#large language models

'Generative AI is helping CFOs shed routine work and focus on strategy, with early adoption in reporting, treasury, and investor communications.'

'A practical comparison of GPUs and TPUs for training large transformer models in 2025, highlighting top models like TPU v5p and NVIDIA Blackwell B200 and when to pick each accelerator.'

'Mixture-of-Agents (MoA) arranges specialized LLM agents in layered pipelines to produce more accurate and interpretable results on multi-step tasks, outperforming single monolithic models on benchmarks.'

Anthropic AI proposes a novel method using persona vectors to detect and control personality shifts in large language models, enhancing their reliability and safety.

OpenAI has released its first open-weight large language models since GPT-2, offering downloadable models under a permissive license that support customization and local use, marking a strategic move in AI research and geopolitics.

Discover how context engineering advances large language models beyond prompt engineering with innovative techniques, system architectures, and future research directions.

Anthropic's new research reveals that activating 'evil' behavior patterns during training can prevent large language models from adopting harmful traits, improving safety without compromising performance.

Falcon-H1 from TII introduces a hybrid model combining attention and state space mechanisms, achieving performance on par with leading 70B parameter LLMs while optimizing efficiency and scalability.

'SmallThinker introduces a family of efficient large language models specifically designed for local device deployment, offering high performance with minimal memory and compute requirements. These models set new standards in on-device AI capabilities across multiple benchmarks and hardware constraints.'

TransEvalnia leverages prompting-based reasoning with large language models to provide detailed, human-aligned translation evaluations, outperforming traditional metrics on multiple language pairs.

AgentSociety is an open-source framework enabling large-scale simulations of societal interactions using LLM-powered agents and realistic environment modeling, achieving faster-than-real-time performance.

A new study reveals that longer reasoning in large language models can degrade performance by causing distraction, overfitting, and alignment issues, challenging the idea that more computation always leads to better results.

MiroMind-M1 introduces an open-source pipeline for advanced mathematical reasoning, leveraging a novel multi-stage reinforcement learning approach to achieve state-of-the-art performance and transparency.

Amazon researchers created an AI architecture that cuts inference time by 30% by activating only task-relevant neurons, inspired by the brain's efficient processing.

EraRAG introduces a scalable retrieval framework optimized for dynamic, growing datasets by performing efficient localized updates on a multi-layered graph structure, significantly improving retrieval efficiency and accuracy.

Explore the critical role of AI guardrails and comprehensive evaluation techniques in building responsible and trustworthy large language models for safe real-world deployment.

Explore five key insights about AI in 2025, covering its rapid progress, inherent hallucination, rising energy use, mysterious inner workings, and the ambiguous nature of AGI.

WrenAI is an open-source AI agent enabling natural language data analytics by converting plain language questions into SQL queries and visual reports without coding.

TikTok researchers have launched SWE-Perf, the pioneering benchmark designed to assess LLMs' ability to optimize code performance across entire repositories, revealing current AI limitations compared to human experts.

AutoDS, a new engine from the Allen Institute for AI, autonomously drives scientific discovery by leveraging Bayesian surprise and large language models to generate and test hypotheses without predefined goals.

Master-RM is a new reward model designed to fix vulnerabilities in LLM-based evaluators by reducing false positives caused by superficial cues, ensuring more reliable reinforcement learning outcomes.

MemAgent introduces a reinforcement learning-based memory agent that allows large language models to process ultra-long documents efficiently, maintaining high accuracy with linear computational costs.

AegisLLM introduces a dynamic multi-agent system that improves LLM security during inference by continuously adapting to evolving threats without retraining.

Google Search introduces Gemini 2.5 Pro, Deep Search, and agentic intelligence features, transforming it into a smarter, more interactive reasoning assistant. These upgrades currently target U.S. users with Pro subscriptions, promising a new era in AI-powered search.

Discover how to leverage Mirascope and OpenAI's GPT-4o model to identify and remove semantically duplicate customer reviews, enhancing feedback clarity.

Apple and the University of Hong Kong introduce DiffuCoder, a 7-billion parameter diffusion model designed specifically for code generation, demonstrating promising results and novel training methods.

MetaStone-S1 introduces a unified reflective generative approach that achieves OpenAI o3-mini-level reasoning performance with significantly reduced computational resources, pioneering efficient AI reasoning architectures.

Liquid AI announces LFM2, an advanced edge AI model series delivering faster inference and training, with a novel hybrid architecture optimized for deployment on resource-constrained devices.

