acceptodds
Under review as a conference paper at ICLR 2027

RideSkill: A Hierarchical Algorithm for Generalized Ride Sharing with LLM-Driven Automatic Evolution

Abstract

Ride-sharing, which allows multiple passengers with different origin-destination (OD) pairs to share a single vehicle, is a challenging operational problem, as it requires orders with different OD pairs to be efficiently bundled and assigned to vehicles under uncertain and varying scenarios. Although multi-agent reinforcement learning (MARL) solutions have achieved promising performance, they suffer from limited generalization (adapting to different environmental scenarios), low transferability (adapting to different platform objectives), and training difficulties in large-scale systems, such as the curse of dimensionality. Recently, motivated by the scaling of large language models (LLMs), several works have incorporated LLMs into ride-hailing systems, either by employing LLMs directly as decision-making agents or using them for automatic algorithm design. However, none of these approaches support vehicle sharing, which complicates the problem by expanding both the state and action spaces exponentially. Moreover, most of them require frequent LLM calls at inference time, making them infeasible for real-time deployment. To address these issues, we propose RideSkill, a hierarchical method for ride-sharing that leverages LLM-assisted automatic algorithmic design. RideSkill consists of a combiner that assigns appropriate skills to each vehicle from a learned skill repository, enabling adaptive dispatch under varying scenarios and objectives, and a repositioner that sequentially relocates idle vehicles to emerging regions, avoiding conflicts among vehicles. Crucially, the skill repository, combiner, and repositioner are all trained by an LLM-based automatic evolutionary method, eliminating the need for LLM calls during deployment and thus ensuring high real-time performance. Evaluated on a real-world ride-hailing dataset, RideSkill achieves the best performance compared to model-based, MARL-based, and LLM-based benchmarks. Most crucially, our method demonstrates strong generalization and flexibility across different fleet sizes, vehicle speeds, passenger capacities, order densities, and objectives, suggesting high practical application value. Our code is provided at https://anonymous.4open.science/r/RideSkill-6AAD .

Primary area:
applications to robotics, autonomy, planning
Keywords:
Ride-Sharing, Order Dispatch, LLM Automatic Algorithm Design

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 4 reviews, above 27% of the others.
Reviewer_ZCuEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RideSkill, a hierarchical framework for generalized ride-sharing with offline LLM-assisted algorithm evolution. Instead of using LLMs directly for online decision making, the framework uses them during training to evolve reusable dispatching and repositioning strategies, which are then deployed without further LLM calls. The method is designed to adapt to changes in operating conditions, such as fleet size, vehicle speed, passenger capacity, and demand patterns, as well as different platform objectives. Experiments on RideGym with NYC taxi data compare RideSkill against classical matching methods, MARL baselines, and prior LLM-based approaches. The reported results show consistently strong performance across these settings, with additional ablations examining objective adaptation, repositioning, and fairness.

Critical strengths
  1. The decomposition into reusable scoring skills, an objective-conditioned combiner, and a sequential repositioner is intuitive and practically motivated. In particular, using the LLM only during offline algorithm search avoids the inference latency and cost associated with LLM-based online dispatch, which is important for large-scale real-time mobility systems.

  2. The paper evaluates the method across multiple operational axes, including fleet size, vehicle speed, capacity, time periods, and objectives, and includes useful ablations on the combiner, repositioner, fairness mechanism, objective conditioning, and the choice of the underlying LLM. In particular, the comparison between an objective aware policy and an otherwise identical objective blind policy is a useful diagnostic for verifying that the combiner actually responds to the supplied objective.

Critical weaknesses
  1. The methodological novelty is somewhat limited, as the proposed framework mainly combines several intuitive and established components rather than introducing a clearly new learning or optimization principle. The proposed framework combines several intuitive and well-established components, including skill-based scoring, bipartite matching, sequential repositioning, evolutionary search, and LLM-based program generation. While these components are integrated coherently for the ride-sharing setting, the overall method feels more like a combination of existing techniques than a substantially new learning framework. In particular, the hierarchical decomposition, objective-conditioned skill selection, and LLM-guided evolution are all reasonable design choices, but individually they are fairly intuitive, and their combination does not clearly introduce a fundamentally new algorithmic principle. As a result, I find the methodological contribution less compelling from a novelty perspective, despite the practical effectiveness of the system.

  2. The empirical setup does not convincingly support the paper's central claim of generalized adaptation across scenarios and objectives. RideSkill is explicitly optimized on a diverse distribution of scenarios and objectives, whereas the MARL baselines are trained only on a single anchor condition (fleet size 1,000, capacity 4, speed 35 km/h, and one anchor objective) and are then evaluated under distribution shifts. Consequently, the reported OOD advantage conflates the proposed architecture with substantially broader training exposure. A much more informative comparison would be against goal-conditioned, multi-task, domain-randomized, or otherwise generalized RL baselines trained with the same scenario/objective distribution. Without such comparisons, the experiments do not establish that the proposed method itself is responsible for the claimed generalization advantage.

Questions for authors
  1. Can the authors provide a comparison in which generalized learning baselines are trained under the same distribution of scenarios and objectives as RideSkill?

  2. What is the precise class of objectives for which the proposed objective probing mechanism is expected to work?

2 comments
AuthorsOpenReview ↗

Response Part 1

Thank you for the review. Substantive manuscript changes are marked in blue.

