✂️ Segmentation

💬 ACL2025 · 4 paper notes

📌 Same area in other venues: 📷 CVPR2026 (122) · 🔬 ICLR2026 (32) · 🧪 ICML2026 (14) · 🤖 AAAI2026 (29) · 🧠 NeurIPS2025 (45) · 📹 ICCV2025 (73)

🔥 Top topics: Segmentation ×4 · Reasoning ×3 · Dialogue ×2

BERT-like Models for Slavic Morpheme Segmentation

This paper explores the use of BERT-like pretrained language models for morpheme segmentation tasks in Slavic languages. By modeling morpheme segmentation as a sequence labeling problem, the approach achieves results superior to traditional methods across multiple Slavic languages.

DEF-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation

Proposed DEF-DTS, a dialogue topic segmentation method based on LLM multi-step deductive reasoning. Through a three-step pipeline of bidirectional context summarization \(\rightarrow\) utterance intent classification (5 classes) \(\rightarrow\) deductive topic shift judgment, it achieves unsupervised/prompt-based SOTA on three datasets: TIAGE, SuperDialseg, and Dialseg711, outperforming supervised methods on Dialseg711.

InstructPart: Task-Oriented Part Segmentation with Instruction Reasoning

This work proposes InstructPart, the first real-world benchmark that combines task-oriented instructions with part-level segmentation, comprising 2,400 images, 48 object categories, 44 part categories, and 9,600 human-annotated task instructions. Evaluation shows that current VLMs are severely inadequate in instruction-driven part segmentation, while a baseline fine-tuned on LISA+DINOv2 achieves an approximate 100% performance gain.

Pixel-Level Reasoning Segmentation via Multi-turn Conversations

Proposes a new task of pixel-level reasoning segmentation (Pixel-level RS) to achieve fine-grained segmentation by progressively understanding user intent through multi-turn conversations. A PRIST dataset containing 24k dialogue turns is constructed, and a MIRAS framework is designed to outperform existing baselines in both segmentation accuracy and reasoning capability.