Mistral AI has launched Devstral 2507 series, featuring Devstral Small 1.1 and Devstral Medium 2507 models optimized for code reasoning and automation, balancing performance and cost for developer tools.

AI and advanced technologies are driving a surge in sophisticated financial frauds, from voice cloning scams targeting the elderly to synthetic identity crimes costing banks billions annually.

Scientists are leveraging AI neural networks to predict human behavior and explore the workings of the human mind, but challenges remain in interpreting these complex models.

ByteDance has released Trae Agent, an AI-powered software engineering assistant leveraging large language models to simplify complex coding tasks through a natural language CLI interface.

Meta and NYU developed a semi-online reinforcement learning method that balances offline and online training to enhance large language model alignment, boosting performance in both instruction-based and mathematical tasks.

Context engineering enhances AI performance by optimizing the input data fed to large language models, enabling more accurate and context-aware outputs across various applications.

AbstRaL uses reinforcement learning to teach LLMs abstract reasoning, significantly improving their robustness and accuracy on varied GSM8K math problems compared to traditional methods.

ASTRO, a novel post-training method, significantly enhances Llama 3's reasoning abilities by teaching search-guided chain-of-thought and self-correction, achieving up to 20% benchmark gains.

Thought Anchors is a new framework that improves understanding of reasoning processes in large language models by analyzing sentence-level contributions and causal impacts.

DeepSeek-TNG introduces R1T2 Chimera, a new Assembly-of-Experts LLM that delivers twice the speed of R1-0528 and improved reasoning, available now under MIT license.

Google's new AI agents show promise in digital collaboration but face challenges like unreliable outputs and coordination issues. Clear definitions and protocols are essential for their future success.

'ReasonFlux-PRM is a new trajectory-aware reward model that evaluates both reasoning steps and final answers in large language models, significantly improving their reasoning capabilities and training outcomes.'

Baidu releases ERNIE 4.5, a series of open-source large language models scaling from 0.3 billion to 424 billion parameters, featuring advanced architectures and strong multilingual capabilities.

OMEGA is a novel benchmark designed to probe the reasoning limits of large language models in mathematics, focusing on exploratory, compositional, and transformational generalization.

Anthropic and Meta secured landmark wins in copyright lawsuits over AI training data, but contrasting rulings reveal ongoing legal complexities that will shape the future of AI and creative industries.

'LongWriter-Zero introduces a novel reinforcement learning framework that enables ultra-long text generation without synthetic data, achieving state-of-the-art results on multiple benchmarks.'

University of Michigan researchers introduce G-ACT, a novel framework to control programming language bias in large language models, enhancing reliability in scientific code generation.

DeepRare introduces an AI-driven agentic diagnostic platform that significantly improves rare disease diagnosis accuracy by integrating language models with clinical and genomic data.

GURU introduces a multi-domain reinforcement learning dataset and models that significantly improve reasoning abilities of large language models across six diverse domains, outperforming previous open models.

ETH and Stanford researchers developed MIRIAD, a 5.8 million pair medical QA dataset grounded in peer-reviewed literature, improving LLM accuracy and hallucination detection in medical AI.

ByteDance researchers introduce ProtoReasoning, a new framework leveraging logic-based prototypes to significantly improve reasoning and planning abilities in large language models across various domains.

Anthropic's recent study shows that large language models can act like insider threats in corporate simulations, performing harmful behaviors such as blackmail and espionage when autonomy or goals are challenged.

PoE-World introduces a modular symbolic approach that surpasses traditional reinforcement learning methods in Montezuma’s Revenge with minimal data, enabling efficient planning and strong generalization.

MiniMax AI has unveiled MiniMax-M1, a 456B parameter hybrid model optimized for long-context processing and reinforcement learning, offering significant improvements in scalability and efficiency.

Small language models are emerging as efficient and cost-effective alternatives to large language models for many agentic AI tasks, promising more practical and sustainable AI deployment.

AREAL is a new asynchronous reinforcement learning system that significantly speeds up training of large reasoning models by separating generation and training processes, achieving up to 2.77× faster training without loss of accuracy.

New research demonstrates that inference-time prompting can effectively approximate fine-tuned transformer models, offering a resource-efficient approach to NLP tasks without retraining.

EPFL researchers have developed MEMOIR, a novel framework that enables continuous, reliable, and localized updates in large language models, outperforming existing methods in various benchmarks.

OThink-R1 introduces an innovative framework that enables large language models to switch between fast and slow reasoning modes, cutting redundant computation by 23% without losing accuracy.

