All published AI news and analysis.

Researchers have introduced RiVER, a novel framework for training Large Language Models (LLMs) using reinforcement learning without requiring ground-truth solutions. This approach leverages score-based optimization tasks and deterministic execution feedback, addressing challenges like scale and frequency dominance in reward calibration.

Researchers have introduced 'progress advantage,' a method that leverages reinforcement learning (RL) post-training to provide step-level evaluation for LLM agents without requiring dedicated reward model training or human annotations. This technique offers a byproduct of standard RL pipelines, enabling fine-grained scoring for applications like test-time scaling, uncertainty quantification, and failure attribution.

TMA1 is a new open-source project providing local-first observability for AI agents. It records all LLM calls and routes relevant information back into the agent's operational loop through hooks and an MCP (Message Communication Protocol). This aims to enhance agent decision-making and debugging capabilities.

Researchers have introduced InSight, a framework designed to enable Vision-Language-Action (VLA) models to autonomously acquire new manipulation skills. This is achieved by making VLAs steerable at the level of primitive actions, allowing them to learn and integrate new skills without direct human demonstrations for those specific skills.

Researchers have introduced Generative Robust Optimisation (GRO), a novel framework that leverages deep generative models to define uncertainty sets for robust optimisation problems. This approach allows for the representation of complex, nonlinear dependencies in real-world data, overcoming limitations of traditional methods that rely on fixed geometric shapes.

A new research paper explores the generalization capabilities of Host Intrusion Detection Systems (HIDS) that use system-call traces. The study investigates whether HIDS trained on normal behavior associated with specific Common Vulnerabilities and Exposures (CVEs) can detect new, unseen CVEs within the same Common Weakness Enumeration (CWE) class.

Researchers have introduced NegAS, a novel framework for improving out-of-distribution (OOD) object detection using vision-language models (VLMs). NegAS addresses challenges in VLM-based detectors by employing negative label guided attention and a new sigmoid-based scoring function to enhance OOD detection performance while maintaining accuracy on in-distribution data.

Researchers have introduced Multi-Anchor Projection Similarity (MAPS), a novel metric for vision-language geo-localization (VLGL). MAPS addresses limitations of point-to-point alignment by treating visual and textual cues as a joint semantic subspace, improving retrieval accuracy.

Researchers have introduced PolicyTrim, a post-training framework designed to improve the intrinsic policy efficiency of Vision-Language-Action (VLA) models. This method addresses limitations in action chunk utilization and physical step redundancy, leading to significant speedups in real-world robotic manipulation tasks.

Researchers have introduced Gazer, a novel framework designed to improve the semantic accuracy of autoregressive visual models (AVMs). Gazer operates without additional training, integrating feedback from large language models to correct errors during the generation process. This approach addresses limitations in existing methods by diagnosing and rectifying semantic inaccuracies in intermediate generation states.