AI Research news and analysis on Web Pulse: 20890 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
An independent benchmark comparing Jev against trained BERT-style classifiers and zero-shot NLI pipelines across Banking77, Yelp reviews, an emotion set, and a phishing email set found that a 22-million-parameter encoder…
Cactus Compute's Needle 3, a foundation model shipping as a single 8-29 MB file, beats models ten times its size on mobile tool-calling accuracy and matches models two to three times larger on structured extraction, acco…
AI compute demand is outpacing infrastructure capacity, forcing labs to run GPU workloads across different clusters, regions, cloud providers, and on-premises facilities, which introduces time, cost, security, reliabilit…
Princeton's Pravesh Kothari, UCLA's Raghu Meka, and UC Berkeley's Prasad Raghavendra proposed a separate "conceptual track" for theoretical computer science conferences including STOC, FOCS, and SODA to adapt to AI theor…
A test of TypeSafe's Jev decision model against five frontier language models — DeepSeek V4.1 Flash, Kimi K3, GLM 5.3, GPT-6 Astra, and Fable 5.1 — found both approaches fallible on three dilemmas drawn from the 1964 fil…
A custom 32-bit CPU gate-level crackme called HellGates, published by its creator on crackmes.one in November, was solved by GPT-6 in under 20-30 minutes in September 2026 on SRE-Bench, after going unsolved for a year by…
A developer contributed a recent-papers section to the About page of harvard-edge/cs249r_book, the open-source textbook for Harvard CS249r covering ML Systems, Agentic AI, and Physical AI. The change was merged as PR #19…
A developer released a browser-based demo of a sparse spiking neural network that classifies hand-drawn digits using only event-driven additions, running entirely client-side via ONNX Runtime Web. The model uses 5% conne…
OpenAI's self-reported results show GPT-6 Astra reaching 95.9% mean voxel IoU on the University of Virginia's BenchCAD benchmark, against 83.3% for GPT-5.6 Sol and 84.3% for Claude Fable 5.1 under a modified evaluation p…
Research reported by Ars Technica found that Google's SynthID-Text watermarking can alter LLM responses to adversarial prompts and increase the risk of safety guardrail bypasses, affecting tool invocation and model adher…
A developer building a multi-agent drug-discovery system tested TypeSafe's Jev decision model as an independent validation gate for catching incorrect literature claims, molecular edits, and property records. Jev decline…
Alicja Ziarko and co-authors posted arXiv paper 2508.13113v2 on 29 September 2025 introducing Combinatorial Representations for Temporal Reasoning (CRTR), a method that uses a negative sampling scheme to remove spurious …
Xiaomi's MiMo team launched a public dashboard at mimo.xiaomi.com/rl on September 16 that streams live training telemetry for two trillion-parameter models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, with the Pro run costing rou…
Artificial Analysis, co-founded by CEO Micah Hill-Smith and Chief Product Officer George Cameron, added safety-refusal reporting to its Coding Agent Index on September 18th alongside version 1.5 of its coding-agent bench…
A developer published an engineering guide detailing how Retrieval-Augmented Generation (RAG) pipelines are vulnerable to knowledge injection and indirect prompt injection attacks, in which adversaries plant malicious in…
A developer built the Multilingual Code and Language Dataset, an open-source collection of nearly 150 GB of uncompressed text and source code covering Central Asian languages including Kyrgyz, Kazakh, Uzbek, and Tajik, n…
A machine learning model trained on five columns from a public CDC dataset to predict a sixth column from the same file achieved an R² of 0.998, a near-perfect score that signals hidden target leakage, according to a fre…
TypeSafe AI emerged from stealth with Jev, a model that returns typed decisions with calibrated probabilities instead of text, reporting 70 to 500 ms end-to-end latency at $0.042 per million input tokens and claiming up …
Researchers at ETH Zurich built a disembodied robotic hand that uses its fingers as legs to walk, righting itself within twenty seconds in 21 of 25 fall-recovery tests and crossing fourteen surfaces ranging from a rubber…
A developer published a public machine learning learning repository covering the stack from NumPy to Transformers, built over 5 months with daily commits. The repo's notebooks span classical ML with scikit-learn and XGBo…