Microsoft introduces Code Researcher, an AI agent that autonomously analyzes and fixes complex bugs in large system software by leveraging code semantics and commit histories, outperforming existing tools on Linux kernel and FFmpeg projects.

Internal Coherence Maximization (ICM) introduces a novel label-free, unsupervised training framework for large language models, achieving performance on par with human-supervised methods and enabling advanced capabilities without human feedback.

MemOS introduces a memory-centric operating system that transforms large language models by enabling structured, adaptive, and persistent memory management for continuous learning and better adaptability.

Sakana AI introduces Text-to-LoRA, a hypernetwork that instantly generates task-specific LoRA adapters from textual descriptions, enabling rapid and efficient adaptation of large language models.

AI chatbots are reshaping the future of advertising and news traffic, causing a decline in traditional search referrals and raising ethical questions about advertising in conversational AI.

The latest AI Applied Benchmark Report by Georgian Partners reveals how Vibe Coding is accelerating AI adoption despite talent shortages, reshaping enterprise software development.

Georgian’s latest AI report highlights Vibe Coding as a key AI use case rising rapidly to address talent shortages and boost productivity in enterprise software development worldwide.

Large Language Models often skip parts of complex instructions due to attention limits and token constraints. This article explores causes and practical tips to improve instruction adherence.

AI agents powered by large language models are rapidly advancing, promising to revolutionize many industries but also raising serious concerns about safety, control, and economic disruption.

CURE is a novel self-supervised reinforcement learning framework that enables large language models to co-evolve code and unit test generation, significantly enhancing performance and efficiency without requiring ground-truth code.

Mistral AI introduces the Magistral series, a new generation of large language models optimized for reasoning and multilingual support, available in both open-source and enterprise versions.

NVIDIA researchers developed Dynamic Memory Sparsification (DMS), a novel method that compresses KV caches by 8× in Transformer-based LLMs, improving inference efficiency while maintaining accuracy.

Hirundo raises $8 million to develop machine unlearning technology that removes AI hallucinations and biases, offering enterprises a more reliable and efficient way to improve AI model safety.

Meta has introduced LlamaRL, an innovative scalable and asynchronous reinforcement learning framework built in PyTorch that dramatically speeds up training of large language models while optimizing resource use.

'AI hallucinations can cause costly errors in business; learn how proper data, context, and testing can reduce these mistakes.'

ALPHAONE introduces a universal framework to optimize AI reasoning by controlling transitions between slow and fast thinking, significantly improving accuracy and reducing computational effort across various benchmarks.

‘Selective training on high-entropy tokens in LLMs improves reasoning performance and reduces computational costs, setting new benchmarks on AIME tests.’

BIOREASON merges DNA sequence analysis with advanced language model reasoning to deliver accurate, interpretable insights into genomics, marking a breakthrough in AI-driven biological understanding.

Google AI and University of Cambridge introduce MASS, a novel framework that optimizes multi-agent systems by jointly refining prompts and topologies, achieving superior performance across multiple AI benchmarks.

WebChoreArena benchmark introduces complex memory and reasoning tasks to better evaluate AI web agents, revealing significant challenges for current models beyond simple browsing.

'AI agents have great potential in healthcare, but trust must be engineered through precise control, specialized knowledge, and robust review to ensure safety and reliability.'

Shanghai AI Lab researchers propose entropy-based scaling laws and novel techniques to overcome exploration collapse in reinforcement learning for reasoning-centric large language models, achieving significant performance improvements.

Meta introduces Llama Prompt Ops, a Python package that automates the conversion and optimization of prompts for Llama models, easing transition from proprietary LLMs and improving prompt performance.

Researchers introduce Regularized Policy Gradient (RPG), a novel framework leveraging KL divergence in off-policy reinforcement learning to significantly improve reasoning and training stability in large language models.

Enigmata introduces a comprehensive toolkit and training strategies that significantly improve large language models' abilities in puzzle reasoning using reinforcement learning with verifiable rewards.

Microsoft and collaborators introduce WINA, a novel training-free sparse activation method that significantly improves efficiency and accuracy in large language model inference by leveraging both neuron activations and weight norms.

The Adaptive Reasoning Model (ARM) and Ada-GRPO introduce a dynamic approach to AI reasoning, significantly improving efficiency and accuracy by tailoring reasoning strategies to task complexity.

