extensible, but weaker orchestrators may stop after task execution and skip criticism, recording, or ingestion.Workflow instructions and an external completeness check require the remaining steps and resume sessions when mandatory artifacts are missing, though safeguards do not remove backbone dependence.Source: §4.7 · pp. 17–18

5 Experiments

Across two held-out traffic-simulation benchmarks, SimSkill improved independently verified completion for two of three tested backbones, including on difficult tasks, but its benefits depended on the model and resource budget. Ablations found contributions from both procedural and semantic memory, while cost savings were not universal.

5.1 Experimental Setup

  • The study uses two frozen 40-task benchmarks and independently verifies completion rather than relying on systems' self-reported success.The setup compares complete SimSkill with vanilla Claude Code across three backbones; the five-condition ablation uses V1 with DeepSeek-V4-Pro and Qwen3.7-Max. Claude Opus 5 judges artifacts in a separate context-free session, and GLM-5.2 additionally assigns pointwise scores to V2 runs for two backbones.Source: §5.1 · pp. 19–20

5.2 Main Results

  • Memory does not uniformly reduce inference cost: DeepSeek-V4-Pro and Qwen3.7-Max reach higher V1 endpoints with higher median costs, while GLM-5.2 saves cost and time without improving completion.On V1, median costs rise from $0.49 to $0.78 for DeepSeek-V4-Pro and from $1.03 to $1.84 for Qwen3.7-Max; GLM-5.2's median cost falls approximately 31% and median time approximately 52%. On V2, DeepSeek-V4-Pro's median cost rises from $2.92 to $3.93 for eight additional completions, while its median elapsed time is slightly lower. Success-at-budget curves show verified coverage by expenditure or time.Source: §5.2 · pp. 20, 23, Figure 5 · p. 21, Figure 6 · p. 22
  • On verified completion, SimSkill raises DeepSeek-V4-Pro from 34 to 38 tasks on V1 and from 19 to 27 on V2, gains of 10 and 20 percentage points.Each benchmark contains 40 tasks. The comparison is between complete SimSkill and vanilla Claude Code.Source: Table 7 · p. 23, §5.2 · p. 23
  • Gains extend beyond direct skill recall: Qwen3.7-Max improves in every V1 tier, and DeepSeek-V4-Pro gains eight verified tasks across V2 Tiers 3 and 4.The V1 Tier 4 gain is four tasks among 11; on V2, DeepSeek-V4-Pro gains two tasks in Tier 3 (11/20 versus 9/20) and six in Tier 4 (16/20 versus 10/20). Both Qwen3.7-Max successes on V2 are Tier 4 tasks.Source: §5.2 · p. 23
  • Completion gains are model-dependent: Qwen3.7-Max improves from 13 to 23 tasks on V1 and from zero to two on V2, while GLM-5.2 shows no gain.On V2, GLM-5.2 ties vanilla; on V1, SimSkill trails by one task. The independent pointwise judge reports DeepSeek-V4-Pro means of 0.940 versus 0.891 and Qwen3.7-Max means of 0.728 versus 0.674 for SimSkill versus vanilla; DeepSeek-V4-Pro medians are 0.955 versus 0.965, and Qwen3.7-Max medians are 0.770 versus 0.685.Source: §5.2 · pp. 23–24, Figure 7 · p. 24

5.3 Ablation Study

  • Both memory types contribute in the V1 ablation, with procedural memory contributing slightly more and their endpoint effects adding rather than interacting super-additively.DeepSeek-V4-Pro endpoints are 34 tasks for inference-framework-only, 35 with semantic memory, 37 with procedural memory, and 38 with both. Qwen3.7-Max endpoints are 16, 19, 20, and 23, respectively; the factorial interaction contrast is zero for both backbones.Source: §5.3 · pp. 24, 26, Figure 8 · p. 25, Table 8 · p. 26
  • Ablation resource effects differ by backbone: DeepSeek-V4-Pro's full system is cheaper and faster than either single-memory variant, whereas Qwen3.7-Max's advantage emerges mainly at long budgets.For DeepSeek-V4-Pro, full-system median cost and time are $0.78 and 1056 s, versus $1.41 and 2423 s with procedural memory alone and $1.31 and 2383 s with semantic memory alone. The endpoint ordering does not hold at every intermediate cost or time threshold.Source: §5.3 · p. 26

5.4 Scope of the Evidence

  • Overall, the evidence supports expanded task coverage rather than universal cost savings: memory benefits depend on the backbone, while both memory forms contribute.The study reports gains for two of three backbones and advantages on hard compositional and novel tasks, alongside retrieval and orchestration overhead that frequently raises total inference cost.Source: §5.4 · p. 26

6 Conclusion · A Supplementary Traces for the Learning Case Studies

SimSkill builds reusable competence from verified interaction with SUMO by consolidating experience into memory without updating its backbone model. Its benchmark gains were substantial but model-dependent, and the supplementary traces document how autonomous learning episodes changed memory.