W1. Methodological contribution

Thank you for your insightful comment. RideSkill contributes an application-inspired discovery-to-composition mechanism for ride-sharing. General LLM-guided evolution supplies the program-search operators; RideSkill uses them to discover reusable vehicle–order scoring skills, learns a combiner that selects and blends these skills for each vehicle according to its state, the operating scenario, and the platform objective, and then repositions idle vehicles. This makes the discovered behaviors reusable across changing dispatch conditions and objectives.

The skill–combiner design follows the structure of ride-sharing. An idle vehicle can prioritize pickup efficiency, a loaded vehicle must protect existing passengers' routes, and a vehicle with little time slack needs a different assignment preference. Platform priorities also change across operating periods. A library of basic skills captures these distinct behaviors, while conditional composition determines which behaviors each vehicle should use in the current setting. General program-evolution methods can generate dispatch policies; the additional mechanism here is the discovery and conditional composition of reusable dispatch skills, rather than evolving a single scoring program for a fixed task. This is the methodological distinction from the general LLM-search methods and the adapted ride-hailing comparator discussed in Appendix A.

Sections 3.2–3.3 specify the skill interface, conditional composition, and sequential repositioning. Appendix E, “Analysis,” reports the controlled module comparisons. The contribution is this transportation-specific representation and adaptive reuse of skills within a complete dispatch algorithm.

W2/Q1. Generalized RL comparison with matched task support

Thank you for your insightful comment. We added REDA-MT, a scenario- and objective-conditioned multi-task REDA policy, on the same finite task support as the added Phase-2 search. On the Table 1 grid under the anchor objective, its mean Reward is 58.6–68.6% lower than anchor-trained REDA across the four scenario groups in Table 9 (Appendix D.10), and RideSkill has higher mean Reward than both REDA policies. This directly evaluates a shared RL policy trained across the varied scenarios and objectives.

REDA-MT shares one network, target network, optimizer, and replay buffer across five objective families, with 136–1,989 vehicles, capacities 1–10, and speeds 25.40–44.96 km/h. It receives reward-probe responses and scenario/preference fields; the supplied callable generates both probe responses and environment rewards. REDA-MT was trained for 126 episodes and anchor REDA for 500; we evaluate the multi-task policy's fixed final checkpoint, whose validation return was non-monotonic during training. RideSkill includes repositioning, while both REDA policies are dispatch-only.

Below we show Hour and Fleet size; Table 9 in Appendix D.10 reports all four scenario groups. Each group first averages three environment seeds per configuration, then reports the mean and population standard deviation across configuration means. Service rate, completion rate, and utilization are fractions; times are minutes.

Scenario groupMethodRewardService rateCompletion rateWait timeRide timeDetour timeUtilization
HourREDA6743±9980.92±0.060.79±0.081.44±0.136.60±0.681.15±0.640.78±0.14
HourREDA-MT2441±1420.75±0.130.53±0.109.12±0.466.34±0.161.23±0.220.91±0.07
HourRideSkill7561±15620.94±0.050.84±0.041.65±0.125.38±0.240.23±0.030.90±0.10
Fleet sizeREDA5681±32450.64±0.250.53±0.231.73±0.298.06±0.912.76±1.020.92±0.06
Fleet sizeREDA-MT1937±10300.43±0.220.31±0.169.87±0.506.35±0.131.33±0.110.96±0.02
Fleet sizeRideSkill7935±29250.69±0.240.63±0.211.54±0.094.24±0.920.22±0.030.95±0.02

In this comparison, broadening REDA’s task support did not recover RideSkill’s performance. Changes in fleet size, capacity, and speed alter feasible assignments and transition dynamics, while changes in objectives alter reward scales. Including context in the state identifies the setting but still requires the network to learn effective behavior throughout that range. REDA-MT's non-monotonic validation and lower anchor-objective return are consistent with this optimization difficulty. This is a comparison on shared task support under the training budgets and deployment components described above. Table 9 in Appendix D.10, “Multi-task reinforcement-learning comparison,” reports the full seven-metric results.

AuthorsOpenReview ↗

Response Part 2

Q2. Objective class and probing

Thank you for your insightful comment. We clarified the supported objective interface and the identifiability condition for objective probing in Appendix F. The interface accepts deterministic functions of the declared per-vehicle step-event fields, fixed within an episode, including nonlinear terms and within-event interactions. Natural-language preferences are translated into this callable before rollout, and the same function supplies event rewards and probe responses. Episode return sums these event rewards; episode-level ratios, accumulated thresholds, and history-dependent rates require sufficient event/state information or an explicit approximation.

The combiner uses reward responses to a finite set of synthetic events to select and blend skills. For a linear feature family,

wθ(e)=θ⊤f(e),w_{\theta}(e)=\theta^{\top}f(e),

fix the context, query events, and linear preprocessing. The observed probe responses then satisfy

y=Bθ,y=B\theta,

where B is the probe-feature matrix after preprocessing. Let the rows of D contain cumulative feature differences between relevant feasible trajectories. Probe-equivalent objectives preserve these return contrasts for every coefficient vector exactly when

ker⁡(B)⊆ker⁡(D).\operatorname{ker}(B)\subseteq\operatorname{ker}(D).

Full column rank of B identifies all coefficients in exact arithmetic. The condition states precisely when the probes identify the return contrasts needed to distinguish those trajectories.

The released program's diagnostics identify confounded assignment-bonus and some party-price directions, together with mixed response channels. Two callables with the same probe transcript in the same context induce the same routing, even if they differ on unqueried events. The router's response to distinguishable objectives then depends on its learned composition logic and the behaviors available in the frozen skill library.