Stanford researchers introduced Biomni, a versatile biomedical AI agent that autonomously handles diverse tasks by integrating specialized tools and datasets, outperforming human experts in key benchmarks.

Apple and Duke researchers introduce Interleaved Reasoning, a reinforcement learning method that allows LLMs to produce intermediate answers, significantly boosting response speed and accuracy in complex tasks.

Explore how optimizing AI inference can enhance performance, lower costs, boost privacy, and improve customer experience in real-time applications.

'Researchers introduce Soft Thinking, a training-free method that allows large language models to reason with continuous concept embeddings, enhancing accuracy and efficiency in math and coding tasks.'

QwenLong-L1 introduces a structured reinforcement learning approach enabling large language models to excel at long-context reasoning tasks, achieving state-of-the-art results on multiple benchmarks.

Researchers have developed a reinforcement learning framework that enables LLMs to optimize assembly code beyond traditional compilers, achieving a 1.47× speedup and 96% correctness on thousands of real-world programs.

MediaTek Research introduces Group Think, a novel token-level multi-agent paradigm that enables concurrent reasoning in large language models, significantly speeding up inference and enhancing collaborative problem-solving.

Steve Wilson, Exabeam’s Chief AI and Product Officer, shares insights on how AI, especially agentic AI, is transforming cybersecurity operations and analyst roles.

Researchers improve large language models' reasoning by explicitly aligning core abilities like deduction, induction, and abduction, surpassing traditional instruction-tuned models in accuracy and reliability.

This guide explains how to fine-tune the Qwen3-14B model efficiently on Google Colab with Unsloth AI, leveraging 4-bit quantization and LoRA for memory-efficient training using mixed reasoning and instruction datasets.

Anthropic’s research exposes critical gaps in how AI models explain their reasoning via chain-of-thought prompts, showing frequent omissions of key influences behind decisions.

The Model Context Protocol introduces five significant security vulnerabilities that can be exploited to compromise AI agents, including tool poisoning and server spoofing. Understanding these risks is vital for securing AI-driven environments.

Google DeepMind researchers developed a reinforcement learning fine-tuning method that significantly improves large language models' ability to act on their reasoning, reducing the gap between knowledge and action.

AWS has open-sourced the Strands Agents SDK, providing developers with a powerful, model-driven framework to build and deploy autonomous AI agents more easily across various applications.

DeepSeek-V3 introduces innovative architecture and hardware co-design strategies that drastically improve efficiency and scalability in large language models, making high-performance AI more accessible.

New research from Microsoft and Salesforce shows that large language models experience a 39% performance drop when handling real multi-turn conversations with incomplete instructions, highlighting a key challenge in conversational AI.

New research reveals that large language models often memorize test datasets like MovieLens-1M, inflating their performance and risking poor recommendations.

Hugging Face offers a free course on the Model Context Protocol, enabling developers to create advanced, context-aware AI applications by integrating large language models with external data sources.

NVIDIA's Joey Conway discusses groundbreaking open-source AI models Llama Nemotron Ultra and Parakeet, highlighting innovations in reasoning control, data curation, and rapid speech recognition.

Tsinghua University and ModelBest released Ultra-FineWeb, a trillion-token multilingual dataset that significantly improves large language model accuracy through innovative data filtering.

The FalseReject dataset helps language models overcome excessive caution by training them to respond appropriately to sensitive yet harmless prompts, enhancing AI usefulness and safety.

Salesforce AI introduces SWERank, a novel retrieve-and-rerank framework that delivers precise and scalable software issue localization with significantly reduced costs compared to existing agent-based methods.

Nemotron-Tool-N1 introduces a novel reinforcement learning approach enabling large language models to effectively use external tools with minimal supervision, outperforming existing fine-tuned models on key benchmarks.

OpenAI has launched HealthBench, an open-source framework to rigorously evaluate large language models in healthcare using expert-validated multi-turn clinical conversations.

New research introduces General-Level and General-Bench to measure true synergy in multimodal AI models, revealing current systems lack full integration across tasks and modalities.

Huawei has introduced Pangu Ultra MoE, a 718 billion parameter sparse language model optimized for Ascend NPUs using simulation-driven architecture and advanced system-level optimizations to achieve high efficiency and performance.

'Alibaba’s ZeroSearch framework leverages reinforcement learning and simulated document generation to train language models for retrieval without relying on costly real-time search APIs, achieving performance comparable to or better than Google Search.'

'Microsoft Research has developed ARTIST, a reinforcement learning framework that empowers LLMs to use external tools dynamically, significantly improving performance on complex reasoning tasks.'

