acceptodds
Under review as a conference paper at ICLR 2027

Beyond Distance: Learning Courier Behavior with Large Language Models

Abstract

Last-mile delivery is a cornerstone of modern urban logistics, and its efficiency depends heavily on high-quality route planning. Conventional approaches typically optimize for a single geometric objective such as minimizing travel distance or time, as in the classical Traveling Salesman Problem (TSP). Solving such problems to optimality can require substantial computation time as the number of delivery stops increases. However, these methods overlook practical factors that shape real-world delivery, including couriers' behavioral preferences. To address these limitations, we propose the Delivery Optimization and Route Advisor (DORA), a route-planning system built on the compact Qwen3-4B-Instruct model and an adapted Group Relative Policy Optimization (GRPO) algorithm. DORA learns couriers' behavioral patterns from real delivery traces and recommends routes that better reflect couriers’ observed delivery patterns. We train and evaluate DORA on LaDe, a publicly available last-mile delivery dataset, and on the Amazon Last Mile Routing Research Challenge (ALMRRC) as well. DORA is designed to reduce inference costs relative to larger language-model-based approaches, supporting cost-effective deployment for last-mile route prediction.

Primary area:
applications to robotics, autonomy, planning
Keywords:
last-mile delivery, route prediction, courier behavior modeling, large language models, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vkhFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents DORA, an LLM-based framework for learning courier routing behavior from historical delivery traces and routing advice. Rather than minimizing geometric distance, DORA seeks to reproduce observed delivery sequences. The framework integrates candidate routes from existing routing models, historical information, supervised fine-tuning, and route-level reinforcement learning. Experiments on LaDe and Amazon demonstrate the potential of combining learned routing advice with executable language-model decisions.

Critical strengths

The authors correctly identify the discrepancy between distance-minimizing routes and observed courier behavior. Learning behavioral routing patterns is an interesting direction beyond classical distance-based optimization.

Critical weaknesses
  1. The framework mainly combines existing components, including Qwen3, LoRA, GRPO-style learning, candidate routing proposals, and action masking. Although the integration is useful, the paper does not introduce a fundamentally new learning architecture or optimization objective.

  2. The overall methodological narrative is difficult to follow. In particular, the LaDe and Amazon settings involve substantially different input representations, output spaces, training procedures, and execution mechanisms, yet the rationale behind these different designs and their connection under a unified framework are not sufficiently explained. Moreover, the method section introduces several technical components, particularly in the Amazon pipeline, without sufficient intuitive examples, making the overall workflow difficult to understand.

  3. The paper emphasizes the potential efficiency of compact LLMs, but does not provide a fair end-to-end runtime comparison with specialized routing methods.Appendix F reports approximately 10.6–11.1 minutes of A100 decoding time for 300 LaDe tasks per model, excluding Graph2Route proposal generation and historical retrieval. Amazon additionally requires candidate generation, multiple LLM samples, voting, and LKH execution. The reported measurements come from different hardware and experimental configurations, so they cannot support a reliable cost-effectiveness conclusion.

Questions for authors

See Weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer vkhF

Thank you for the review. We address each point below.

Novelty

We agree. The contribution we claim is a formulation of route prediction in which a policy reads the outputs of existing routing systems as advice and acts through an interface that makes every output an executable route, with two such interfaces evaluated on two real delivery datasets. On LaDe, the policy restructures its advice. 253 to 274 of its 300 routes match none of the nine proposals it reads (Section 5.2), and its rewrites cut returns to an already visited area by 0.87 per route (95% CI [0.73, 1.02]). On Amazon, the relational interface improves on a full rerun of the challenge winner (Section 5.4). The revised introduction will state the contributions at this level.

One framework for both settings

Both settings follow the formulation in Section 3 (Eq. 1). Fixed advisers produce advice, a Qwen3-4B-Instruct policy with a LoRA adapter emits an action, and a fixed executor turns it into a complete route. The two datasets come from delivery operations in two countries, and the settings differ as follows.

