Paper deep dive
SimTIO: A Simulation-Grounded Multi-Agent LLM Framework for Compositional Traffic Intervention Optimization
Shuyang Li, Ruimin Ke
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/9/2026, 3:55:18 AM
Summary
The paper introduces SimTIO, a simulation-grounded multi-agent Large Language Model (LLM) framework designed to optimize traffic interventions by composing and selecting signal-control, corridor-speed, and routing actions. Using the SUMO microscopic simulator, SimTIO identifies baseline bottlenecks and employs specialist LLM agents to perform feedback-guided local search within a strict simulation budget. The framework ensures safety by constraining network-wide delay, spillover, and throughput, ultimately reducing Top-10 bottleneck time loss by an average of 9.18% across 15 test cases compared to baselines.
Entities (11)
Relation Signals (9)
SimTIO ā uses ā SUMO
confidence 100% Ā· SimTIO first simulates an unmodified SUMO scenario to identify a baseline-frozen set of ten bottleneck edges.
SimTIO ā uses ā LLM
confidence 100% Ā· This study presents SimTIO, a simulation-grounded multi-agent large language model framework...
Ruimin Ke ā affiliatedwith ā Rensselaer Polytechnic Institute
confidence 95% Ā· Ruimin KeAssistant Professor Civil and Environmental Engineering, Rensselaer Polytechnic Institute
Shuyang Li ā affiliatedwith ā Rensselaer Polytechnic Institute
confidence 95% Ā· Shuyang LiGraduate Student Civil and Environmental Engineering, Rensselaer Polytechnic Institute
SimTIO ā constrains ā Network-Wide Delay
confidence 95% Ā· Final selection minimizes bottleneck time loss while constraining network-wide delay...
SimTIO ā optimizes ā Bottleneck Time Loss
confidence 95% Ā· Final selection minimizes bottleneck time loss while constraining network-wide delay...
SimTIO ā generates ā Signal-Control
confidence 90% Ā· A grounded sampler then initializes signal-control, corridor-speed, and demand-preserving routing actions...
SimTIO ā generates ā Corridor-Speed
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2609.05740v1
- Canonical: https://arxiv.org/abs/2609.05740v1
Trouble viewing inline? Open PDF directly ā
Full Text
55,152 characters extracted from source content.
Expand or collapse full text
Shuyang LiGraduate Student Civil and Environmental Engineering, Rensselaer Polytechnic Institutelis36@rpi.edu[Troy, New York, 12180][0009-0006-8588-2947] KeAssistant Professor Civil and Environmental Engineering, Rensselaer Polytechnic Instituteker@rpi.edu[Troy, New York, 12180] Shuyang Li, Ruimin Ke SimTIO: A Simulation-Grounded Multi-Agent LLM Framework for Compositional Traffic Intervention Optimization 1 Abstract Objectives: Traffic analysts must translate diagnosed bottlenecks into executable interventions while avoiding local improvements that degrade network-wide performance. This study develops SimTIO, a simulation-grounded multi-agent large language model framework for generating, composing, and selecting traffic interventions under explicit operational constraints. Methods: An unmodified SUMO scenario was simulated to derive vehicle- and edge-level traffic measures and a baseline-frozen set of ten bottleneck edges. A grounded sampler initialized signal-control, corridor-speed, and demand-preserving routing interventions, and three specialist agents used measured simulation feedback to select one-parameter refinements from validator-confirmed mutation catalogs. Each case used at most seven candidate simulations: three initial actions, up to two feedback mutations, and up to two compatible two-action bundles. The evaluation covered five U.S. urban networks, three synthetic-demand seeds, and 2,400 originādestination trips per scenario; final selection bounded global delay, neighboring-road spillover, throughput, and teleport events. Findings: Across 15 cityāseed cases, the LLM-feedback optimizer reduced Top-10 bottleneck time loss by 9.18%Ā±10.18%9.18\%Ā± 10.18\% and network-wide delay by 2.78%. It found a feasible improving plan in 86.7% of cases, compared with 73.3% for grounded random search and 80.0% for a deterministic heuristic under the same simulation budget. Only three of the 13 applied LLM plans were unmodified initial actions; six were feedback-refined singles and four were bundles. Paired Wilcoxon tests found no statistically significant Top-10 advantage over either baseline (p=0.407p=0.407 and p=0.639p=0.639). Novelty: SimTIO separates LLM-guided feedback refinement from simulation-based decision authority and evaluates heterogeneous intervention compositions rather than relying on predicted or additive effects. Its baseline-frozen objective and no-operation guard prevent unsafe local gains from determining the final plan. Practical Applications: The framework can support traffic analysts in screening signal, speed-limit, and routing interventions before field deployment. Practical use requires locally calibrated demand, verified network data, and agency review of the simulation-selected intervention. 2 Introduction Microscopic traffic simulation is widely used to evaluate congestion, signal control, routing, emissions, and network-level mobility. Simulators such as SUMO provide detailed measurements and allow interventions to be tested without affecting a real transportation system. However, simulation is primarily an evaluation tool: converting a diagnosed bottleneck into an executable action still requires an analyst to select a control strategy, modify interdependent simulation artifacts, and evaluate possible network-wide side effects. This gap between diagnosis and action limits the automation of simulation-based traffic optimization. Large language models (LLMs) offer a flexible way to reason across heterogeneous interventions, including signal retiming, corridor control, and demand rerouting. Nevertheless, plausible language is not evidence of traffic benefit. An LLM may reference nonexistent network entities, request unsupported operations, omit required parameters, or propose an intervention that improves a local bottleneck by transferring congestion elsewhere. Traffic actions may also interact: two individually reasonable interventions can reinforce or interfere with one another. Because exhaustive simulation of all combinations is impractical, an LLM-based optimizer must ground proposals in executable capabilities, search combinations within a limited budget, and reserve final selection for measured simulation outcomes. This study develops a simulation-grounded multi-agent LLM traffic optimizer. The framework first simulates the unmodified scenario and extracts compact bottleneck-centered evidence containing traffic measurements, local connectivity, and relevant signal identifiers. A grounded sampler initializes one validator-confirmed action from each available signal-control, corridor-speed, and demand-routing