acceptodds
Under review as a conference paper at ICLR 2027

TAXIGPT: DECOUPLING MOBILITY PRIORS AND ORDER-DERIVED FEEDBACK FOR PASSENGER SEEKING

Abstract

Passenger seeking is central to taxi and autonomous ride-service operations as a vacant vehicle need to decide where to move before its next request is assigned. An effective policy must generate realistic routes and identify movements that lead to passengers. Learning both from a single reward signal is difficult be- cause explicitly modeling both movement and pickup/service dynamics is costly, and coarse approximations can admit unrealistic reward-seeking behavior. We propose TAXIGPT, a knowledge-decoupled framework that separates mobility learning, order-environment reconstruction, and seeking-policy adaptation. A decoder-only Transformer learns a mobility prior from serving trajectories through next-location prediction. An order-derived environment reconstructs passenger availability and joint service outcomes from historical orders, returning rewards and state transitions for variable-length route proposals. Group-relative policy optimization adapts the prior across successive seeking and serving cycles, while locality constraints and reference regularization help preserve movement structure. In Shenzhen simulations, TAXIGPT outperforms five passenger-seeking baselines, increasing cumulative episode profit by 160.7% and reducing average seeking distance by 63.3% relative to the respective best baselines. Applying locality during adaptation further improves profit by 5.6% over inference-only filtering under the same deployment constraint. These results support learning movement priors from real trajectories and seeking preferences from order-derived interaction, enabling effective passenger seeking with less reliance on explicitly modeled urban movement dynamics.

Primary area:
reinforcement learning
Keywords:
Trajectory Modeling, Reinforcement Learning, Sequential Decision Making, Urban Mobility, Autonomous Ride Services, Mobility Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HAoFOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TaxiGPT, a framework for learning passenger-seeking policies for ride-hailing. The key idea is to decouple the knowledge required for passenger seeking into different parts.

Critical strengths

The problem analysis is good, clearly pointing out the significant challenge in the field.

Critical weaknesses
  1. The problem setup is not clear enough. Providing a detailed MDP formulation would be beneficial.

  2. The task of this paper is similar to relocation in ride-hailing. However, the proposed method only considers a single vehicle, neglecting system-level interactions.

  3. The modeling in this paper lacks practicality. In fact, thousands of on-demand vehicles operate simultaneously in a city. An order near a driver does not mean the platform will dispatch it to that driver. It also depends on the spatiotemporal demand-supply relationship. Without a ride-hailing simulator evaluation or AB test, the current results lack convincing evidence.

  4. The compared methods lack recent baselines.

Questions for authors

My major concern is weakness 3, regarding the validation and correctness of the story. If the authors could provide a clear explanation, I would like to increase my score.

3 comments

Thank you for your precious time and review. We hope this could address your concern. We added both frozen-policy evaluation and adaptation in a finite-request, multi-vehicle environment, and clarify the individual-policy formulation below.

W1. Thank you for mentioning this. We will add a detailed MDP formulation in revised version. The process in Sections 3.1-3.3 is a leg-level semi-Markov decision process. Each seeking leg starts from sk=(ℓk,tk,wk)s_k=(\ell_k,t_k,w_k) (location, time and day type). The action is a variable-length route τk=(ℓk,1,…,ℓk,Lk,EOT)\tau_k=(\ell_{k,1},\ldots,\ell_{k,L_k},\mathrm{EOT}), executed until pickup or route termination. A pickup leads to the sampled service destination after seeking plus service time. Without pickup, the next leg starts at the route endpoint after a 15-minute wait. Equation (4) rewards service value and subtracts seeking cost. Equation (1) maximizes cumulative discounted reward.

In the competitive environment, the full state also includes the request pool and release schedule, plus other vehicles' locations and service status. TAXIGPT retains sks_k as its local observation, so the fleet-level state is partially observed.