Appendix F.1, “Objective interface and finite-probe identifiability,” gives these conditions and diagnostics. Appendix D, “Objective Response,” reports observed responses to pickup-time and detour-time rewards and natural-language long-trip and detour objectives.

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

This paper proposes RideSkill, a hierarchical ride-sharing dispatcher trained offline by LLM-assisted evolutionary search. It evolves vehicle-order scoring skills, an objective-conditioned combiner, and a sequential idle-vehicle repositioner. At deployment, it blends selected skills, solves a matching program, and repositions unmatched idle vehicles without LLM calls. Experiments in RideGym use Manhattan data and compare rule-based, MARL, and LLM-based baselines across several scenario axes.

Critical strengths
  1. The task is important and challenging: ride-sharing dispatch must jointly consider vehicle capacity, onboard passengers, current orders, future demand, and empty-vehicle allocation.
  2. The central idea is appealing. The method integrates order dispatch and idle-vehicle repositioning while using LLM assistance offline rather than making LLM calls at deployment.
  3. The empirical results are consistently strong across several scenario axes and rule-based, MARL, and LLM-based baselines. Objective-response and fairness analyses also go beyond a single anchor-task result.
Critical weaknesses
  1. Eq. (2) does not encode its claimed one-new-order rule. Constraint (2c), sum_j u_i,j >= 1 (including the no-op), supplies only a lower bound. When u_i,empty = 0, (2d) is automatically satisfied, and no constraint stops vehicle i from receiving multiple real orders; capacity (2e) is not a cardinality bound. Further, (2d) contains a product of decision variables, so the displayed problem is not an ILP. The released matching.py instead enforces one-to-one feasibility via score-sorted greedy matching, not the displayed global ILP. The authors should clarify which decoder produced the results and align Eq. (2), Algorithm 4, and code; if an ILP was used, release it and compare it with the greedy decoder.
  2. The appendix gives useful objective-response evidence over four tested objectives, and Table 1 tests several scenario parameters. However, the generalization scope remains narrow: experiments use one Manhattan data source, temporally adjacent train/validation/test dates, and only a limited set of objective/test windows. These results support adaptation in the evaluated setting, but do not establish broad cross-city or cross-data-distribution generalization. Related-work and baseline coverage should also discuss Triple-BERT (ICLR 2026), a directly relevant large-scale ride-sharing dispatch method. If protocols are compatible, a comparison would be valuable; otherwise, incompatibility and claims relative to recent RL dispatch methods should be clarified.
  3. Avoiding runtime LLM calls makes stronger real-time performance plausible, but the claim is not directly measured. Online decision latency and scaling with fleet and pending-order count should be reported. Current ablations isolate repositioning and the single-skill comparison, but not task generation, objective-probing/combiner training, or audit/dedup components. More granular ablations are needed to support these training-design contributions.
Questions for authors
  1. Why does Eq. (2) omit a per-vehicle upper bound on new orders? With u_i,empty = 0, (2c)-(2d) permit multiple real orders. Which decoder generated the results: the released score-sorted greedy matcher or the stated ILP? Please align the paper, algorithm, and code.
  2. Can the authors more precisely qualify the generalization claim and discuss Triple-BERT, a recent directly relevant ride-sharing dispatch method? If protocols are compatible, please compare; otherwise explain incompatibility and qualify claims relative to recent RL dispatch methods.
  3. Can the authors evaluate on more geographically or temporally distinct demand settings, or otherwise qualify the scope of the generalization claim?
  4. Please report online latency/scaling, and ablate task generation, objective-probing/combiner training, and audit/dedup components where feasible.
3 comments
AuthorsOpenReview ↗

Response Part 1

Thank you for the review. Substantive manuscript changes are marked in blue.

W1/Q1. Matching formulation and decoder

Thank you for your insightful comment. We corrected Equation (2) to require exactly one eligible real edge or a private no-op per vehicle, with per-order exclusivity and capacity masks. This supplies the missing upper bound and removes the nonlinear product. The equation is a linear maximum-weight reference; the re-evaluated main results use the released score-sorted greedy decoder.

The decoder applies row softmax, retains real edges scoring above the vehicle's no-op, sorts them by fairness-weighted priority, and accepts an edge only if both endpoints are free. We used this decoder for the re-evaluated full RideSkill, no-repositioning, and independent-skill rows of Table 1, and for the new Queens and training-control evaluations. The archived objective-response, fairness, and GLM-backend values have unresolved decoder provenance. Equation (2) gives the linear matching reference, while Algorithm 4 specifies the greedy implementation.

The corrected formulation and executed rule are in Section 3.2, Equations (2) and (4)–(6), and Appendix C, Algorithm 4.

W2/Q2. Generalization scope and recent dispatch baselines

Thank you for your insightful comment. We added comparisons with RideGym adaptations of Triple-BERT (ICLR 2026) and OSPO (NeurIPS 2026), and a Manhattan-to-Western-Queens transfer test using a different month (Q3). RideSkill has the highest mean Reward across the four compatible Table 1 scenario groups and transfers effectively in the Queens test. The table below shows Hour and Fleet size; Table 1 in Section 4.3 gives the complete comparison. The Queens test extends the geographical and temporal evaluation within NYC.

The comparison shares demand windows, environment seeds, reward, dynamics, and seven metrics, while retaining each method's decoder. Triple-BERT is a centralized single-agent method that optimizes the global matching plan; OSPO updates a shared dispatch policy using one-step rewards and fleet-relative advantages. Section 4.1 and Appendix D.2 describe the adaptations.

