Independent AI signal, once a day
Anthropic has published a new article on its engineering blog, detailing how Claude's R&D capabilities are built through the parallel collaboration of multiple agents. The article shares the successes, challenges, and engineering difficulties encountered during the development of this multi-agent research system, providing valuable insights for AI system design and optimization.
The tweet explores the possibility of Large Language Models (LLMs) achieving self-editing through Reinforcement Learning (RL). Jyo Pari introduces a framework called SEAL, which allows an LLM to generate its own training data (self-editing) and update its weights based on new inputs. This self-editing capability is learned via RL, using the model's downstream performance as a reward signal, heralding a new paradigm for LLM self-improvement.
TuringPost introduces Reinforced Meta-thinking Agents (ReMA), a framework combining meta-learning and reinforcement learning to significantly enhance Large Language Models' (LLM) effectiveness, especially in multi-agent collaboration. ReMA divides problem-solving into meta-thinking (planning) and reasoning (execution). It utilizes a Multi-Agent Meta-thinking Reasoning Process (MAMRP) with high-level agents for meta-planning and low-level agents for reasoning, allowing dynamic plan adjustments. ReMA demonstrates notable performance improvements on both math and LLM-as-a-Judge benchmarks.
Jerry Liu announced the launch of 'Use-Case Presets' for LlamaParse, a series of specialized parsing agents that can render different document types (like forms, technical documents, invoices) into predefined formats. For example, forms can be output as structured 2D tables, technical documents as XML, and invoices as line items. Additionally, LlamaParse has updated its documentation page to detail the general presets (fast, balanced, advanced) and the new use-case presets.
Prominent technologist Vipul Ved Prakash boldly predicts that within the next 12 months, the era of hand-written code will come to an end. This statement suggests the rapid advancement of artificial intelligence in code generation, indicating that AI-assisted programming and fully automated code generation are poised to become mainstream. This shift is expected to fundamentally transform software development, prompting significant industry discussion and reflection on future programming paradigms.
Google DeepMind announced the debut of "ANCESTRA," the first film from its partnership with Darren Aronofsky's Primordial Soup, at the Tribeca Film Festival. Directed by Eliza McNitt, the film integrates traditional filmmaking with DeepMind's generative video model, Veo, showcasing the potential of AI in cinematic art creation.
This project presents a robust proof-of-concept for an AI-powered hedge fund, meticulously designed to explore the advanced application of artificial intelligence in sophisticated trading decision-making. The system boasts a highly collaborative multi-agent architecture, integrating specialized AI agents focused on critical financial aspects such as disciplined valuation, market sentiment analysis, in-depth fundamental data interpretation, precise technical indicator analysis, and comprehensive risk management. Furthermore, it simulates the distinct investment philosophies of twelve renowned financial masters, offering diverse strategic perspectives. The project provides flexible deployment options via Poetry and Docker, features seamless integration with various Large Language Models (LLMs) for enhanced analytical capabilities, and includes robust backtesting functionalities for performance evaluation. Primarily intended for educational and research purposes, this sophisticated system rigorously simulates trading decisions without engaging in any actual financial transactions, emphasizing its role as a learning and development tool.
DeepEval is a robust, open-source LLM evaluation framework, akin to Pytest but specialized for large language model outputs. It incorporates the latest research to assess LLM performance across various metrics, including G-Eval, RAGAS, hallucination, answer relevancy, and conversational metrics, with evaluations running locally on your machine. This framework is crucial for validating RAG pipelines, chatbots, and AI agents, enabling developers to determine optimal models and prompts, prevent prompt drifting, and ensure application reliability. DeepEval also supports building custom metrics, generating synthetic datasets, seamless CI/CD integration, and red-teaming for over 40 safety vulnerabilities. Furthermore, it allows easy benchmarking of any LLM on popular benchmarks like MMLU and HumanEval, providing a comprehensive solution for LLM quality assurance and performance optimization throughout the development lifecycle.
This GitHub repository offers a practical, hands-on guide to Retrieval-Augmented Generation (RAG) techniques. It meticulously breaks down over 20 advanced RAG methods, such as semantic chunking, reranking, Graph RAG, Multimodal RAG, and Corrective RAG, all implemented from scratch using pure Python libraries like OpenAI and NumPy, without relying on frameworks like LangChain. The project aims to provide readable, modifiable, and educational code, enabling developers to grasp the fundamental workings of RAG. It also covers handling large datasets with knowledge graphs, serving as a valuable resource for building efficient and interpretable RAG systems.
