Source-linked AI summary
ORBITER: Conflict-Aware Decision-Making for Agentic Last-Mile Delivery
Mingzhao Li, Chenxi Liu, Yan Zhao, Hao Miao
TL;DR
Next-order delivery decisions involve competing spatial, temporal, and behavioral cues that existing approaches leave implicit, while direct LLM predictions can be unreliable. ORBITER makes these decisions explicit, adjudicates proposer disagreements with evidence and independent review, and achieves the best ACC@1 and MRR across four cities.
Problem
Existing delivery models leave comparisons among available orders and their local spatiotemporal cues implicit, while LLM decisions remain sensitive to task presentation and often unreliable.
Method
ORBITER represents service as local decision points, combines heterogeneous proposer rankings with structured conflict reports and evidence gathering, then uses an independent critic to review decisions.
Results
9.2% and 4.9% average gains over the strongest baseline in ACC@1 and MRR, respectively, with ORBITER ranking first in both metrics across all four cities.
Takeaways & Limitations
ORBITER’s advantage spans both next-order accuracy and true-order ranking rather than a single city or metric.
Abstract
from arXiv · showhide
Last-mile delivery aims to handle dynamically arriving orders with couriers while modeling complex spatial and temporal correlations. Recent learning-based methods model spatiotemporal dependencies among orders to predict courier service sequences, but leave next-order decision making unexplained. Describing the current delivery state in language allows LLMs to reason explicitly about the spatial, temporal, and behavioral cues behind an individual decision. As direct predictors, however, LLMs remain sensitive to task presentation and often produce unreliable decisions. To address these challenges, we introduce ORBITER, an agentic Order Arbiter for next-order decision-making in last-mile delivery. ORBITER models courier service through decision points, each containing the courier's spatiotemporal state and visible orders and exposing local trade-offs for modeling and verification. Fixed proposers rank the candidates, and a structured report identifies where their rankings disagree. The LLM uses task-specific tools to gather evidence on the leading alternatives, while an independent critic checks the resulting decision against that evidence. We conduct extensive evaluations on data in four cities, where ORBITER outperforms existing state-of-the-art baselines by up to 9.2% on average showing its effectiveness.
Introduction
Last-mile delivery requires next-order decisions under changing spatial and temporal constraints, but prior sequence and ranking models obscure local trade-offs and leave explanations unverified. ORBITER addresses this by modeling inference-safe decision points and combining fixed proposals, LLM evidence gathering, and independent critique.
- Motivation: Last-mile delivery must coordinate rising order volumes across dispersed locations under tight deadlines and changing conditions.These demands affect operational efficiency, courier experience, and customer satisfaction.
- Challenges: Prior models predict service sequences or rank candidates, but end-to-end outputs obscure comparisons among available orders and local spatiotemporal cues.Distance, deadlines, waiting times, location, and service history can favor different candidates, while explanations may remain unverified post-hoc justifications.
- ORBITER: ORBITER shifts the modeling unit from complete routes to individual next-order decisions represented as inference-safe decision points.Each point is generated after a completed service event, retains orders visible at that time, and excludes information from later events.
- ORBITER: The framework preserves each courier’s local spatiotemporal state and available orders for explicit modeling and verification.This formulation makes individual courier decisions explicit rather than folding candidate trade-offs into a final route.
- ORBITER: ORBITER combines trained predictors with explicit LLM reasoning in a conflict-aware agentic framework for spatiotemporal next-order decision making.The supplied context describes fixed proposers, structured disagreement reports, task-specific evidence gathering, and independent critique.
Related Work
Prior mobility-prediction work uses deep learning for location and route prediction, while LLM agents primarily serve as predictors or behavior simulators. ORBITER instead directs LLM reasoning toward conflicts among heterogeneous predictors and checks decisions against task evidence.
- Deep Learning Based Mobility Prediction: Deep learning mobility prediction covers location and route prediction from trajectories and spatiotemporal context.Location prediction is dominated by deep sequential, attention, and graph models, while related work reaches delivery route planning.
- LLM-Based Behavior Simulation: LLM agents simulate behavior as urban residents or delivery riders, retaining the role of predictor or simulator.These designs do not explicitly check LLM outputs against task evidence.
- Conflict-Aware Reasoning: ORBITER directs LLM reasoning toward conflicts among heterogeneous predictors rather than adding another predictor.Its framing treats disagreement as a signal for reasoning instead of noise.
Preliminaries
The preliminaries represent last-mile delivery through locations, orders, and courier trajectories, then define reproducible decision points for next-order selection. At each point, the agent observes an inference-safe state and aims to rank the next served order as highly as possible.
- Definitions: Locations are pickup or drop-off stops, while orders carry locations, acceptance times, deadlines, and observed completion times.A location is represented by coordinates and an area of interest.
- Definitions: A courier trajectory is the time-ordered sequence of served locations and corresponding service times, without the outstanding orders behind each choice.The trajectory records the realized route but not the candidate alternatives available before each next stop.
- Decision Points: A decision point occurs after service at (p_t, τ_t), with candidates accepted by τ_t but not yet served and next order y_t ∈ C_t when |C_t| ≥ 2.Each candidate is described by observable attributes, including location and temporal constraints.