W2. The learning objective and the presence of competing vehicles are separate questions. Our control decision is where a vacant vehicle should seek next, with the objective of improving that vehicle’s cumulative return. This individual-policy scope is shared with CAC, Guo and Chen (2024), and the focal RL taxi in O’Keeffe et al.’s (2022) hybrid fleet. A centralized dispatcher controls joint assignments across vehicles, which is a different decision scope. W3 explicitly adds competing supply and exclusive allocation while preserving our seeking-policy objective.

W3. Thank you for pointing this out. We added a multi-vehicle simulator in which a nearby request can be assigned to another vehicle. Historical completed trips serve as proxies for passenger requests. Each reconstructed request has a release time, pickup cell, destination and service duration, and remains available for up to 15 minutes. Vacant vehicles in its cell compete for assignment, and each request can be assigned only once. Once matched, the request is removed and the vehicle remains occupied until service completion, then resumes seeking at the destination. Pickup opportunities therefore depend on both the remaining requests and the competing vehicles' current locations and availability.

Following the hybrid-fleet setup of [1], background vehicles use either random-neighbor or training-demand-greedy policies. We test both at four fleet densities, giving eight settings with different levels and patterns of competition.

Evaluation without retraining. We freeze the original TAXIGPT checkpoint and test whether it still performs well when other vehicles can take nearby requests. The table compares learned policies across the eight settings:

MethodProfit (Yuan/episode)TAXIGPT minus method [95% CI]
TAXIGPT454.7--
Pretrained Transformer-PPO358.4+96.3 [53.9, 144.3]
CAC353.9+100.8 [42.3, 157.8]
SL-1.0 (adapted)333.5+121.2 [70.9, 176.4]

All methods use the same 50 test scenarios. We average the eight settings within each scenario before calculating the profit differences and their 95% confidence intervals. TAXIGPT outperforms all three controls under competition, with all three intervals above zero.

Learning with competition. We also adapt the same mobility prior in the competitive environment. GRPO increases profit from 395.1 to 474.8 Yuan/episode, a paired gain of +79.7 [38.0, 126.7] Yuan. This result complements the frozen-policy evaluation: the original policy performs well when requests are contested, and adaptation also improves the prior when competition is included during learning.

W4. We added Guo et al.'s (2024) SARSA(lambda) seeking method [2] in its published fixed-price SL-1.0 setting, matching our monetary protocol. We also added pretrained Transformer-PPO with the same architecture, trajectory representation and R=10 support. W3 reports the results, providing recent domain coverage and a shared-interface learning comparison.

[1] O'Keeffe et al. (2022). Using reinforcement learning to minimize taxi idle times. [2] Guo et al. (2024). Seeking in Ride-on-Demand Service: A Reinforcement Learning Model With Dynamic Price Prediction.

Thank you for your effort and response. Most of my concerns have been addressed, and I have increased my score to 2. However, I am still not convinced by the story. Intuitively, if too many vacant vehicles choose the same seek region, it may not be good for the overall system.

Thank you for taking the time to review and respond to our reply and increasing your score. We agree that excessive supply concentration can reduce system-wide efficiency. However, individual passenger seeking and fleet-level coordination are two well-established research problems with distinct optimization objectives, both of which continue to receive substantial research attention.

Individual passenger seeking optimizes a vacant vehicle's routing decisions to improve its own long-term return. This problem has been studied by O'Keeffe et al. (2022), Guo et al. (2024), and more recently Liu et al. (2025). Notably, O'Keeffe et al. explicitly evaluate a single RL taxi competing with other vehicles. In contrast, fleet-level repositioning addresses the coordinated spatial allocation of multiple vehicles to balance supply and demand, as studied by Beojone et al. (2024), Lyu et al. (2026), and Jusup et al. (2026).

TAXIGPT follows the established individual-seeking formulation, optimizing cumulative return across successive seek–serve cycles (Eq. 1). Our new competitive evaluation further accounts for interactions with other vehicles through finite, exclusively assigned requests. The original policy is evaluated across eight settings with different background policies and fleet densities, demonstrating its effectiveness under competing supply.