6 Conclusion

  • After approximately 80 hours, SimSkill had accumulated 150 procedural skills and 153 semantic-memory pages through autonomous curriculum generation, verified execution, and episodic, procedural, and semantic memory.The case studies show successful and unsuccessful experience creating, revising, and maintaining reusable memory.Source: §6 · p. 26
  • 10 percentage points on Benchmark V1 and 20 points on Benchmark V2 for DeepSeek-V4-Pro, plus 25 points on V1 for Qwen3.7-Max, were verified-completion gains; Qwen3.7-Max also completed two V2 tasks versus none for vanilla.Both benchmarks contained 40 tasks.Source: §6 · p. 26
  • Procedural and semantic memory contributed additively, but gains were not universal: GLM-5.2 did not improve, and broader task coverage did not consistently reduce inference cost.The reported benchmarks each contained 40 tasks.Source: §6 · p. 26
  • Future work should test repeated runs across successive memory snapshots, improve curriculum selection, consolidation, and large-scale retrieval, and evaluate transfer and forgetting across simulators and scientific domains.The paper also proposes combining explicit memory with parameter adaptation to preserve new knowledge inspectably while accelerating mature capabilities.Source: §6 · p. 26
  • SimSkill illustrates agent workflows in which an LLM composes natural-language goals, principles, knowledge, and procedures with executable tools, while algorithms and code provide precise, reproducible execution.The paper presents natural language as a medium for accumulating and reorganizing computational capability, not only as an interface.Source: §6 · pp. 26–27

A Supplementary Traces for the Learning Case Studies

  • The supplementary listings condense and lightly normalize autonomous learning logs, preserving decision sequences and memory operations while omitting command output, repeated status messages, and long numerical tables.Unlike the benchmark appendices, these traces come from learning episodes where accepted experience could modify procedural and semantic memory.Source: A Supplementary Traces for the… · p. 31

A.1 Retrieval and Task Decomposition for PP-T4-3-V2 · A.2 Curriculum Revision After a Failed Premise

For PP-T4-3-V2, bounded retrieval provides reusable components that the agent composes into a task-specific workflow. A failed shared-micromobility premise leads to adversarial capability checks, corrected memory, and a curriculum principle for establishing novelty.

A.1 Retrieval and Task Decomposition for PP-T4-3-V2

  • Bounded retrieval supplies heterogeneous procedural and semantic components, but the action agent composes a task-specific end-to-end workflow rather than retrieving a fixed one.The trace identifies eight procedural-memory items and two semantic-memory items, including discrete network design and accessibility measurement.Source: §A.1 · p. 31
  • The PP-T4-3-V2 decomposition covers network and demand construction, portfolio selection and simulation, surrogate modeling, Pareto analysis, stress tests, and equity and benefit reporting.It specifies a 5 x 5 grid with a river constraint and five projects, selecting 14 portfolios by D-optimal design.Source: §A.1 · p. 31

A.2 Curriculum Revision After a Failed Premise

  • SUMO lacks the required station-based shared-asset primitive, but tests refute claims that fleet rationing, runtime bicycle assignment, or inventory waiting and later riding are impossible.The trace reports person-triggered stops, lines="ANY", and corrected capacity and stop-extension settings as counterexamples.Source: §A.2 · p. 31
  • After rejecting the premise, the agent creates no skill based on it, adds a station-based shared-micromobility semantic page, and corrects stopinfo and parking-occupancy knowledge.Source: §A.2 · p. 31
  • The curriculum learns that missing memory does not establish that SUMO lacks a capability; absence should be established adversarially before proposing a replacement.Characterization, validation, comparison, and integration can also be sources of novelty.Source: §A.2 · p. 32
  • The next verified gap is reconstructing SUMO routes and an OD matrix from sparse GPS probes after checking tool availability and existing memory coverage.The trace directs the agent to verify tracemapper.py and route2OD.py first and notes that memory covers only fixed-point count-based reconstruction.Source: §A.2 · p. 32

A.3 Predictive-Control Skill Acquisition … C Memory-Guided Congestion–Safety Cross-Check (OA-T3)

The appendices trace predictive-control and GPS map-matching skill acquisition, summarize the procedural-memory inventory, and describe memory-guided task verification. The predictive-control episode required critic-driven revisions, while the reported memory-guided episodes passed independent artifact-based checks.