family. After those actions are simulated, three specialist agents use the measured outcomes to select local refinements from program-generated legal mutation catalogs. A deterministic capability registry rejects unsupported actions, invalid identifiers, and incomplete parameters before execution. The LLM therefore serves as a feedback-guided local search operator rather than as the final traffic-control decision maker. The system further performs a simulation-feedback beam search. Measured KPIs and constraint violations are returned to the responsible specialist, which produces an auditable one-parameter refinement of a retained parent. The remaining budget is used for pairwise search over compatible interventions; if no unseen legal refinement or compatible bundle exists, that portion of the budget remains unused. Candidate actions are applied sequentially and re-simulated, so their combined effect is measured rather than inferred from textual predictions. Final selection minimizes time loss on ten baseline-frozen bottleneck edges while allowing at most a 1% network-wide delay increase and 5% neighboring-road spillover, while prohibiting throughput loss or additional teleport events. The unmodified scenario remains available as a no-operation guard. An offline Pareto audit additionally checks whether the delay-selected plan is dominated on measured delay, emissions, or mobility outcomes. The primary evaluation covers five urban networks, three demand seeds, and 2,400 originādestination trips per scenario. Across 15 runs, the selected outcomes reduce Top-10 bottleneck time loss by 9.18%Ā±10.18%9.18\%Ā± 10.18\% and network-wide delay by 2.78%. The LLM-feedback search finds a feasible improving plan in 86.7% of cases, compared with 73.3% for grounded random search and 80.0% for a deterministic heuristic under the same simulation cap. Only three of the 13 applied LLM plans are unmodified initial actions; six are feedback-refined singles and four are bundles. This work contributes a simulation-grounded multi-agent framework that generates heterogeneous traffic interventions from compact diagnostic and topological evidence while validating every proposal against executable capabilities. It introduces a budgeted composition mechanism that applies and re-simulates compatible intervention pairs, allowing measured interactions to influence selection without exhaustive search. The framework further combines a baseline-frozen bottleneck objective with constraints on global delay, spillover, throughput, and simulation integrity, together with a no-op guard and an offline Pareto dominance audit. Finally, the multi-network, multi-seed evaluation quantifies both achievable improvements and the rejection of locally promising but globally unsafe candidates. Together, these components position the specialist LLMs as role-specific, feedback-guided local search operators while reserving final authority for executable tools, simulation evidence, and explicit optimization constraints. 3 LITERATURE REVIEW 3.1 Execution-Grounded LLM Algorithm Design Large language models have increasingly moved from one-shot solution generation toward feedback-driven optimization. OPRO treats an LLM as a black-box optimizer by conditioning new proposals on previously evaluated solutionāscore pairs Yang et al. (2024), while Self-Refine demonstrates how verbal feedback can support iterative revision without additional training Madaan et al. (2023). For executable algorithm discovery, FunSearch couples an LLM with deterministic evaluation and an evolutionary program database Romera-Paredes et al. (2024). Evolution of Heuristics evolves both heuristic ideas and executable implementations Liu et al. (2024), and ReEvo converts performance comparisons into reflective feedback for subsequent generations Ye et al. (2024). MCTS-AHD organizes candidates in a search tree to improve exploration Zheng et al. (2025), whereas AlphaEvolve scales the same generateāevaluateāretain principle through an evolutionary coding-agent pipeline with automated evaluators Novikov et al. (2025). Collectively, these studies show that LLM proposals become more reliable when validity and quality are determined by external execution rather than linguistic judgment alone. Our framework adopts this principle, but its search objects are structured traffic interventions and its evaluator is a microscopic traffic simulator. 3.2 Multi-objective, Compositional, and Multi-agent Search Most LLM-based algorithm-design systems search for a single program or heuristic under a scalar score. MEoH extends this setting by maintaining non-dominated heuristics and accounting for both objective-space dominance and program diversity Yao et al. (2025). E2OC addresses interdependent operator design by representing composition as a sequential decision process and using Monte Carlo tree search Qiu et al. (2026), while OptiMUS illustrates a complementary form of execution grounding in which LLM-generated optimization models are checked and solved by mathematical-programming solvers Ahmaditeshnizi et al. (2024). Multi-agent debate further suggests that role separation and critique can improve proposal diversity and reasoning quality Du et al. (2024). In our framework, signal, network, and routing agents therefore select feedback-guided refinements for complementary intervention families, but neither the agents nor their predicted effects determine the final plan. Valid single interventions are evaluated independently, promising candidates are combined subject to compatibility and evaluation-budget constraints, and the baseline, singles, and bundles form a common candidate pool. A constraint-aware selector is used for the primary experiment, while an offline Pareto dominance audit checks secondary traffic outcomes without changing the selected plan, following the general non-dominated-selection principle used in multi-objective optimization Deb et al. (2002). 3.3 LLMs for Traffic Control and Simulation-Grounded Optimization Before LLM-based control, learning-based traffic-signal research developed simulator-trained policies such as the max-pressure-inspired PressLight Wei et al. (2019a) and the graph-attention-based CoLight Wei et al. (2019b). Recent LLM approaches instead use language models for traffic reasoning and decision support. LLM-Assisted Light augments an LLM with traffic perception and control tools Wang et al. (2024); LLMLight directly selects signal phases from structured observations Lai et al. (2025); and CoLLMLight extends this paradigm to cooperative, network-wide control Yuan et al. (2026). Hybrid methods such as iLLM-TSC inspect and correct reinforcement-learning actions under abnormal or partially observed conditions Pang et al. (2024), whereas EvolveSignal searches for executable signal-control programs using simulator feedback Wang et al. (2025). A related line of work uses LLM agents to automate traffic-simulation workflows: ChatSUMO translates natural-language requests into SUMO scenarios Li et al. (2024), ChatSUMO-Agent adds closed-loop planning, tool execution, and refinement Li et al. (2026), and the concurrent TrafficSimAgent preprint studies hierarchical agents for general traffic-simulation tasks Du et al. (2025). In contrast, our focus is not scenario generation or stepwise signal control, but simulation-grounded search over heterogeneous network, routing, and signal interventions. SUMO provides the executable microscopic evaluator Lopez et al. (2018), while the cross-layer formulation is motivated by the established coupling between signal control and route assignment Meneguzzer (1997) and between network design and travelersā route choices Yang and Bell (1998). 