Rui Pan, Dylan Zhang, Hanning Zhang, Xingyuan Pan, Minrui Xu, Jipeng Zhang, Renjie Pi, Xiaoyu Wang, Tong Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

High-quality supervised finetuning (SFT) data are essential for unlocking pretrained LLMs’ capabilities. Typically, instructions are paired with responses from various sources—by human annotators or other LMs—which are often out of the distribution of the target model to be finetuned. At scale, this mismatch can lead to diminishing returns and even hurt model performance and robustness. We hypothesize that SFT is most effective when the data is aligned with the model’s pretrained distribution, and propose **GRAPE**—a novel SFT framework that tailors supervision to the target model. For each instruction, it **g**athers **r**esponses from various sources and selects the one that **a**ligns most closely to the model’s **pre**trained distribution, as measured by the normalized probability. Standard SFT is then performed on these selected responses. We first evaluate GRAPE in a controlled experiment, sampling multiple responses per question in the UltraInteract dataset from diverse models. We finetune using GRAPE-selected data on LMs from different families, including LLaMA-1-8B, Mistral-7B, and Qwen2.5-7B. GRAPE significantly outperforms strong baselines—including distilling from the strongest model—with absolute gains up to **13.8%** averaged across benchmarks, and outperforms a 3× larger data baseline with improvements up to **17.3%**. GRAPE's benefits generalize to off-the-shelf SFT data. When used to subsample from the post-training data of Tulu3 and Olmo-2, GRAPE surpasses strong baselines trained on 4.5× more data by **6.1%**, and outperforms state-of-the-art selection methods by **3.9%** on average. Notably, with only **1/3 the data** and **half the training epochs**, GRAPE enables LLaMA-1-8B to **exceed Tulu3-SFT performance by 3.5%**. Our findings highlight that aligning supervision with the pretrained distribution provides a simple yet powerful strategy to improve both the **efficiency** and **effectiveness** of SFT.

Vision Large Language Models (VLLMs) are transforming the intersection of computer vision and natural language processing.Nonetheless, the potential of using visual prompts for emotion recognition in these models remains largely unexplored and untapped.Traditional methods in VLLMs struggle with spatial localization and often discard valuable global context.To address this problem, we propose a Set-of-Vision prompting (SoV) approach that enhances zero-shot emotion recognition by using spatial information, such as bounding boxes and facial landmarks, to mark targets precisely.SoV improves accuracy in face count and emotion categorization while preserving the enriched image context.Through a battery of experimentation and analysis of recent commercial or open-source VLLMs, we evaluate the SoV model's ability to comprehend facial expressions in natural environments.Our findings demonstrate the effectiveness of integrating spatial visual prompts into VLLMs for improving emotion recognition performance. Question:How many visible faces are there in the image?Analyze each face's emotion and group the face's emotions. Answer:In the image, there are 22 visible faces.Emotions can be grouped into three main categories. Neutral Emotion: Approximately 16Faces.This neutral expression is typical in settings requiring attention to a speaker or event, reflecting an engaged yet reserved emotional state. Mildly Positive Emotion: Approximately 5Faces.This could be a response to something interesting or mildly amusing happening at the event, but their reactions are subdued, in keeping with the formal setting.Happy: 1 Face.This expression stands out as more joyful or enthusiastic about the proceedings.