This project serves as a dedicated open-source large language model tutorial for beginners in China, primarily focusing on Linux platforms. It provides comprehensive, full-process guidance covering environment setup, local deployment, and efficient fine-tuning for a wide range of open-source LLMs. The initiative aims to simplify the complex procedures of deploying, using, and applying these models, including mainstream ones like LLaMA, ChatGLM, and InternLM. Technical aspects covered include command-line invocation, online demo deployment, and integration with frameworks like LangChain. The tutorial's objective is to enable more students and researchers to effectively leverage open-source large models, fostering their adoption in academic and practical settings, and facilitating the development of customized, domain-specific LLMs.
The YouTube Transcript API is a Python library designed for efficient retrieval of YouTube video transcripts and subtitles, including automatically generated ones. This API eliminates the need for headless browsers, supports subtitle translation, and offers flexible API interfaces and command-line tools. It addresses IP blocking issues by supporting proxy configurations to ensure service stability and provides various output formats. The project offers a convenient solution for video content analysis, multilingual processing, and automated data extraction.
This GitHub repository presents "AI Agents for Beginners - A Course," comprising 11 lessons designed to teach the fundamentals of building AI agents. The curriculum covers AI agent concepts, exploration of agentic frameworks, various agentic design patterns (such as tool use, RAG, planning, multi-agent, and metacognition), and deploying AI agents in production environments. The course provides Python code samples compatible with Azure AI Foundry and GitHub Model Catalogs, leveraging Microsoft's Azure AI Agent Service, Semantic Kernel, and AutoGen frameworks. With multi-language support, this course is ideal for beginners seeking a systematic approach to AI agent development.
Milvus 2.6 has been officially open-sourced, marking a significant milestone in its 2025 product roadmap, aiming to deliver a more efficient, powerful, and economical vector database solution. The new version introduces key improvements in cost reduction and efficiency, notably through the RaBitQ 1-bit quantization technology, achieving a substantial 72% memory reduction and significant QPS improvements. Regarding search capabilities, Milvus 2.6 enhances Analyzer/Tokenizer functionalities, adds Phrase Match, and introduces Decay Function for time-decay re-ranking, thereby improving hybrid search accuracy and timeliness. Architecturally, the update includes tiered storage for hot and cold data, a Streaming Service for real-time data processing, and support for 100k collections, significantly boosting system stability, scalability, and overall performance. Milvus 2.6 is dedicated to addressing the efficiency and stability challenges of data retrieval in the AI era, providing a robust foundation for large-scale AI applications.
At its latest press conference, iFlytek unveiled a new paradigm for intelligent interaction, signaling a shift from basic command-based human-computer interaction to a deep intelligent collaboration era. Through AIUI platform upgrades and the empowerment of the Spark Interaction Large Model, iFlytek has achieved full-duplex human-like dialogue, emotional perception and empathy, human-like memory systems, and rapid response times, significantly enhancing user experience, particularly with a surge in interaction frequency in children's scenarios. Furthermore, iFlytek launched its Robot Superbrain Platform, integrating the Spark Large Model to enable multimodal perception and intelligent action, and introduced a smart voice backpack to empower existing robots. This represents more than just product feature upgrades; it's a comprehensive re-shaping of the hardware industry ecosystem with AI interaction at its core, transforming human-machine relationships from "command-execution" to "collaborative partnership," thereby accelerating the creation of a smarter intelligent world.
Developed by researchers from UC Santa Cruz, UC Berkeley, Cisco Research, and Yale University, the SafeKey framework addresses the limited generalization of Large Reasoning Models (LRMs) against “jailbreak” attacks. By identifying the “Key Sentence” phenomenon and “Dormant Safety Signals,” SafeKey introduces two innovative mechanisms: the “Dual-Path Safety Head” to amplify internal safety signals before key sentence generation, and “Query-Mask Modeling” to compel models to rely on their self-generated understanding for safe decisions. This framework aims to reinforce the “safety aha-moment” during key sentence generation, significantly enhancing LRM safety and robustness without impacting core capabilities. Experimental results demonstrate SafeKey's effectiveness in reducing the risk rate by 9.6%, offering an efficient and generalizable solution for secure LRM deployment.
