PaperlayerView exact revision on arXiv

Source-linked AI summary

SimTIO: A Simulation-Grounded Multi-Agent LLM Framework for Compositional Traffic Intervention Optimization

Shuyang Li, Ruimin Ke

arXiv:2609.05740v1cs.MAcs.SYeess.SY

TL;DR

Traffic analysts must turn diagnosed bottlenecks into executable interventions without degrading network-wide performance, yet simulation diagnosis alone does not produce those actions. SimTIO guides a constrained search with measured simulation feedback and selects among executable candidates. Across 15 runs, it reduced bottleneck and network-wide losses, while finding feasible improving plans more often than the two comparators.

  • Problem

    Simulation evaluates congestion and interventions, but translating a diagnosed bottleneck into an executable action requires choosing controls, modifying interdependent artifacts, and assessing network-wide side effects.

    Source: INTRODUCTION · p. 3

  • Method

    SimTIO uses specialist LLM agents to refine validated traffic interventions from measured SUMO feedback, re-simulates compatible pairs, and selects measured feasible plans under operational constraints.

    Source: INTRODUCTION · p. 3

  • Results

    Across 15 runs, Top-10 bottleneck time loss fell 9.18%±10.18%, network-wide delay fell 2.78%, and feasible improving plans occurred in 86.7% of cases versus 73.3% for grounded random search and 80.0% for the deterministic heuristic.

    Source: CONCLUSIONS · p. 16

  • Takeaways & Limitations

    The results position LLMs as constrained, feedback-guided local search operators, with executable tools, microscopic simulation, and explicit safety constraints retaining final authority.

    Source: CONCLUSIONS · p. 17

    The findings are limited by synthetic demand, three demand seeds, one LLM model and delay-oriented prompt, and a small simulation budget; the comparison also lacks matched selector ablations and an established traffic-control baseline.

    Source: CONCLUSIONS · pp. 16–17

Abstract

from arXiv · show

Traffic analysts must translate diagnosed bottlenecks into executable interventions without allowing local improvements to degrade network-wide performance. This study presents SimTIO, a simulation-grounded multi-agent large language model framework for composing and selecting traffic interventions under explicit operational constraints. SimTIO first simulates an unmodified SUMO scenario to identify a baseline-frozen set of ten bottleneck edges. A grounded sampler then initializes signal-control, corridor-speed, and demand-preserving routing actions, while three specialist agents use measured simulation feedback to select one-parameter refinements from validator-confirmed mutation catalogs. Compatible actions are combined and re-simulated so that their interaction effects are measured rather than inferred. Final selection minimizes bottleneck time loss while constraining network-wide delay, neighboring-road spillover, throughput loss, and teleport events, with the unmodified scenario retained as a no-operation guard. Across 15 cases covering five U.S. urban networks, three synthetic-demand seeds, and 2,400 origin-destination trips per scenario, SimTIO reduced Top-10 bottleneck time loss by an average of 9.18 percent and network-wide delay by 2.78 percent. It found a feasible improving plan in 86.7 percent of cases, compared with 73.3 percent for grounded random search and 80.0 percent for a deterministic heuristic under the same seven-simulation budget, although the differences in Top-10 improvement were not statistically significant. These results support using LLMs as constrained, feedback-guided local search operators while reserving final decision authority for executable tools, microscopic simulation, and explicit safety constraints.

Shuyang Li

SimTIO addresses the gap between diagnosing congestion and producing executable interventions that do not shift harm elsewhere. It uses measured simulation outcomes and explicit constraints to guide search and select plans.

