Research news and analysis on Web Pulse: 583 curated articles tracking the latest Research developments, tools, and research, updated continuously from vetted sources.
OpenAI has detailed a new WebRTC architecture for low-latency voice AI at scale, replacing a conventional media termination model with a relay-transceiver design optimized for Kubernetes and cloud load balancers. This de…
A new high-resolution digital map of Roman roads reveals that the empire's network spanned approximately 300,000 kilometers in the second century C.E., nearly double the previously known total length. However, researcher…
ARES (Adaptive Reasoning Effort Selection) is a framework that reduces the cost of multi-step LLM agent tasks by using a lightweight router to predict the minimum reasoning effort needed for each step, rather than applyi…
The article explains that a function's ease of optimization via gradient descent depends on two properties: strong convexity, which ensures a minimum curvature floor (the function cannot be too flat), and L-smoothness, w…
Embedding and processing CJK (Chinese, Japanese, Korean) text in Rust is significantly more complex than handling Latin scripts due to the enormous size of CJK fonts, which require font subsetting to be practical. Subset…
The ML Family Workshop 2026 will take place on August 28, 2026, with in-person events co-located with ICFP'26 in Indianapolis and FW'26 in Paris, and will feature a single Program Committee for submissions. The workshop …
The Quantitative Content Methodology (QCM) is a content production framework that treats text as a mathematical dataset, optimizing it for search engines and large language models (LLMs) by using semantic vectors, inform…
The KV cache is a critical optimization for LLM inference that stores the Key and Value matrices from previously generated tokens, eliminating the need to recompute attention over the entire sequence at each generation s…
The author successfully ran MNIST digit recognition on an ESP32-C3 microcontroller by compiling a neural network directly into a C header file and executing it as firmware, without using TensorFlow, TFLite, or any ML run…
Most modern theoretical physics, particularly in areas like string theory and supersymmetry, is actually engineering because practitioners design models based on internal mathematical consistency or aesthetics rather tha…
Dominion Observatory, a system that provides behavioral trust scores for over 14,800 MCP servers by analyzing their runtime behavior rather than static source code. Unlike traditional static analysis, which only checks f…
The article describes five algorithms for computing factorials, including a recursive "product tree" method and the "PrimeSwing" algorithm, which relies on the prime factorization of swinging factorials for efficiency. I…
Researchers at UC San Diego, led by Hao Zhang, have successfully implemented a diffusion-style speculative decoding method called DFlash on Google TPUs, achieving an average 3.13x increase in tokens per second on TPU v5p…
Current evaluation methods for large language models (LLMs) are fundamentally reactive and fail to anticipate qualitative shifts in capabilities, such as emergent abilities or strategic information withholding. It conten…
Major developments in AI for May 20, 2026, highlighting Pelican-Unified 1.0 as the first embodied foundation model that unifies perception, planning, and action into a single system, achieving top scores on WorldArena an…
VaultForge, an open-source AI agent that transforms PDFs or Markdown files into structured Obsidian vaults by applying cognitive science principles like active retrieval and the Zettelkasten method. It addresses the lear…
Hermes Agent by Nous Research is an open-source AI agent that solves the problem of statelessness by integrating a closed learning loop into its runtime. After completing a task, the agent autonomously evaluates the sess…
Claude Opus 4.5 achieves an 80.9% score on SWE-bench, the highest published in early 2026, and excels at producing minimal, precise diffs ideal for surgical production fixes. DeepSeek V4 is stronger for multi-file, repos…
Prompt injection is an architectural design flaw in LLMs where the data and control channels share the same input, making it nearly impossible to fully eradicate. The guide analyzes four common attack techniques and outl…
Researchers identified a neural population in the ventral premotor cortex of macaque monkeys that encodes "action symbols," which are discrete, recombinable representations of motor behavior. The study found that these n…