We respectfully hope that the contribution of TAXIGPT will not be discounted simply because it addresses individual passenger seeking rather than the distinct problem of fleet-wide coordination. Both are meaningful and actively studied research objectives. Requiring a single-vehicle seeking method to also optimize fleet-wide allocation would introduce an additional objective beyond its stated research problem.

We sincerely hope that the established literature, together with our additional competitive experiments, will help you reconsider your assessment of TAXIGPT within its intended research scope. Thank you!

Reviewer_nACDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses passenger seeking for taxis and autonomous ride services: a vacant vehicle must decide where to move before its next request. It proposes TaxiGPT, a "knowledge-decoupled" framework with three parts: a decoder-only Transformer pretrained by next-location prediction on serving trajectories (the mobility prior); an environment reconstructed from historical orders that returns pickup outcomes, service transitions and rewards for variable-length route proposals; and a locality-constrained, GRPO-style adaptation of the prior over successive seek–serve cycles, with reference and stopping regularization to limit drift from pretrained movement. In Shenzhen simulations, TaxiGPT is reported to outperform five passenger-seeking baselines on all operational metrics (+160.7% episode profit and −63.3% seeking distance vs. the respective best baselines), ablations attribute the gains to pretraining, order-derived adaptation and the full reward horizon, and applying locality during adaptation is reported to improve profit by 5.6% over inference-only filtering. Mobility-fidelity diagnostics (held-out NLL, transition JSD, step-length W1, large-jump rate) are reported alongside the operational results.

Critical strengths

1. A complete, explicitly specified pretrain-then-constrained-RL pipeline for a real operational decision problem, evaluated with both utility and movement-fidelity metrics. The method is described with enough detail (tokenization, environment semantics, normalized policy, surrogate, regularizers, full hyperparameter table) that it could be re-implemented, and the paper is candid about several limitations (no competing vehicles, modeled rather than observed fares, asymmetric diagnostic protocols). The dual evaluation — operational metrics plus held-out NLL / transition JSD / step-W1 — is more careful than most applied-RL papers in this area. This matters to my assessment because it makes the paper's problems diagnosable rather than hidden; it is the reason I do not recommend a clear reject.

2. The observation that a frozen next-location model trained on serving trajectories is itself a strong seeking policy in this environment. This is the most interesting empirical fact in the paper and suggests a cheap, label-free prior for this class of problems. Its consequence for my assessment is limited, however, because the paper does not analyze why (see Q2), and because the same fact also exposes how weak the baselines are.

Critical weaknesses

1. The central claim, that a learned mobility prior plus order-derived feedback reduces the movement dynamics that must be explicitly engineered into the simulator, is not supported by the method as designed, and the remaining technical contents are mostly applications of existing recipes. The environment advances a fixed amount of time per token regardless of displacement, so the locality mask bounds teleportation rather than enforcing physical plausibility, and the paper's own ablations show the simulator remains exploitable: the unconstrained variant is the most profitable precisely by jumping, and the randomly initialized policy reaches the highest pickup rate by moving implausibly fast. Movement plausibility in the full model is therefore supplied by a hand-set clock, a hand-set mask, two regularizers that are not ablated, and a prior whose input context already contains pickup-volume features — i.e., by engineered constraints of the kind the paper claims to avoid, which also degrades the "decoupling" framing. For technical novelty, most components: autoregressive pretraining on grid tokens, invalid-action masking, GRPO with a KL-to-reference penalty — is standard, and the order-derived environment is the historical-order pickup-probability simulator this literature has used for over a decade. Some prior work is not discussed: MobilityGPT (Haydari et al., 2024), which pretrains a GPT on tokenized mobility and fine-tunes it with RL, and MDP approaches that solve the seeking policy directly on the same order statistics. And UniTraj (NeurIPS 2025) also explores vehicle trajectory foundation models which aligns with the “mobility prior” claimed by the authors.

