AIAny
Icon for item

YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family

Turns adapter placement for PEFT on YOLO-family real-time detectors into an auditable constraint-planning problem that emits budgeted target-module plans or calibrated refusals; shows planner-selected RS-LoRA improves mAP and cuts peak training memory in evaluated detectors.

Introduction

Real-time object detectors use heterogeneous operators and detection-specific components that break assumptions behind PEFT recipes borrowed from language models. YOLO-PEFT treats adapter placement as an auditable constraint-planning task: given a detector graph, a PEFT request and a resource budget, it assigns operator and semantic roles, evaluates operator-validity, detector-semantic, graph-interface and deployment predicates, records reason codes for excluded modules, and either emits a budgeted training plan or returns Refuse before training.

Key Findings
  • Planner-selected RS-LoRA on VOC07+12 (trainval→test) reaches mAP50-95 of 0.7138 on YOLO11s and 0.7307 on YOLO12s, compared with 0.6428 and 0.6662 for Full-SFT — concrete gains in detection accuracy under the evaluated protocol.
  • On a evaluated RT-DETR-L setup, multiple LoRA-family configurations triggered a predefined catastrophic threshold, enabling a calibrated Refuse-to-Full-SFT decision to avoid blind failure modes.
  • A controlled YOLO11 audit shows LoRA reduces peak training memory by 43.9% while increasing wall-clock training time by ~1.72×, illustrating a memory/time trade-off common to PEFT in detectors.
  • The planner records explicit reason codes for excluded modules and enforces operator/semantic predicates, replacing manual trial-and-error with inspectable, budgeted placement plans and preserving verified train-save-merge-export paths.
Who it's for and trade-offs

Great fit if you fine-tune real-time detectors under tight resource or deployment constraints and need auditable placement decisions rather than heuristic trial-and-error. It is valuable for engineering teams that require an explicit acceptance/refusal signal before spending compute on training. Look elsewhere if your detector architecture is outside the planner's validated coverage (the paper flags refusal on unseen architectures as an open validation problem) or if minimizing wall-clock training time is the top priority, since PEFT choices here increased training duration in exchange for memory savings.

Information

  • Websitearxiv.org
  • AuthorsXu Lin, WenJie Nie, Jinlong Peng, Weifu Fu, YueXiao Ma, Xiawu Zheng, Yong Liu
  • Published date2026/08/07

More Items

Uses video generation only as a training signal to co-train a pretrained video expert and a lightweight action expert, then discards the video branch at inference to produce a low-latency end-to-end driving planner; enhanced with RL for compositional driving rewards.

Analyzes how to build effective training environment distributions for multimodal agents and proposes Ability-aware Environment Selection (AES) and Hierarchical Difficulty Curriculum (HDC) to improve diversity and difficulty scheduling, yielding large relative gains in experiments.

Evaluates VLMs' ability to form global spatial awareness from long-horizon egocentric video. Introduces GST-Bench: a VQA benchmark with human-verified questions from 6,790 minutes of synthetic video, reveals a large gap (best zero-shot 42.68 vs human 79.08) and provides GST-Train dataset.