AI Research news and analysis on Web Pulse: 22649 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
Shrijith Venkatramana, an engineer building the AI code review tool LiveReview, traces modern LLM tokenization back to Byte Pair Encoding, a 1994 data-compression algorithm by Philip Gage. The writeup explains how BPE-st…
Researchers Ilyas, Santurkar, Tsipras, Engstrom, Tran and Madry published a 2019 paper arguing that adversarial examples are not bugs but features of machine learning models, challenging the prevailing view that such per…
A team released Drone-Bench, a benchmark measuring how well AI models navigate drones and identify targets, reporting that the frontier model Astra now beats the human baseline on every task, whereas frontier models stru…
A paper submitted on 28 Sep 2025 to arXiv's Computer Science > Artificial Intelligence section formalizes the correspondence between fibring of modal logics and fibring of neural networks, a gap the authors say was never…
A developer has published a learning repository containing 38 runnable notebooks that teach the Hugging Face stack by progressively removing abstractions, from high-level APIs through fine-tuning, PEFT, reasoning/RL, and…
A team led by Yi Duan at Shanghai Jiao Tong University published a survey on recursive self-improvement (RSI) that introduces a Headroom-Closed Index (HCI) to quantify how far existing large language models fall short as…
Cognition released SWE-2, a coding model scoring 50% on FrontierCode 1.1 Main — within one point of Fable 5.1 — at 64% lower cost, built on a 2.8-trillion-parameter Kimi K3 base with RL post-training. Separately, the Rub…
A developer ran six Codex coding-agent executions of a Java String Calculator kata, varying the underlying model (Sol, Terra, Luna) and reasoning effort, to study how agents, a repository skill, and the Seed4J CLI intera…
GPT-6 Astra completed 7 of 100 tasks on the StationeryBench robotics benchmark using dual-arm robots, a 7% success rate that far exceeds MolmoAct2, which finished zero tasks on the same set. The results suggest the model…
A multi-agent framework reached 68% accuracy on the QUBOBench benchmark for automatically translating business problems into Quadratic Unconstrained Binary Optimization (QUBO) formulations, a 22% improvement over a stand…
OpenAI's GPT-6-Astra model can lay out a printed circuit board in KiCad, build a 3D city scene in Unity, create an animated automobile transmission in FreeCAD and Blender, and draft a tax return from a W-2, according to …
An unnamed OpenAI model solved the forced version of the Navier-Stokes Millennium Prize problem, and one day later OpenAI told The New York Times it had made "substantial progress on another Millennium Prize problem," ac…
High bandwidth memory (HBM) is hitting scaling limits driven primarily by thermal gradients across the stack, according to a technical deep dive that cites JEDEC standards and Samsung's HBM4 architecture presented at ISS…
NVIDIA Research's Efficient AI Team and Singapore Lab reported that a two-stage pipeline specialized for a single NVIDIA DGX Spark generates a 384p draft and refines it to 768p, producing a 5-second 768p video in 56 seco…
Researchers submitted a paper to arXiv on 24 Feb 2026 introducing OpenGCRAM, a memory compiler that supports both SRAM and Gain Cell RAM (GCRAM) and generates macro-level designs and layouts for commercial CMOS processes…
An independent developer has released Prism, a language model with an 842KB weight file and 107,776 parameters that runs in a browser and generates text from bigram subword tokens, spelling words rather than linking whol…
Vals.ai's new International Olympiad in Informatics benchmark found that GPT-6 Astra solved every problem across all three years tested, while GPT-5.6 Sol scored 91.17%, Claude Fable 5.1 scored 90.78% and GPT-5.6 Terra s…
Developer Sudharsanselvaraj built TokenPrint, an open-source 3D visual debugger that lets users inspect the internal computation of transformer and LLM inference layer by layer. The tool treats inference as a journey, ex…
GPT-6 Astra completed 7 of 100 tasks with dual-arm robots on the StationeryBench robotics benchmark, while competitor MolmoAct2 finished none, according to early benchmark results reported by The Decoder. A researcher ch…
A technical post argues that matrix calculus is unnecessary for differentiating functions of vector or matrix variables in machine learning, advocating instead for writing out tensor indices and applying ordinary differe…