CVPR 2026 Workshop (MEIS) · Denver, Colorado

Fine-Tuning Large Language Models for Cooperative Tactical Deconfliction of Small Unmanned Aerial Systems

Simulation-to-language data and parameter-efficient fine-tuning for multi-agent drone traffic

Iman Sharifi, Alex Zongo, Peng Wei

Intelligent Aerospace Systems Lab · The George Washington University

Paper Preprint Poster Cite

Can a language model act as a deconfliction policy?

Pre-trained LLMs reason well but are not grounded in airspace rules, and they answer inconsistently. We generate rule-consistent training data from simulation, fine-tune a 7B model with supervised LoRA and with GRPO, and test the result in closed-loop multi-agent simulation.

Simulation-to-language fine-tuning pipelineBlueSky scenarios are labeled by human-aligned rule-based deconfliction, converted into text prompts with target actions, and used to fine-tune a language model with supervised LoRA or GRPO. The fine-tuned model is then evaluated in closed-loop simulation alongside rule-based agents. 1 · BlueSky scenarios 20–30 sUAS · Frisco, TX 2 · human-aligned labels rule-based deconfliction ~38k samples 3 · prompt and target ownship: hdg 090, 40 kt intruder: 0.2 nm, right reason step by step… target: slow down 4 · fine-tune Qwen-Math-7B SFTLoRA GRPOLoRA + preference 5 · closed-loop evaluation LLM + rule-basedagents in BlueSky Metrics: decision accuracy on held-out data, near mid-air collisions and operation time in simulation.
Schematic of the simulation-to-language pipeline: labeled BlueSky scenarios become prompt and target pairs, which fine-tune the model before closed-loop evaluation.
decision accuracy, base → SFT
27% → 88%
NMACs in scenario A, base → SFT
3.5 → 1.0
rule-consistent training samples
~38k

Held-out test set and BlueSky closed-loop simulations reported in the workshop poster. Simulation results, not a safety guarantee.

01 / Research questions

Zero-shot is not enough for safety-critical airspace.

Dense low-altitude drone traffic needs short-horizon, cooperative decisions under partial observability and heterogeneous vehicles. We ask three questions: can pre-trained LLMs handle cooperative tactical deconfliction, do they need fine-tuning and how can they be fine-tuned without real-world data, and how do they handle multiple agents in a centralized manner?

Base model behavior

The pre-trained model acts conservatively, congests the airspace, and occasionally proposes unsafe actions.

Latency

At least 3 s per agent for variable-length answers; forcing short answers makes them unreasonable.

Our approach

Fine-tune on rule-consistent data that encodes human tactical reasoning, instead of prompting zero-shot.

02 / Simulation-to-language

Turn simulated encounters into supervised reasoning data.

Scenarios

BlueSky multi-agent simulation of 15–25 minute delivery operations over Frisco, Texas, with 20–30 sUAS per scenario.

Vehicles

Heterogeneous drone configurations inspired by Google Wing and Amazon Prime Air aircraft.

Labels

Human-aligned rule-based deconfliction provides about 38k labeled samples, with prompts that ask for structured reasoning.

03 / Two strategies

Supervised LoRA vs. preference-based GRPO.

A pre-trained Qwen-Math-7B model is fine-tuned with supervised fine-tuning using low-rank adaptation (LoRA), and with LoRA combined with Group-Relative Policy Optimization (GRPO).

Decision quality on the held-out test set (%)
ModelAccuracyPrecisionRecallF1
Base27752031
SFT (LoRA)88756669
GRPO53754050

From the MEIS@CVPR 2026 poster.

04 / Closed-loop simulation

Fewer near mid-air collisions with supervised fine-tuning.

In BlueSky with mixed LLM and rule-based agents, supervised LoRA fine-tuning substantially reduces near mid-air collisions compared with the base model. GRPO adds coordination benefits but is less robust when interacting with heterogeneous agent policies.

Near mid-air collisions, all interactions (mean ± std)
ScenarioBaseSFTGRPO
A3.5 ± 1.11.0 ± 0.81.7 ± 0.5
B3.4 ± 1.11.9 ± 1.23.0 ± 0.7
C4.0 ± 0.91.9 ± 0.72.3 ± 0.8

Ten LLM agents with rule-based agents; per-interaction breakdowns and operation times are in the paper.

05 / Limits and next steps

Promising, not yet deployable.

Inference latency in centralized multi-agent settings, the compute and memory cost of GRPO, and the absence of safety guarantees remain open. Next steps include data augmentation with real-world traffic, a safety module around the model, and comparisons with rule-based and reinforcement learning methods. Related: LLM-guided RL with fallback control.

Reference

Cite this work

Sharifi, I., Zongo, A., & Wei, P. (2026). " Fine-Tuning Large Language Models for Cooperative Tactical Deconfliction of Small Unmanned Aerial Systems. " Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (MEIS) (pp. 1067-1076).

BibTeX 
@inproceedings{sharifi2026finetuning,
  title         = {Fine-Tuning Large Language Models for Cooperative Tactical
                   Deconfliction of Small Unmanned Aerial Systems},
  author        = {Iman Sharifi and Alex Zongo and Peng Wei},
  booktitle     = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages.        = {1067--1076},
  year          = {2026},
  note          = {MEIS @ CVPR 2026 Workshop}
}