Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A new arXiv paper (2609.10863v1) identifies a duality between continuous and discrete flow matching, showing that projecting continuous convex-interpolant paths with one-hot targets through a position-wise argmax yields …
Researchers introduced MHE-Former, a Transformer-based multi-hypothesis framework that uses entropy maximization to generate diverse 3D hand and body mesh recovery predictions from monocular input, according to a paper p…
A Kubernetes Dynamic Resource Allocation (DRA) driver makes composable CXL memory a schedulable cluster resource for cross-node KV-cache reuse in LLM serving, according to an arXiv paper (2609.10790v1). On a two-node clu…
A production skill router covering 34,396 skills shows that fine-tuning on synthetic data improves in-distribution skill retrieval for LLM agents but causes catastrophic forgetting on real and out-of-distribution data, a…
A new arXiv paper (2609.10613v1) proposes a posterior reweighting framework that models safety-aligned multimodal large language models (MLLMs) as implicitly operating over competing behavioral modes, explaining in-conte…
Researchers introduced SolCloudLLM, an LLM-based multimodal forecasting framework that fuses ground-based sky images with historical time-series data through bidirectional multimodal fusion to predict short-term photovol…
A new arXiv paper (2609.11699v1) introduces Negative Self-Distillation (NSD), a framework that improves large language model reasoning by diverging from self-generated flawed reasoning rather than imitating privileged gr…
Adobe has patented a self-correcting AI coding system that trains a large language model by pairing correct and incorrect code responses and fine-tuning on the resulting preference dataset, according to a patent filing t…
Salesforce has filed a patent for a server-side method that makes AI agents retrieve live business data before taking any action, grounding their decisions in current CRM records rather than stale or missing information.…
Datasette released versions 1.0a39 and 0.65.4 to patch security flaws found in an audit run with Claude Fable 5.1, GPT-5.6, and GPT-6 Astra, according to the project's release notes. The audit was prompted by issues repo…
A developer built a Graph RAG retrieval layer over a 50 million-row supply chain knowledge graph, capping subgraph expansion at 50 nodes and three hops after finding that four hops and larger context windows caused the m…
A post by lumpenspace and 99+ others on X argues that Anthropic's "hacking snafu" is as airtight a case as possible against the "misalignment" interpretation, and notes the release of 4 (FOUR) new Claudes. The post credi…
OpenAI Chief Executive Sam Altman told employees this week that the company is open to slowing development of its AI systems, according to sources familiar with the matter, a shift from its rapid frontier-model deploymen…
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…
OpenAI Chief Executive Sam Altman told employees at a company-wide meeting this week that OpenAI is open to slowing the development of its AI systems, Bloomberg News reported on Thursday, saying the ChatGPT maker could p…
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…
A developer who switched from Anthropic's Claude Code to OpenAI's Codex (Terra/Sol/Astra) reported that Codex's Astra model generated a one-line git commit message — "Fix YouTube resolver playback and isolate proxy traff…
Anthropic said Thursday it shut down multiple state-sponsored surveillance operations that used its AI models between January and July, originating in China, Iran and west Africa. The campaigns targeted pro-democracy fig…
A developer built the F-072 Rule Engine, a hybrid keyword-capture system that scans an LLM trading advisor's full output—including its internal chain-of-thought—for risk tokens such as '风险' (risk) and '接刀' (catching a fa…
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…