2. The headline comparison is confounded by the action interface, the natural reference points and statistical support are missing, and the mobility-preservation evidence is partly circular. The domain baselines move one grid cell per time step whereas TaxiGPT may move up to ten cells per step under the same clock, so much of the reported gap in seeking time, distance and profit reflects a difference in action spaces rather than in policies — which the paper itself concedes. The adaptation and tuning of CAC, cGAIL and local-action PPO are not described, and the reported seeking time and distance of one baseline are mutually inconsistent with its action set. The natural reference points are absent: a model-based or demand-greedy planner on the same order statistics, and real drivers' vacant trajectories replayed through the environment, so neither the simulator nor the absolute numbers can be calibrated. Results are five-seed means without variance, and the number of test episodes is not stated, so the small differences that carry the paper's secondary claims (+5.6%, +2.8%) cannot be interpreted. The fidelity evidence is partly circular: the Jump threshold coincides with the locality bound, the Table 3 diagnostics are computed under different supports across variants, and NLL is measured on serving trajectories that a seeking policy need not reproduce. Dataset scale is not reported, and evaluation covers a single city and month.

Questions for authors

Q1 time model vs. the "less simulator engineering" claim. The environment advances a fixed 20 s per token regardless of how far the token moves, and the locality mask still allows a single token to travel about 1.5 km (≈75 m/s). (a) Do the authors regard this fixed clock and the hand-set mask as part of the "explicitly modeled movement dynamics" the paper aims to reduce? If so, what has actually been removed from the simulator relative to prior grid-based online-RL approaches? (b) The ablations show that policies exploit this time model whenever they can (the no-locality variant is the most profitable by jumping; the scratch policy attains the highest pickup rate by moving implausibly fast). Is there anything in the environment that penalizes unrealistic movement, or does the full model's plausibility rest entirely on the pretrained prior and the regularizers? If the latter, how to reconcile this with the drift of the prior under reward optimization (ΔNLL ≈ +1.5 nats)?

Q2 does the decoupling hold?. The mobility prior's input context includes per-cell/time pickup volume, and pretraining uses serving rather than vacant trajectories without a stated reason. The frozen prior already reaches 435.80 Yuan. How do the authors distinguish whether this comes from learned movement structure, from the demand features in the input, or from serving trajectories naturally ending near demand hotspots? Does this affect the "serving trajectories supply movement, orders supply demand" formulation?

Q3 pickup model. Is a Bernoulli(p_pick) draw made at every executed token, including repeated visits to the same cell within one 15-min window, and is demand ever consumed? Since historical pickups also reflect where supply was, how do the authors assess the sensitivity of the main conclusions, in particular the 2.9-min average seeking time, to these choices and to the absence of competing vehicles?

Q4 interface confound. To what extent do the authors attribute the Table 1 gaps, especially Seek and Dist, to the policy rather than to the action interface? Would they report either the domain baselines with the same R=10 multi-cell action set, or TaxiGPT restricted to single-cell moves? Please also explain how RL-Idle-A2C's Seek (137.43 min) and Dist (8.331 km) are consistent with an 8-neighbour action set with no stay action, and provide some details on the adaptation and tuning for CAC, cGAIL and local-action PPO.

Q5 reference points and statistics. (a) What are the Profit/SR/Seek of real drivers' vacant trajectories from the test dates replayed through the order environment, and of a model-based or demand-greedy planner using the same λ statistics? These are necessary for interpreting 540.91 Yuan and 2.90 min. (b) Please provide seed standard deviations for Tables 1–3 and the number of test episodes, so that the +5.6% and +2.8% differences can be assessed against seed noise.

Q6 fidelity evidence. Given that Jump's threshold equals the locality bound and that the Table 3 diagnostics are computed under different supports, what evidence beyond marginal statistics (e.g. road connectivity of consecutive cells, or route visualizations) supports the claim that generated routes are realistic? If the Test-only-locality diagnostics were recomputed under its deployment mask, would the "locality during RL attains the lowest JSD" conclusion still hold?

5 comments

Thank you for recognizing the explicit methodology, complementary operational and mobility diagnostics, and strong frozen prior. We would like to respectfully clarify two distinctions central to TAXIGPT: learning movement preferences does not require removing every simulator assumption, and knowledge decoupling concerns learning objectives rather than statistical independence of inputs. We also report additional controls for the empirical questions.

