{"id":"36d549d2-a5ef-4f60-95fe-f2b948df7a3d","arxiv_id":"2607.18286","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A single preference-conditioned reinforcement-learning policy can be tuned at runtime to trade bus priority against general traffic delay at signalized intersections, and beats fixed-time and rule-based baselines at moderate settings.","lead":"This paper trains a single reinforcement-learning controller for transit signal priority that can be tuned at runtime to trade off bus delays against general traffic delays without retraining. It provides a standardized benchmark on simulated intersections and shows the tuned controller often beats fixed-time and rule-based baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic bus insertion distribution is unvalidated; the reported trade-off frontier and baseline dominance may not transfer to real transit operations.","rationale":"The reader's weakest assumption — that the synthetic timetable-based bus insertion is representative of real transit priority events — is the most load-bearing uncertainty in the paper. The central claim is an operational one: a single preference-conditioned controller gives a useful runtime knob for transit signal priority. That claim is supported only inside a simulated regime whose bus arrival process is generated from uniform distributions chosen for experimental convenience, not from real transit data. If real bus operations exhibit different headway correlations, route structures, or demand patterns, the learned Q-function and the reported dominance over baselines may not transfer, undermining the practical relevance of the contribution. The paper's distribution-shift experiment only varies bus prevalence, leaving headway correlation and route structure untouched, so it does not address this concern. The abstract's unqualified claim is also internally inconsistent with Table III (CHI758 shows near-zero dominance), but that is a presentation issue that reinforces the need for qualification rather than a separate technical flaw. Therefore the reader's conditional verdict stands: the method is promising and the benchmark is reproducible, but the generality of the trade-off frontier and baseline dominance is not established beyond the synthetic regime. No change to the CONDITIONAL verdict is warranted.","tokens_in":12967,"tokens_out":13839,"duration_ms":122957,"concrete_test":"Obtain real bus timetable/AVI data (e.g., GTFS for Chicago or New York), fit headway, departure-offset, and route-duration distributions, and generate evaluation scenarios for the same 13 intersections using these fitted distributions instead of Sec. II-D uniforms. Run the already-trained MORL policies and the FixedTime/RuleTSP baselines at w_bus in {0.3, 0.5, 0.7} with the same 10-seed protocol. If MORL's mean (J_b, J_nb) no longer Pareto-dominates RuleTSP on a majority of intersections, the synthetic-regime dependence is confirmed and the abstract's outperformance claim should be scoped to the synthetic regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an empirical one: a single preference-conditioned policy yields a runtime-tunable bus-priority/delay trade-off that beats fixed-time and rule-based TSP. The evidence for this claim is generated entirely under the synthetic timetable-insertion regime of Sec. II-D: 4 concurrent routes per episode, headways U[300,900] s, departure offsets U[-120,600] s, route durations U[10,800,36,000] s, with total demand rescaled. No validation shows that these distributions are representative of real transit operations (e.g., GTFS/AVI headway distributions, route structures, dwell times, or diurnal demand cycles). The distribution-shift experiment in Sec. V-B only changes bus prevalence, not headway correlation or route-structure realism. Because the learned policy's action-value estimates and the reported dominance over baselines are fitted to this synthetic regime, the central operational claim may fail outside it. The paper acknowledges 'dependence on scenario realism' (Sec. VI) but does not test it. Additionally, the abstract's unqualified 'outperforms fixed-time and rule-based baselines' is contradicted by Table III on CHI758 (FT-dom 1.0, TSP-dom 0.0) and is weak on SEA543, so the claim needs qualification regardless.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a preference-conditioned multi-objective reinforcement learning controller for transit signal priority (TSP). The controller π(a|s,w) is trained once per intersection and can be tuned at runtime through a scalar preference w to trade off bus delay against all-vehicle delay, under a shared feasibility layer enforcing min/max green and transition constraints. The authors build on IntersectionZoo and SUMO-RL, add bus-prevalence and timetable-insertion wrappers, and compare against fixed-time, a rule-based TSP overlay, and fixed-weight PPO specialists. Evaluation is carried out on three case-study intersections plus ten additional intersections, with 10 seeds per setting, pinned simulator versions, and released code. The paper reports empirical Pareto fronts, hypervolume, dominance