Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

MobTA: Bus-Conditioned Zero-Shot Trajectory Generation via Task Arithmetic

NeurIPS 2026 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Robust fine-tuning of zero-shot models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MobTA, which transfers a mobility-trajectory generation shift from a source city to a target city via task arithmetic on LoRA/embedding/head task vectors, so target-city mobility data are never needed (only bus timetables).

Problem

Generating human mobility trajectories for a target city when no real mobility trajectories from that city are accessible, using only source-city trajectories and public bus timetables from both cities.

Approach

Builds three task vectors (source bus, source mobility, target bus) by fine-tuning an LLM on serialized trajectories with grid-based location tokens, then adds the source-city bus-to-mobility shift to the target-city bus task vector; also gives a first-order projection argument for stability across base and instruction-tuned LLMs.

Key results

• On WX->SH, MobTA (WX->) reaches distance JSD 0.0969 and location JSD 0.0487 versus COLA 0.1831 and 0.1544 with the same source-target pairing.
• MobTA (SG->SH) achieves the best SH distance JSD of 0.0749, approaching Geollama+ fine-tuned on real SH mobility at 0.0515.
• Base and instruction-tuned LLaMA-3.2-3B and Qwen3-1.7B variants give nearly identical WX->SH results (e.g., distance JSD 0.0926 to 0.0969).

Surprising findings

• Base and instruction-tuned LLMs produce nearly identical task-arithmetic results, consistent with the stated theoretical argument.
• Bus timetables alone differ from mobility trajectories in distance and radius distributions, so bus-only training is insufficient.

Limitations

• Relies on the availability of public bus timetables in the target city, so it does not apply to cities without public bus systems.
• Mixing coefficients mu_E, mu_L, mu_H are set by hand and their relation to city discrepancy is left open.
• The stability theory rests on assumptions (small projection of instruction-tuning difference onto the trajectory subspace) that are not directly verified.

Experimental setup

Datasets: Shanghai mobility trajectories Wuxi mobility trajectories Singapore mobility trajectories Moovit bus timetables Shanghai 2015 bus routes

Benchmarks: Shanghai (SH) Wuxi (WX) Singapore (SG)

Metrics: JSD on distance JSD on radius JSD on duration JSD on location stay duration JSD reachability visit frequency

Baselines: TimeGeo MoveSim TSG TrajGDM DiffTraj TrajGen Traveller COLA Geollama (bus-only)

Builds on: task arithmetic Geollama

Closest papers by meaning

• GTG: Generalizable Trajectory Generation Model for Urban Mobility (AAAI 2025 · similarity 0.84)
• TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation (NeurIPS 2026 · similarity 0.83)
• BO4Mob: Bayesian Optimization Benchmarks for High-Dimensional Urban Mobility Problem (NeurIPS 2025 · similarity 0.82)
• SMART: Scalable Multi-agent Real-time Motion Generation via Next-token Prediction (NeurIPS 2024 · similarity 0.82)
• PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models (NeurIPS 2025 · similarity 0.82)
• SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark (NeurIPS 2026 · similarity 0.82)