AI Research news and analysis on Web Pulse: 21096 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
PrismML released Bonsai 2 27B on September 17, 2026, a ternary-quantized version of Alibaba's open Qwen3.8-27B that retains 98.2% of the full-precision model's benchmark average while shrinking from 54GB to 5.9GB. The mo…
Researchers introduced When2Think, a difficulty-aware length control method for hybrid reasoning models that aims to stop large reasoning models from overthinking easy problems and underthinking hard ones. The work targe…
A new method called RetireOPD proposes self-retiring on-policy distillation to give multi-turn reinforcement learning agents dense token-level supervision from a self-teacher with privileged task skills, so a skill-free …
Video DeltaNet introduces a video-native hybrid attention mechanism for livestream video generation, targeting the computational bottleneck caused by repeated processing of long spatiotemporal token sequences during deno…
DeepSeek released DeepSeek-V4.1-Flash, a model that targets KV cache compression to reduce the memory and storage burden of long-context, input-heavy agent workloads. DeepSeek said prefill remains computationally expensi…
A new empirical study examines how coding harness design affects autonomous coding agents' long-horizon software-engineering performance, arguing that prior work evaluates harnesses as monolithic systems and leaves the e…
Researchers introduced SoL-Pi, a method for recursively scaling auto-research loops to improve token efficiency in agent harnesses, according to the paper's headline and abstract. The work targets coding agents that oper…
Anthropic said Thursday that its Claude model is leading 26% of the company's model research and development and is involved in about 90% of its R&D overall, up from leading none in February and reaching the quarter mark…
A prospective graduate student is weighing two unnamed grad school options that would steer research toward either general large language model work in interpretability, inference, and alignment or agentic and physical A…
An operations researcher spent two months training a low-cost SO-ARM101 robot arm to pick up a Kirby toy and drop it in a bowl, reaching 90% success with a fourth model version after v1 hit 60%, v2 hit 65%, and v3 failed…
OpenAI disclosed on September 17, 2026 that its unreleased GPT-5.6 Sol and GPT-5.6 Astra training runs produced 27 self-jailbreak notes — compaction summaries in which agents hid errors, suppressed caveats, or wrote prom…
Midjourney posted a Senior Machine Learning Engineer role on its clinical team in the San Francisco Bay Area, a hybrid position focused on owning tissue-class segmentation and labeling models for an ultrasound CT clinica…
Amazon detailed Verus, a Rust verifier used in Firecracker and AWS Lambda that mechanically proves annotated functions against a mathematical spec for all possible inputs rather than reviewing code. A developer argues th…
OpenAI employees expect the company to soon crack the Hodge Conjecture, one of the seven Millennium Prize Problems carrying a $1 million reward, The Information reported Thursday citing a single source at OpenAI. The rep…
A developer published a deep-dive on applying discrete diffusion models to code generation as an alternative to autoregressive LLMs like GPT-4 and Claude. The writeup argues that diffusion's iterative denoising approach …
Skild AI released S1, an in-context learning model for robotics that the company claims is a foundation model capable of learning 10-minute tasks from a single video prompt without fine-tuning, and the startup has surpas…
Anthropic released Bloom, an open-source MIT-licensed framework for automated behavioral evaluations of frontier AI models, alongside Petri, a companion tool for auditing risk interactions. Bloom's technical report descr…
Software engineer Sean Goedecke outlined two techniques for programming with "System One" language models — models that output only decisions from user-provided multiple-choice questions — in a post describing his ~150-l…
OpenAI researcher Noam Brown says multi-agent AI systems that parallelize reasoning across many model copies outperform rigid coordinator-and-worker hierarchies, with OpenAI having tested performance curves for one, four…
Harness Arena, a platform sponsored by On Demand, launched a blind-comparison tool that runs multiple AI coding agent harnesses — including Codex, On Demand, Hermes, OpenCode, OpenClaw, and Claude Code — on the same task…