Q1. The review conflates movement preferences with timing and action support. The Introduction (lines 108–112) and Secs. 3.2–3.4 explicitly distinguish them: serving-trajectory pretraining learns next-location and stopping preferences; the order environment provides pickup/service outcomes; the nominal clock and locality rule define time and allowed moves. Equation (6) preserves relative probabilities among allowed actions. Thus, the mask alone cannot supply the learned mobility policy. Indeed, under the same operational locality, scratch RL earns 96.03 Yuan/episode, versus 540.91 with pretraining (Table 2).

The higher simulated profit without locality illustrates the risk of exploiting a permissive movement interface, which motivates applying locality during adaptation. The 20-second clock does not enforce physical speed, and we do not claim road-level travel-time calibration. NLL rising from 1.523 to 3.032 measures drift on the serving-prediction task as the model learns to seek passengers; it is not a measure of physical movement. The periodically refreshed reference and fixed stopping reference moderate this shift, while Q6 evaluates generated movement under matched support.

Q2. The stated absence of a rationale for using serving trajectories overlooks the Introduction (lines 99–100) and Appendix A. Serving trips record realized movement without requiring demonstrations labeled by successful seeking. The decomposition assigns distinct objectives: pretraining predicts locations/EOT, while order interaction evaluates the pickup, service, and subsequent returns of newly proposed searches. Appendix A explicitly permits shared training-date records and context; therefore, including pickup-volume covariates does not invalidate this learning-role decomposition.

The frozen prior's 435.80 Yuan/episode reflects the complete contextualized representation. The core transfer question is tested directly: reward-driven adaptation raises profit to 540.91 Yuan (+24.1%), increases SR from 67.38% to 75.84%, and reduces Seek from 5.04 to 2.90 min (Table 2).

Q3. Yes. The original Bernoulli model queries availability at each executed token, including revisits, without depleting individual orders. Its 15-minute window represents availability at arrival, not new arrivals every 20 seconds (Sec. 3.3). We additionally evaluated finite, completed-trip-derived requests with release/expiry times, one-time exclusive assignment, and competing vehicles that enter service after matching and later rejoin vacant supply. Two background policies at four fleet densities produce eight competitive settings.

The original TAXIGPT checkpoint is frozen in this evaluation:

MethodCompetitive Profit (Yuan/episode)TAXIGPT advantage [95% CI]
TAXIGPT454.7—
Pretrained Transformer-PPO358.4+96.3 [53.9, 144.3]
CAC353.9+100.8 [42.3, 157.8]
SL-1.0 (adapted)333.5+121.2 [70.9, 176.4]

These intervals use 50 matched test scenarios, averaging the eight settings within each scenario before paired bootstrap. TAXIGPT maintains a positive profit advantage over all three controls when requests cannot be reused and competitors can take them first. Moreover, adapting the same mobility prior within the competitive environment increases profit from 395.1 to 474.8 Yuan, a paired gain of +79.7 [38.0, 126.7]. These experiments test sensitivity to request depletion and competition.

Q4. Different action interfaces make Table 1 an end-to-end method comparison, as Sec. 4.2 already states. They do not establish that most of its gap is caused by the interface. We added a pretrained Transformer-PPO control sharing the Transformer representation, pretrained initialization, route/stopping interface, and R=10 locality support. TAXIGPT exceeds it by +83.9 [22.1, 148.1] Yuan in finite single-vehicle evaluation and +96.3 [53.9, 144.3] Yuan under competition. These are same-interface policy comparisons: the advantage in these settings cannot be explained by giving only TAXIGPT multi-cell actions.

RL-Idle-A2C's apparently inconsistent 137.43-min Seek / 8.331-km Dist has an execution-level explanation: instrumented replay found repeated outward actions clamped at a grid corner. Time advances while displacement stays zero, despite no explicit stay action.