A.3 Predictive-Control Skill Acquisition

  • The predictive-control curriculum targets future-decision optimization absent from offline-static or reactive controllers, comparing rolling-horizon DP and simulation-rollout MPC across prediction, demand, arrival, and coordination conditions.The planned design includes validated prediction, null and oracle controls, replication intervals, and a two-signal coordination arm.Source: §A.3 · p. 32
  • After the first attempt failed, the predictive-control agent checked controller constraints, DP against exhaustive enumeration, predictor and optimizer effects, and critical-cell replication; nine independent checks yielded success=true.The first attempt consolidated 1,016 existing SUMO runs without re-simulating completed cells.Source: §A.3 · p. 32
  • The predictive-control episode produced updated skills and pages, and memory lint checked ten changed artifacts, parsed three new scripts, resolved links, and required no repair.The ingestion and maintenance sequence is summarized as part of the predictive-control learning trace.Source: §A.3 · p. 32

A.4 Compositional Map-Matching Acquisition

  • The map-matching task reconstructs SUMO routes and an OD matrix from sparse GPS probes, varying ping interval, location noise, fleet penetration, and dropout and validating edge, route, and OD recovery.It composes existing network-import, demand-reconstruction, replication, and output-analysis skills with new probe-generation, matching, scoring, and aggregation scripts.Source: §A.4 · pp. 32–33
  • The map-matching episode exposed a dependency failure: load-osm-network passed negative longitude as an option, causing tested Americas bounding boxes to fail.The repaired skill uses the single-token --bbox=... form.Source: §A.4 · p. 33

B Complete Categorized Procedural-Memory Inventory

  • The procedural-memory inventory contains 150 learned skills across traffic-simulation functions; categories denote primary function, while cross-links and composition connect skills across categories.The inventory includes, among others, 30 signal and intersection-control skills and 21 calibration, estimation, and experimental-design skills.Source: B Complete Categorized… · pp. 34–35

C Memory-Guided Congestion–Safety Cross-Check (OA-T3)

  • In the three memory-guided tasks, DeepSeek-V4-Pro retrieved from a fixed memory snapshot with ingestion disabled; each episode passed in-loop critic acceptance and independent artifact-based verification.The episodes used the complete SimSkill condition and produced reproducible artifacts.Source: C Memory-Guided Congestion–Safety… · p. 36

C.1 Task and Memory Use · C.2 Memory-Guided Inference and Critic Revision

OA-T3 tests whether SimSkill can combine traffic simulation, congestion visualization, and safety analysis in one consistent scenario. Retrieved skills guide execution and analysis, while critic revision checks both artifacts and the claims derived from them.

C.1 Task and Memory Use

  • OA-T3 combines a 4 × 4-block signalized grid, 200 m spacing, and 1,000 veh/h inbound-edge demand to compare congestion and SSM-conflict hotspots.The task requires network and demand generation, simulation, visualization, safety-device configuration, spatial matching, and statistical comparison within one consistent scenario.Source: §C.1 · p. 36
  • Six procedural skills supplied network construction, trip generation, simulation, heat-map creation, SSM analysis, and output-integrity checks.The retrieved skills were create-grid-network, generate-random-trips, run-simulation, visualize-network-congestion-heatmap, analyze-intersection-safety, and analyze-simulation-outputs.Source: §C.1 · p. 36
  • Two semantic pages guided analytical choices: lane occupancy as the congestion metric and the meaning of safety indicators, thresholds, and conflict types.The pages were spatial-congestion-heatmap-with-plot-net-dump and surrogate-safety-measures; they shaped decisions rather than execution.Source: §C.1 · p. 36

