AI Research news and analysis on Web Pulse: 22647 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
A developer documented four context window management patterns for long-running AI agents after observing that a four-hour agent session spent roughly three quarters of its token budget on its own transcript—tool outputs…
Anthropic published a scenario explorer and technical report modeling how transformative AI could reshape the US economy by 2030, projecting GDP gains of 1.6% to 32.4% across modest, substantial, and extreme scenarios. T…
Specific Labs' Real-SWE benchmark, published in September 2026, found that the best frontier coding agent resolves only 38.8% of tasks on private enterprise codebases, meaning roughly six in ten agent-generated changes a…
A developer found that swapping a hashed vectoriser for a sentence-transformer in a schema retriever produced a strict recall-at-10 gain from 82.6% to 88.1% on a pooled Spider 1.0 catalog (n=1,034), but the effect disapp…
DeepSeek R1, an open-source reasoning model, combines a Mixture of Experts architecture with Reinforcement Learning from Reasoning Feedback (RLRF) to improve multi-step logic, math, and coding performance. Benchmarks cit…
A developer outlined a six-layer "safety pyramid" for building trustworthy AI agents, spanning technical robustness, interpretability, content safety, instruction following, value alignment, and adversarial robustness. T…
An engineer's experiment found that a zero-parameter count table over the current document overtakes a 1.43M-parameter transformer with a fixed 64-token window as documents grow longer. At L=1000 tokens the cache reached…
Parth Nirgide has published an open technical specification and defensive prior art disclosure for Beneficial Intelligence (BI), a framework that deliberately withholds complete AI-generated answers and instead delivers …
A three-stage rule-chaining framework described in arXiv:2609.10654v1 solved 230 of 240 tasks on the ARC-AGI-2 benchmark, according to the paper's reported results. The system uses a fallback hierarchy that first attempt…
Hyundai Motor Group's autonomous driving subsidiary 42dot will debut an Nvidia-powered Level 2+ vehicle fleet in 2028 and launch its own Atria AI-powered Level 2++ fleet in the second half of 2029, CEO Park Min-woo said …
Researchers David Ha and Jürgen Schmidhuber built the first AI to learn by rehearsing inside its own generated "dream" simulation, solving the CarRacing benchmark with a score of 906 against a 900 passing bar and survivi…
A field note analyzing Michels et al.'s preprint "Vibe Coding: Practice, Performance, Productivity, and Risk" argues that AI code review effort should be allocated by code age and coupling rather than diff size, since AI…
Yifan Zhang published a 2026 technical report describing the Recurrent Looped Transformer (RLT), an architecture pairing a causal encoder with a recurrent decoder that carries its final hidden state and layerwise sliding…
OpenAI said in a blog post last week that it aims to build an automated AI researcher capable of advancing deep learning research, though it remains unsure "how to safely get all the way to aligned, full RSI," and descri…
Engineering teams deploying reinforcement learning with verifiable rewards (RLVR) post-training have found that building the training harness accounts for only 20% of the work, with the remaining 80% going to verifier de…
Enterprise teams including Shopify, Meta, Intercom, Ramp, Snowflake, and ONESTRUCTION have replaced frontier LLM API prompting with compact open models of 8B to 35B parameters trained via reinforcement learning against a…
Token Rhythm's NeoHorse-1-4B, a 4 billion parameter model built on Qwen 3.5, was trained on live router decision logs rather than a static dataset, with a stronger teacher model correcting its attempts in real time and t…
Edge0-35B-A3B, a 35B-parameter sparse mixture-of-experts model built on Qwen3.5-MoE, loses an average of 3.9 points across five benchmarks when quantized to 4-bit through its edge0 inference pipeline versus its fp16 base…
Qwen released Qwen-Drive-1.0-4B, a 4-billion-parameter vision-language model built on Qwen3.5-4B that fuses perception and planning into a single model with a bird's-eye-view perception head and two planner variants, pub…
Can Bölük forked Mario Zechner's minimalist Pi agent in late 2025 to build omp (Oh My Pi), a maximalist Rust-and-Bun harness that embeds ripgrep, AST parsing, hash-anchored editing, LSP, DAP, structured memory, and typed…