Q5. We added train-demand greedy references. Under competition, TAXIGPT exceeds neighboring-cell greedy by +135.0 [86.1, 183.5] Yuan. With the same R=10 support, greedy earns 461.0 versus TAXIGPT's 454.7 Yuan/episode (paired difference −6.2 [−51.5, 47.6]). In the finite single-vehicle movement audit, 39% of R=10 greedy's leg-start moves span at least eight cells, versus 7% for TAXIGPT. This illustrates why simulated profit should be interpreted alongside movement behavior.

We also replayed observed vacant-driver routes for the first seeking leg. Under competition, first-pickup incidence is 0.80, or 0.53 after removing each recorded route's own matched request. These observation-derived, success-conditioned horizons provide a behavioral reference. They cannot directly provide unbiased full-episode Profit/SR/Seek: after a simulated service, the vehicle's location and time may diverge from the recorded continuation.

Five-seed variability. Combined with the original paper's reported means, the full TAXIGPT results are:

Operational metricMean ± 5-seed SD
Profit (Yuan/episode)540.91 ± 23.57
Mean Profit (Yuan/leg)21.93 ± 0.63
EE (Yuan/h)117.53 ± 8.46
SR (%)75.84 ± 2.82
Seek (min/leg)2.90 ± 0.42
Dist. (km/leg)1.272 ± 0.016

Q6. The JSD advantage over Test-only locality persists when the generation supports are matched. Under the identical R=10 deployment mask and matched start contexts, JSD is 0.0169 for TAXIGPT versus 0.0262 for Test-only locality, a Test-only-minus-Full difference of +0.0093 [0.0028, 0.0180] (95% context-cluster CI). Step-W1 does not distinguish them. Crucially, Table 3's operational comparison already used identical test-time locality.

Jump is a locality-bound compliance metric, so its reduction under a mask is expected. NLL probes serving-prediction retention, whereas JSD and Step-W1 measure displacement statistics. Appendix C already states that these marginal metrics do not establish road connectivity or complete physical realism. The matched-mask JSD result therefore supports the specific claim that learning with locality improves displacement-distribution fidelity over inference-only filtering, independently of Jump's threshold.

Thank you for your time and review. We hope this could clairfy our work and address your concerns.

Thank you for the detailed rebuttal and the additional experiments. The finite-request, competitive evaluation largely answers my concern about the pickup model (Q3), the matched-support JSD result resolves the diagnostic asymmetry (Q6), and the same-interface PPO control and the explanation of the A2C numbers are helpful (Q4). I also accept that the paper states its rationale for using serving trajectories, my wording there was imprecise.

However, several concerns remain unresolved, and some of the new results raise questions of their own. I summarize them in four points.

1. A simple demand-greedy rule matches TaxiGPT under the same action support (new, from the rebuttal).

  • With R=10 support, greedy earns 461.0 vs. 454.7 Yuan/episode (paired difference −6.2 [−51.5, 47.6]), so a rule with no learning is indistinguishable from the full method on the main metric. You argue that TaxiGPT moves more plausibly, but the environment assigns no cost to implausible moves, so the evaluation cannot credit it.
  • Please re-score the existing checkpoints (TaxiGPT, the frozen prior, R=10 greedy) with travel time proportional to displacement rather than a fixed 20 s per token. This needs no retraining, and a clear win for TaxiGPT would be direct evidence for the paper's thesis.

2. The claim of realistic movement still lacks direct evidence (unresolved, Q1/Q6). The rebuttal confirms that the clock does not enforce speed and that Jump is satisfied by construction, leaving only marginal statistics that Appendix C says do not establish road connectivity. Your own audit shows 7% of TaxiGPT's leg-start moves spanning at least eight cells, i.e. roughly 0.8 km in one 20-s step, and Q1(a), on what has actually been removed relative to earlier grid simulators, was not answered. Please either provide direct evidence that routes are drivable (e.g. per-step speed distribution, road connectivity) or narrow the claims in the abstract and conclusion.

