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 self-supervised method called Reflective Recovery, which turns failed reasoning trajectories into recovery training data, raised accuracy on DeepSeek-R1-Distill-Qwen-7B from 30.0% to 37.5% on AIME 2025 and from 37.6% t…
Researchers introduced ReaLMem, described as the first benchmark built from authentic multi-year personal visual archives with first-person subjective annotations, to evaluate AI models on long-term multimodal personal m…
A frontend-backend architecture that lets full-duplex speech-to-speech models delegate tool calls to a text-based backend LLM achieved 92-97% tool-call recall and 81.2% accuracy in rejecting irrelevant calls in a single-…
A new arXiv paper (2609.19366v1) reports that category-specific harm representations in large language models carry safety-relevant information beyond a shared general harmfulness direction. Using activation steering wit…
A Retrieval-Augmented Generation framework for scientific image quality assessment took 1st place in the SIQA-U understanding track of the SIQA challenge at the ICME 2026 Grand Challenges, according to arXiv paper 2609.1…
A new arXiv paper (2609.19425v1) reports that tool-augmented LLM agents hallucinate calls to tools that do not exist, and that no existing tool-selection or gating defense can reject them because a fabricated call is by …
A new arXiv paper (2609.19934v1) proposes the Depth Control Protocol (DCP), a diagnostic suite that disentangles three factors conflated by depth truncation when evaluating whether depth-recurrent language models genuine…
A study of 67,200 responses from GPT, Claude and Gemini found that the language of a question changes how the models assess the war in Ukraine, with the balance of Russia-leaning and Ukraine-leaning answers varying acros…
Researchers behind the arXiv paper DART: Distillation-Aware Reparameterization introduced a training-free method that combines low-rank coordinate transport with target-schedule response calibration to reuse LoRA adapter…
A cooperative framework called UnifiedPlayers, comprising a Planning Player, an Execution Player, and an Evaluation Player, outperformed the strongest prior baseline by at least 3.5% on mathematical reasoning and 3.9% on…
A new arXiv paper (2609.20068v1) unifies marginal utility, matrix factorization, and the transformer Key-Value cache under a single allocation rule that retains the top dimensions whose eigenvalue exceeds the shadow pric…
A new arXiv paper (2609.20152v1) introduces MTVA-Bench, a benchmark that evaluates the language model inside cascaded voice agents under the same conditions it faces in production, covering 49 agents, 490 reviewed scenar…
A decomposed confidence layer combining perception, layout, and validation channels with conformal risk control raised AUROC for financial-document key-value extraction from 0.54-0.74 with native Vision Language Model ve…
Anthropic announced that future Claude models will embed an invisible watermark based on Google DeepMind's SynthID-Text, a deployment with regulatory relevance under Article 50(2) of the EU AI Act, which requires provide…
VIDRAFT AI Research published a paper (arXiv:2609.20269) testing whether Transformer attention mechanisms need to repeat identically at every layer, using a 7×7 Latin square to spread seven sequence mixers evenly across …
Mozilla and Mistral announced a partnership bringing the 119-billion-parameter Mistral Small 4 model to Firefox Smart Window, replacing OpenAI's gpt-oss-120B in the feature's "Personalized" category. Firefox 155 opened S…
A developer has released @coder12-z/llm-shield, an MIT-licensed, zero-dependency Node.js library that wraps LLM API calls to handle intermittent failures such as transient 400 "could not parse JSON body" errors, truncate…
A developer has published a from-scratch Python walkthrough for building a retrieval-augmented generation (RAG) system, covering chunking, embeddings, a minimal vector store, hybrid retrieval, re-ranking, and cited gener…
Cooley built a proprietary AI product called GO Public on top of ChatGPT Work to compress the most labor-intensive phases of an initial public offering, the international law firm announced September 18, 2026. Cooley, wh…
OpenAI's InstructGPT/RLHF work showed that GPT-2-sized models (over 100x smaller than GPT-3) trained on the right task beat GPT-3, according to an essay by the author of The Bitterest Lesson. The essay argues that the pr…