A collaborative research team, including members from Southeast University, Max Planck Institute for Informatics, and UC Berkeley, has introduced KRIS-Bench, a groundbreaking benchmark designed to systematically evaluate the reasoning capabilities of image editing models. Uniquely, KRIS-Bench assesses models from a knowledge type perspective, categorizing challenges into factual, conceptual, and procedural knowledge. It features 22 diverse editing tasks, ranging from basic to advanced, with output scoring facilitated by multimodal large models and human calibration. Crucially, the evaluations conducted using KRIS-Bench indicate that existing image editing models, both closed and open-source, exhibit significant limitations, particularly in "procedural reasoning." This research endeavors to propel image editing models beyond simple pixel manipulation, fostering their evolution into "visual sages" capable of human-like cognitive understanding, thereby enabling them to grasp the underlying physical, chemical, and causal principles inherent in image editing tasks.
DeepSeek engineer Yu Xingkai has recently open-sourced Nano-vLLM, a lightweight large language model (LLM) inference engine, demonstrating a significant breakthrough in efficiency. This project remarkably achieves throughput performance comparable to the high-performance vLLM, all while being implemented with just 1200 lines of Python code. Nano-vLLM is designed to offer an easily readable and highly efficient alternative for LLM deployment, boasting core features like rapid offline inference, a remarkably concise codebase, and advanced optimization functionalities including Prefix caching, Torch compilation, and CUDA graphs. Benchmark tests, conducted on an RTX 4070 using the Qwen3-0.6B model, clearly indicate that Nano-vLLM's performance is nearly on par with the original vLLM. This innovative development provides a promising new approach and a valuable tool for optimizing LLM inference, underscoring the potential for streamlined and highly performant deep learning systems.
The Unstructured Data Meetup Japan edition, initiated by leading vector database provider Zilliz, is scheduled for July 5th at the AWS office in Tokyo. This technical sharing event focuses on unstructured data and Generative AI, targeting developers, AI engineers, data scientists, and AI product managers. The meetup aims to foster collaboration and explore cutting-edge advancements in vector databases, large language models, AI applications, and big data. Key presentations include Zilliz's introduction to Milvus and Zilliz Cloud, SmartNews's insights into Milvus applications for search, recommendation, and content understanding, and Mercari's deep dive into AI and LLM integration within their retrieval systems, covering offline data labeling, query expansion, and embedding technologies. This gathering will facilitate the exchange of the latest developments in unstructured data and generative AI fields.
Though reasoning-based large language models (LLMs) have excelled in mathematics and programming, their capabilities in knowledge-intensive medical question answering remain underexplored. To address this, we introduce ReasonMed, the largest medical reasoning dataset, comprising 370k high-quality examples distilled from 1.7 million initial reasoning paths generated by various LLMs. ReasonMed is constructed through a multi-agent verification and refinement process, where we design an Error Refiner to enhance the reasoning paths by identifying and correcting error-prone steps flagged by a verifier. Leveraging ReasonMed, we systematically investigate best practices for training medical reasoning models and find that combining detailed Chain-of-Thought (CoT) reasoning with concise answer summaries yields the most effective fine-tuning strategy. Based on this strategy, we train ReasonMed-7B, which sets a new benchmark for sub-10B models, outperforming the prior best by 4.17\% and even exceeding LLaMA3.1-70B on PubMedQA by 4.60\%.
We introduce Magistral, Mistral's first reasoning model and our own scalable reinforcement learning (RL) pipeline. Instead of relying on existing implementations and RL traces distilled from prior models, we follow a ground up approach, relying solely on our own models and infrastructure. Notably, we demonstrate a stack that enabled us to explore the limits of pure RL training of LLMs, present a simple method to force the reasoning language of the model, and show that RL on text data alone maintains most of the initial checkpoint's capabilities. We find that RL on text maintains or improves multimodal understanding, instruction following and function calling. We present Magistral Medium, trained for reasoning on top of Mistral Medium 3 with RL alone, and we open-source Magistral Small (Apache 2.0) which further includes cold-start data from Magistral Medium.
