Open Responses is a new open inference standard developed by the AI community and Hugging Face, designed to replace outdated chatbot formats with a more flexible system for autonomous agents. It addresses limitations of the Chat Completion format, which is still widely used despite being ill-suited for modern agent-based workflows.

This study compares two training paradigms for multi-domain reinforcement learning in large language models—task mixing and model merging—finding minimal interference between domains and synergistic effects in reasoning-intensive tasks. The research uses open-source datasets to analyze performance across math, coding, science, and instruction-following domains.

SciAgentGym is a new benchmark for evaluating LLM agents’ ability to perform multi-step scientific tasks using domain-specific tools, revealing that even advanced models like GPT-5 struggle with complex workflows. The study introduces SciAgentBench and SciForge to address these challenges in tool orchestration and training data synthesis.

Google’s new AI, Gemini 3, has outperformed competitors in reasoning and code generation, as highlighted by recent benchmarks, marking a significant response to the release of GPT-5. The article emphasizes the rapid advancements in AI capabilities and the competitive landscape in the field.

The article explores the origin of the name ‘Nano Banana,’ likely tied to Google’s AI initiatives, though specific details are sparse. It highlights Google’s focus on innovation, AI research, and product development through platforms like Gemini models and Google Cloud.

The article highlights the advancements in large language models (LLMs) in 2026, emphasizing their integration into real-world applications like AI chatbots, automation tools, and enterprise solutions. It also discusses developments such as reasoning models and multimodal models that handle diverse input/output formats.

The article highlights that soft contamination in LLM training data—via semantic duplicates of benchmark test data—leads to inflated benchmark performance metrics, as these duplicates improve model performance on benchmarks but may not reflect true out-of-distribution generalization. Experiments show widespread contamination and suggest recent benchmark gains may be confounded by both genuine improvements and data leakage.

ReFilter is a novel latent-based fusion framework that enhances the robustness of Retrieval-Augmented Generation (RAG) by filtering and fusing tokens at the token level, improving performance on both general and biomedical QA benchmarks. It addresses scalability issues in existing methods by reducing irrelevant content and inference costs through a gated filter and context encoder.

The article introduces a scalable pipeline for generating high-quality training data for web agents using automatic data generation and a constraint-based evaluation framework. The method leverages partially successful trajectories and is evaluated on a new benchmark, BookingArena, where it outperforms open-source approaches and matches commercial systems. The work addresses challenges in creating diverse web interaction datasets and evaluating complex tasks.

DiffuRank introduces a document reranking framework using diffusion language models (dLLMs) to overcome inefficiencies and inflexibility of autoregressive models. It proposes three strategies leveraging dLLMs’ parallel decoding capabilities for improved efficiency and controllability, validated on multiple benchmarks.