Loading
Life sciences · Preprint
arXiv · September 10, 2026
Early or partial results. Treat as a signal, not a conclusion.
This preprint proposes soft prompting—optimizing a small number of continuous prompt tokens while freezing pretrained vision-language model backbones—for few-shot object detection in out-of-domain settings. Using only 7,168 parameters on average, soft prompting matches LoRA baselines on Roboflow20-VL (14.2 mAP) while avoiding catastrophic forgetting and transferring across model versions, but exhibits higher optimization variance and has not undergone peer review.
Comparative methods study. Out-of-domain object detection tasks including aerial, industrial, and medical imagery; robotic manipulation (RoboCasa) with frozen vision-language-action policies.. Intervention: Soft prompting: optimization of 1–3 continuous prompt tokens placed at cross-modal boundary, initialized from empty space token, while backbone remains frozen (7,168 parameters on average).. Compared with: LoRA fine-tuning; discrete prompt optimization; other existing adaptation methods (DetPO, GEPA, prompt-search)..
Soft prompting with tokens placed at cross-modal boundary outperforms other placements (10.0 vs. 8.4 mAP) One to three learned tokens (7,168 parameters on average) match best LoRA configuration on Roboflow20-VL (14.2 mAP, 10-shot) while training 20,000x fewer parameters LoRA rank matching soft prompt accuracy reduced NaturalBench VQA accuracy by 35% relative; soft prompting left pretrained performance unchanged
Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.
This is a machine learning methods preprint with no direct clinical application. Professionals in computer vision and AI may find the parameter-efficient adaptation approach and absence of catastrophic forgetting relevant to model deployment, but the work requires peer review and validation in applied domains.
This is an unreviewed preprint presenting a methodological proposal (soft prompting for vision-language models) evaluated on multiple computer vision benchmarks with small sample sizes (10-shot) and no peer review; the work is technically sound but lacks the validation and clinical/practical significance needed for stronger evidence grades.
As stated by the source record.
Quoted from the source exactly as published.
This is a machine learning methods preprint with no direct clinical application. Professionals in computer vision and AI may find the parameter-efficient adaptation approach and absence of catastrophic forgetting relevant to model deployment, but the work requires peer review and validation in applied domains.
Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.
We address few-shot object detection with vision-language models (VLMs) in out-of-domain settings such as aerial, industrial, and medical imagery, using only ten annotated images for supervision. Existing adaptation methods are discrete prompt optimization and LoRA fine-tuning. We revisit a third option: soft prompting, where a small number of continuous prompt tokens are optimized while the pretrained backbone remains frozen. We identify two key design choices. First, placing prompt tokens at the cross-modal boundary between visual and text tokens outperforms other placements (10.0 vs. 8.4 mAP). Second, initializing prompts from the empty space token outperforms semantic and random initialization. With these choices, one to three learned tokens (7,168 parameters on average) match the best LoRA configuration on Roboflow20-VL (14.2 mAP, 10-shot) while training over 20,000x fewer parameters. Soft prompting remains harder to optimize, exhibiting higher variance across random seeds. Unlike LoRA, however, it causes no forgetting: the LoRA rank matching our accuracy reduces NaturalBench VQA accuracy by 35% relative, rising to 56% at the largest rank, whereas soft prompting leaves pretrained performance unchanged. The learned tokens behave like prompts rather than weights. They transfer to a newer model without retraining (+0.8 mAP on Qwen3.5-9B) and can be verbalized into readable prompts competitive with prompt-search methods (matching DetPO and outperforming GEPA). The approach also extends beyond detection. On RoboCasa manipulation tasks, the frozen $π_{0.5}$ vision-language-action policy benefits from soft prompting, matching the LoRA baseline on two of three tasks when tokens are placed at the gradient bottleneck. These results suggest modern VLMs already encode much of what is needed for specialized domains; the challenge is learning how to ask.
Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.