MazelTovy/thinking-on-the-move-nyc

Thinking on the Move — NYC Metro LoRA adapter

LoRA fine-tune of meta-llama/Llama-3.1-8B-Instruct on lunch-time restaurant choice across the NYC Metropolitan Area, scaled up from the Downtown Brooklyn pilot to ~50× the geography.

Code: https://github.com/MazelTovy/thinking-on-the-move Companion DTBK adapter: MazelTovy/thinking-on-the-move-dtbk

Headline result

On 2022 out-of-sample evaluation (500 k visits, k = 30 candidates each), this adapter posts the best Top-1 (0.7447) and best Top-5 (0.8031) of any of the nine methods we tested — beating the strongest classical ranker (LambdaRank/LightGBM, Top-1 = 0.7300) by +1.47 pp.

Method	Top-1 ↑	Top-5 ↑
Frequency prior	0.2034	0.4572
Gravity	0.2025	0.4522
Huff / MCI	0.2034	0.4572
MNL-grid	0.2034	0.4572
MNL-rich (conditional logit)	0.6495	0.7603
Deep Gravity (lite MLP)	0.7237	0.7986
LambdaRank (LightGBM)	0.7300	0.8017
XGBRanker (rank:ndcg)	0.7282	0.8025
LoRA-Llama-3.1-8B (this adapter)	0.7447	0.8031

Trade-off: aggregate ranking metrics (NDCG, Spearman) are lower because the supervision target gives the true POI prob = 0.6 (decisive top-1 prior), flattening the per-POI distribution. See the project's methodology doc for a full discussion.

Training

Base	meta-llama/Llama-3.1-8B-Instruct
Adapter	LoRA on all attention + MLP linear layers
LoRA r / α / dropout	16 / 32 / 0.05
Trainable params	~80 M (1.0 % of base)
Optimiser	AdamW
Learning rate	2e-4, cosine, 100 warmup steps
Batch	4 × grad-accum 8 = effective 32
Max length	2048 tokens
Epochs	1
Train visits	1 M visits sampled from 2021 (pre-Nov-2021 hold-out)
Validation	last-60-day temporal hold-out + 500 k 2022 OOS visits
Compute	1 × NVIDIA H200, ~36 h

Choice-set construction

Inputs at inference time are pre-built inference records rather than raw trajectories — see tom.infer_vllm for the canonical builder. Per record:

• 30 candidate POIs sampled from the worker's demographic-cluster pool
• Distance kernel: 1.5 km primary / 3 km fallback / 5 km hard cap
• Origin = worker's work-CBG centroid (not raw stop), to test destination prediction from workplace alone
• Polygon policy: exclude_shared_building (default)

How to use

For batched inference (recommended), use vLLM:

huggingface-cli download MazelTovy/thinking-on-the-move-nyc --local-dir ./adapter
python -m tom.infer_vllm \
    --records_in your_inference_records.jsonl \
    --model_path meta-llama/Llama-3.1-8B-Instruct \
    --lora_path ./adapter \
    --out predictions.jsonl \
    --temperature 0.2 --max_model_len 4096

For single prompts:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "meta-llama/Llama-3.1-8B-Instruct"
adapter = "MazelTovy/thinking-on-the-move-nyc"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Intended use & limitations

Designed for: research into urban consumption demand at metro scale, retail planning, neighborhood-substitution analysis, comparing LLM-agent behavior against classical spatial-choice models.

Not designed for: identifying or re-identifying individual mobility users, real-time personalised recommendations, surveillance, or any decision affecting a specific named worker.

Known limitations:

• Trained on the 5-borough NYC Metropolitan area. We do not test cross-city transfer (e.g., to Chicago or LA).
• Cuebiq users skew younger and more smartphone-active; cluster-level demographic re-weighting only partially corrects.
• The choice-set construction pre-filters by distance, which can mask true cross-borough substitution patterns.

Citation

@unpublished{xu2026tom,
  title  = {Thinking on the Move: An LLM-Agent Framework for Modeling
            Urban Consumption Choices from Mobility Traces},
  author = {Xu, Sizhe (Alex) and Natekar, Divya},
  year   = {2026},
  note   = {NYU CUSP capstone, P38; in collaboration with Downtown
            Brooklyn Partnership.}
}

License

MIT for the adapter weights. Base model is governed by the Meta Llama 3.1 Community License Agreement.