counts, operating-point delays, tail-delay diagnostics, and a distribution-shift experiment that varies bus prevalence.","tokens_in":13307,"tokens_out":6527,"duration_ms":63605,"significance":"If the reported results hold, the paper makes a useful contribution: it provides a reproducible benchmark for preference-conditioned TSP and demonstrates that a single conditioned policy can approximate the Pareto front between bus priority and general-traffic delay, with a runtime knob that does not require retraining. Strengths include the open-source implementation with pinned IntersectionZoo/SUMO versions, the use of 10 seeds and 13 intersections, a shared feasibility layer for fair baseline comparison, and the inclusion of tail-delay and fairness diagnostics. These are concrete assets for the RL-for-traffic-signal-control community. However, the central claim is currently overstated in the abstract and the evaluation rests on a synthetic timetable-insertion regime whose representativeness is not validated; both issues need to be addressed before the operational claims can be taken at face value.","major_comments":[{"comment":"The abstract states that the controller 'outperforms fixed-time and rule-based baselines' without qualification, but Table III shows CHI758 with FT-dom=1.0 and TSP-dom=0.0, and SEA543 with TSP-dom=3.5±4.1 — meaning the learned policy does not dominate either baseline for most preference settings at these intersections. The body text is more careful, saying 'often dominates' and 'at moderate preferences.' The abstract and the concluding claims should carry the same qualifier, or the dominance claim should be restricted to the case-study intersections at moderate preferences.","section":"Abstract and Section V-D / Table III"},{"comment":"The entire empirical claim is evaluated under the synthetic timetable-insertion regime described in Section II-D: headways U[300,900] s, departure offsets U[−120,600] s, route durations U[10,800,36,000] s, and 4 concurrent routes per episode. The distribution-shift experiment in Section V-B changes only bus prevalence, not headway correlations, route structures, dwell times, or demand patterns. Since the learned policy and baselines are all measured in this constructed regime, the reported trade-off frontier and dominance numbers are not shown to transfer to real transit operations. The limitation paragraph in Section VI acknowledges 'dependence on scenario realism' but does not test it. To support the operational 'runtime-tunable TSP' claim, the authors should validate the timetable model against real GTFS/AVI data, or at minimum add a sensitivity analysis sweeping the headway, route-co","section":"Section II-D and Section V-B"},{"comment":"Section IV-B states that hyperparameters and the preference curriculum were tuned on CHI758 and then held fixed for all remaining intersections. However, Table III and Section V-D include CHI758 in the ten additional intersections used to demonstrate cross-intersection reproducibility, and Table IV uses all 13 intersections for the structural correlation analysis. This violates the holdout principle: CHI758 is not an independent test site. The paper should either exclude CHI758 from the additional-intersection evaluation and the correlation analysis, or explicitly label it as a development intersection whose results are not out-of-sample. The same concern applies to the RuleTSP parameter pilot sweep on CHI758.","section":"Section IV-B and Table III"}],"minor_comments":[{"comment":"The interval '[10,800,36,000]s' is ambiguous because of the comma separator. Use 10 800–36 000 s or write 10800–36000 to avoid confusion with a three-element list.","section":"Section II-D"},{"comment":"The sentence 'We apply a homotopy schedule with λ increasing from 0.0 to 0.3 over 200k steps' introduces λ without defining it. Please specify what λ controls (e.g., reward-shaping weight, Envelope Q-learning parameter) and how it interacts with the preference curriculum.","section":"Section III-D"},{"comment":"The cell color encoding described in the caption may be invisible in grayscale print. Add explicit numeric labels or hatching so the table is readable without color.","section":"Table IV"},{"comment":"The statement 'MORL has slightly higher HV on all three case studies' is supported by the table, but the standard deviations overlap substantially (e.g., CHI2412: 83.7±1.3 vs 83.1±2.2). Consider adding a significance test or explicitly stating that the difference is not statistically assessed.","section":"Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with good reproducibility practices. The main risks are the unqualified abstract claim, the unvalidated synthetic timetable regime, and the CHI758 tuning/evaluation overlap. These are fixable within the scope of the paper, but they are load-bearing for the central claims, so I recommend major revision rather than minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid take: this is a genuinely useful empirical contribution, not a breakthrough. The new thing is the systematic application of preference-conditioned MORL (Envelope Q-learning) to constrained transit signal priority, plus an IntersectionZoo wrapper and a baseline pack that makes results comparable. They ship code, pin SUMO/IntersectionZoo versions, run 10 seeds across 13 intersections, and report Pareto frontiers, hypervolume, and tail-delay diagnostics. That is real work, and it is mostly done carefully. The single-policy runtime tuning claim is supported in the simulated regime: the preference sweep produces a smooth trade-off, and at moderate weights MORL often beats fixed-time and rule-based TSP.