C.2 Memory-Guided Inference and Critic Revision

  • The action agent composed retrieved artifacts into one workflow that generated a common simulation, twelve interval-level congestion maps, directed-edge SSM assignments, and comparative rankings.The LLM connected network, simulation, visualization, safety-analysis, and validation procedures rather than copying a single skill.Source: §C.2 · p. 36
  • The critic required three outer attempts, with the final attempt reconciling the report with artifacts and stating insertion and ranking limitations.The first attempt had an incorrect demand calculation and silent conflict-parser failure; the second repaired both but still misreported outputs and omitted limitations.Source: §C.2 · p. 36
  • SimSkill inference continued after program execution until the critic accepted both the artifacts and the claims derived from them.After repairing demand and parsing defects, the second attempt still had a report-output mismatch and omitted limitations, prompting a third revision.Source: §C.2 · p. 36

C.3 Verified Result and Output … D.2 Memory-Guided Inference and New Construction

The verified OA-T3 result distinguishes broad rank association from agreement on the most critical locations and shows that congestion maps cannot replace direct safety-hotspot analysis in the realized scenario. DG-T4-3-V2 combines retrieved foundations with newly constructed inference under a fixed simulation budget and independent artifact verification.

C.3 Verified Result and Output

  • 0.748 was the Spearman correlation across all 120 edges, but no edge appeared in both top-ten lists.Table 9 summarizes the accepted OA-T3 congestion–safety cross-check; its caption identifies raw ego–foe SSM counts.Source: §C.3 · p. 36, Table 9 · p. 37
  • The verified conclusion is that a congestion heat map cannot substitute for direct safety-hotspot analysis in this scenario.The conclusion is scoped to the realized simulation: only 3,619 of 20,000 scheduled vehicles entered, leading occupancy values were narrowly separated, and each interaction appeared twice in SSM output. The output included reproducible scripts, analysis files, a combined ranking, a corrected report, and twelve heat maps generated using retrieved visualization guidance.Source: §C.3 · pp. 36–37, Figure 9 · p. 38

D.1 Task and Memory Use

  • DG-T4-3-V2 required posterior estimates for 240 nonnegative OD flows across eight intervals from 24 screenline counts and sparse Bluetooth travel times, with 15 designated SUMO runs.Holdout rows had to be excluded from fitting and model selection, and the result had to quantify non-identifiability; the stored library lacked an end-to-end procedure for likelihood-free, time-dependent OD estimation.Source: §D.1 · p. 37
  • Retrieved skills supplied network and execution foundations, ODME logic and diagnostics, fixed-budget optimization, and run-variability guidance, while knowledge pages prompted uncertainty reporting and count-fit caveats.No retrieved skill implemented likelihood-free, time-dependent OD inference, so the action agent constructed that task-specific component in new code.Source: §D.1 · p. 37

D.2 Memory-Guided Inference and New Construction

  • The agent used retrieved ODME and identifiability knowledge to define a low-dimensional model for all 240 flows, apply one route set to both methods, exclude holdout data, and build an ABC–SMC workflow.The holdout observations were also excluded from stopping decisions.Source: §D.2 · p. 37
  • The 15-run ledger allocated 12 SUMO runs to surrogate training, three surrogate-only posterior populations, and the final three runs to two posterior representatives plus the ODME comparator.Posterior inference operated on the cached physics-informed surrogate without SUMO calls; unusable executions were kept in a separate discarded-run ledger.Source: §D.2 · pp. 37, 39
  • The first outer action–critic attempt was accepted after independent checks of holdout masks, the run ledger, surrogate-only inference, comparator, and required diagnostics.The critic found five minor report-to-artifact discrepancies, which were corrected without code changes or rerunning the experiment.Source: §D.2 · p. 39

D.3 Verified Result and Output · E Procedural Transfer to Shared E-Scooter Operations (MT-T4-4-V2)

The accepted output met the prescribed simulation budget, convergence checks, and holdout criterion, while retaining limitations around OD-cell identification and simulator–surrogate discrepancy. Its artifacts include diagnostic outputs that show both inferred results and underdetermination.

D.3 Verified Result and Output

  • The accepted output satisfied the prescribed simulation budget, convergence checks, and holdout criterion.Table 10 summarizes the verified result.Source: §D.3 · p. 39
  • Aggregate demand was better identified than individual OD cells, and good count fit did not eliminate the OD null space or simulator–surrogate discrepancy.These limitations were highlighted by semantic memory.Source: §D.3 · p. 39
  • The final artifacts included network and signal files, ten task scripts, run ledgers, a fitted surrogate, posterior samples, credible intervals for 240 flows, a deterministic ODME comparator, summaries, and five diagnostic figures.Figure 10 combines demand and posterior-predictive checks, parameter correlations, and non-identifiable contrasts, making the inferred result and underdetermination visible.Source: §D.3 · p. 39

E.1 Task and Selective Memory Use

In MT-T4-4-V2, SimSkill compared three shared e-scooter rebalancing strategies under operational, environmental, and equity constraints, transferring procedural memory while coding missing domain logic.

