On August 12, 2025, Perplexity AI made an unsolicited $34.5 billion cash offer to Google to acquire Chrome, aiming to challenge Google's search dominance by integrating AI into the browser. Google has no plans to sell Chrome.

A study submitted to arXiv on August 11, 2025, evaluates GPT-5's performance in multimodal medical reasoning, indicating that GPT-5 surpasses all baselines and pre-licensed human experts, achieving state-of-the-art accuracy.

AI companion apps are projected to generate $120 million in revenue in 2025, with 220 million downloads globally and $82 million in revenue generated during the first half of the year.

GitHub Copilot Pro now supports GPT-5 in VS Code, offering users a 30-day free trial.

Context engineering is transforming AI applications across industries, with Five Sigma Insurance seeing an 80% reduction in claim processing errors and Microsoft seeing a 26% increase in completed software tasks.

Researchers proposed MCPToolBench++, a large-scale AI Agent tool use benchmark built upon a marketplace of over 4k MCP servers from more than 40 categories, consisting of single-step and multi-step tool calls.

Submitted on August 11, 2025, Mem4D is a novel framework for dynamic scene reconstruction that decouples static and dynamic memory using a dual-memory architecture.

Researchers propose Direct Multi-Preference Optimization (DMPO), a data selection principle for aligning Large Language Models (LLMs) with diverse human values, achieving over 10% relative improvement.

Researchers introduce Griffon v2, a unified high-resolution generalist model for multimodal perception, enabling flexible object referring with visual and textual prompts, achieving state-of-the-art performance on REC and phrase grounding.

Researchers introduce X-evolve, a method that evolves solution spaces powered by large language models, generating tunable programs and efficiently exploring the space with a score-based search algorithm.

Researchers propose a new task called Power Battery Detection (PBD) and release PBD5K, a large-scale benchmark with 5,000 X-ray images for quality inspection.

Researchers introduce Grove MoE, a novel MoE architecture with adjugate experts, enabling model capacity expansion while maintaining manageable computational overhead, achieving performance comparable to SOTA open-source models.

Researchers introduce InterChart, a diagnostic benchmark evaluating vision-language models' reasoning across multiple related charts, revealing accuracy declines as chart complexity increases.

On August 10-11, 2025, researchers submitted papers addressing audio-visual deepfake detection, with one achieving top performance in temporal localization and a top-four ranking in classification for the ACM 1M Deepfakes Detection Challenge.

Researchers propose a method using topological data analysis to identify which parts of GPT-2 contribute to bias towards specific groups, finding biases concentrated in certain attention heads.

Researchers propose a novel Semantic-Enhanced LLM (SE-LLM) that enhances token embedding by exploring inherent periodicity and anomalous characteristics of time series, improving interpretability for LLMs in temporal sequence analysis.

Researchers submit a paper detailing techniques for efficient speculative decoding for Llama models at production scale, achieving a new state-of-the-art inference latency of about 4 ms per token on 8 NVIDIA H100 GPUs.

On August 11, 2025, arXiv announced the v5 revision of a paper introducing characterization and enforcement techniques to address fairness concerns in generative AI (GenAI).

Wentao Jiang and colleagues introduced REX-RAG, a framework enhancing LLM reasoning via Reinforcement Learning with Retrieval-Augmented Generation (RAG), improving performance by 5.1% on Qwen2.5-3B and 3.6% on Qwen2.5-7B over baselines.

Researchers propose a novel contrastive learning objective to make transformers symmetry-aware for automated planning, addressing the limitations of PlanGPT.

Researchers investigate how in-context information influences model behavior and whether LLMs can identify their unreliable responses, finding that misleading context often induces confidently incorrect responses.

Researchers propose a new approach for learning user preferences in image generation models using Multimodal Large Language Models, outperforming others in preference prediction accuracy.

Researchers propose Optimal Transport Depth Up-scaling (OpT-DeUS) for Large Language Models (LLMs), achieving better performance and training efficiency than existing methods.

A paper submitted on August 11, 2025, explores user privacy perceptions towards LLM's Retrieval Augmented Generation (RAG)-based memory, revealing diverse mental models and significant concerns regarding privacy, control, and accuracy.

Published on August 11, 2025, a paper introduces a simulated dataset designed to support research on proactive robots that infer human needs from natural language conversations within workplace environments.

On August 11, 2025, a paper was submitted addressing posterior sampling in score-based generative models, introducing a method to sample from a distribution close to both the noised prior posterior and the true posterior.

On August 11, 2025, arXiv announced the revised version of a paper introducing Constraint Satisfaction Visual Grounding (CSVG), a zero-shot method that reformulates 3D visual grounding (3DVG) as a Constraint Satisfaction Problem (CSP).

Tahoe Therapeutics, a Palo Alto biotech startup, raised $30 million in funding led by Amplify Partners to develop AI models of living cells for cancer research, bringing total funding to $42 million and a valuation of $120 million.

As reported by Digitimes on August 12, 2025, Chinese tech companies are actively recruiting AI experts, mirroring a broader trend of young people learning AI skills to stay competitive in the job market.

A Workday study from August 12, 2025, found that 75% of employees are comfortable working with AI agents, but only 30% want AI as their manager, with 82% of organizations expanding AI agent use.

Alon Goren, writing for Forbes Technology Council, argues that AI agents, which possess autonomy and can make independent decisions, are a better option for achieving transformative enterprise value than Retrieval Augmented Generation (RAG).

Researchers introduced FEAT, a multi-agent AI framework for automated cause-of-death analysis, which uses a domain-adapted large language model to standardize death investigations and outperformed state-of-the-art AI systems.

Researchers developed an improved segmentation model for automated multi-species and damage plant semantic segmentation in herbicide trials, which significantly improved species identification and damage classification and is now deployed in BASF's phenotyping pipeline.

AIS-LLM, a novel framework, integrates time-series AIS data with a large language model (LLM) for maritime traffic analysis, enabling simultaneous execution of trajectory prediction, anomaly detection, and collision risk assessment.

Researchers from Cornell University present an approach integrating vision-based localization with large language model (LLM)-based navigation for indoor environments, achieving 96% accuracy across waypoints in an office corridor.

A new paper, NeuroDx-LM, proposes a large-scale model for EEG-based neurological disorder detection, showing state-of-the-art performance in seizure and schizophrenia detection.

On August 11, 2025, Yunze Luo et al. submitted a paper proposing CURec, a framework for enhancing recommender systems using large language model (LLM) fine-tuning by generating collaborative-aligned content features.