Unintended Misalignment from Agentic Fine-Tuning: Risks and Mitigation

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.

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.

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.

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.

The article introduces DeepEL, a new framework for Entity Linking (EL) that integrates Large Language Models (LLMs) at every stage of the EL process. This approach aims to enhance natural language understanding by improving entity disambiguation and input representation. Previous methods often applied LLMs in isolation, limiting their effectiveness. DeepEL addresses this by proposing a self-validation mechanism that leverages global context, thus aiming for greater accuracy and robustness in entity linking tasks.

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.

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.