Whispers

A new benchmark reveals that frontier large language models degrade approximately 25% of documents during extended workflows. Separately, a Fields Medal winner has reported that ChatGPT 5.5 Pro is capable of solving complex PhD-level mathematics problems. AI

The author details the practical implementation of the A2A Protocol, an open standard for agent discovery and task delegation. This second part focuses on the code, outlining the architecture where the orchestrator acts as both a server and a client. It highlights the importance of the orchestrator being an A2A service to receive structured tasks and emit failure events, contrasting this with a simpler client-only script. The project structure and setup for the shared agent and customer-specific orchestrators are also provided. AI

A new report indicates that no single AI model consistently leads across all benchmarks, with different models excelling in specific areas like coding or math. The evaluation process itself is also complex, as multiple frontier models provide divergent reasoning for their scores when judging agent performance. This suggests that developers need to employ continuous, multi-model evaluation strategies rather than relying on a single leaderboard for model selection. AI

This article highlights five foundational AI papers that are considered essential reading for AI engineers. It aims to explain the core contributions of each paper and their lasting significance in the field. The selection focuses on works that have fundamentally shaped current AI development and understanding. AI

A new study published on arXiv investigates emergent misalignment (EM) in large language models, finding it is not a universal phenomenon but rather an artifact of overtraining. Researchers tested 12 open-source models across four families and discovered that EM is more prevalent in larger models and emerges late in the training process. The study suggests practical mitigation strategies, such as early stopping during fine-tuning, which can eliminate EM while retaining most task performance. AI

This article argues that fine-tuning a vision-language model (VLM) is less about the technical training process and more about strategic decisions made beforehand. The author highlights four key choices that significantly impact the outcome of fine-tuning, suggesting that focusing on these decisions yields better results than solely optimizing training parameters. AI

Meitu's AI research arm, MT Lab, has had six papers accepted into major international conferences including ICLR, CVPR, and ICML. One paper on scene text editing, accepted by ICML 2026, has already been integrated into Meitu Design Room and Meitu Xiuxiu PC as a 'seamless text modification' feature. This new functionality supports multiple languages and maintains visual consistency without obvious editing marks. AI

Researchers have introduced a new framework for open-world object counting, addressing the brittleness of current vision-language models in accurately identifying and counting objects based on user intent. They propose redefining counting as a multi-grained problem, where both visual examples and detailed text prompts, including negative prompts, specify the target appearance and semantic granularity. To overcome the data limitations for this approach, they developed an automated pipeline using 3D synthesis and VLM filtering to create KubriCount, the largest dataset for counting tasks. Their new model, HieraCount, leverages both text and visual exemplars to significantly improve multi-grained counting accuracy and generalize to real-world scenarios. AI

DBSCAN is a clustering algorithm that identifies dense regions of data points to discover arbitrary shapes. It groups together points that are closely packed, marking outliers as noise. This method is particularly effective for finding clusters of varying densities and complex structures within datasets. AI

Researchers have introduced Key-Value Means (KVM), a new attention mechanism for transformers that can handle both fixed-size and growing states. When implemented with a fixed-size cache, KVM functions as an O(N) chunked RNN with minimal parameter additions. A growable KVM cache version demonstrates competitive performance on long-context tasks, offering subquadratic prefill time and sublinear state growth. This approach is compatible with standard operations, supports chunk-wise parallelizable training, and provides a flexible trade-off between prefill time complexity and memory usage. AI

A developer fine-tuned Google's Gemma 4 E4B model into a bias judge for approximately $30, a process that took two weeks with most of the effort focused on data pipeline construction rather than GPU time. The resulting model, capable of running locally in 30 seconds, evaluates pairs of responses to identify social bias using the Bias Benchmark for QA (BBQ) dataset. The developer encountered challenges with classification leaks, data ceilings imposed by the BBQ dataset, and disagreements among different LLMs used for labeling, ultimately leading to a refined data construction strategy. AI

This article explores the concept of Recursive Self-Improvement (RSI) by proposing a novel metric, the RSI for RSI. The author details the development and application of this metric, aiming to provide a quantitative measure for assessing the effectiveness of self-improving AI systems. The work contributes to the theoretical understanding of AI advancement and its potential for accelerated progress. AI

The 2050 AGI 4 Science conference featured 17 young scholars discussing the evolving landscape of AI in scientific research. The event highlighted a shift from general AI models to deep integration within specific scientific fields, with a focus on problem-driven, interdisciplinary collaboration. Discussions explored AI's potential to tackle high-cost experimentation, reshape technical routes through first principles, and bridge the gap between academic research and industrial application. AI

A developer fed 15 scientific papers into Google's Gemma 4 model to test its hypothesis generation capabilities. The model produced a hypothesis that was not explicitly stated in any of the provided papers. Interestingly, when the model was asked to review its own generated hypothesis, its confidence score decreased. AI

Researchers have developed a novel method to deliver nanobodies, which are small antibody fragments, directly inside cells to treat cystic fibrosis. These nanobodies are fused with a cell-penetrating peptide, enabling them to cross the cell membrane and stabilize the misfolded CFTR protein from within. When combined with existing therapies, this approach has shown potential to restore nearly normal protein function in patient-derived cells. AI

A technique for estimating population medians with minimal data is explored, drawing from Douglas Hubbard's "How to Measure Anything." The method leverages the probability that a set of independent samples will all fall above or below the population median. By calculating the complement probability, it's possible to determine the likelihood that the median lies within the range of the sampled data. AI