Despite rapid advancements in video generation models, generating coherent storytelling videos that span multiple scenes and characters remains challenging. Current methods often rigidly convert pre-generated keyframes into fixed-length clips, resulting in disjointed narratives and pacing issues. Furthermore, the inherent instability of video generation models means that even a single low-quality clip can significantly degrade the entire output animation's logical coherence and visual continuity. To overcome these obstacles, we introduce AniMaker, a multi-agent framework enabling efficient multi-candidate clip generation and storytelling-aware clip selection, thus creating globally consistent and story-coherent animation solely from text input. The framework is structured around specialized agents, including the Director Agent for storyboard generation, the Photography Agent for video clip generation, the Reviewer Agent for evaluation, and the Post-Production Agent for editing and voiceover. Central to AniMaker's approach are two key technical components: MCTS-Gen in Photography Agent, an efficient Monte Carlo Tree Search (MCTS)-inspired strategy that intelligently navigates the candidate space to generate high-potential clips while optimizing resource usage; and AniEval in Reviewer Agent, the first framework specifically designed for multi-shot animation evaluation, which assesses critical aspects such as story-level consistency, action completion, and animation-specific features by considering each clip in the context of its preceding and succeeding clips. Experiments demonstrate that AniMaker achieves superior quality as measured by popular metrics including VBench and our proposed AniEval framework, while significantly improving the efficiency of multi-candidate generation, pushing AI-generated storytelling animation closer to production standards.
We present VRBench, the first long narrative video benchmark crafted for evaluating large models' multi-step reasoning capabilities, addressing limitations in existing evaluations that overlook temporal reasoning and procedural validity. It comprises 1,010 long videos (with an average duration of 1.6 hours), along with 9,468 human-labeled multi-step question-answering pairs and 30,292 reasoning steps with timestamps. These videos are curated via a multi-stage filtering process including expert inter-rater reviewing to prioritize plot coherence. We develop a human-AI collaborative framework that generates coherent reasoning chains, each requiring multiple temporally grounded steps, spanning seven types (e.g., event attribution, implicit inference). VRBench designs a multi-phase evaluation pipeline that assesses models at both the outcome and process levels. Apart from the MCQs for the final results, we propose a progress-level LLM-guided scoring metric to evaluate the quality of the reasoning chain from multiple dimensions comprehensively. Through extensive evaluations of 12 LLMs and 16 VLMs on VRBench, we undertake a thorough analysis and provide valuable insights that advance the field of multi-step reasoning.
We propose Ming-Omni, a unified multimodal model capable of processing images, text, audio, and video, while demonstrating strong proficiency in both speech and image generation. Ming-Omni employs dedicated encoders to extract tokens from different modalities, which are then processed by Ling, an MoE architecture equipped with newly proposed modality-specific routers. This design enables a single model to efficiently process and fuse multimodal inputs within a unified framework, thereby facilitating diverse tasks without requiring separate models, task-specific fine-tuning, or structural redesign. Importantly, Ming-Omni extends beyond conventional multimodal models by supporting audio and image generation. This is achieved through the integration of an advanced audio decoder for natural-sounding speech and Ming-Lite-Uni for high-quality image generation, which also allow the model to engage in context-aware chatting, perform text-to-speech conversion, and conduct versatile image editing. Our experimental results showcase Ming-Omni offers a powerful solution for unified perception and generation across all modalities. Notably, our proposed Ming-Omni is the first open-source model we are aware of to match GPT-4o in modality support, and we release all code and model weights to encourage further research and development in the community.
Recently, agents based on multimodal large language models (MLLMs) have achieved remarkable progress across various domains. However, building a generalist agent with capabilities such as perception, planning, action, grounding, and reflection in open-world environments like Minecraft remains challenges: insufficient domain-specific data, interference among heterogeneous tasks, and visual diversity in open-world settings. In this paper, we address these challenges through three key contributions. 1) We propose a knowledge-enhanced data generation pipeline to provide scalable and high-quality training data for agent development. 2) To mitigate interference among heterogeneous tasks, we introduce a Mixture-of-Experts (MoE) architecture with task-level routing. 3) We develop a Multimodal Reasoning-Augmented Reinforcement Learning approach to enhance the agent's reasoning ability for visual diversity in Minecraft. Built upon these innovations, we present Optimus-3, a general-purpose agent for Minecraft. Extensive experimental results demonstrate that Optimus-3 surpasses both generalist multimodal large language models and existing state-of-the-art agents across a wide range of tasks in the Minecraft environment. Project page: https://cybertronagent.github.io/Optimus-3.github.io/