INTRODUCTION

  • Traffic simulation measures congestion and tests interventions safely, but translating a diagnosed bottleneck into executable action still requires analysts to choose strategies, modify artifacts, and assess network-wide side effects.This diagnosis-to-action gap limits automation of simulation-based traffic optimization.Source: INTRODUCTION · p. 3
  • Local improvements can shift congestion elsewhere, and interventions may reinforce or interfere with one another, making measured evaluation of combinations important.Exhaustively simulating all combinations is impractical.Source: INTRODUCTION · p. 3
  • A seeded sampler initializes validator-confirmed signal, corridor-speed, and demand-routing actions; three specialist agents use measured SUMO outcomes to choose local refinements from legal mutation catalogs.A deterministic registry rejects unsupported actions, invalid identifiers, and incomplete parameters before execution.Source: INTRODUCTION · p. 3
  • Compatible intervention pairs are re-simulated, and final selection minimizes time loss on ten baseline-frozen bottleneck edges subject to operational constraints and a no-operation guard.Constraints cover global delay, neighboring-road spillover, throughput, demand preservation, and teleport events.Source: INTRODUCTION · p. 3
  • Across 15 runs, Top-10 bottleneck time loss fell 9.18%±10.18% and network-wide delay fell 2.78%; feasible improving plans occurred in 86.7% of cases versus 73.3% for grounded random search and 80.0% for a deterministic heuristic.The comparisons used the same simulation cap.Source: INTRODUCTION · p. 3
  • The framework combines capability validation, budgeted re-simulation of compatible pairs, constrained selection, a no-op guard, and an offline Pareto audit to assess measured traffic outcomes.Specialist LLMs provide feedback-guided local search, while executable tools, simulation evidence, and explicit constraints retain final authority.Source: INTRODUCTION · pp. 3–4

Execution-Grounded LLM Algorithm Design

The paper situates SimTIO among execution-grounded LLM optimization, compositional and multi-agent search, and traffic-control systems. Its framework uses baseline evidence to guide legal intervention search, then relies on microscopic simulation for decision evidence.

Execution-Grounded LLM Algorithm Design

  • Execution-grounded LLM systems use external evaluation rather than linguistic judgment to assess proposals; SimTIO applies this principle to structured traffic interventions evaluated by a microscopic simulator.Examples include OPRO, FunSearch, and AlphaEvolve (Novikov et al. 2025).Source: Execution-Grounded LLM Algorithm Design · p. 4

Multi-objective, Compositional, and Multi-agent Search

  • Related work spans non-dominated heuristic search, sequential operator composition (Qiu et al. 2026), solver-checked models, and multi-agent debate.SimTIO evaluates valid singles and compatible bundles in a common pool, with an offline Pareto audit that does not change the selected plan.Source: Multi-objective, Compositional, and… · p. 4

LLMs for Traffic Control and Simulation-Grounded Optimization

  • Unlike prior traffic LLM work focused on scenario generation or stepwise signal control, SimTIO searches across network, routing, and signal interventions using SUMO as its evaluator.The related work also notes established coupling between signal control and route assignment and between network design and route choices.Source: LLMs for Traffic Control and… · p. 5

Framework and Traffic Evidence

  • The workflow establishes a baseline, diagnoses bottlenecks, evaluates one seeded intervention per control family, refines retained parents using specialist feedback, and searches compatible two-action bundles within the budget.Unused budget remains unused when no legal refinement or compatible bundle is available; SUMO supplies decision evidence.Source: Framework and Traffic Evidence · p. 5
  • Baseline diagnosis aggregates vehicle and edge measurements into an edge score using speed ratio, density, occupancy, time loss, and waiting time; the score ranks where to search but is not the optimization objective.Edges shorter than 10 m are excluded before ranking, and the target set is fixed from the baseline.Source: Framework and Traffic Evidence · pp. 5–6
  • The fixed target edges also seed a one-hop local subgraph capped at 80 edges, preserving nearby topology and controller identifiers while limiting prompt length.Freezing the target set prevents apparent gains caused merely by shifting congestion to different edges.Source: Framework and Traffic Evidence · p. 6

Optimization Problem Formulation

The optimization targets executable intervention plans rather than LLM recommendations or predicted effects. It minimizes measured loss on fixed bottleneck edges among evaluated plans while applying compatibility rules and operational constraints.

