Solving Semi-Supervised Few-Shot Learning from an Auto-Annotation Perspective

A team of researchers has introduced Stage-Wise Finetuning with Temperatures (SWIFT), a method designed to improve semi-supervised few-shot learning (SSFSL) by harnessing open-source vision-language models and addressing a core failure in existing techniques [1]. The work, posted on arXiv by Tian Liu and colleagues, targets a scenario where a model must learn from only a handful of labeled examples alongside a larger pool of unlabeled data, a setup the authors liken to real-world auto-annotation [1][3]. While the related field of few-shot learning (FSL) has already tapped powerful open-source Vision-Language Models (VLMs) and their pretraining data to boost results, the SSFSL literature has largely neglected these resources [1][2]. When the researchers applied established semi-supervised learning (SSL) methods to finetune a VLM, the models unexpectedly performed worse than FSL baselines that ignored the unlabeled data entirely [1][3]. An in-depth analysis traced the problem to the VLM's output layer. The models produced "flat" distributions of softmax probabilities, which led to zero utilization of the unlabeled examples and weak supervision signals [2][3]. The proposed fix involves using temperatures to sharpen the softmax output. This adjustment increases the confidence scores of pseudo-labels, improving the rate at which unlabeled data is used and strengthening the training signal [1][3]. The method also incorporates task-relevant open data retrieved from the VLM's publicly available pretraining set. To handle the imbalanced class distributions, noisy labels, and domain gaps present in that retrieved data, the team employed a stage-wise training strategy [2][3]. The resulting framework, SWIFT, was evaluated across five SSFSL benchmarks. It outperformed recent FSL and SSL methods by roughly 5 accuracy points [1][3]. In a striking result, SWIFT's performance rivaled that of supervised learning, a setting where the VLM is finetuned assuming the unlabeled data comes with ground-truth labels [1][2]. The paper, initially submitted in December 2025 and revised in June 2026, details three training stages where the temperature tuning consistently yields improvements [1][3]. Deep learning techniques, including the transformer architectures that underpin modern VLMs, have been applied to fields such as computer vision and natural language processing for years, often producing results that surpass human expert performance on specific benchmarks [5][6]. The SWIFT approach extends this trajectory by showing how to effectively leverage unlabeled data that previous methods could not utilize [2][3].