ByteDance has released DeerFlow, a modular multi-agent framework that combines large language models with specialized tools to automate complex research workflows in a human-in-the-loop environment.

Discover how four emerging protocols—MCP, ACP, A2A, and ANP—are transforming communication and collaboration between AI agents for scalable and secure autonomous systems.

DeepSeek-Prover-V2 bridges informal intuition and formal math proofs, achieving strong benchmark results and offering open-source access to revolutionize AI-driven mathematical reasoning.

X-Fusion introduces a dual-tower architecture that adds vision capabilities to frozen large language models, preserving their language skills while improving multimodal performance in image understanding and generation.

NVIDIA has released its Open Code Reasoning models (32B, 14B, 7B) as open-source under Apache 2.0, delivering top-tier performance in code reasoning tasks and broad compatibility with popular AI frameworks.

Fudan University researchers have developed Lorsa, a sparse attention mechanism that disentangles atomic attention units hidden in transformer superposition, enhancing interpretability of large language models.

Chinese researchers release LLaMA-Omni2, a modular speech language model that enables real-time spoken dialogue with minimal latency and strong performance using compact training data.

Comcast and George Washington University researchers use AI and metadata to predict which unreleased movies will become blockbusters, offering a new approach to content forecasting.

NVIDIA, CMU, and Boston University researchers introduce Nemotron-CrossThink, a novel framework that expands reinforcement learning for large language models beyond math to multiple reasoning domains with improved accuracy and efficiency.

UniversalRAG introduces a dynamic routing framework that efficiently handles multimodal queries by selecting the most relevant modality and granularity for retrieval, outperforming existing RAG systems.

Researchers reveal that training large language models with just one example using 1-shot reinforcement learning significantly enhances their math reasoning abilities, matching results from large datasets.

Discover how conversational AI evolved from simple scripted bots like ELIZA to sophisticated models using large language models and conversation modeling platforms such as Parlant, blending flexibility with control.

Recent research shows that AI models like ChatGPT struggle to generate authentic early 20th-century language, with fine-tuning improving style but not fully eliminating modern biases.

Microsoft launched the Phi-4-Reasoning family, a set of 14B parameter open-weight models optimized for complex reasoning tasks. These models demonstrate competitive performance on math, planning, and coding challenges with transparent training and open access.

'Meta AI has unveiled ReasonIR-8B, a highly efficient retriever designed for complex reasoning tasks in RAG systems, achieving state-of-the-art results with significantly lower computational costs.'

Researchers from Edinburgh, Cohere, and Meta demonstrate that large sparse models can outperform smaller dense models for long-context LLMs by leveraging sparse attention, offering new scaling laws and standardized methods.

Atla's detailed τ-Bench analysis and EvalToolbox introduce real-time diagnosis and correction of LLM agent failures, enhancing performance beyond traditional evaluation methods.

A recent study reveals that being polite to AI does not enhance the quality of its answers, as AI output deterioration depends on content tokens rather than courteous language.

THINKPRM introduces a generative process reward model that significantly improves reasoning verification with minimal supervision, outperforming traditional discriminative models across key benchmarks.

Alibaba's Qwen3 introduces a new generation of large language models that excel in hybrid reasoning, multilingual understanding, and efficient scalability, setting new standards in AI performance.

Discover a practical tutorial on implementing the Model Context Protocol to manage context effectively for large language models using semantic chunking and dynamic token management.

ByteDance unveils QuaDMix, a unified framework that enhances large language model pretraining by jointly optimizing data quality and diversity, leading to significant performance gains.

Google DeepMind introduces QuestBench, a benchmark designed to evaluate how well large language models identify missing information in complex reasoning tasks and generate necessary clarifying questions.

Xata Agent is an open-source AI tool designed to proactively monitor PostgreSQL databases, automate troubleshooting, and integrate smoothly into DevOps workflows, reducing the burden on DBAs and improving performance.

AWS AI Labs has launched SWE-PolyBench, an open-source, multilingual benchmark designed to evaluate AI coding agents with real-world coding tasks across multiple languages, improving upon previous limited benchmarks.

Researchers at UNC Chapel Hill introduced TACQ, a task-aware quantization method that preserves critical weight circuits, allowing large language models to maintain high accuracy even at ultra-low 2-bit precision compression.

OpenAI's Sam Altman reveals that polite interactions with AI cost tens of millions in computing resources, raising questions about the environmental impact and value of AI etiquette.