\n\nThe biggest gap is the abstract, which says 'outperforms fixed-time and rule-based baselines' without qualification. The paper's own Table III shows CHI758 where MORL dominates FixedTime 1.0/11 and RuleTSP 0.0/11, and SEA543 is also weak. The Discussion rightly limits the claim to moderate preferences, so the abstract needs to match.\n\nI also share the reader's concern about CHI758 being both the tuning site and an evaluation site. It is not fatal — the parameters were held fixed afterward — but it should be stated more clearly. And the three case-study intersections (NYC10802, LA2114, CHI2412) are presented without any selection criterion; a reader will wonder whether they were picked for favorable behavior. The distribution-shift experiment only changes bus prevalence, not route structure or headway correlations, which is a genuine limitation for the operational claim. The paper acknowledges this in Section VI but does not test it.\n\nOne thing the reader flagged that I do not weight as heavily is the synthetic bus insertion distribution. It is unvalidated, yes, but this is a benchmark paper; the synthetic regime is clearly specified and reproducible. The problem is only when they make an operational claim. That is more about framing than the method.\n\nWhat is missing is an empirical comparison with [6] (Zhou et al.), the closest prior constrained-TSP RL method. They cite it but never run it. Given the paper's own emphasis on benchmarking, this is a clear soft spot.\n\nBottom line: the paper deserves a serious referee. It is a solid empirical study with released code, honest limitations, and a defensible central claim in the simulated regime. I would send it to review, with the expectation that the abstract be fixed, the case-study selection justified, and a baseline against the most relevant prior method added.","headline":"A reproducible, well-executed empirical MORL-for-TSP paper that is let down mostly by an overbroad abstract and a few structural evaluation choices; the central runtime-tunability claim holds up in the simulated regime.","tokens_in":13753,"tokens_out":2694,"would_cite":true,"duration_ms":24818,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single trained traffic-signal controller can be tuned at runtime, through one preference dial, to emphasize bus priority or general traffic delay, without retraining.","keywords":["transit signal priority","preference-conditioned reinforcement learning","multi-objective reinforcement learning","runtime tuning","traffic signal control","tail-delay fairness","trade-off frontier","constrained phase control"],"falsifier":"Run the same 11-setting preference sweep on the same intersections but replace the synthetic bus timetables with recorded arrival and headway data from a real transit agency; the claim survives only if the single conditioned policy still traces a smooth bus-delay versus non-bus-delay frontier and beats the fixed-time and rule-based baselines at moderate priorities. A sharper version: if any realistic headway distribution collapses the frontier to a single point or makes non-bus tail delay explode at moderate settings, runtime tunability as described is not portable.","tokens_in":12858,"feed_emoji":"🚦","tokens_out":5770,"duration_ms":55141,"temperature":0.7,"pith_summary":"Transit signal priority needs to balance two goals that pull against each other: moving buses through an intersection quickly and not punishing everyone else. Most learning-based controllers hard-wire that balance into a fixed reward, so changing the priority later means retraining. This paper claims that one trained policy, parameterized by a single preference weight, can be tuned at runtime to sweep continuously between bus-priority emphasis and general-traffic delay, while still respecting real signal constraints like minimum and maximum green times. In simulation across thirteen intersections, that single policy traces a smooth trade-off frontier, often beating fixed-time and rule-based baselines at moderate settings, while exposing that aggressive bus-priority settings can push large delays onto a small tail of non-bus vehicles. If true, agencies could swap many trained controllers for one tunable controller whose dial they can turn at run time.","feed_headline":"One trained policy becomes a runtime dial for transit priority","feed_subtitle":"Preference-conditioned learning lets traffic operators shift the bus-vs-car trade-off without retraining.","key_machinery":"The load-bearing object is the preference-conditioned action-value function Q(s,a,w), from which the controller