Entries are mean±population standard deviation across configuration means, after averaging environment seeds at each configuration. Rates and utilization are fractions; times are minutes. Bold marks the best displayed mean per metric within each scenario, including ties.

Scenario groupMethodRewardService rateCompletion rateWait time (min)Ride time (min)Detour time (min)Utilization
HourOSPO6670±9970.85±0.130.74±0.122.29±0.165.77±0.240.16±0.000.87±0.13
HourTriple-BERT7107±15150.93±0.030.80±0.032.22±0.165.51±0.060.50±0.170.78±0.15
HourRideSkill7561±15620.94±0.050.84±0.041.65±0.125.38±0.240.23±0.030.90±0.10
Fleet sizeOSPO5924±31030.49±0.270.42±0.242.12±0.116.80±1.020.16±0.000.95±0.02
Fleet sizeTriple-BERT6988±30440.68±0.240.59±0.223.00±0.685.23±0.341.06±0.410.93±0.06
Fleet sizeRideSkill7935±29250.69±0.240.63±0.211.54±0.094.24±0.920.22±0.030.95±0.02

RideSkill has the highest mean Reward in the complete Table 1 comparison, while OSPO and Triple-BERT lead some individual operational metrics.

W3/Q4. Online latency and training-design evidence

Thank you for your insightful comment. We measured online computation across vehicle and order counts to test RideSkill's real-time operation. Mean decision time remains below one second throughout the tested range, well within the 1-minute dispatch interval. We also added Phase-2 training controls and a semantic-repair comparison: the proposed training design improves default-objective Reward over both controls, and correcting a generated semantic error improves external performance. The objective-access comparison below shows how supplying an objective changes dispatch behavior.

The complete decision takes 69.55–645.34 ms in cell means across 250–2,000 vehicles and 100–2,000 pending orders. This includes objective probing, skill composition, matching, and repositioning. The heatmap in Appendix D.8, “Online decision time,” reports the scaling of the implemented dispatcher.

For the new controlled experiments, the original Claude and GLM endpoints were no longer available. Within the available time and budget, we used DeepSeek-V4.1-Flash with the dispatch interface and evaluation protocol fixed. Appendix E, “Analysis,” presents training controls and semantic validation to examine the role of the corresponding modules. Their entries are mean±sample standard deviation across nine common scenarios (100–2,000 vehicles, capacity 2–10, speed 25–45 km/h).

AuthorsOpenReview ↗

Response Part 2

We compare two controls with the proposed design:

  • w/o normalization substitutes raw-return fitness and feedback in raw units for baseline-relative, scale-normalized/clipped fitness and its corresponding normalized feedback. This tests normalization in selection and feedback jointly.
  • w/o configuration diversity fixes training at 800 vehicles, capacity 6, and 35 km/h while retaining objective diversity. This tests the role of varying the operating configuration during training.

The skill library, initial candidates, objective families, search operators, and search budget are matched. Repositioning and fairness are held off. Rates and utilization are percentages in the training and semantic tables; times are minutes.

Variant (default objective)RewardService rateCompletion rateWait timeRide timeDetour timeUtilization
proposed5511.06±3660.9165.05±38.6959.80±35.601.631±0.5934.185±0.5260.261±0.10671.67±29.97
w/o normalization5398.24±3563.9265.46±38.9759.60±35.441.900±0.3844.341±0.6020.346±0.03770.32±30.26
w/o configuration diversity5334.99±3519.9964.93±38.6659.15±35.142.006±0.3154.320±0.5240.345±0.04672.20±29.36

Under the default objective, the proposed design raises mean Reward by 2.09% versus w/o normalization and by 3.30% versus w/o configuration diversity, winning 6/9 and 8/9 paired scenarios. It also reduces mean waiting, riding, and detour times relative to both controls. Service rate and utilization show mixed changes, while the proposed design improves the overall dispatch objective. Appendix E.2 reports this comparison.

In a retrospective semantic-validation case, the audit found two missing negative signs in a generated mutation’s service-time and detour-cost aversion terms. A generated repair restored them. Mean training Reward rose 4123.56→4289.57 (+4.03%), improving all three original scenes. External results are:

Semantic auditRewardService rateCompletion rateWait timeRide timeDetour timeUtilization
Before repair4535.39±2949.9266.27±40.2256.83±33.882.358±0.2335.663±1.2031.484±1.29970.23±30.53
After repair5238.32±3489.5665.70±39.6259.13±35.552.266±0.2384.652±0.5510.489±0.04269.34±30.30

External mean Reward rises 15.50%, improving eight of nine scenes (the remaining scene decreases 0.71%). Completion improves and all three time means fall, while service rate and utilization decline slightly. This case measures the operational effect of diagnosing and repairing a concrete semantic error.

The objective-access control holds the frozen dispatcher fixed and supplies or hides the objective. The six recorded operational metrics below expose the resulting behavior and tradeoffs (rates/utilization are fractions; times are minutes). Within each hidden/supplied pair, bold marks the best mean for each metric: higher rates and utilization, and lower times except for ride time under the long-trip objective.