- Decision Points: Anchoring decisions at service time τ_t makes them reproducible from operational logs while preserving information available before the next service event.The state includes the courier’s completed trajectory prefix and observable candidate attributes.
- Problem Definition: The problem models next-order selection at individual service steps, where the agent observes an inference-safe state Φ(s_t) and seeks to place y_t as high as possible in ρ̂_t.This represents courier service as a sequence of step-level agent decisions.
Methodology
ORBITER reconstructs leakage-controlled decision points with explicit candidate sets, then combines heterogeneous proposer rankings with conflict-aware evidence collection and independent criticism. It returns an evidence-grounded next-order decision while preserving the leading alternatives and rationale.
- Decision Point Generation: ORBITER reconstructs each service transition as a decision point with an explicit candidate set, excluding states with fewer than two visible orders or unavailable completed orders.Each candidate set is formed from orders visible at the service time, ensuring labels are conditioned on live alternatives.
- Decision Point Generation: Each candidate is represented by waiting time, deadline slack, courier–order distance, location, AOI, and time-window attributes in an 8-dimensional feature row.Rows are stacked into a candidate matrix shared by every component.
- Heterogeneous Proposal Generator: ORBITER ranks candidates with fixed proposers spanning heuristics, statistical machine learning, and deep learning, preserving full rank vectors and their disagreements.The disagreement report identifies Top-1 proposals, near-top alternatives, panel splits, pairwise ranking correlations, and candidate-specific disputes.
- Conflict-Aware Agentic Reasoning: The agentic loop ties each tool call to an unresolved candidate comparison, requiring time-safe, candidate-directed evidence with traceable provenance before critic review.Tools test current trade-offs, retrieve comparable historical decisions, or probe plausible perturbations; the critic can approve, reject, or request another observation.
- Evidence-Grounded Decision: ORBITER returns the agent decision and two leading rivals as a Top-3 ranking, retaining supporting evidence, leading counterevidence, and critic review in its record.Sharper proposer distributions receive greater fusion weight, while an unweighted mean is used when all concentration values are zero.
Experiments
ORBITER is evaluated on four LaDe-P city subsets against thirteen baselines, achieving the best ACC@1 and MRR in every city. Ablations, backbone comparisons, and a Shanghai case study show the contributions of proposer guidance, evidence-based reasoning, conflict reports, and verification.
- Datasets: The evaluation uses Shanghai, Chongqing, Jilin, and Yantai subsets of LaDe-P, containing 10.677 million packages served by 21,000 couriers over six months.The data include locations, service time windows, event times, AOI attributes, and courier information.
- Baselines and Evaluation: ORBITER is compared with thirteen heuristic, machine-learning, deep-learning, and LLM-based baselines using ACC@1 and MRR.ACC@1 measures the committed next-order decision, while MRR measures the rank of the true order.
- Main Results: 9.2% and 4.9%: ORBITER outperforms the strongest baseline by these average margins on ACC@1 and MRR, respectively, while ranking first in every city.The results indicate gains in both next-order accuracy and the rank of the true order.
- Ablation Studies: 37.98%: removing heterogeneous proposer guidance lowers mean ACC@1, showing that proposer rankings provide primary task guidance for subsequent reasoning.Candidate states and tools alone do not reliably identify the next order from the full candidate set.
- Ablation Studies: 6.14%, 4.68%, and 1.06%: removing evidence-based reasoning, the disagreement report, and the independent critic lowers mean ACC@1 or the three metrics by these amounts.The components respectively support conflict resolution, focus on relevant proposer disagreements, and final corrective review.
- Backbone Sensitivity: DeepSeek-V4-Flash performs best across backbones, while Qwen3.5-Flash trails by 1.4% in mean ACC@1 and GPT-4o-mini and Qwen3-8B trail by 3.4% and 5.0%.ACC@3 and MRR follow the same trend under fixed proposers, tools, and reasoning budget.
- Case Study: In a Shanghai decision with 18 visible orders, ORBITER resolves competing deadline and spatial-clustering hypotheses by testing whether Order 2 can precede urgent Order 1.Order 1 has 8 min of slack, while Order 2 is in a denser same-AOI cluster; the counterfactual route check adds 12.68 m of detour and reaches Order 1 in 3.62 min.
Conclusion
ORBITER is presented as a conflict-aware agentic framework for next-order decision making in last-mile delivery. It combines decision-point representations, heterogeneous ranking, disagreement reports, evidence-based reasoning, and independent criticism.
- Framework: ORBITER makes next-order decisions through decision points that retain courier state and visible orders at each service event.Decision points are generated from spatiotemporal order logs.
- Conflict modeling: Heterogeneous proposers rank visible orders using different decision cues, while structured disagreement reports expose competing hypotheses.The disagreement report identifies conflicts among proposer rankings.
- Verification: Conflict-Aware Agentic Reasoning tests competing hypotheses with task-specific evidence, followed by review from an independent critic.The supplied passage introduces the independent critic but ends before specifying its full reviewing procedure.