Harnessing Deep LLM Participation for Robust Entity Linking

OptScale introduces a probabilistic framework for inference-time scaling in Large Language Models (LLMs), addressing the limitations of heuristic strategies in parallel sampling. This framework formalizes optimality under the assumption of independent and identically distributed samples, allowing for a theoretical lower bound on the number of samples needed to achieve desired performance levels. The practical algorithm developed, OptScale, dynamically determines the optimal number of sampled responses, enhancing compute efficiency.

The introduction of ATLAS (AGI-Oriented Testbed for Logical Application in Science) marks a significant advancement in evaluating Large Language Models (LLMs). This new benchmark addresses the limitations of existing high-difficulty assessments, which often lack interdisciplinary focus and are prone to data contamination. Comprising around 800 original problems across seven scientific fields, ATLAS aims to enhance the fidelity of evaluations in real-world scientific reasoning.

The integration of Large Language Models (LLMs) with 3D vision is revolutionizing robotic perception and autonomy. This approach enhances robotic sensing technologies, allowing machines to understand and interact with complex environments using natural language and spatial awareness. The review discusses the foundational principles of LLMs and 3D data, examines critical 3D sensing technologies, and highlights advancements in scene understanding, text-to-3D generation, and embodied agents, while addressing the challenges faced in this evolving field.

LoopTool is a new framework designed to enhance the training of Large Language Models (LLMs) by integrating data synthesis and model training into a cohesive process. This approach addresses the limitations of traditional static data pipelines, which often fail to adapt to a model's weaknesses and allow for noisy labels that hinder training efficiency. LoopTool employs three modules: Greedy Capability Probing for diagnosing model capabilities, Judgement-Guided Label Verification for correcting annotation errors, and Error-Driven Data Evolution for refining datasets.

Supervised Fine-Tuning (SFT) is essential for adapting Large Language Models (LLMs) to specialized fields like medical reasoning. Current SFT methods often utilize unfiltered datasets, which can be redundant and of low quality, leading to high computational costs and poor performance. This study introduces a new data selection strategy called Difficulty-Influence Quadrant (DIQ), which aims to optimize sample selection based on both difficulty and optimization utility, enhancing the efficiency of medical reasoning applications.

Large Language Models (LLMs) are being enhanced for autonomous driving with the introduction of TLS-Assist, a modular layer that improves traffic light and sign recognition. This innovation addresses the current limitations of LLM-based driving agents, which often struggle to detect critical safety objects. TLS-Assist translates detections into structured natural language messages, ensuring that safety cues are prioritized. The framework is adaptable to various camera setups and has been evaluated in a closed-loop environment using the LangAuto benchmark in CARLA.

This study analyzes the transformative role of Large Language Models (LLMs) in research and development (R&D) processes. By automating knowledge discovery, enhancing hypothesis generation, and fostering collaboration within innovation ecosystems, LLMs significantly improve research efficiency and effectiveness. The research highlights how LLMs facilitate more adaptable and informed R&D workflows, ultimately accelerating innovation cycles and reducing time-to-market for groundbreaking ideas.

ConInstruct is a newly introduced benchmark aimed at evaluating the conflict detection and resolution capabilities of Large Language Models (LLMs). While previous studies have focused on how well LLMs follow user instructions, they often neglect scenarios with conflicting constraints. The benchmark assesses LLMs' performance in detecting and resolving such conflicts, revealing that proprietary models generally perform well, with DeepSeek-R1 and Claude-4.5-Sonnet achieving the highest F1-scores.