MetricPickup-time objective hiddenPickup-time objective suppliedDetour-time objective hiddenDetour-time objective suppliedNatural-language long-trip hiddenNatural-language long-trip suppliedNatural-language detour-time hiddenNatural-language detour-time supplied
Service rate0.93±0.040.92±0.070.93±0.040.91±0.050.93±0.040.89±0.100.93±0.040.90±0.05
Completion rate0.82±0.030.82±0.060.82±0.030.81±0.040.82±0.030.72±0.070.82±0.030.80±0.04
Wait time (min)1.66±0.131.39±0.211.66±0.131.80±0.481.66±0.133.37±0.561.66±0.132.07±0.76
Ride time (min)5.33±0.205.27±0.325.33±0.205.10±0.275.33±0.207.07±0.445.33±0.205.10±0.24
Detour time (min)0.30±0.040.15±0.010.30±0.040.17±0.010.30±0.041.75±0.190.30±0.040.16±0.01
Utilization0.84±0.140.91±0.090.84±0.140.87±0.110.84±0.140.89±0.090.84±0.140.88±0.09

Supplying the pickup-time objective lowers mean wait from 1.66 to 1.39 minutes; supplying the detour-time objective lowers detour from 0.30 to 0.17 minutes. The natural-language long-trip objective increases mean ride time from 5.33 to 7.07 minutes, accompanied by higher detour and lower completion. This comparison measures the joint response of objective probing and conditional composition to objective access (Appendix D, “Objective Response”).

AuthorsOpenReview ↗

Response Part 3

Skill diversity works together with repository capacity and conditional composition. In a small repository, redundant skills consume scarce slots without extending the available behaviors. In a larger repository, distinguishing skills by their objectives and decision mechanisms gives the combiner more distinct behaviors to select and combine, analogous to a functional basis for dispatch preferences. Its effect on dispatch performance depends jointly on library size, skill quality, validation, and combiner learning: changing redundancy also changes the behaviors available to the combiner and its learning problem. The mechanism is described in Section 3.3 and Appendix C, Algorithm 1: behavioral signatures use shared fixed scenarios and reference scaling; similarity guides proposals while the library is being filled, without rejecting them at a similarity threshold. Once the library is full, replacement targets its most redundant unprotected member and requires the new skill's maximum similarity to decrease by the prescribed margin. These rules manage redundancy as part of the complete discovery-and-composition process.

The independent skills in Table 1 are competitive, but different scenario groups favor different skills. The strongest skill in each group is identified by evaluation after the fact; those comparisons do not supply a rule for choosing the best skill for a new vehicle state, scenario, or objective. The combiner addresses this selection problem by choosing and blending skills according to each vehicle's state, the scenario, and the platform objective. Table 1 in Section 4.3 compares this dispatch-only combination with the strongest independent skill and the complete system, whose repositioner further reallocates idle vehicles. The controls examine this ride-sharing-specific design; general LLM evolution provides its search operators.

Q3. Geographically and temporally distinct demand

Thank you for your insightful comment. We evaluated frozen policies on Western Queens demand from August 3, 2026, using its road network and 18 zones, with 08:00–20:00 demand windows. This tests transfer across NYC boroughs and months. RideSkill has the highest mean Reward across all five scenario groups in the complete comparison, showing that it remains effective in the new demand setting. We show Hour and Fleet size below; Table 8 in Appendix D.9 gives all scenario groups.

Each scenario group follows Table 1's configurations, with three environment seeds per configuration and 60-minute episodes. The table gives mean±population standard deviation across configuration means. Rates/utilization are fractions and times are minutes.

Scenario groupMethodRewardService rateCompletion rateWait time (min)Ride time (min)Detour time (min)Utilization
HourNearest matching1,517±2600.98±0.000.89±0.010.78±0.045.07±0.130.50±0.080.14±0.02
HourKM1,515±2570.98±0.000.89±0.010.78±0.045.08±0.130.52±0.090.14±0.02
HourGS1,513±2580.98±0.000.89±0.010.79±0.045.10±0.130.53±0.090.14±0.02
HourREDA1,501±2630.97±0.000.88±0.011.25±0.044.90±0.120.27±0.030.15±0.03
HourMFRL1,536±2700.98±0.000.89±0.011.09±0.044.78±0.120.19±0.030.15±0.03
HourBMG-Q1,506±2650.98±0.000.89±0.011.44±0.044.88±0.130.26±0.030.16±0.03
HourOSPO1,436±2640.98±0.000.86±0.012.73±0.124.78±0.140.16±0.000.19±0.03
HourTriple-BERT1,184±2210.92±0.010.78±0.025.62±0.264.53±0.140.21±0.020.22±0.03
HourRideSkill1,562±2730.98±0.000.90±0.010.86±0.164.70±0.130.12±0.020.48±0.28
Fleet sizeNearest matching1,406±1220.97±0.030.86±0.041.08±0.415.40±0.380.76±0.390.24±0.18
Fleet sizeKM1,398±1330.97±0.030.86±0.051.06±0.395.44±0.380.81±0.420.24±0.17
Fleet sizeGS1,393±1400.96±0.040.85±0.051.11±0.485.42±0.340.80±0.390.24±0.17
Fleet sizeREDA1,432±690.97±0.020.86±0.031.44±0.345.15±0.230.42±0.210.28±0.22
Fleet sizeMFRL1,472±720.98±0.010.87±0.021.25±0.315.04±0.310.34±0.290.27±0.21
Fleet sizeBMG-Q1,446±590.98±0.010.87±0.021.56±0.115.14±0.270.43±0.260.28±0.21
Fleet sizeOSPO1,394±410.98±0.020.85±0.022.89±0.164.88±0.020.16±0.000.34±0.25
Fleet sizeTriple-BERT1,128±600.91±0.030.75±0.035.47±0.484.61±0.040.28±0.140.35±0.21
Fleet sizeRideSkill1,521±290.99±0.000.89±0.011.12±0.434.82±0.060.13±0.030.68±0.31

Service and completion are close for several methods, while utilization differs more substantially; utilization counts pickup travel and empty repositioning as non-idle activity. Table 8 in Appendix D.9 reports the full comparison.

