Agint: Agentic Graph Compilation for Software Engineering Agents

A new benchmark called CaptionQA has been introduced to evaluate the utility of model-generated captions in supporting downstream tasks across various domains, including Natural, Document, E-commerce, and Embodied AI. This benchmark consists of 33,027 annotated multiple-choice questions that require visual information to answer, aiming to assess whether captions can effectively replace images in multimodal systems.

Inferix has been introduced as a next-generation inference engine that utilizes a block-diffusion decoding paradigm, merging diffusion and autoregressive methods to enhance video generation capabilities. This innovation aims to create long, interactive, and high-quality videos, which are essential for applications in agentic AI, embodied AI, and gaming.

MUSE, a new framework for emotional synthesis in images, has been introduced, addressing inefficiencies in current Image Emotional Synthesis (IES) methods by integrating emotional generation and editing tasks. This approach leverages Test-Time Scaling, allowing for stable synthesis guidance without the need for additional model updates or specialized datasets.

Recent advancements in Large Language Models (LLMs) have led to the development of a multi-reward Group Relative Policy Optimization (GRPO) framework aimed at enhancing the stability and prosody of single-codebook text-to-speech (TTS) systems. This framework integrates various rule-based rewards to optimize token generation policies, addressing issues such as unstable prosody and speaker drift that have plagued existing models.

A recent study evaluates the ability of models to generalize attribute knowledge across unrelated categories, such as identifying shared attributes between dogs and chairs. This research introduces new train-test split strategies to assess the robustness of attribute prediction tasks under conditions of reduced correlation between training and test sets.

The REFLEX paradigm has been introduced as a self-refining approach to automated fact-checking, addressing the challenges of misinformation on social media by leveraging internal knowledge from large language models (LLMs) to enhance both accuracy and explanation quality. This innovative method reformulates fact-checking into a role-play dialogue, allowing for joint training of verdict prediction and explanation generation.

A recent study examined how AI-mediated communication influences group dynamics in discussions on controversial political topics. In an online experiment with 557 participants, it was found that those receiving personalized AI assistance tended to cluster based on their stances, while those with relational assistance formed more diverse connections. This indicates that AI can significantly affect group composition and interaction patterns.

A systematic study has been conducted on cross-LLM behavioral backdoor detection, revealing significant vulnerabilities in AI agent supply chains. The research evaluated six production LLMs, including GPT-5.1 and Claude Sonnet 4.5, highlighting a stark generalization gap in detection accuracy across different models.