Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Google released GLM-5.2 on 2026-06-13 under the model identifier zai-org/glm-5.2-maas, offering a 1000k-token context window and 64k maximum output, with no retirement date announced. The model listing on models.dev poin…
SK hynix showcased its next-generation AI memory portfolio under the slogan 'New Spectrum' at the AI Infra Summit 2026, held September 15th-17th in Santa Clara, California, where the company doubled the size of its booth…
A developer surveyed five coding agents — Claude Code, Cursor, Codex, VS Code, and Windsurf — and found that all of them already support resuming a prior conversation, so the common complaint that agents "forget" history…
A proposed zeroth-order paradigm for LLM preference alignment addresses likelihood displacement, a problem that motivates alternative ways to extract information from preference pairs with small likelihood margin. The ap…
A method called SpectralShift extends the effective context window of Gated DeltaNet, a linear attention architecture used in place of softmax attention for long-context modeling, by reparameterizing the model's spectral…
Google launched the Gemini 3.8 Live and Gemini 3.8 Live Thinking AI models, adding real-time audio transcription, reasoning and translation plus multi-step task handling, the company said on 17 September 2026. Google rep…
A developer's analysis of LLM agent orchestration found that idle parent agents on Codex burn tokens polling for worker completion because the runtime lacks event-driven wake-ups, with one run issuing 89 waits that consu…
Users of the Assistant product reported that reasoning models disappeared from the model picker, with one ultimate subscriber saying they could no longer select Sonnet with reasoning. A respondent said the reasoning and …
HypoEvolve applies genetic algorithms to multi-agent large language models to discover scientific hypotheses, according to the research. The system combines scientific agents with evolutionary search through critique, co…
A new arXiv paper (arXiv:2609.13009) by researchers including Ali Ansari, Haoran Sun, Andy Zeyi Liu, Mark Jabbour, Lucas Baker and Arman Cohan, with audits by Yale physics faculty and graduate researchers, finds that lea…
A new arXiv paper, GraphEcho (arXiv:2609.17695v1), introduces a benchmark showing that large language model agents can mistake repeated graph paths for independent corroboration, with redundant supporting paths increasin…
A new arXiv paper (2609.17804v1) reports that large language models solve grade-school math word problems through a four-stage sequential pipeline — Schema Abstraction, Operation Planning, Operand Binding, and Computatio…
SAGE, a governed multi-stage LLM pipeline built around a shared versioned rule store, cut enterprise guideline-to-artifact turnaround from two to three days of manual work to 20-100 minutes across 120 documents, accordin…
A new arXiv paper (2609.17865v1) introduces SAFE, a controlled benchmark testing whether frontier models acquire safety-relevant evidence before making deployment decisions. Across GPT-5.5, o3, Claude Opus 4.8, and Claud…
A study measuring 54 configurations of Qwen2.5-7B-Instruct on vLLM 0.12 across L4, A100, and H100 GPUs found that 18 of 36 configurations reach the cost, quality, and latency Pareto frontier, with combined optimization m…
A training-free routing policy called the Tri-Metric Router achieved 0% out-of-memory failures and 88.5 ± 4.4% oracle alignment on out-of-distribution holdouts for retrieval-augmented generation on commodity GPUs, accord…
Researchers proposed OBC-Prune, an outcome-based calibration method for pruning large reasoning models that weights reasoning tokens by their causal contribution to correct answers rather than by uniform statistical sali…
A study posted to arXiv (2609.17532v1) uses large language model-derived features from free-text respiratory therapy notes plus a logistic regression pipeline to improve prediction of extubation failure (EF) in invasivel…
Researchers released DantinoX, an open-source JAX/Flax library that runs three language-generation paradigms — autoregressive decoding, discrete masked diffusion, and continuous flow-matching — on a single modular Transf…
A systematic benchmark of open-source instruction-tuned LLMs for key-value pair extraction found that models including Gemma, Mistral, Qwen2.5, LLaMA 3, and DeepSeek approach supervised layout-aware systems on clean text…