LLM Papers Survey (2026-02-26)

Introduction

This article summarizes notable LLM-related papers as of 2026-02-26. Papers are automatically collected from arXiv, Semantic Scholar, and Hugging Face Daily Papers, with Japanese summaries generated using the Claude API.

1. Query-focused and Memory-aware Reranker for Long Context Processing

• Authors: Yuqing Li, Jiangnan Li, Mo Yu, Guoxuan Ding, Zheng Lin et al.
• Published: 2026-02-12
• Source: huggingface
• arXiv ID: 2602.12192

Summary

Based on analysis of retrieval heads in large language models, this paper proposes a reranking framework that trains models to estimate passage-query relevance using the attention scores of selected heads. This listwise approach leverages holistic information within the entire candidate shortlist and naturally produces continuous relevance scores, enabling training on arbitrary retrieval datasets without requiring Likert-scale supervision. The framework is lightweight and effective with small-scale models (e.g., 4B parameters), outperforming state-of-the-art pointwise and listwise rerankers across multiple domains including Wikipedia and long narrative datasets. It further establishes a new state-of-the-art on the LoCoMo benchmark for dialogue understanding and memory usage, and supports flexible extensions such as augmenting candidate passages with contextual information and training attention heads from middle layers.

2. PyVision-RL: Forging Open Agentic Vision Models via RL

• Authors: Shitian Zhao, Shaoheng Lin, Ming Li, Haoquan Zhang, Wenshuo Peng et al.
• Published: 2026-02-24
• Source: huggingface
• arXiv ID: 2602.20739

Summary

In reinforcement learning for agentic multimodal models, “interaction collapse” is a challenge where models learn to reduce tool usage and multi-turn reasoning. This paper introduces PyVision-RL, a reinforcement learning framework for open-weight multimodal models that combines an oversampling-filtering-ranking rollout strategy with accumulative tool rewards to prevent collapse and encourage multi-turn tool use. Using a unified training pipeline, the authors develop PyVision-Image and PyVision-Video for image and video understanding. For video reasoning in particular, PyVision-Video employs on-demand context construction, selectively sampling task-relevant frames during reasoning to significantly reduce visual token usage. Experiments demonstrate strong performance and improved efficiency, showing that sustained interaction and on-demand visual processing are critical for scalable multimodal agents.

3. From Perception to Action: An Interactive Benchmark for Vision Reasoning

• Authors: Yuhao Wu, Maojia Song, Yihuai Lan, Lei Wang, Zhiqiang Hu et al.
• Published: 2026-02-24
• Source: huggingface
• arXiv ID: 2602.21015

Summary

While understanding physical structure is essential for real-world applications such as embodied agents, interactive design, and long-horizon manipulation, existing Vision-Language Model (VLM) evaluations center on structure-agnostic, single-turn setups (e.g., VQA) that fail to assess agents’ ability to reason about geometry, contact, and support relations in dynamic environments. To address this gap, the authors propose the CHAIN (Causal Hierarchy of Actions and Interactions) benchmark, an interactive 3D physics-driven testbed that evaluates whether models can understand, plan, and execute structured action sequences grounded in physical constraints. CHAIN shifts evaluation from passive perception to active problem solving, spanning tasks such as interlocking mechanical puzzles and 3D stacking and packing. A comprehensive study of state-of-the-art VLMs and diffusion-based models reveals that top-performing models still struggle to internalize physical structure and causal constraints, often failing to produce reliable long-horizon plans.

4. Multi-Vector Index Compression in Any Modality

• Authors: Hanxiang Qin, Alexander Martin, Rohan Jha, Chunsheng Zuo, Reno Kriz et al.
• Published: 2026-02-24
• Source: huggingface
• arXiv ID: 2602.21202

Summary

This paper proposes efficient multi-vector index compression methods for late interaction retrieval in any modality including text, images, and video. While late interaction has emerged as a dominant paradigm for information retrieval, computation and storage costs grow linearly with document length. To address this, the paper proposes four approaches for query-agnostic multi-vector document representation compression: sequence resizing, memory tokens, hierarchical pooling, and a novel attention-guided clustering (AGC). AGC uses an attention mechanism to identify the most semantically salient regions of a document as cluster centroids and weights token aggregation. Evaluation on retrieval tasks spanning text (BEIR), visual documents (ViDoRe), and video (MSR-VTT, MultiVENT 2.0) shows that AGC consistently outperforms other parameterized compression methods and achieves competitive or improved performance compared to uncompressed indexes.

5. Learning Cross-View Object Correspondence via Cycle-Consistent Mask Prediction

• Authors: Shannan Yan, Leqi Zheng, Keyu Lv, Jingchen Ni, Hongyang Wei et al.
• Published: 2026-02-22
• Source: huggingface
• arXiv ID: 2602.18996

Summary

This paper addresses the task of establishing object-level visual correspondence across different viewpoints in videos, focusing on egocentric-to-exocentric and exocentric-to-egocentric scenarios. The proposed method is a simple yet effective framework based on conditional binary segmentation, where an object query mask is encoded into a latent representation to guide localization of the corresponding object in a target video. To learn robust, view-invariant representations, a cycle-consistency training objective is introduced: the predicted mask in the target view is projected back to the source view to reconstruct the original query mask. This bidirectional constraint provides a strong self-supervisory signal without requiring ground-truth annotations and enables test-time training (TTT) at inference. Experiments on Ego-Exo4D and HANDAL-X benchmarks demonstrate the effectiveness of the optimization objective and TTT strategy, achieving state-of-the-art performance.

This article is auto-generated. Please refer to each source URL for details about the papers.