4 Methodology 4.1 Framework and Traffic Evidence The proposed framework separates grounded exploration, LLM-guided refinement, and simulation-based decision making. As shown in Figure 1, an unmodified SUMO scenario is first simulated to establish a baseline. A deterministic diagnosis module extracts traffic measurements and a compact bottleneck-centered subgraph. A seeded grounded sampler initializes one intervention from each control family, which is validated and evaluated independently. A feedback-guided beam then sends measured outcomes back to the responsible specialists for one LLM-selected local refinement. The remaining budget is used for compatible two-action bundles. If no unseen legal refinement or compatible bundle is available, the budget remains unused rather than being filled by a deterministic backfill policy. The best measured candidate is selected subject to explicit safety constraints. The LLM supplies grounded search directions, whereas the orchestrator enumerates legal local alternatives and SUMO provides the decision evidence. Figure 1: Simulation-grounded multi-agent traffic optimization framework. Grounded sampling initializes diverse interventions; specialist LLM agents refine measured parents, after which compatible actions may be composed before constrained final selection. The input consists of a SUMO network, an OD-demand file, and a simulation configuration. The baseline run produces vehicle-level tripinfo and 60-second edge measurements. The diagnosis aggregates speed, density, occupancy, waiting time, time loss, and traffic exposure. For each edge e, it computes Ce=0.35ā(1āre)+0.25āĻā(Ļe,35)+0.20āĻā(oe,30)+0.10āĻā(Le,300)+0.10āĻā(We,300), splitC_e=&0.35(1-r_e)+0.25\,Ļ( _e;35)+0.20\,Ļ(o_e;30)\\ &+0.10\,Ļ(L_e;300)+0.10\,Ļ(W_e;300), split (1) where rer_e is the observed-to-permitted speed ratio, Ļe _e is density, oeo_e is occupancy, LeL_e and WeW_e are accumulated time loss and waiting time, and Ļā”(x,h)=minā”(1,maxā”(0,x/h))Ļ(x;h)= (1, (0,x/h)). The score CeC_e is computed once from the unmodified baseline and is used only as a diagnostic ranking index. It is not the objective optimized by SimTIO and is not recomputed after an intervention. To reduce sensitivity to microscopic network segmentation, non-internal edges shorter than 10 m are excluded before ranking. Let ā°eligibleE_eligible denote the remaining edges; the fixed target set is =Top10eāā°eligibleā”(Ce).T=Top10_e _eligible(C_e). (2) Thus, CeC_e determines where the optimizer searches, whereas candidate quality is determined by the simulated total time loss L_T on the fixed set T. Up to five highest-ranked edges also seed a one-hop local subgraph, limited to 80 edges, that provides agents with valid neighbor, junction, and traffic-light identifiers without inserting the full network XML into their prompts. Freezing T prevents a moving target in which an intervention appears effective merely because congestion is shifted to different edges. The local subgraph retains enough topology to identify contiguous corridors and nearby controllers while limiting prompt length. 4.2 Optimization Problem Formulation The optimization object is an executable intervention plan, not an LLMās prose recommendation or predicted KPI value. We represent a plan as =(xsig,xnet,xroute)ācap,x= (x^sig,x^net,x^route ) _cap, (3) where an inactive component is denoted by ā . xsigx^sig specifies a traffic-light controller and a fixed-time phase plan derived using Websterās signal-timing method Webster (1958); xnetx^net specifies either a corridor-speed change, an edge removal, or a one-way conversion together with its target and discrete parameter; and xroutex^route specifies an avoided edge and the rerouted fraction. The capability-grounded domain capX_cap contains only schema-valid actions, scenario-valid identifiers, and supported discrete values. A single plan activates one component, while a compatible bundle activates at most two components. This cardinality limit and the compatibility rules prevent the optimizer from combining conflicting edits or changing the same control family twice. Let the deterministic SUMO evaluation map be ā”()=(L,L,D,Q,H,E,V),S(x)= (L_T,L_N,D,Q,H,E,V ), (4) where L_T is total time loss on the baseline-frozen Top-10 edges, L_N is time loss on their one-hop neighbors, D is network-wide mean vehicle time loss, Q is throughput, H is the teleport count, E is total CO2CO_2, and V is mean speed. Superscript 00 below denotes the response of the unmodified scenario 0x_0. We also define Aā”()=1A(x)=1 when all demand records and edge-level OD pairs are preserved. Let evalācapC_eval _cap denote the plans actually evaluated within the simulation budget. The optimizer returns the best measured feasible incumbent: ā±eval=āeval: _eval=\x _eval: A()=1,L()ā¤0.99L0, A(x)=1, L_T(x)ā¤ 0.99L_T^0, (5) D()ā¤1.01D0,L()ā¤1.05L0, D(x)ā¤ 1.01D^0, L_N(x)ā¤ 1.05L_N^0, Q()ā„Q0,H()ā¤H0, Q(x)ā„ Q^0, H(x)ā¤ H^0\, x =argā”mināā±evalāLā(),ā±evalā ā ,0,ā±eval=ā . = cases _x _evalL_T(x),&F_evalā ,\\ x_0,&F_eval= . cases Thus, the optimized quantity is frozen-bottleneck time loss. The 1% Top-10 requirement is an acceptance threshold; global delay and neighboring time loss are bounded spillover constraints; and demand integrity, throughput, and teleports are operational constraints. Emissions and speed are reported as secondary outcomes and enter only the offline Pareto dominance audit. This separation prevents an LLM-predicted effect from being mistaken for the objective used by the primary optimizer. Because evalC_eval is budget limited, x is the best evaluated incumbent rather than a claim of global optimality over capX_cap. 4.3 Grounded Random Initialization The diagnosis defines a discrete search space over the baseline-frozen objective edges. Before sampling, the executable capability validator removes non-optimizable controllers, missing reverse-edge pairs, detectable local topology violations, and malformed controls. This pre-execution check confirms that a