AI Research news and analysis on Web Pulse: 22672 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
More than 100 researchers using AI coding agents cut the ECDSA.Fail resource score for secp256k1 point-addition circuits by 86.1%, from 10.75 billion to 1.496 billion, according to a paper published September 9. The top …
Researchers introduced CARDEA, an AI system for end-to-end coronary angiography (CAG) interpretation that grounds its reasoning in spatial evidence to make decisions auditable, according to the paper's description. The w…
A developer's quantization sweep of a small transformer found that the token embedding table accounts for 71.7% of the model's 1,024,000 parameters, making embedding precision the dominant lever over network precision. A…
Torafirma Systems published a public benchmark report for its Orbit system dated 2026-09-11, claiming a CPU memory query latency of 217 ns/token and a perplexity of 236.9 on WikiText-2 versus 317.1 for a comparison basel…
ByteDance's Seed team, working with Singapore University of Technology and Design, Georgia Tech, and other institutions, released HarnessDev, a research project showing that LLMs can generate and iteratively refine their…
Researchers Yuxing Lu, Yicheng Chen, and Shanchan Wu published a paper on Procedural Graphs, a self-evolving execution structure that lets LLM agents rewrite their own procedural knowledge as a directed graph of steps, e…
A developer built Arb3a Baloot, a reinforcement-learning agent that plays the Saudi trick-taking card game Baloot, and shipped it as a free, ad-free Android app this week. The bot uses PPO with self-play on a simulator r…
Researchers from Duke University and Tsinghua University published a paper on September 3, 2026 introducing PlaidQ, a continuous Gaussian latent-diffusion language model for code generation. Distilling the diffusion proc…
A Mixture-of-Experts model called T1, with 122B total parameters, has been trained with reinforcement learning to operate a real shell in a cloud sandbox for up to 300+ tool calls on long-horizon terminal tasks such as c…
A 29 July 2026 arXiv paper by Aleksandr Berdnikov reports that most open-source language models of roughly 100B parameters hold a decodable representation of the night sky map in their residual stream, with the represent…
A developer testing four Fable 5 AI agents in a Werewolf social deduction game found that the agents produced plausible strategic accusations based on events that never occurred, such as claiming a player was suspiciousl…
A developer argues that Mojo, which reached 1.0 in August 2026 and whose compiler and toolchain Modular open sourced under Apache 2.0 shortly after, is unlikely to replace Python for AI development but may complement it …
Z.ai released GLM-5.3, a 743-billion-parameter model that matches its predecessor GLM-5.2 in size but delivers a 50% gain in programming capability and tops global cybersecurity benchmarks through post-training scaling r…
YuE2 (best-of-8) scored 6.9632 on SongBench, the highest observed mean among 15 evaluated settings on WildSongBench's 192 prompts, according to a September 5, 2026 evaluation, edging out Suno v5 at 6.8721. The same resul…
A developer argues that looped transformer architectures, such as the described GPT-6 Astra which runs the same blocks through roughly 44 recurrent passes while reusing weights, decouple a model's visible chain of though…
Warren McCulloch, a psychiatrist and neurophysiologist in his mid-forties, and Walter Pitts, a largely self-taught logician barely twenty, co-authored the 1943 paper "A Logical Calculus of the Ideas Immanent in Nervous A…
Independent developer croqaz released three new vintage LLM models, a benchmark, and an evaluation pipeline trained only on pre-1900 texts, following the earlier Llama-340M base model that produced Victorian prose but co…
Long-horizon AI agents fail at an exponential rate because a 95% per-step success probability yields only a 35.8% chance of completing 20 turns and 7.7% at 50 steps, and the chance of a second error jumps from 5% to over…
A controlled pilot on a dedicated NVIDIA H200 GPU found ISO-AdamW scored 758 correct answers versus 754 for baseline AdamW on a 1,000-question held-out test, a +0.4% delta with a 95% confidence interval of [−1.8%, +2.7%]…
InclusionAI is offering free API access to its Ling-3.0-Flash-VL multimodal model and releasing the weights under the MIT license, the company said. The model uses a mixture-of-experts design with 124 billion total param…