Reviewer_BMB2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RideSkill, a hierarchical framework for ride-sharing dispatch. An LLM-based evolutionary process generates reusable vehicle–order scoring functions, a combiner that selects and blends them according to vehicle states and platform objectives, and a repositioner for idle vehicles. The resulting programs operate without runtime LLM calls and are evaluated under varying operating conditions and objectives.

Critical strengths

Generating executable dispatch programs offline avoids repeated LLM calls during deployment, which is useful for large-scale dispatch.

The modular design makes the dispatch rules easier to inspect and allows different vehicles to use different combinations of skills.

The experiments cover several operating conditions, including fleet size, vehicle capacity, speed, and demand, and report promising results against multiple types of baselines.

Critical weaknesses

My main concern is the novelty beyond existing LLM-based ride-hailing methods. Such methods could potentially be adapted to ride-sharing by incorporating capacity constraints, ongoing trips, and additional detour costs. The paper does not sufficiently explain what makes this adaptation difficult or establish the advantage of its skill framework over a carefully adapted existing approach. Applying LLM techniques to ride-sharing alone does not establish a substantial methodological contribution.

Although Appendix Table 4 provides the search budgets and stopping settings, the paper does not report convergence curves, actual training time, or LLM usage and cost. It is therefore difficult to assess how much computation is required to obtain a useful skill and whether performance has stabilized within the reported search budget.

The main ablations focus on the single-skill variant and removing the repositioner. These comparisons, together with the supplementary analyses, do not isolate the contributions of several key designs, including self-checking, diverse task generation, and group-relative fitness. More targeted ablations are needed to establish which components account for the improvements.

The authors should discuss how the framework could extend to other order-matching or sharing applications, including ride-hailing. It remains unclear which components could be reused and which would require new skills or additional training. Demonstrating transfer beyond one application would strengthen the paper’s generality and broaden its contribution.

Questions for authors

see weaknesses.

3 comments
AuthorsOpenReview ↗

Response Part 1

Thank you for the careful review. Substantive manuscript changes are marked in blue.

W1. Ride-sharing contribution and LLM-based comparison

Thank you for your insightful comment. We adapted Zhang et al.’s ride-hailing method to ride-sharing for Table 1. RideSkill contributes an application-inspired dispatch algorithm built around skill discovery and conditional composition. It discovers reusable vehicle–order scoring skills and learns a combiner that selects and blends them for each vehicle according to its state, operating scenario, and platform objective. General LLM evolutionary methods provide effective tools for discovering programs. RideSkill uses those tools to construct a dispatch architecture that adapts the behavior of a frozen skill library to changing traffic conditions and platform priorities.

This design follows the structure of ride-sharing. Idle vehicles, vehicles carrying passengers, and vehicles with tight service commitments need different assignment preferences. The combiner uses residual capacity and committed routes to select and combine skills for these different states, while the supplied objective expresses the platform's priorities. A platform can therefore change its tradeoff between service, completion, and passenger costs without rerunning program search.

The adapted Zhang et al. comparator evolves a single dispatch program. The general evolutionary search operators are usable for this task; the additional mechanism in RideSkill is the reusable meta-skill library and its state- and objective-conditioned composition, which the earlier comparators do not provide. This mechanism connects skill discovery to the heterogeneous needs of vehicles and to scenario and objective adaptation.

The interfaces are described in Sections 3.2–3.3, the general program-search distinction in Appendix A, “LLM-guided evolutionary program search,” and the adapted comparator in Appendix D, “Comparative Baselines.”

W2. Search progress and computational cost

Thank you for your insightful comment. We added the three-phase search-progress curves in Figure 3, Section 4.2, “Training process.” They track the retained best-so-far policies across generations and mark the best-found and early-stopping generations. Each plotted point averages six monitoring scenarios spanning 6,877–11,252 orders, 308–1,291 vehicles, capacities 1–8, and speeds 29.43–41.03 km/h.

Our complete training took 23.75 hours with Claude and 40.62 hours with GLM. These times include local evaluation and API response and thinking time; the difference mainly reflects model response speed and API service conditions, which vary with the time of access. The API costs are around USD 80 for Claude and USD 5.25 for GLM. Appendix D.1, Table 4, reports the search budgets and stopping settings.

W3. Training-design controls

Thank you for your insightful comment. The new controls in Appendix E, “Analysis,” test normalization and training-configuration diversity, alongside a separate semantic-repair case. Under the default objective, the proposed design improves Reward and reduces waiting, riding, and detour times relative to both training controls; the repair case shows how correcting a generated error improves dispatch performance.

The original Claude and GLM endpoints are no longer callable, so we used DeepSeek-v4.1-flash for these controlled tests within the available time and budget.

The compared configurations share the frozen skill library, initial candidates, objective families, demand windows, operators, and budget of an initial generation plus three updates. We test two controls:

  • w/o normalization replaces baseline-relative, scale-normalized/clipped fitness and its corresponding feedback with raw-return fitness and feedback in raw units. It tests normalization in selection and feedback jointly.
  • w/o configuration diversity fixes training at 800 vehicles, capacity 6, and 35 km/h. Objectives and demand windows remain matched, so this tests variation in operating configurations while retaining objective diversity.