candidate is supported and locally well formed; complete demand and OD preservation are checked again after the modification is materialized. Using the recorded case seed, the primary optimizer samples one validator-confirmed proposal from each available specialist family in round-robin order. The signal family selects an evidenced controller for fixed-time green-split optimization based on Websterās method while preserving phase states and yellow/all-red transition intervals. The network family samples a legal corridor-speed, edge-removal, or one-way action; corridor speed factors belong to 0.7,0.8,0.9\0.7,0.8,0.9\. The initial routing family uses a one-edge trust region and chooses Ī±ā0.25,0.50,0.75,1.00Ī±ā\0.25,0.50,0.75,1.00\ (6) as the rerouted fraction. The first three samples are exactly the prefix of the grounded-random baseline generated with the same seed. Thus, the LLM-feedback and grounded-random conditions share the same first three actions and differ only in how the remaining simulation budget is allocated. Only vehicles whose reference routes traverse a selected edge are eligible, and avoidance is soft so that endpoints and unavoidable connections remain valid. Unselected vehicles retain their reference routes, and every modified demand file must preserve all records and OD pairs. Specialization prevents an agent from editing the entire scenario. As illustrated in Figure 2, signal proposals change green allocation without altering phase states or transitions; corridor edges are resolved deterministically from an evidenced anchor; and the routing utility, rather than the LLM, selects eligible vehicles and verifies route connectivity. The model therefore supplies high-level search choices, while deterministic code performs the network edits. Figure 2: Intervention mechanisms supported by the optimizer: fixed-time signal green-split adjustment, corridor-level speed control, and partial rerouting around a diagnosed bottleneck. Grounded initialization selects the initial targets and parameters, while an LLM feedback specialist selects a validator-confirmed one-parameter refinement. Deterministic utilities apply the corresponding SUMO modifications. Each initialized action is encoded as a structured proposal containing an executable action, target, parameters, and evidence references. Before execution, a deterministic capability registry checks that the action is implemented, required artifacts and parameters are present, and every edge or traffic-light identifier exists in the scenario. Invalid fractions, placeholders, unsupported operations, and incomplete proposals are rejected. Validation is performed in two stages. Schema validation first checks the proposal type, required fields, numeric ranges, and argument names. Scenario validation then binds those arguments to the current SUMO files and checks identifier existence and action-specific preconditions. This design prevents an unsupported proposal or mutation from entering the expensive simulation stage and provides a trace from every executable modification back to baseline evidence. 4.4 Simulation-Grounded Candidate Search Every validated random-initial proposal is applied to an isolated copy of the baseline scenario and evaluated through a complete SUMO simulation. All candidates start from identical baseline artifacts, and a failed candidate does not affect the remaining evaluations. The resulting vehicle- and edge-level outputs are processed using the same definitions as the baseline, producing measured KPIs and explicit constraint violations for each intervention. The feedback-guided configuration then retains a feasible-first beam ā¬tB_t of width b. Ties are resolved by measured objective value, while role diversity prevents minor variants from one intervention family from occupying the entire beam. For every parent cāā¬tc _t, its specialist receives the measured KPI vector, Top-10 improvement, global and neighbor changes, and failed constraints. Deterministic code first enumerates a finite catalog ā³ā”(c)M(c) of unseen, validator-confirmed mutations that retain the action and change exactly one decision argument. The specialist then selects one mutation identifier mā”(c)āā³ā”(c)m(c) (c) using the measured feedback; it cannot introduce a new identifier or numeric value. The update is t+1=tāŖmā”(c):cāā¬t,ā¬t+1=Topbfeas,divā”(t+1),C_t+1=C_tāŖ\m(c):c _t\, _t+1=Top_b^feas,div(C_t+1), (7) where every selected mutation is revalidated and re-simulated before it can enter the next beam. An invalid catalog selection receives one LLM repair attempt. Action changes, multi-argument rewrites, and previously tested duplicates are rejected without consuming the SUMO budget. The initial configuration uses one grounded-initialization round followed by one feedback-refinement round with beam width b=2b=2; disabling the refinement round recovers a one-shot search. Candidate simulation budget that cannot be used because a refinement is invalid or duplicated remains unused; all feedback-guided mutations are chosen by the responsible LLM specialist from the program-generated legal catalog. After feedback refinement, the system retains the top K single candidates and enumerates compatible pairs within a remaining simulation budget B. A bundle contains at most one action from each intervention family, and actions that conflict on the same edge or controller are rejected. Compatible actions are applied in the order topologyāspeedāroutingāsignal,topology , (8) so each action is rebound to the artifacts produced by the preceding step. The complete bundle is then re-simulated rather than scored by adding its constituentsā predicted or isolated effects. Successfully evaluated singles and bundles form the final candidate pool evalC_eval. The beam width, Top-K filter, and bundle budget control the number of additional SUMO runs. Singles are ordered using their measured performance, after which the compatibility checker excludes repeated intervention families and conflicting targets, such as speed modification and soft avoidance of the same edge. The sum of single-candidate scores is used only to prioritize which compatible pairs receive the limited simulation budget. It is not treated as the bundleās performance because two individually useful actions can reinforce, cancel, or shift the effects of one another. The bundle run therefore captures interaction effects that cannot be inferred reliably from the isolated evaluations. In the reported experiments, K=4K=4, at most two bundles are simulated, and the total candidate budget is capped at seven SUMO evaluations per case; these settings apply equally to the LLM and non-LLM proposal-source conditions. 