Optimization Problem Formulation

  • An intervention plan is an executable object with signal-control, network-control, and routing components; inactive components are denoted by ∅.Signal actions specify a controller and fixed-time phase plan, network actions specify a corridor or edge operation and parameter, and routing actions specify an avoided edge and rerouted fraction.Source: Optimization Problem Formulation · p. 6
  • A plan activates one component, while a compatible bundle activates at most two; compatibility rules prevent conflicting edits or duplicate changes to a control family.Source: Optimization Problem Formulation · p. 7
  • The optimizer minimizes measured time loss on baseline-frozen Top-10 edges among evaluated feasible plans, with the unmodified scenario available when no feasible plan exists.Global delay and neighboring-road loss are bounded spillover constraints; demand integrity, throughput, and teleports are operational constraints.Source: Optimization Problem Formulation · p. 7

Grounded Random Initialization

Grounded initialization samples validator-confirmed interventions from a search space defined by diagnosed bottlenecks. Deterministic checks and utilities constrain proposals before execution and apply the resulting SUMO modifications.

Grounded Random Initialization

  • The optimizer samples validator-confirmed proposals in round-robin order from available specialist families, using the recorded case seed and baseline-frozen objective edges.The validator removes unsupported controllers, missing reverse-edge pairs, detectable local topology violations, and malformed controls; demand and OD preservation are checked again after materialization.Source: Grounded Random Initialization · p. 7
  • Signal proposals optimize fixed-time green splits using Webster’s method while preserving phase states and yellow/all-red transitions.The signal family selects an evidenced controller; deterministic utilities apply the SUMO changes.Source: Grounded Random Initialization · p. 7, Figure 2 · p. 8
  • Network proposals sample legal corridor-speed, edge-removal, or one-way actions, while initial routing uses a one-edge trust region and a rerouted fraction.Corridor speed factors are selected from {0.7, 0.8, 0.9}.Source: Grounded Random Initialization · p. 7
  • Routing soft-avoids selected edges only for vehicles whose reference routes traverse them, while retaining other routes and preserving every demand record and OD pair.The routing utility selects eligible vehicles and verifies route connectivity; deterministic code performs network edits.Source: Grounded Random Initialization · p. 8
  • Structured proposals are checked for supported actions, required artifacts and parameters, valid identifiers, numeric ranges, and scenario-specific preconditions before simulation.The two-stage validation traces executable modifications back to baseline evidence and prevents unsupported mutations from entering simulation.Source: Grounded Random Initialization · p. 8

Simulation-Grounded Candidate Search

Candidate search evaluates isolated interventions in SUMO, uses measured outcomes to guide constrained specialist refinements, and simulates compatible bundles to capture their interactions. LLMs select only from program-generated legal mutations; bundle construction and final selection remain deterministic.

Simulation-Grounded Candidate Search

  • Each candidate runs through a complete SUMO simulation on an isolated copy of the baseline, producing measured KPIs and explicit constraint violations.All candidates start from identical baseline artifacts, and a failed candidate does not affect other evaluations.Source: Simulation-Grounded Candidate Search · p. 9
  • Specialists use measured outcomes and failed constraints to select one-argument mutations from finite, unseen, validator-confirmed catalogs for retained beam candidates.The beam is feasible-first, width b, and uses role diversity; each selected mutation is revalidated and re-simulated before entering the next beam.Source: Simulation-Grounded Candidate Search · p. 9
  • Compatible intervention bundles are re-simulated to measure interaction effects, since individually useful actions can reinforce, cancel, or shift one another’s effects.The experiments cap each case at seven SUMO evaluations, with K=4 and at most two bundles; these settings are shared across proposal-source conditions.Source: Simulation-Grounded Candidate Search · p. 9

LLM Feedback Prompts

  • The LLM may select a catalog identifier but cannot invent identifiers or values; deterministic code handles bundle construction, compatibility filtering, simulation, Pareto auditing, and final selection.An invalid selection gets one repair attempt, and invalid or duplicate refinements do not consume the SUMO budget.Source: Simulation-Grounded Candidate Search · p. 9, LLM Feedback Prompts · p. 10

Signal-Control Specialist

The specialist prompts constrain feedback to measured outcomes and legal mutations, with intervention-specific risks guiding the selection. Road-network speed control is applied through a deterministic corridor-resolution tool.