selects the phase that maximizes the inner product of the preference vector and the Q-values. A single scalar weight on bus priority blends a two-dimensional per-step reward vector: negative instantaneous bus delay on one entry, and negative all-vehicle delay plus a tail-delay penalty on the other. Around this is a feasibility layer that enforces minimum and maximum green times and transition rules, so every controller, learned or rule-based, is judged under the same operational constraints. To make sparse bus-priority events learnable, the training scenarios insert timetable-based","core_discovery":"The central claim is that a preference-conditioned controller, trained once per intersection, makes the bus-versus-all-vehicle delay trade-off a continuous runtime parameter rather than a training-time choice. Learning a single action-value function over state, action, and preference lets the same controller arbitrate conflicting priority requests differently as the dial moves. The paper reports that the conditioned policy spans a smooth empirical trade-off frontier, dominates both fixed-time and rule-based transit-priority overlays for most preferences at moderate weights, and stays within minimum and maximum green and transition constraints. At high bus-priority weights, the same diagnosti","pith_inferences":["Beyond the paper's stated results, the practical product is likely to be a bounded dial: exposing only moderate weights as the normal operating range and treating high bus-priority weights as an emergency mode with tail-delay guardrails.","The same conditioning idea should extend to richer preference inputs than one scalar, since the paper deliberately left bus occupancy and schedule deviation out of the observation; adding them would let the dial control headway adherence or passenger-weighted delay instead of raw bus delay.","Because the paper's strongest gains appear on low-complexity intersections and degrade with phase count and cycle length, agencies should expect runtime tunability to be worth less at complex intersections and should validate per site rather than citywide."],"forward_implications":["A single trained policy can replace a set of fixed-weight specialists while matching or exceeding them in frontier quality on the tested intersections, removing the need to train and store many controllers.","Operators can adjust the bus-versus-traffic emphasis at runtime for time-of-day shifts, incidents, or special events without retraining.","Moderate preference settings capture most bus-delay gains with limited non-bus and tail-delay costs, suggesting a safe operating range for normal use.","The approach shows qualitative robustness when bus prevalence drops, indicating that runtime tunability may transfer to less transit-saturated settings with scenario-dependent offsets.","The constrained benchmark and shared feasibility layer give future controllers a common ground for comparison under realistic signal constraints."],"fun_headline_variants":["Tune bus vs car delay on the fly with one RL policy","Runtime dial for transit priority: one trained controller","Single policy, continuous trade-off: transit signal priority","Preference-conditioned RL: tune transit priority at runtime","Flip the bus priority switch without retraining traffic signals"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire demonstration leans on the assumption that the synthetic timetable-based bus insertion (headways uniformly drawn from 300-900 seconds over four concurrent routes) produces a representative distribution of real transit-priority events; if actual bus operations differ in headway correlation, route structure, or demand, the learned policy's trade-off frontier and its dominance over the baselines may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Tune bus vs car delay on the fly with one RL policy","Runtime dial for transit priority: one trained controller","Single policy, continuous trade-off: transit signal priority","Preference-conditioned RL: tune transit priority at runtime","Flip the bus priority switch without retraining traffic signals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000628,"raw_usage":{"total_tokens":2760,"prompt_tokens":780,"completion_tokens":1980,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":1901}},"tokens_in":524,"tokens_out":1980,"duration_ms":14552,"temperature":1.0,"reasoning_tokens":1901,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T09:14:06.667865+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 11-setting preference sweep on the same intersections but replace the synthetic bus timetables with recorded arrival and headway data from a real transit agency; the claim survives only if the single conditioned policy still traces a smooth bus-delay versus non-bus-delay frontier and beats the fixed-time and rule-based baselines at moderate priorities. A sharper version: if any realistic headway distribution collapses the frontier to a single point or makes non-bus tail delay explode at moderate settings, runtime tunability as described is not portable.","supporting_citations":[],"review_version":1}