3. The headline comparison is still not supported as stated (partly addressed; the new results sharpen it).

  • Action support: neighbouring-cell greedy is 135.0 Yuan below TaxiGPT while R=10 greedy is on par, so the action range alone explains a large share of profit; and TaxiGPT's margin over CAC in the new evaluation (454.7 vs. 353.9) is far smaller than the one behind the abstract's +160.7%. The original baselines were not re-run with equal support, and a policy clamped at a grid corner is not a meaningful baseline.
  • Statistics: seed variability is given only for TaxiGPT, and the differences behind the +5.6% and +2.8% claims (28.7 and 14.9 Yuan) are comparable to or smaller than the reported SD of 23.57.
  • Please provide a consistent table in a single environment with equal action support for all learned methods, including the frozen prior, greedy and the original baselines, with seed SDs, and a revised abstract. Please also clarify why the Transformer-PPO control (358.4) appears to fall below the unadapted prior (395.1), and what "SL-1.0" is.

4. The contribution and its positioning remain unclear (unresolved, Q2 and novelty). If "decoupling" only means that pretraining and adaptation use different objectives, the claim is much weaker than the abstract suggests, and the source of the frozen prior's 435.80 Yuan, movement structure or the pickup-volume covariates, is still untested.

I appreciate the additional work, but the new results confirm rather than remove my main concerns, and the overall story and contributions are still not convincing. I don't think it can be addressed in this discussion period, so I am keeping my rating unchanged for now.

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

This paper proposes TAXI GPT, which learns a mobility prior through a pretrained Transformer and then optimizes the taxi passenger-seeking policy using GRPO in a simulated environment constructed from historical trip orders.

Critical strengths

The idea of separately modeling the mobility prior and order-derived feedback is reasonable. The method achieves noticeable performance improvements over the evaluated baselines in simulation experiments on the Shenzhen dataset.

Critical weaknesses

Weaknesses

W1. Concerns about simulator fidelity and realism.

I am concerned that the proposed simulator may not be sufficiently realistic. Have the authors compared it against a higher-fidelity traffic simulator or real-world data on a small subset to validate its fidelity? Currently, the authors mainly validate the statistical properties of the generated trajectories, but this does not necessarily demonstrate that the simulated passenger pickup probabilities, waiting times, and policy returns are accurate.

I believe the fidelity and realism of the simulator need further validation. At a minimum, a small-scale comparison experiment would help demonstrate that the feedback provided by the simulator is reasonable. Since all RL training and evaluation in this paper rely on this simulator, inaccuracies in the simulator could undermine the credibility of the reported performance improvements.

W2. Insufficient and potentially unfair baseline comparisons.

I am not fully convinced that the baseline comparisons are sufficiently comprehensive. Although the authors compare against several passenger-seeking methods, some relevant works on vacant taxi routing and passenger seeking appear to be missing.

Moreover, the evaluated baselines seem to differ in their action spaces and route termination mechanisms. This makes it difficult to determine whether the observed performance improvements come from the proposed method or simply from a more flexible action space.

Since the techniques used in this paper are not particularly novel individually, I believe stronger and fairer baseline comparisons are especially important to establish the effectiveness of the proposed approach.

W3. Potential reward hacking in the order simulator.

Have the authors implemented any basic safeguards against reward hacking in the simulator? For example, in the current order simulator, each token corresponds to 20 seconds. If a vehicle repeatedly visits the same grid cell and each visit independently samples from that cell's passenger pickup probability, then attempting pickup five times in a high-demand cell could result in a very high cumulative probability of picking up at least one passenger.

However, these five attempts may occur within a relatively short period and could effectively correspond to the same batch of historical passenger demand. This could allow the model to learn a strategy that performs well in the simulator but would not be valid in the real world.

Have the authors introduced any safeguards to prevent the same batch of orders from being sampled repeatedly? Alternatively, have they analyzed whether the learned policy exhibits behaviors such as repeatedly visiting high-demand grid cells to obtain rewards?

W4. Insufficient ablation studies to isolate the contributions of individual components.