LaDeAmazon
Route size9.2 stops on averageabout 148 stops
Prompttask table, nine G2R stop orders, two earlier routes as area (AOI) sequencesup to twelve candidate zone orders, no coordinates
Actionstop order under a remaining-stop maskzone and cluster relations
Executornone, the decoded order is the routeLKH with the relations as soft penalties
TrainingSFT, then route-level RLSFT

The policy acts at the granularity of its advice and leaves the rest to the executor. A LaDe route is short and its proposals are complete stop orders, so the policy writes the whole route. An Amazon route is about sixteen times longer and its candidates are zone orders, so the policy writes relations among zones and LKH places the stops. The task's geometry enters through the executor, so the Amazon prompt carries no coordinates. The revision will state this rule in Section 3 with this table.

An example of the Amazon pipeline

We agree that an example was missing. Schematically, take an Amazon route with 150 stops in 18 zones.

  1. The prompt lists up to twelve candidate zone orders, nine from a zone-order model, up to two retrieved routes, and the champion's route.
  2. The policy is sampled seven times, and each sample writes relations such as "zone 4 before zone 7" or "cluster 2 before cluster 3".
  3. A relation is kept if at least four of the seven samples contain it.
  4. LKH orders all 150 stops with the kept relations as soft penalties, which it trades off against travel time.

The revision will add this example to Section 4.

End-to-end runtime

We agree. At inference, DORA keeps its advisers and adds its own step on top of them. On LaDe, that step is about 2.1 to 2.2 A100-seconds of decoding per task, on top of the nine G2R proposals it reads, which take about 0.08 CPU-seconds per task to generate. On Amazon, it adds about 6.7 A100-seconds per route for eight policy samples and about 0.9 for generating the candidate zone orders, both in batched inference, and a two-second LKH solve to the champion pipeline (about 42.8 CPU-seconds per route). We will remove the abstract's sentence on cost-effective deployment. The revision will report these stage costs, with their hardware, in one table in Appendix F.

We are happy to discuss any remaining concern.

Reviewer_jn5TOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DORA, an LLM-based framework for courier route prediction that learns to aggregate existing routing proposals and historical delivery traces. It employs supervised fine-tuning and route-level GRPO for direct route generation on LaDe, while using supervised structured constraint prediction combined with an LKH solver on Amazon. Experiments evaluate the approach on two real-world delivery datasets.

Critical strengths
  1. The paper addresses a practical problem in last-mile logistics and evaluates the approach on two real-world delivery datasets.

  2. The proposed constrained decoding and structured constraint execution ensure valid and executable route outputs.

  3. The experiments include relevant routing baselines, multiple training seeds on LaDe, and statistical uncertainty estimates, providing useful evidence for assessing the method.

Critical weaknesses

W1. Insufficient task motivation and practical relevance. Although predicting actual courier routes can support downstream tasks such as ETA estimation and dispatching, the paper does not sufficiently establish these practical benefits. Moreover, matching historical courier behavior does not necessarily imply better routing efficiency, as observed routes may be suboptimal.

W2. Lack of a unified methodology. The two datasets use substantially different approaches: LaDe employs direct route generation with GRPO, whereas Amazon uses supervised relation prediction followed by LKH optimization. This dataset-specific design weakens the methodological coherence and leaves the generalizability of the proposed framework unclear.

W3. Limited methodological novelty and insufficient justification for LLMs. The method primarily aggregates existing routing proposals using standard SFT/GRPO, without introducing a clearly novel learning mechanism. More importantly, the necessity of LLMs is not established, as comparisons with lightweight learned aggregation models using the same routing advice are missing.

W4. Statistically inconclusive improvements on LaDe. DORA's improvements over Graph2Route and the simple Borda-of-nine baseline are not statistically significant, with both 95% confidence intervals including zero. Similarly, the additional improvement from GRPO over SFT is statistically inconclusive, weakening the claimed effectiveness of route-level RL.