Signal-Control Specialist

  • The signal specialist compares legal mutations against measured outcomes and baseline evidence, preferring repairs to global-delay or spillover violations while preserving Top-10 benefit.The prompt also calls for preserving yellow/all-red states and considering minor-movement starvation and downstream queue transfer.Source: Signal-Control Specialist · p. 10

Road-Network Intervention Specialist

  • set_corridor_speed is a static corridor-level speed-limit adjustment, with the deterministic tool resolving up to five same-road edges from one anchor.The specialist selects a legal mutation rather than rewriting the action or numeric values.Source: Road-Network Intervention Specialist · p. 11
  • The road-network specialist weighs measured outcomes and baseline evidence against topology risks such as OD disconnection, route loss, detours, and congestion transfer.For speed actions, the prompt flags queue smoothing and downstream spillback, and favors repairing global-delay or neighbor-spillover violations without giving up measured Top-10 benefit.Source: Road-Network Intervention Specialist · p. 11

Demand-Routing Specialist

The routing specialist refines soft-avoidance actions within a one-edge trust region, selecting legal mutations using measured outcomes and preserving demand records and edge-level OD pairs.

Demand-Routing Specialist

  • The routing trust region contains exactly one avoided edge, and legal rerouting fractions are 0.25, 0.5, 0.75, and 1.0.Soft avoidance may retain an unavoidable endpoint edge.Source: Demand-Routing Specialist · p. 11
  • The specialist compares legal mutations using measured outcomes and baseline evidence, favoring reduced displaced congestion, excessive detour, or failed global-delay and neighbor-spillover constraints while retaining local benefit.Every demand record and its exact edge-level OD must remain unchanged.Source: Demand-Routing Specialist · p. 11

Constrained Final Selection

Final selection preserves demand and chooses measured interventions that improve the frozen bottlenecks without violating network-wide safeguards. The baseline remains available as a no-operation guard.

Constrained Final Selection

  • Demand preservation is mandatory: candidates changing demand records or OD pairs are discarded, and throughput is measured against the original source demand.This prevents apparent gains from removing difficult trips or changing the throughput denominator.Source: Constrained Final Selection · p. 11
  • A candidate must improve bottleneck time loss by at least 1%, limit global-delay growth to 1% and neighbor time-loss growth to 5%, and preserve throughput and teleports.The unmodified scenario is retained as a no-operation guard.Source: Constrained Final Selection · p. 12
  • The constraints separate target improvement from network acceptability: bottleneck loss measures the local objective, while neighbor loss, global delay, throughput, and teleports guard against harmful trade-offs.Among feasible candidates, selection minimizes frozen Top-10 time loss; the 1% threshold filters changes deemed too small to be operationally meaningful.Source: Constrained Final Selection · p. 12

Pareto Dominance Audit

The audit checks whether the selected plan lies on the measured Pareto front, while the evaluation compares proposal sources under a shared simulation budget. Results show strong LLM-feedback feasibility coverage, but not statistically significant superiority in Top-10 reduction.

Pareto Dominance Audit

  • The offline audit tests whether the primary Top-10 objective selects a feasible candidate on the Pareto front across four baseline-relative gains, without adding LLM calls or SUMO runs.It leaves the selected plan unchanged and introduces no manually weighted preference profiles.Source: Pareto Dominance Audit · p. 12

EXPERIMENTAL EVALUATION

  • The evaluation includes no-operation outcomes as zero improvement, assessing the complete guarded optimizer rather than only successful interventions.The main comparison measures LLM feedback against grounded random search and a deterministic heuristic under the same microscopic-simulation budget.Source: EXPERIMENTAL EVALUATION · p. 12

Experimental Setup

  • The experiments span five OSM-derived urban networks and three synthetic-demand seeds, with 2,400 passenger trips per scenario; demand is not calibrated to observed counts.The setup therefore tests optimization behavior across network structures rather than real-world traffic fidelity.Source: Experimental Setup · p. 12