4.5 LLM Feedback Prompts The three role-specific system prompts used during feedback refinement are reported below. Grounded initialization is seeded and programmatic and therefore has no LLM prompt. Each specialist receives the retained parent action, measured SUMO outcomes, constraint violations, baseline evidence, and a program-generated catalog of unseen validator-confirmed one-argument mutations. The model may select one catalog identifier but cannot rewrite an action, invent an identifier, or introduce a numeric value. Responses use JSON mode and are limited to 500 output tokens; an invalid selection receives one repair attempt. Bundle construction, compatibility filtering, simulation, Pareto auditing, and final selection remain deterministic. 4.5.1 Signal-Control Specialist ā¬ You are the signal-control feedback specialist in a traffic optimization team. The input contains: - parent_proposal: the evaluated signal action; - measured_outcomes: baseline-relative SUMO KPIs; - constraint_violations: failed acceptance or safety constraints; - legal_mutations: unseen validator-confirmed one-argument mutations. Return EXACTLY ONE JSON object: "selected_mutation_id": "...", "reason": "...", "evidence_refs": ["..."], "risk": "..." Rules: - Select exactly one ID that appears in legal_mutations. - Do not rewrite the action or arguments and do not invent an ID, TLS target, phase state, cycle length, split, or numeric value. - Compare all catalog entries using the measured parent outcome and supplied baseline evidence; do not automatically choose the first entry. - Prefer a mutation that addresses failed global-delay or spillover constraints while preserving the measured Top-10 benefit. - Preserve yellow/all-red states and consider minor-movement starvation and downstream queue transfer. - reason and risk must be substantive. - If legal_mutations is empty, return selected_mutation_id="" and explain why. 4.5.2 Road-Network Intervention Specialist ā¬ You are the road-network feedback specialist in a traffic optimization team. The input contains: - parent_proposal: the evaluated network action; - measured_outcomes: baseline-relative SUMO KPIs; - constraint_violations: failed acceptance or safety constraints; - legal_mutations: unseen validator-confirmed one-argument mutations. Return EXACTLY ONE JSON object: "selected_mutation_id": "...", "reason": "...", "evidence_refs": ["..."], "risk": "..." Rules: - Select exactly one ID that appears in legal_mutations. - Do not rewrite the action or arguments and do not invent an ID, edge target, corridor member, speed factor, or numeric value. - Compare all catalog entries using the measured parent outcome and supplied baseline evidence; do not automatically choose the first entry. - set_corridor_speed is a static corridor-level speed-limit adjustment; the deterministic tool resolves up to five same-road edges from one anchor. - Prefer a mutation that repairs global-delay or neighbor-spillover violations without giving up the measured Top-10 benefit. - For topology actions, consider OD disconnection, route loss, detour, and congestion transfer. For speed actions, consider queue smoothing and downstream spillback. - reason and risk must be substantive. - If legal_mutations is empty, return selected_mutation_id="" and explain why. 4.5.3 Demand-Routing Specialist ā¬ You are the demand-routing feedback specialist in a traffic optimization team. The input contains: - parent_proposal: the evaluated single-edge rerouting action; - measured_outcomes: baseline-relative SUMO KPIs; - constraint_violations: failed acceptance or safety constraints; - legal_mutations: unseen validator-confirmed one-argument mutations. Return EXACTLY ONE JSON object: "selected_mutation_id": "...", "reason": "...", "evidence_refs": ["..."], "risk": "..." Rules: - Select exactly one ID that appears in legal_mutations. - Do not rewrite the action or arguments and do not invent an ID, avoided edge, rerouting fraction, or numeric value. - The trust region contains exactly one avoided edge. Legal rerouting fractions are 0.25, 0.5, 0.75, and 1.0. - Compare all catalog entries using the measured parent outcome and supplied baseline evidence; do not automatically choose the first entry. - Prefer a mutation that reduces displaced congestion, excessive detour, or a failed global-delay or neighbor-spillover constraint while retaining local benefit. - Every demand record and its exact edge-level OD must remain unchanged; rerouting is soft avoidance and may retain an unavoidable endpoint edge. - reason and risk must be substantive. - If legal_mutations is empty, return selected_mutation_id="" and explain why. 4.6 Constrained Final Selection Demand preservation is a hard feasibility requirement. A candidate is discarded if it changes the number of demand records or any OD pair. Throughput is measured against the original source demand as Q=Ncompleted/NsourceQ=N_completed/N_source, preventing an intervention from appearing beneficial by silently removing difficult trips. This formulation avoids assigning an arbitrary fixed delay penalty to unfinished vehicles. Instead, incomplete service is exposed directly through throughput and teleport constraints. Delay and time loss are computed from the vehicles and edge measurements actually reported by SUMO, while demand preservation ensures that the denominator cannot be changed by an intervention. The same aggregation code and simulation horizon are used for the baseline and every candidate. Equation 5 is applied only to measured SUMO outcomes. An intervention must therefore produce at least a 1% bottleneck improvement, may increase global delay by at most 1%, may increase neighboring time loss by at most 5%, and cannot reduce throughput or increase teleports. The unmodified scenario 0x_0 acts as a no-operation guard. The rule distinguishes target improvement from network-wide acceptability. L_T tests whether the diagnosed bottlenecks improve, L_N detects immediate spillover, and D prevents a local gain from being purchased through worse overall travel conditions. Throughput and teleports serve as operational safeguards. The 1% threshold suppresses changes too small to be operationally meaningful, while the neighbor tolerance permits limited redistribution around a bottleneck. If several candidates pass all requirements, minimizing frozen Top-10 time loss keeps the final decision aligned with the stated local optimization objective. 4.7 Pareto Dominance Audit To check whether the delay-oriented selector discards an unambiguously better candidate, an offline analysis examines each saved candidate pool using four baseline-relative gains: gTā(c) g_T(c) =1āLā(c)Lā(c0), =1- L_T(c)L_T(c_0), gDā(c) g_D(c) =1āDā”(c)Dā”(c0), =1- D(c)D(c_0), (9) gEā(c) g_E(c) =1āEā”(c)Eā”(c0), =1- E(c)E(c_0), gVā(c) g_V(c) =Vā”(c)Vā”(c0)ā1. = V(c)V(c_0)-1. Candidates must preserve complete demand, throughput, teleports, and the global and neighbor bounds in Equation 5. A candidate is Pareto-optimal when no other feasible candidate is at least as good in all four gains and strictly better in one. The baseline remains in every pool. This audit requires no additional LLM calls or SUMO runs and does not alter the selected plan: it reports whether the plan chosen by the primary Top-10 delay objective lies on the measured Pareto front. Consequently, no manually weighted preference profiles are introduced into the optimizer. 