W5. The Amazon experiments fail to demonstrate the value of the learned policy. The deterministic COPY baseline, which requires no LLM, significantly outperforms DORA (AS: 0.023681 vs. 0.023932). This suggests that the improvements over the champion primarily arise from relational re-optimization rather than LLM-based learning.

W6. Unsupported inference-cost claims. The abstract emphasizes the efficiency of compact LLMs, but the paper provides neither direct comparisons with larger LLMs nor comprehensive end-to-end inference-cost measurements. Thus, the claimed deployment and cost advantages remain unsubstantiated.

Questions for authors

Please see the weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer jn5T (1/2)

Thank you for the detailed review. We answer W1 to W6 in order.

W1. Task motivation and practical relevance

Our task is route prediction, the order in which a courier will visit a given set of stops, and we agree that the submission did not explain why this order matters. In prior work on the task, the predicted route feeds other platform systems. DRL4Route (Mao et al., 2023) notes that arrival-time estimation takes the predicted route as input, because a task's arrival time depends on its position in the route, and that dispatching can assign new orders near a worker's predicted route. A survey of service route and time prediction (Wen et al., 2023) makes the same point about dispatching. For these uses, the target is the route the courier will actually drive.

We agree that observed routes may be suboptimal by distance, and we do not claim that DORA's routes are more efficient. Experienced drivers route by knowledge of their areas that distance does not capture (Merchán et al., 2024). On the 300 LaDe tasks, courier routes are on average 39.4% longer than the shortest open path (straight-line distance) and follow a habitual order of areas, which predicts them better than distance-based OR-Tools (KRC 0.218 for the area rule, and 0.263 when LKH orders the stops within each area, against 0.133). The revision will present DORA as route prediction throughout, add this motivation to the introduction, and state in the limitations that the paper does not measure ETA or dispatching gains.

W2. Unified methodology

Your description is accurate. What the two instances share is the formulation in Section 3 (Eq. 1). Fixed advisers produce advice, a Qwen3-4B-Instruct policy with a LoRA adapter emits an action, and a fixed executor turns it into a complete route. The two datasets come from delivery operations in two countries, and the action follows their route size and advice granularity. LaDe tasks average 9.2 stops and their proposals are complete stop orders, so the policy writes the stop order. Amazon routes average about 148 stops and their proposals are zone orders, so the policy emits zone and cluster relations and LKH places the stops using travel times.

Both policies start with supervised training on observed routes, the LaDe policy then adds route-level RL, and adapter settings were chosen per dataset (Appendix B). The paper shows one formulation improving on its strongest adviser on both datasets. We do not claim that a trained policy transfers across datasets. The revision will state the interface rule in Section 3 with a per-dataset table.

W3. Novelty and the need for a language model

We agree that DORA introduces no new learning mechanism, and its training uses standard SFT and a GRPO-style objective. The contribution we claim is a formulation of route prediction in which a policy reads the outputs of existing routing systems as advice and acts through an interface that makes every output an executable route, with two such interfaces evaluated on two real delivery datasets. The revised introduction will state the contributions at this level.

We ran the comparison you describe, with its design fixed before scoring. A LightGBM next-stop ranker that receives the same information as DORA's prompt (the task table, each stop's positions in the nine G2R proposals, and the area order of the two retrieved routes) and is trained on the same 512 tasks reaches KRC 0.249 on the 300 tasks, against 0.256 for DORA (difference +0.007, 95% CI [-0.027, 0.042]). On LaDe, a learned aggregator of the same advice therefore does about as well without a language model, and the revision will not claim that one is necessary; it adds this ranker to Table 1. What the paper does show is what DORA adds to the routing methods it reads, which is to restructure their routes. It rewrites the greedy G2R proposal on 275 to 287 of 300 tasks, by seed, and 253 to 274 of its routes match none of the nine proposals (Section 5.2). Over all 300 tasks, these rewrites cut returns to an already visited area (AOI) by 0.87 per route (95% CI [0.73, 1.02]) and raise KRC on the rewritten routes by 0.054 to 0.079, while Borda-of-nine, which combines the same proposals without a language model, returns 1.31 times.