Task and Selective Memory Use

  • MT-T4-4-V2 compares no rebalancing, scheduled truck rebalancing, and incentive-based user rebalancing for a docked shared e-scooter system.No shared-scooter or micromobility semantic page was available, so task-owned code supplied the missing domain-specific logic.Source: §E.1 · p. 39
  • Eight procedural skills were retrieved, six shaped the solution, and inspected vehicle-state tools were unnecessary for its event-driven formulation.The skills supplied network construction and SUMO validation, lane representation, truck routing and VKT accounting, emissions reporting, and accessibility and equity reporting patterns.Source: §E.1 · pp. 39, 41
  • Table 11 reports MT-T4-4-V2 base-scenario results averaged over four paired perturbation seeds.Source: Table 11 · p. 41
  • Retrieval provided candidate capabilities rather than commands that had to be applied mechanically.The selected event-driven formulation required no TraCI state-transfer loop.Source: §E.1 · p. 41

E.2 Memory-Guided Inference and Task-Specific Extension

SimSkill combined retrieved procedures with task-specific mechanisms not represented in memory. It then ran and independently verified the complete policy-comparison workflow.

E.2 Memory-Guided Inference and Task-Specific Extension

  • Retrieved procedures handled SUMO network construction and route-time validation, while the agent built scooter, battery, pedestrian, docking, truck, and incentive logic in a task-owned discrete-event engine.Stored procedures supplied reliable components; the LLM constructed the missing domain-specific mechanism.Source: §E.2 · p. 41
  • The agent completed 18 policy–scenario–seed runs across four base and two rain perturbations, corrected four local defects, and the critic accepted the first outer attempt after byte-identical reruns and requirement checks.Non-blocking modeling qualifications were retained in the report rather than treated as reasons to discard the verified result.Source: §E.2 · p. 41

E.3 Verified Result and Output · E.4 Cross-Case Interpretation

The results favor truck rebalancing for served share, including at low-income-origin stations, while incentives offer a truck-free complement; across three cases, memory was adapted into task-specific workflows and verified artifacts. All benchmark tasks ran in test mode, so none changed memory for later tasks.

E.3 Verified Result and Output

  • Truck rebalancing added 10.7 km of combustion-vehicle travel and 2.13 kg of CO2, whereas incentives used no truck travel and cost $94.0 per simulated day.Neither full-dock failures nor sidewalk conflicts occurred in the abstract network.Source: §E.3 · p. 41
  • In rain, served shares were 87.6% with no rebalancing, 98.9% with trucks, and 97.8% with incentives; the agent recommended trucks primarily and incentives as a truck-free complement.The policy ordering remained stable in the rain sensitivity case when considering availability, equity, VKT, emissions, and cost together.Source: §E.3 · p. 41, Figure 11 · p. 42
  • The final output packaged five task scripts, the SUMO network and specification, route-time validation, aggregate and per-seed data, request-level outcomes, a report, and two figures.Figure 11 combines the network derived from retrieved infrastructure procedures with policy, equity, and VKT outcomes; the report retained non-blocking critic qualifications.Source: §E.3 · p. 41, Figure 11 · p. 42

E.4 Cross-Case Interpretation

  • Across three cases, memory supported artifact composition and critic-guided correction, methodological scaffolding for a task-owned likelihood-free estimator, and selective procedural transfer without a matching semantic page.MT-T4-4-V2 also set aside retrieved skills that were unnecessary.Source: §E.4 · p. 41
  • In every case, the action agent selected, adapted, and composed memory into a task-specific workflow, while the critic verified artifacts and claims and outputs preserved reproducible materials.The preserved outputs included scripts, data, results, and figures.Source: §E.4 · p. 41

E.3 Verified Result and Output

  • Served share rose from 58.8% without rebalancing to 95.8% with trucks, while low-income-origin served share increased from 40.9% to 95.7%.The incentive policy raised overall served share to 75.1%.Source: §E.3 · p. 41, Figure 11 · p. 42

E.4 Cross-Case Interpretation

  • All tasks ran in test mode, so none could alter memory available to subsequent benchmark tasks.Source: §E.4 · p. 42

Related papers

Loading 2609.03753v1…