I also find the ablation studies somewhat insufficient. Although the authors provide ablations on pretraining, RL, and locality, the current experiments do not clearly establish which components are responsible for the performance improvements.

For example, it would be useful to include a controlled comparison using exactly the same pretrained Transformer, action space, and locality constraints, while replacing GRPO with PPO or another RL optimization method.

Without such a comparison, it is difficult to determine how much of the improvement can be attributed specifically to GRPO, or whether the combination of these existing techniques is necessary.

Questions for authors

N/A

1 comment

Thank you for recognizing the rationale of separating mobility learning from order-derived feedback. We appreciate your precious time and hope this could address your concern. We distinguish the existing component evidence from the new tests of competition and shared-interface learning.

W1. We added a finite-request, multi-vehicle evaluation to test whether the performance advantage persists when requests are limited and vehicles compete for them. Historical completed trips supply request proxies with release times, pickup cells, 15-minute availability windows, destinations and service durations. Eligible co-located vehicles compete for exclusive assignment. Each request can be served once, and occupied vehicles resume seeking only after service. Two background policies (random-neighbor and training-demand-greedy) at four fleet densities give eight settings, following the hybrid-fleet approach of O'Keeffe et al. [1].

The original TAXIGPT checkpoint is frozen throughout. Results average the eight competitive settings:

MethodProfit (Yuan/episode)TAXIGPT minus method [95% CI]
TAXIGPT454.7--
Pretrained Transformer-PPO358.4+96.3 [53.9, 144.3]
CAC353.9+100.8 [42.3, 157.8]
SL-1.0 (adapted)333.5+121.2 [70.9, 176.4]

All methods use the same 50 test scenarios. We average the eight settings within each scenario before calculating the profit differences and their 95% confidence intervals.

TAXIGPT earns more than all three controls even when other vehicles can take available requests. Absolute pickup times and profits have not been validated against real-world observations.

W2. Section 4.2 explicitly defines Table 1 as a comparison of complete seeking methods. The new pretrained Transformer-PPO comparison addresses the additional attribution question by sharing TAXIGPT's architecture, trajectory representation and R=10 support. Its positive paired differences above establish an advantage within the same action interface. We therefore provide both end-to-end and shared-interface evidence, without treating Table 1 as an optimizer-only comparison.

For recent domain coverage, we added Guo et al.'s (2024) SARSA(lambda) seeking policy [2] in its published SL-1.0 setting. Fixing the price multiplier to 1.0 aligns this variant with our fixed-price evaluation. The table reports its competitive result.

W3. The original environment independently queries Bernoulli availability at executed tokens, including revisits. In the new experiment, matched or expired requests cannot be obtained again. No request was assigned twice in 4,000 audited evaluation episodes. Without retraining, TAXIGPT earns 494.7 Yuan/episode with finite single-vehicle demand and 454.7 under competition. Its paired advantage over Transformer-PPO is +83.9 [22.1, 148.1] Yuan in the former and +96.3 [53.9, 144.3] in the latter. Thus its advantage over this control persists after the repeated-allocation mechanism is removed.

W4. Tables 2-3 already separate the principal learning roles through four within-framework comparisons. Full TAXIGPT earns 540.91 Yuan/episode:

Role testedControlProfit
Mobility pretrainingScratch RL96.03
Order-derived adaptationFrozen prior435.80
Remaining-episode feedbackImmediate-leg return, H=1526.03
Adaptation-time localityTest-only locality512.18

These are separate controls, with identical deployment locality for the last comparison. They support the contributions of pretraining, task feedback, downstream seek-serve consequences and learning within deployment constraints. Replacing GRPO with PPO is a further optimizer-level test, which we now provide above. Our contribution is the division of learning roles and their realization in a constrained generative seeking policy. The component-attribution question therefore already has direct evidence in the submission, now complemented by the PPO comparison.

[1] O'Keeffe et al. (2022). Using reinforcement learning to minimize taxi idle times. [2] Guo et al. (2024). Seeking in Ride-on-Demand Service: A Reinforcement Learning Model With Dynamic Price Prediction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.