AuthorsOpenReview ↗

Response to Reviewer jn5T (2/2)

W4. Statistical significance on LaDe

We agree, and the corrected LaDe numbers (see our general comment) change which gains are significant. On the 300 development tasks, DORA improves significantly on the routing proposals it reads: +0.056 over the saved-full G2R reference (95% CI [0.018, 0.093]) and +0.047 over Borda-of-nine ([0.009, 0.085]), with all five training seeds above both. Its differences from the historical area-order rule (+0.038, [-0.004, 0.079]) and DRL4Route (+0.030, [-0.021, 0.083]) are not significant. Route RL minus SFT is +0.003 ([-0.003, 0.009]); the significant RL comparison is with continued supervised training at the native learning rate (+0.018, [0.002, 0.033]), which lowers four of the five SFT anchors. The revision will describe DORA as improving on the proposals it reads, retitle Section 5.2 accordingly, and describe route RL as continuing training without that decline, with no significant gain over SFT.

W5. Value of the learned policy on Amazon

We agree with this reading. COPY is DORA's formulation with the learned policy replaced by a fixed rule that writes the champion's route as the same relations, which the same compiler and two-second LKH solve then execute. It improves on the champion by more (AS -0.001190, [-0.001365, -0.001024], lower is better) than the system with the learned policy (-0.000939, [-0.001174, -0.000704]). Our student-minus-COPY test, fixed before scoring, gives +0.000231 ([0.000064, 0.000404]) on the 2,752 routes not used for adapter selection. On Amazon, the gain over the challenge winner therefore comes from executing its route as soft relations, which adds one two-second solve to the champion's 42.8 CPU-seconds per route and needs no language model. The learned policy does not add to this gain. The revision will say so in the abstract, introduction, and conclusion, and will present the Amazon experiment as evidence for the relational interface.

W6. Inference-cost claims

We will remove the abstract's sentence on cost-effective deployment. We have not compared DORA with larger language models under the same recipe, so the revision will make no claim about model size and cost. At inference, DORA keeps its advisers and adds its own step on top of them. On LaDe, that step is about 2.1 to 2.2 A100-seconds of decoding per task on top of G2R. On Amazon, it adds about 6.7 A100-seconds per route for eight policy samples and about 0.9 for generating the candidate zone orders, both in batched inference, and a two-second LKH solve to the champion pipeline (about 42.8 CPU-seconds per route). The revision will report these stage costs, with their hardware, in one table in Appendix F.

We are happy to discuss any remaining concern.

Reviewer_4d4eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DORA, an LLM-based framework for predicting courier delivery sequences from historical behavior and routing advice. DORA employs Qwen3-4B-Instruct with LoRA adaptation and introduces two route-generation interfaces. For LaDe, the model directly generates valid stop permutations using supervised fine-tuning and route-level GRPO. For Amazon, it predicts structural routing relations that are incorporated into an LKH solver. Experiments on LaDe and Amazon evaluate the framework against traditional routing methods, learned route predictors, and routing-advice aggregation baselines.

Critical strengths

The paper addresses courier route prediction beyond geometric distance optimization by learning delivery sequences from observed courier behavior. This formulation reflects the difference between distance-minimizing routes and actual delivery decisions.

Critical weaknesses
  1. The paper primarily relies on LLMs to learn historical courier routing patterns, without sufficiently explaining the underlying behavioral preferences. Moreover, the independent contribution of the LLM component over existing routing methods remains unclear.

  2. The paper adopts different strategies across the two datasets without sufficient justification. Furthermore, the performance improvements are not consistently significant, and the additional computational costs are not adequately evaluated.