Evaluation uses the default objective on nine 60-minute scenarios spanning three demand windows and configurations from 100–2,000 vehicles, capacity 2–10, and speed 25–45 km/h, with repositioning and fairness held off. Values are mean±sample standard deviation across scenarios; Service rate, completion rate, and utilization are percentages, while wait, ride, and detour are minutes.

VariantRewardService rateCompletion rateWait timeRide timeDetour timeUtilization
proposed5511.06±3660.9165.05±38.6959.80±35.601.631±0.5934.185±0.5260.261±0.10671.67±29.97
w/o normalization5398.24±3563.9265.46±38.9759.60±35.441.900±0.3844.341±0.6020.346±0.03770.32±30.26
w/o configuration diversity5334.99±3519.9964.93±38.6659.15±35.142.006±0.3154.320±0.5240.345±0.04672.20±29.36
AuthorsOpenReview ↗

Response Part 2

The proposed design improves mean Reward by 2.09% over w/o normalization and 3.30% over w/o configuration diversity. Mean waiting, riding, and detour times decrease relative to both controls. Service rate is slightly higher without normalization, and utilization is slightly higher without configuration diversity. Overall, normalization and varied training configurations improve the default-objective dispatch result. Appendix E.2 reports this comparison.

For semantic validation, a retrospective audit of one generated mutation identified two missing negative signs in service-time and detour-cost aversion channels. One generated repair restored them. Mean training Reward rose from 4123.56 to 4289.57 (+4.03%), improving all three original scenes. The original and repaired programs were then evaluated on the same nine external scenes:

Semantic auditRewardService rateCompletion rateWait timeRide timeDetour timeUtilization
Before repair4535.39±2949.9266.27±40.2256.83±33.882.358±0.2335.663±1.2031.484±1.29970.23±30.53
After repair5238.32±3489.5665.70±39.6259.13±35.552.266±0.2384.652±0.5510.489±0.04269.34±30.30

External mean Reward improves by 15.50%, with gains in eight scenes and a decline in one (−0.71%). Completion increases and time metrics decrease; Service rate and Utilization decrease slightly. This case demonstrates the operational effect of diagnosing and repairing a concrete semantic error. The results are in Appendix E, “Analysis,” under “Training-design controls” and “Semantic validation”; Section 3.3 describes the self-checking procedure.

We also added the strongest independent skill to Table 1 in Section 4.3. These skills are competitive, but the best skill differs across scenario groups. The table selects each group's strongest skill retrospectively by mean anchor Reward and holds it fixed across that group's configurations. In online dispatch, this best skill is not known in advance. The combiner performs the selection and composition using each vehicle's state, the operating scenario, and the supplied objective. Below we show Hour and Fleet size, including RideSkill with and without repositioning; Table 1 reports all scenario groups. Cells are mean±population standard deviation across configuration means, after averaging environment seeds at each configuration; rates and utilization are fractions, and times are minutes.

Scenario group / variantRewardService rateCompletion rateWait timeRide timeDetour timeUtilization
Hour / strongest independent skill7,405±1,5290.95±0.040.83±0.041.88±0.225.52±0.200.36±0.060.76±0.15
Hour / without repositioning7,409±1,5400.93±0.050.83±0.041.78±0.085.32±0.240.24±0.030.77±0.14
Hour / full RideSkill7,561±1,5620.94±0.050.84±0.041.65±0.125.38±0.240.23±0.030.90±0.10
Fleet size / strongest independent skill7,833±2,8070.70±0.240.63±0.211.80±0.114.39±0.970.37±0.040.91±0.07
Fleet size / without repositioning7,806±2,8150.68±0.240.62±0.211.59±0.094.20±0.880.22±0.030.91±0.07
Fleet size / full RideSkill7,935±2,9250.69±0.240.63±0.211.54±0.094.24±0.920.22±0.030.95±0.02

Across all five scenario groups in Table 1, the complete system has about 1.1–2.1% higher mean Reward than the strongest independent skill. Conditional composition brings these behaviors into one adaptive dispatcher, and repositioning further improves the spatial allocation of idle vehicles. The rows with and without repositioning show the contribution of this final stage.

W4. Transfer to other matching applications

Thank you for your insightful comment. Non-pooling ride-hailing is the capacity-one specialization of the present interface, reusing the skills, combiner, matcher, and repositioner. Capacity one is included in the capacity sweep summarized in Table 1. Food delivery and pickup-and-delivery logistics can reuse the discovery-and-composition framework, with task-specific state, scoring, and feasibility contracts.

For delivery, the interface must represent preparation times, deadlines, parcel loads, and pickup-before-delivery constraints. Skills whose inputs retain the same meaning may be reused; new features or actions can require new skills and retraining or revalidation of the combiner and repositioner. These component-level extensions are discussed in Appendix F, “Discussions,” under “Future Directions.”

Thanks for your rebuttal, I still have concerns on the novelty of this paper, and I don't think your response on W1 convinced me. So I maintain my score.

Reviewer_ahTZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses LLMs to generate a dispatch policy in a simulated rideshare system. Crucially, the rideshare policy itself is not an LLM, it is a collection of heuristics, which has been optimized using an evolutionary search process that incorporates an LLM, both to evolve the algorithms and to propose training tasks. An additional source of novelty is their reward estimator, inspired by GRPO, which allows training when tasks have different reward scales. The simulated environment, RideGym, is based on the well-known NYC taxi dataset. The authors show better performance than a greedy baseline and an ML-based method.

Critical strengths

The technique is interesting and does in fact appear to work well, better than baselines. The experiments are well done and convincing.

