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 (2611.07372v1) introduces EnvOpt, a method for automated end-to-end optimization of stationary agent environments using real or synthetic tasks without graders or reference answers. On AutomationBench, …
Researchers posted arXiv:2611.07482v1, introducing SAVeR$^2$, a safety alignment method for large reasoning models that decomposes the safety objective into four fine-grained verifiable reward signals and adds a reasonin…
A new arXiv paper (2611.07063v1) introduces SD-LoRA, a structural distillation framework that trains a heterogeneous graph teacher over high-resolution ViT tokens and transfers its structural bias into a lightweight LoRA…
Vis2State, a training-free agentic framework introduced in arXiv paper 2611.07397v1, abstracts visual puzzles into explicit, revisable problem states that link visual inspection and program execution, and achieves state-…
Nace AI released Drex, a family of four open-source decision models with 8B (DLM), 9B (v1.5), 35B with 3B active (v2.0), and 27B (Max) parameters that return a probability distribution over user-supplied options without …
Researchers introduced TB-MathModel, a terminal-style benchmark that makes 18 real CUMCM and MCM contest problems from 2023–2025 agentic and verifiable, scoring agents' single JSON artifacts against reproductions of Outs…
Researchers introduced Agent-R$^2$, a real-source data synthesis framework that distills 4,000 task instances from 10,000 web-corpus candidates to train agents on realistic productivity work, according to the arXiv paper…
A new arXiv paper (2608.11676v2) proposes XBRIDGE, a decode-free communication protocol that lets heterogeneous multi-agent LLM systems share information across different model families without losing entity identity. Th…
A new arXiv paper (2611.07216v1) introduces CMaG, a training-free un-alignment method that reduces compositional safety scores on Qwen2.5-VL-7B and LLaVA-1.5-7B by up to 65 points while over 98% of outputs remain coheren…
A preregistered mechanistic study of six open models from 14B to 675B parameters found that wrong sign tokens in multi-step LLM mathematics are committed late, by MLP layers in the last quarter of the network (70-100% pe…
A new paper (arXiv 2611.07071) proposes "mechanistic shaping," a framework that uses a model's internal representations to specify which concepts it should rely on and which it should express during training, rather than…
Anthropic has introduced message threads for its Messages API, an opt-in change that lets clients send only new messages plus a reference to the prior response instead of resending full conversation history. Anthropic re…
The IndiaAI Mission, run under MeitY, released three indigenous foundation models on November 3, 2026: Soket Labs' EKA-1, a 26.3B-parameter sparse hybrid model trained on roughly 7.5 trillion tokens, and ZenteiQ's Dhir 2…
Researchers Edward Misback, Eunice Jun, Erik Vank, Zachary Tatlock and Steven Tanimoto presented Codetations, a VSCode extension that attaches persistent, interactive annotations to arbitrary code spans without modifying…
A new LangGraph example in the open-source schemagate project cuts schema tokens sent to the model by 68% to 78% on a bundled 42-table demo database, sending 489 to 706 tokens per question instead of 2,199, with all thre…
A developer behind Pythos Oracle, a mathematics and physics AI system that pairs language models with deterministic math engines and independent verification, reported that the system solved all 50,000 problems in its 50…
Google has filed a patent application describing a method for ranking which input data signals actually improve an AI model by hiding one signal at a time — replacing its values with a null marker — and measuring the per…
A single-seed replication on Qwen3-1.7B found that the reverse-engineered Dimuon LoRA optimizer posted the lowest held-out eval loss at every adapter rank tested, reaching 0.4749 at rank 256 versus 0.4766 for tuned AdamW…
Microsoft's Foundry Local release lets developers run Qwen3.8-27B, a dense 27B vision-language model with 262K context and tool-calling, as a local OpenAI-compatible endpoint that GitHub Copilot can use via a bring-your-…
Open-weight models are the best-value option at just 3% of budgets on Village's LLM Atlas price-performance frontier as of October 31, 2026, down from 41% in June, with Google's Gemma 4 E4B and Xiaomi's MiMo-V2.6-Flash h…