Questions for authors
  1. The paper shows that distance-based optimization cannot fully reproduce actual courier routes. However, it does not explain why couriers deviate from shortest-distance routes. What specific behavioral patterns does DORA learn beyond simply fitting historical delivery sequences?
  2. The paper adopts substantially different architectures and training strategies for LaDe and Amazon. Are these choices specifically optimized for each dataset, and how can the generalizability of the proposed framework be demonstrated?
  3. The framework introduces additional LLM inference. How much computational overhead does this introduce, and is the improvement in prediction performance sufficient to justify the additional inference cost?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 4d4e

Thank you for the review. We address each point below.

Why couriers deviate from shortest routes, and what DORA learns beyond fitting history

Couriers deviate from the shortest route because they route by experience of their areas, not by distance alone. The paper introducing the Amazon data describes seasoned drivers' routes as reflecting tacit knowledge of their environment and criteria such as convenience, efficiency, and safety (Merchán et al., 2024). LaDe records routes, not reasons, so we read couriers' revealed preferences from the regularities in their routes. The main regularity is a habitual order of areas (AOIs). Sorting stops by the point at which the courier usually reached each area in earlier routes, then by acceptance time, predicts courier routes with KRC 0.218, and 0.263 when LKH orders the stops within each area, against 0.133 for distance-based OR-Tools (Table 1). Like the area rule (24.7%), couriers seldom start at the stop nearest the station (31.3%, against 80.0% for the shortest open path). A habitual order is not the shortest path for each day's stops, and courier routes are on average 39.4% longer than that path.

The clearest pattern DORA learns beyond fitting history is to finish an area before leaving it. Its G2R adviser is fitted to the full pool of historical routes, yet the greedy G2R proposal returns to an already visited area 1.00 times per route, against 0.66 for couriers and 0.13 for DORA.

On Amazon, the learned preferences are explicit rules. Trained on the relations that observed routes satisfy, the policy writes rules such as "zone 4 before zone 7", which LKH applies as soft penalties while settling the rest by travel time. The result improves on the challenge winner (AS -0.000939, 95% CI [-0.001174, -0.000704], lower is better).

What DORA adds over existing routing methods

Against the routing methods it reads, DORA's contribution is to restructure their routes. It rewrites the greedy G2R proposal on 275 to 287 of 300 tasks, by seed, and 253 to 274 of its routes match none of the nine proposals (Section 5.2). Over all 300 tasks, these rewrites cut returns to visited areas by 0.87 per route (95% CI [0.73, 1.02]) and raise KRC on the rewritten routes by 0.054 to 0.079, while Borda-of-nine, which combines the same proposals without a language model, returns 1.31 times.

Different designs for LaDe and Amazon, and generalizability

Both instances follow the formulation in Section 3 (Eq. 1). Fixed advisers produce advice, a Qwen3-4B-Instruct policy with a LoRA adapter emits an action, and a fixed executor turns it into a complete route. The two datasets come from delivery operations in two countries, and the action follows their route size and advice granularity. LaDe tasks average 9.2 stops with stop-level proposals, so the policy writes the stop order. Amazon routes average about 148 stops with zone-level proposals, so the policy emits zone and cluster relations for an LKH solve.

Both policies start with supervised training on observed routes. Adapter settings were chosen per dataset (Appendix B). The paper shows one formulation improving on its strongest adviser on both datasets. We do not claim that a trained policy transfers across datasets. The revision will state the interface rule in Section 3 with a per-dataset table.

Significance and inference cost

The revision will state each claim at the strength of its interval. The significant gains are DORA over the routing proposals it reads on LaDe (+0.056 over the saved-full G2R reference, 95% CI [0.018, 0.093]) and, on Amazon, the structured system over the champion, which also holds on the 2,752 routes not used for adapter selection.