Critical weaknesses

There is a substantial literature on LLM-guided evolutionary program synthesis, just outside the rideshare domain. It is largely not cited. This is an important weakness in related work. This related work should be added, and the evolutionary search approach distinguished at a technical level from those techniques.

Questions for authors

See concern about existing LLM/evolutionary search work, how is it different?

3 comments

Thank you for reviewing our paper. Substantive manuscript changes are marked in blue.

W1/Q1. Related work and methodological distinction

Thank you for your insightful comment. We expanded Appendix A.3 to cover FunSearch, EoH, ReEvo, and LLaMEA and distinguish their general program-search methods from RideSkill’s application-inspired discovery and conditional composition of reusable dispatch skills. RideSkill contributes the transportation-specific discovery-to-composition mechanism: Phase 1 discovers reusable vehicle–order scoring skills; Phase 2 selects and blends them for each vehicle from residual capacity, committed routes, and the current platform objective; matching resolves order competition; and Phase 3 repositions idle vehicles. The same mechanism adapts the dispatch policy when the objective or operating condition changes.

The skill–combiner design follows the dispatch needs of ride-sharing: idle vehicles, vehicles carrying passengers, and vehicles with little remaining service slack benefit from different scoring behaviors. Conditional composition selects and blends these behaviors as vehicle states, scenarios, and platform objectives change. General program-evolution methods supply effective tools for searching functions, thought–program pairs, or metaheuristics under a task score. RideSkill searches a library of feasible scoring skills and a state- and objective-conditioned composition policy whose outputs are directly consumed by shared-capacity matching. Zhang et al. has been adapted to the same RideGym interface for comparison, but it evolves a single scoring/objective program for the adapted dispatch pipeline; it does not discover reusable dispatch skills and conditionally compose them for heterogeneous vehicles. Thus the LLM is the construction mechanism, while the ride-sharing-specific representation, conditional reuse, and operational phases are the contribution.

Appendix A.3 reviews general LLM-guided program search and explains this distinction; Appendix A.2 and Table 2 compare transportation-specific LLM methods. Sections 3.2–3.3 define the skill and composition interface, and Appendix E, “Analysis,” reports the controlled module comparisons.

References:

[1] Bernardino Romera-Paredes et al., Mathematical discoveries from program search with large language models, Nature, 2024.

[2] Fei Liu et al., Evolution of Heuristics: Towards Efficient Automatic Algorithm Design Using Large Language Model, ICML, 2024.

[3] Haoran Ye et al., ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution, NeurIPS, 2024.

[4] Niki van Stein and Thomas Bäck, LLaMEA: A Large Language Model Evolutionary Algorithm for Automatically Generating Metaheuristics, IEEE Transactions on Evolutionary Computation, 2025.

Thanks for your reply. Although I have read your comment here, and the short summaries in appendix A.3, I still have a hard time understanding how your technique fits into the larger context of LLM-guided evolutionary program synthesis.

Thank you for your prompt follow-up. RideSkill follows the same basic approach as LLM-guided evolutionary program synthesis: the LLM supplies program-level mutation and crossover within a standard evolutionary strategy, and candidate programs are executed and selected by measured performance. Ride-sharing additionally requires candidates to perform across changing scenarios and objectives with substantially different reward scales. A fixed-task evaluation does not expose candidates to these changes, and aggregating raw returns across tasks can let high-scale objectives dominate selection. We therefore sample diverse scenario–objective pairs and use group-relative fitness in the combiner and repositioner searches to balance task contributions.

Our contribution is an application-inspired ride-sharing dispatch method. Vehicles with different passenger loads and service commitments need different assignment preferences. We address these requirements by discovering reusable scoring skills and evolving a combiner that selects and blends them according to vehicle state, scenario, and objective. This design allows the frozen dispatcher to adapt across the evaluated scenarios and objectives without retraining or repeating program search. The contribution lies in this practical method for discovering and composing ride-sharing skills, rather than a new general evolutionary search algorithm.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank all reviewers, the AC, and the PC for the careful evaluation and constructive discussion. With one day remaining in the rebuttal period, we summarize the main contribution of RideSkill and the changes made in response to the reviews.

Contributions.

  • An application-inspired ride-sharing framework. RideSkill addresses the operational structure of ride-sharing: vehicles with different passenger loads and service commitments require different assignment preferences. It discovers reusable vehicle–order scoring skills, uses a state- and objective-conditioned combiner to select and blend them, and sequentially repositions idle vehicles.
  • Adaptation across scenarios and objectives. The dispatcher responds to fleet size, speed, capacity, demand conditions, and platform objectives without retraining or repeating program search at deployment. Multi-scenario and multi-objective training, together with group-relative fitness, addresses the heterogeneous task support and reward scales that make fixed-task dispatch policies difficult to transfer.
  • Practical and real-time dispatch. The discovered skills, combiner, and repositioner form an executable dispatcher that operates without LLM calls in the dispatch loop, supports inspection and reuse of learned behaviors, and adapts vehicle decisions to current state, scenario, and objective.

Changes made during the rebuttal.

  • We clarified how RideSkill differs from general LLM-guided evolutionary methods: the evolutionary search is the construction mechanism, while the contribution is adapting it to ride-sharing, where a dispatcher must respond to changing scenarios and platform objectives. The skill repository and conditional combiner address this task-specific need.
  • We expanded the empirical evidence with more recent baselines, additional ablations, and additional geographic and temporal settings, including the Western Queens transfer and a different demand month.
  • We corrected formula typos and added analyses and visualizations of online decision time, search progress, training time, and training behavior.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.