5 Experimental Evaluation We first compare the LLM-feedback optimizer with grounded random search and a deterministic heuristic under the same microscopic-simulation budget. We then reuse the measured LLM candidate pools for an offline Pareto dominance audit. All aggregates include no-operation outcomes as zero improvement and therefore evaluate the complete guarded optimizer rather than only its successful interventions. 5.1 Experimental Setup We use five urban road networks derived from OpenStreetMap (OSM) data (Ā© OpenStreetMap contributors) OpenStreetMap contributors (n.d.) and three demand seeds. Each frozen demand file was generated with randomTrips.py 1.24.0 and contains 2,400 passenger trips departing over one hour. Baselines and interventions are evaluated with Eclipse SUMO 1.27.1. All feedback-refinement calls use the gpt-5.6-luna model. The demand is synthetic and is not calibrated to observed counts; the experiments therefore test optimization behavior across network structures rather than real-world traffic fidelity. Table 1 summarizes the scenarios and Table 2 summarizes the model, search-budget, candidate-space, SUMO, and pairing settings used in the main comparison. Table 1: Equal-demand SUMO scenarios. Delay is baseline mean vehicle time loss, reported as mean Ā± sample standard deviation across three seeds. City Edges Junctions TLS Trips Delay (s) Albany 6,193 4,309 107 2,400 112.80Ā±0.26112.80Ā± 0.26 Troy 6,107 4,355 57 2,400 99.79Ā±0.8499.79Ā± 0.84 Ann Arbor 10,151 7,522 81 2,400 138.58Ā±3.13138.58Ā± 3.13 Los Angeles 6,969 4,954 172 2,400 175.76Ā±1.26175.76Ā± 1.26 Manhattan 5,993 4,065 256 2,400 245.31Ā±13.03245.31Ā± 13.03 Table 2: Reproducibility settings for the main comparison. Component Setting Network source Five urban road networks derived from OpenStreetMap data (Ā© OpenStreetMap contributors). LLM OpenAI Chat Completions API; gpt-5.6-luna; model-default sampling; JSON response mode; 500 output tokens; 60-s timeout and at most two retries after the initial request, including one structured repair attempt for an invalid catalog selection. Search budget At most seven candidate SUMO evaluations per case. LLM search uses three grounded initial proposals, up to two feedback mutations, and up to two compatible bundles; each baseline uses five proposals and up to two bundles. Candidate space Signal timing, five-edge corridor speed control, single-edge rerouting, and validated topology actions. Numeric choices are discrete, and Top-10 ranking excludes edge fragments shorter than 10 m. Signal timing Fixed-time cycle and green-split plans derived using Websterās method; original phase states and yellow/all-red transition intervals are preserved. SUMO Version 1.27.1; 1-s step; complete execution until all vehicles arrive; trip-level statistics and emissions plus 60-s edge measurements. Pairing For each city and seed, all methods use the same demand, OD records, baseline diagnosis, objective, safety constraints, and SUMO seed. Only intervention-targeted artifacts are changed. Final selection minimizes time loss on the baseline-frozen Top-10 edges subject to Equation 5. A selected plan must preserve demand and OD pairs, must not reduce throughput or increase teleports, and may increase global mean delay and one-hop neighboring time loss by at most 1% and 5%, respectively. The unmodified baseline is always an explicit candidate. 5.2 Budget-Matched Proposal-Source Comparison The LLM feedback condition begins with one validator-confirmed random proposal per intervention family. After SUMO evaluation, the model selects a single mutation from a finite legal catalog for each retained beam parent. The grounded random baseline samples supported actions, diagnosed edge and TLS identifiers, speed factors in 0.7,0.8,0.9\0.7,0.8,0.9\, and rerouting fractions in 0.25,0.5,0.75,1.0\0.25,0.5,0.75,1.0\. The deterministic heuristic targets the highest-ranked bottleneck for corridor and routing actions and the nearest TLS in the diagnosed subgraph. The downstream validator, bundle search, simulator, and selector are shared by all three methods. In addition to final Top-10 and global-delay reductions, we report proposal validity and optimization success, defined as selecting a feasible plan with at least 1% Top-10 reduction. Candidate-level simulation efficiency is Ī·sim=Nfeasible,ĪTop10ā„1%NSUMO. _sim= N_feasible,\, _Top10ā„ 1\%N_SUMO. (10) Table 3: Budget-matched proposal-source comparison over five networks and three demand seeds. Success denotes a feasible selected plan with at least 1% Top-10 reduction. Method Valid (%) Success (%) Best nets. Sim. eff. (%) Top-10 (%) Global (%) LLM feedback 100.0 86.7 3/5 35.1 9.18 2.78 Grounded random 100.0 73.3 1/5 34.0 9.00 3.03 Heuristic 97.3 80.0 1/5 39.0 9.60 2.69 Figure 3: Proposal-source comparison over three demand seeds. Bars in (a) show mean Top-10 reduction and sample-standard-deviation error bars. Panel (b) reports case-level optimization success and candidate-level simulation efficiency under the common seven-evaluation cap. Table 3 compares the three proposal sources under the same simulation budget. Mean Top-10 reductions were 9.18%Ā±10.18%9.18\%Ā± 10.18\%, 9.00%Ā±10.88%9.00\%Ā± 10.88\%, and 9.60%Ā±10.32%9.60\%Ā± 10.32\% for LLM feedback, grounded random, and the heuristic, respectively. Across the 15 cityāseed cases, LLM feedback attained the highest optimization success rate (86.7%) and the lowest no-op rate (13.3%). It also achieved the best seed-averaged Top-10 result in Troy, Ann Arbor, and Los Angeles, whereas the heuristic led in Albany and grounded random led in Manhattan. Their global-delay reductions were 2.78%, 3.03%, and 2.69%. The LLMās simulation efficiency was 35.1%, between grounded random (34.0%) and the heuristic (39.0%). Six final LLM plans were feedback-refined singles; together with four selected bundles, these results show that measured outcomes changed the search even though the aggregate advantage was not significant. Paired Wilcoxon tests on Top-10 reduction gave p=0.407p=0.407 against random and p=0.639p=0.639 against the heuristic. Thus, the evidence supports broader cross-network success rather than universal superiority over specialized heuristics. Performance also varies with the demand realization. Averaged over the five networks, LLM feedback led seed 42 (12.38% Top-10 reduction) and seed 44 (10.05%), whereas the heuristic led seed 43 (9.03%). At the paired cityāseed level, LLM