At inference, DORA keeps its advisers and adds its own step on top of them. On LaDe, that step is about 2.1 to 2.2 A100-seconds of decoding per task on top of G2R, and it raises mean KRC from 0.197 for the G2R proposal it reads to 0.256. On Amazon, it adds about 6.7 A100-seconds per route for eight policy samples and about 0.9 for generating the candidate zone orders, both in batched inference, and a two-second LKH solve to the champion pipeline (about 42.8 CPU-seconds per route). It lowers the champion's AS of 0.024871 by 0.000939. Training is a one-time cost of about 11.5 A100-minutes of SFT and 95 of RL per LaDe model (Appendix F). We do not claim that this accuracy justifies the cost. We will remove the abstract's sentence on cost-effective deployment and report these stage costs, with their hardware, in one table in Appendix F.

We are happy to discuss any remaining concerns.

Thank you for your response, which addresses some of my concerns. However, I still have a few questions:

  1. While the current results demonstrate the effectiveness of the DORA routing approach, they do not necessarily prove that the use of an LLM is essential. Could a lightweight network be used as an alternative?
  2. I acknowledge that a trained model may not be directly transferable across different datasets. However, the two datasets in question utilize fundamentally different strategies—not merely different training configurations. One involves direct prediction, while the other requires an LKH-based solver to determine the complete route. Does this imply that a bespoke strategy must be designed for every new dataset?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Correction to the LaDe results

While preparing the experiments for this discussion, we found an error in our LaDe preprocessing. We report it here with corrected numbers. The Amazon experiments use a separate pipeline and are not affected.

What went wrong. Our export sorted each task's stops by delivery time and then, stably, by the minute the order was accepted. Stops accepted in the same minute, about 29% of stop pairs, were therefore stored in their true delivery order. Methods that read the stored order could use it: FIFO and historical AOI + FIFO directly, and G2R, whose encoder reads the stored slots. Borda-of-nine, Medoid-of-nine and DORA read G2R's proposals and inherited the effect. DORA's prompt shuffles the stops and their ids, so it saw the order only through those proposals. Methods that never read the stored order (OR-Tools, nearest neighbour, LightGBM, DeepRoute, DRL4Route and the LKH variants) are unchanged.

Fix. Same-minute ties are now broken by a hash of the order id, so the stored order agrees with the delivery order on about half of the same-minute pairs instead of all of them. We retrained G2R on the corrected export with its official code and hyperparameters. Its predictions no longer depend on the stored order: on 1,789 held-out tasks, re-shuffling the stored order changes its KRC by +0.004 (95% CI [−0.008, 0.016]), while the same check on the original G2R gives +0.346.

Corrected LaDe results (dev300, KRC, higher is better)

MethodSubmittedCorrected
Historical AOI rank + LKH0.2630.263
DORA, retrained on the corrected data (five seeds fixed before training)0.4680.256
DRL4Route0.2260.226
Historical AOI rank + FIFO0.3770.218
Borda-of-nine0.4550.209
Medoid-of-nine0.4470.199
G2R (greedy)0.4460.197
FIFO0.183−0.111

All numbers are KRC over the complete stop order of each static delivery task, as in the paper; they are not comparable with the official LaDe pickup benchmark, which scores the next few pickups from the courier's current position.

Retrained on the corrected data, DORA improves on the proposals it reads: +0.060 over the greedy G2R proposal (95% CI [0.022, 0.097]), +0.047 over Borda-of-nine ([0.009, 0.085]) and +0.058 over Medoid-of-nine ([0.019, 0.096]), with all five seeds above each.

The submitted models on corrected inputs. Without retraining, the three submitted DORA models, reading the corrected G2R proposals, score 0.243 instead of 0.468. They still improve on the proposals they read, by 0.034 to 0.046 KRC over the greedy G2R proposal, Borda-of-nine and Medoid-of-nine (route-paired 95% CIs all above zero).

Revision. The revised paper replaces the LaDe table and every statement that depends on it, including the description of G2R as the strongest learned predictor. All LaDe numbers in our replies in this thread are the corrected ones. We apologize for the error.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.