Budget-Matched Proposal-Source Comparison

  • 86.7% of cases yielded a feasible improving plan with LLM feedback, versus 73.3% for grounded random search and 80.0% for the heuristic; mean Top-10 reductions were 9.18%, 9.00%, and 9.60%, respectively.The paired Top-10 comparisons were not statistically significant: p=0.407 against random and p=0.639 against the heuristic.Source: Table 3 · p. 14, Budget-Matched Proposal-Source… · pp. 14–15
  • Across the 15 city–seed cases, LLM feedback led seed-averaged Top-10 results in Troy, Ann Arbor, and Los Angeles, while the heuristic led in Albany and grounded random led in Manhattan.LLM feedback beat, tied, and lost to grounded random in six, six, and three paired cases, respectively; it beat the heuristic in six and lost in nine.Source: Budget-Matched Proposal-Source… · pp. 14–15
  • For the LLM-feedback optimizer, Ann Arbor, Los Angeles, and Manhattan accepted feasible improvements for all three seeds, while Albany and Troy each retained the baseline once.Manhattan had the largest mean reduction and highest candidate yield; the baseline guard could retain no operation when legal changes were marginal or unsafe.Source: Table 4 · p. 15, Budget-Matched Proposal-Source… · p. 15

Search Behavior and Intervention Composition

LLM search often moved beyond its initial actions, while final intervention choices differed across proposal sources. Routing and topology candidates did not win constrained selection because their local benefits failed to offset global or operational costs.

Search Behavior and Intervention Composition

  • LLM feedback selected five TLS plans, four corridor-speed plans, four bundles, and two no-ops; routing and topology candidates did not win constrained selection.Their local benefits did not compensate for global-delay, spillover, or demand-preservation failures.Source: Search Behavior and Intervention… · p. 15
  • Grounded random search retained the baseline in four cases, while the heuristic’s 12 applied plans comprised nine bundles and three corridor-speed plans.Grounded random selected the same number of TLS plans and bundles as LLM feedback.Source: Search Behavior and Intervention… · p. 15
  • Only three of 13 applied LLM plans were initial singles; six were feedback-refined and four were bundles.Ten applied decisions therefore used a second-stage search operation rather than accepting the grounded random initialization.Source: Search Behavior and Intervention… · p. 15

Pareto Dominance Audit

The audit found most primary-rule selections nondominated, but some bottleneck-focused choices sacrificed secondary metrics. The conclusions report improvements and feasible-plan coverage while bounding the evidence by the study’s evaluation scope.

Pareto Dominance Audit

  • Eleven of 15 primary-rule selections were nondominated, while the four dominated choices show that bottleneck-delay optimization can sacrifice a secondary metric.Pareto fronts averaged 1.53 candidates (range 1–4); the audit is diagnostic and does not change the primary objective.Source: Pareto Dominance Audit · pp. 15–16

CONCLUSIONS

  • The framework returns the best evaluated plan meeting operational constraints and retains the unmodified scenario when no plan qualifies.Constraints cover global delay, neighboring-road spillover, throughput, demand preservation, and teleport events.Source: CONCLUSIONS · p. 16
  • Top-10 bottleneck time loss fell by 9.18% ± 10.18% and network-wide delay by 2.78%; feasible improving plans occurred in 86.7% of cases versus 73.3% for grounded random and 80.0% for the heuristic.Top-10 advantages over grounded random and the heuristic were not statistically significant; the evidence supports competitive search and improved feasible-plan coverage, not universal superiority.Source: CONCLUSIONS · p. 16
  • Findings are bounded by synthetic demand, three seeds, one LLM and delay-oriented prompt, and a small simulation budget; the comparison lacks an established traffic-control or simulation-optimization baseline.The study also does not isolate LLM mutation selection from the surrounding beam-search structure using matched random and rule-based selectors.Source: CONCLUSIONS · pp. 16–17

GENERATIVE AI DISCLOSURE

  • The authors report reviewing and verifying the manuscript text and disclose gpt-5.6-luna as an experimental component of SimTIO in the Methodology section.Source: GENERATIVE AI DISCLOSURE · p. 17
Loading 2609.05740v1…