feedback beat, tied, and lost to grounded random in six, six, and three cases, respectively. It beat the heuristic in six cases and lost in nine. The high cross-case standard deviations in Figure 3 therefore reflect both demand-seed sensitivity and genuine differences between network structures, particularly the large attainable improvements in Manhattan. Table 4 reports the proposed optimizer separately for each network. Ann Arbor, Los Angeles, and Manhattan accepted a feasible improving plan for all three demand seeds; Albany and Troy each retained the baseline once. Manhattan provides the largest mean reduction and the highest candidate yield, while the smaller Albany improvements illustrate that the baseline guard does not force an intervention when the legal search space offers only marginal or unsafe changes. Table 4: Five-network performance of the LLM-feedback optimizer over three demand seeds. Reductions are mean Ā± sample standard deviation in percentage points and include no-op outcomes. Network Top-10 (%) Global delay (%) Sim. eff. (%) Success Albany 2.25 Ā± 2.11 0.17 Ā± 0.52 21.1 2/3 Troy 8.55 Ā± 11.17 0.75 Ā± 1.27 28.6 2/3 Ann Arbor 8.25 Ā± 5.83 1.75 Ā± 0.46 30.0 3/3 Los Angeles 5.58 Ā± 6.84 0.51 Ā± 1.28 31.6 3/3 Manhattan 21.27 Ā± 14.23 10.71 Ā± 7.55 66.7 3/3 Overall 9.18 Ā± 10.18 2.78 Ā± 5.08 35.1 13/15 5.3 Search Behavior and Intervention Composition Figure 4 examines how each method reaches its final decision. LLM feedback selected five TLS plans, four corridor-speed plans, four bundles, and two no-ops. Grounded random selected the same number of TLS plans and bundles but retained the baseline in four cases, consistent with its lower optimization-success rate. The deterministic heuristic was more concentrated: nine of its 12 applied plans were bundles and the remaining three used corridor-speed control. Routing and topology candidates were evaluated, but none won the final constrained selection because their local benefits did not compensate for global-delay, spillover, or demand-preservation failures. Only three of the 13 applied LLM plans were unmodified initial singles. Six were produced by simulation-feedback refinement and four were selected by bundle search. Thus, ten applied decisions required a second-stage search operation rather than merely accepting the grounded random initialization. This distinction is important because the aggregate comparison otherwise understates how often measured simulator feedback changes the executable plan. Figure 4: Search outcomes over 15 cityāseed cases. Panel (a) reports the intervention family selected by each proposal source. Panel (b) separates initial, feedback-refined, bundled, and no-op outcomes for the LLM search. 5.4 Pareto Dominance Audit We audit the same 15 LLM candidate pools without additional model calls or SUMO runs. Dominance is evaluated jointly over Top-10 delay, global delay, CO2CO_2, and mean speed after applying the operational constraints. The resulting fronts contain 1.53 candidates on average (range 1ā4), and 11 of the 15 plans selected by the primary Top-10 rule are nondominated. The four dominated selections show that optimizing bottleneck delay can sacrifice a secondary metric, while the generally small fronts indicate limited multi-objective diversity in the current delay-oriented candidate pools. We therefore report Pareto status as a diagnostic rather than introducing manually weighted profiles or changing the optimizerās primary objective. 6 Conclusions This study presented a simulation-grounded multi-agent LLM framework for traffic intervention optimization in SUMO. The framework separates grounded initialization, feedback-guided local refinement, and final decision making. A seeded sampler constructs initial signal-control, corridor-speed, and routing actions, after which role-specific LLM agents use measured SUMO outcomes to select one-parameter refinements from validator-confirmed catalogs. Deterministic utilities materialize the actions, and every valid single action and compatible pairwise bundle is evaluated through a complete simulation. The returned plan is the best evaluated candidate satisfying constraints on global delay, neighboring-road spillover, throughput, demand preservation, and teleport events; otherwise, the unmodified scenario is retained. Across five road networks and three demand seeds, the guarded optimizer reduced Top-10 bottleneck time loss by 9.18%Ā±10.18%9.18\%Ā± 10.18\% and network-wide delay by 2.78%. It selected a feasible improving plan in 86.7% of cases, compared with 73.3% for grounded random search and 80.0% for the deterministic heuristic under the same seven-simulation cap. Only three of the 13 applied LLM plans were unmodified initial actions; six were feedback-refined singles and four were bundles, showing that the second-stage search frequently changed the executable decision. However, paired Wilcoxon tests found no statistically significant Top-10 advantage over grounded random search (p=0.407p=0.407) or the deterministic heuristic (p=0.639p=0.639); the heuristic also produced a slightly larger mean Top-10 reduction. The evidence therefore supports improved feasible-plan coverage and competitive search performance rather than universal superiority of LLM-guided refinement. The current findings are limited by synthetic demand, three demand seeds, one LLM model and delay-oriented feedback prompt, and a small simulation budget. The comparison also does not isolate LLM mutation selection from the surrounding beam-search structure using identical-parent random and rule-based selectors, and it does not include an established traffic-control or simulation-optimization baseline. Future work should address these limitations using calibrated or observed demand, additional LLM and demand realizations, matched mutation-selector ablations, and conventional traffic-control benchmarks. Larger bundles, coordinated controllers, and time-varying policies also remain open extensions. Overall, the results position LLMs as constrained, feedback-guided local search operators rather than autonomous traffic controllers: executable tools, microscopic simulation, and explicit safety constraints retain authority over the final intervention. AUTHOR CONTRIBUTIONS The authors confirm their contributions to the paper as follows: Shuyang Li: conceptualization, methodology, software development, investigation, formal analysis, visualization, and writingāoriginal draft. Ruimin Ke: supervision and writingāreview and editing. Both authors reviewed and approved the final version of the manuscript. GENERATIVE AI DISCLOSURE The authors reviewed and verified the resulting text and take full responsibility for the manuscriptās content. The use of gpt-5.6-luna as an experimental component of SimTIO is reported separately in the Methodology section. References Ahmaditeshnizi et al. (2024) A. Ahmaditeshnizi, W. Gao, and M. Udell OptiMUS: scalable optimization modeling with (MI)LP solvers and large language models. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235, p. 577ā596. External Links: Link Cited by: Ā§3.2. Deb et al. (2002) K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan A fast and elitist multiobjective genetic algorithm: NSGA-I. IEEE Transactions on Evolutionary Computation 6 (2), p. 182ā197. External Links: Document Cited by: Ā§3.2. Du et al. (2024) Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235, p. 11733ā11763. External Links: Link Cited by: Ā§3.2. Du et al. (2025) Y. Du, J. Zhang, J. Feng, Z. Liu, J. Yuan, and Y. Li TrafficSimAgent: a hierarchical agent framework for autonomous traffic simulation with MCP control. arXiv preprint arXiv:2512.20996. External Links: 2512.20996, Link Cited by: Ā§3.3. Lai et al. (2025) S. Lai, Z. Xu, W. Zhang, H. Liu, and H. Xiong LLMLight: large language models as traffic signal control agents. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 2335ā2346. External Links: Document Cited by: Ā§3.3. Li et al. (2024) S. Li, T. Azfar, and R. Ke ChatSUMO: large language model for automating traffic scenario generation in simulation of urban mobility. IEEE Transactions on Intelligent Vehicles. External Links: Document Cited by: Ā§3.3. Li et al. (2026) S. Li, M. Ma, T. Azfar, and R. Ke ChatSUMO-Agent: an LLM-based agent for conversational traffic simulation in SUMO. Transportation Research Part C: Emerging Technologies. External Links: Document Cited by: Ā§3.3. Liu et al. (2024) F. Liu, T. Xialiang, M. Yuan, X. Lin, F. Luo, Z. Wang, Z. Lu, and Q. Zhang Evolution of heuristics: towards efficient automatic algorithm design using large language model. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235, p. 32201ā32223. External Links: Link Cited by: Ā§3.1. Lopez et al. (2018) P. A. Lopez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y. Flƶtterƶd, R. Hilbrich, L. LĆ¼cken, J. Rummel, P. Wagner, and E. Wiessner Microscopic traffic simulation using SUMO. In 2018 21st International Conference on Intelligent Transportation Systems, p. 2575ā2582. External Links: Document Cited by: Ā§3.3. Madaan et al. (2023) A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, S. Gupta, B. P. Majumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P. Clark Self-Refine: iterative refinement with self-feedback. In Advances in Neural Information Processing Systems, Vol. 36. External Links: Link Cited by: Ā§3.1. Meneguzzer (1997) C. Meneguzzer Review of models combining traffic assignment and signal control. Journal of Transportation Engineering 123 (2), p. 148ā155. External Links: Document Cited by: Ā§3.3. Novikov et al. (2025) A. Novikov, N. VÅ©, M. Eisenberger, E. Dupont, P. Huang, A. Z. Wagner, S. Shirobokov, B. Kozlovskii, F. J. R. Ruiz, A. Mehrabian, M. P. Kumar, A. See, S. Chaudhuri, G. Holland, A. Davies, S. Nowozin, P. Kohli, and M. Balog AlphaEvolve: a coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131. External Links: 2506.13131, Link Cited by: Ā§3.1. OpenStreetMap contributors (n.d.) OpenStreetMap contributors OpenStreetMap. Note: OpenStreetMap data External Links: Link Cited by: Ā§5.1. Pang et al. (2024) A. Pang, M. Wang, M. Pun, C. S. Chen, and X. Xiong iLLM-TSC: integration reinforcement learning and large language model for traffic signal control policy improvement. arXiv preprint arXiv:2407.06025. External Links: 2407.06025, Link Cited by: Ā§3.3. Qiu et al. (2026) J. Qiu, X. Chen, L. Ge, L. Lin, Z. Lu, and Q. Zhang Evolving interdependent operators with large language models for multi-objective combinatorial optimization. arXiv preprint arXiv:2601.17899. External Links: 2601.17899, Link Cited by: Ā§3.2. Romera-Paredes et al. (2024) B. Romera-Paredes, M. Barekatain, A. Novikov, M. Balog, M. P. Kumar, E. Dupont, F. J. R. Ruiz, J. S. Ellenberg, P. Wang, O. Fawzi, P. Kohli, and A. Fawzi Mathematical discoveries from program search with large language models. Nature 625 (7995), p. 468ā475. External Links: Document Cited by: Ā§3.1. Wang et al. (2025) L. Wang, P. Duan, H. Wang, Y. Wang, J. Xu, N. Zheng, and Z. Ma EvolveSignal: a large language model powered coding agent for discovering traffic signal control algorithms. arXiv preprint arXiv:2509.03335. External Links: 2509.03335, Link Cited by: Ā§3.3. Wang et al. (2024) M. Wang, A. Pang, Y. Kan, M. Pun, C. S. Chen, and B. Huang LLM-Assisted Light: leveraging large language model capabilities for human-mimetic traffic signal control in complex urban environments. arXiv preprint arXiv:2403.08337. External Links: 2403.08337, Link Cited by: Ā§3.3. Webster (1958) F. V. Webster Traffic signal settings. Road Research Technical Paper Technical Report 39, Her Majestyās Stationery Office, Road Research Laboratory, London, United Kingdom. Cited by: Ā§4.2. Wei et al. (2019a) H. Wei, C. Chen, G. Zheng, K. Wu, V. Gayah, K. Xu, and Z. Li PressLight: learning max pressure control to coordinate traffic signals in arterial network. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p. 1290ā1298. External Links: Document Cited by: Ā§3.3. Wei et al. (2019b) H. Wei, N. Xu, H. Zhang, G. Zheng, X. Zang, C. Chen, W. Zhang, Y. Zhu, K. Xu, and Z. Li CoLight: learning network-level cooperation for traffic signal control. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, p. 1913ā1922. External Links: Document Cited by: Ā§3.3. Yang et al. (2024) C. Yang, X. Wang, Y. Lu, H. Liu, Q. V. Le, D. Zhou, and X. Chen Large language models as optimizers. In The Twelfth International Conference on Learning Representations, External Links: Link Cited by: Ā§3.1. Yang and Bell (1998) H. Yang and M. G. H. Bell Models and algorithms for road network design: a review and some new developments. Transport Reviews 18 (3), p. 257ā278. External Links: Document Cited by: Ā§3.3. Yao et al. (2025) S. Yao, F. Liu, X. Lin, Z. Lu, Z. Wang, and Q. Zhang Multi-objective evolution of heuristic using large language model. Proceedings of the AAAI Conference on Artificial Intelligence 39 (25), p. 27144ā27152. External Links: Document Cited by: Ā§3.2. Ye et al. (2024) H. Ye, J. Wang, Z. Cao, F. Berto, C. Hua, H. Kim, J. Park, and G. Song ReEvo: large language models as hyper-heuristics with reflective evolution. In Advances in Neural Information Processing Systems, Vol. 37. External Links: Link Cited by: Ā§3.1. Yuan et al. (2026) Z. Yuan, S. Lai, and H. Liu CoLLMLight: cooperative large language model agents for network-wide traffic signal control. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: Ā§3.3. Zheng et al. (2025) Z. Zheng, Z. Xie, Z. Wang, and B. Hooi Monte carlo tree search for comprehensive exploration in LLM-based automatic heuristic design. In Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 267, p. 78338ā78373. External Links: Link Cited by: Ā§3.1.