Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Veilus Digital released the cryptographic core of its Phantom Chat messenger as source-available code on GitHub, allowing independent verification of its post-quantum security claims. The package includes implementations…
Google fired a 7-year veteran engineer, Justin Poehnelt, after he built an open-source, agent-ready CLI for Google Workspace that went viral on Hacker News. The tool, now hosted under Google's own GitHub org, dynamically…
AI engineering faces a contradiction between paying premium for frontier models' reasoning capabilities and aggressively compressing context to reduce costs, leading to a 'fallacy of context compaction' that starves agen…
A developer compared API pricing versus self-hosting costs for open-source AI models, finding that for small projects with 1 million tokens per day, using an API is 32 times cheaper than self-hosting. The analysis reveal…
At the Cannes Lions festival, marketers are shifting focus from creative campaigns to brand coherence as AI systems like LLMs synthesize inconsistent brand messages across channels. The real crisis is not about SEO but a…
Chinese AI firm Zhipu (Knowledge Atlas Technology) is considering a multibillion-dollar share sale in Hong Kong after its stock surged nearly 2,000% since its January 2026 IPO. The company, which went public at HK$116.20…
Researchers introduced EXPO-SQL, a reinforcement learning method that assigns clause-level rewards to improve text-to-SQL query generation. By analyzing execution results and error messages, the method outperforms existi…
Researchers introduced the Normalized Context Utilization (NCU) metric to quantify contextual information gain in Retrieval-Augmented Generation (RAG) systems, revealing that small language models (SLMs) match or outperf…
Researchers at arXiv found that self-generated text recognition (SGTR) finetuning can prevent and reverse emergent misalignment in large language models, outperforming other interventions. The study across GPT-4.1, Qwen2…
Researchers developed a framework using large language models to quantify product desirability from qualitative feedback, achieving up to 0.97 Pearson correlation and 94% accuracy on two Product Desirability Toolkit data…
Researchers propose a training-free IBA framework for Knowledge-Based VQA that decouples entity identification from evidence re-ranking, outperforming fine-tuned baselines on Encyclopedic-VQA and InfoSeek while reducing …
A new study auditing eight automatic attribution metrics across multiple datasets finds that no single metric consistently performs best, with rankings inverting across datasets (Kendall tau = -0.64). The instability lea…
Researchers introduced QuechuaTok, a benchmark evaluating tokenization strategies for Southern Quechua, a low-resource agglutinative language. They found that BPE achieved the lowest fertility rate but only 6.67% morphol…
Researchers introduced RASC+, a retrieval-constrained LLM adjudication method for clinical value set authoring, achieving a candidate-pool recall of 0.730 on the RASC benchmark. Using GPT-5 as a constrained adjudicator, …
Researchers compared six offline reinforcement-learning methods for distilling reasoning from large language models into smaller ones, finding that SFT, RFT, and RIFT produce nearly identical weight updates and accuracy,…
Researchers propose Prompt-Vision Token Activation Map (PV-TAM) to improve vision-language consistency in large vision-language models by using prompt-side semantics and filtering systematic bias from modality boundary m…
Researchers introduced HANCLIP, a family of vision-language models that restructure the embedding space to handle negation by combining hyperbolic and angular objectives. Trained on 20,000 image-text quadruplets, HANCLIP…
Researchers propose Head-Wise Representation Alignment (HeRA), a method that aligns individual attention heads in multimodal large language models to improve cross-modal representation. HeRA uses a contrastive objective …
Researchers introduced spec learning, a framework that uses a brief user instruction and a small set of preference judgments to compile natural-language specifications for LLMs. These specifications condition models at i…
Researchers introduced REALM, the first unified red-teaming benchmark for physical-world vision-language models (VLMs), integrating 12 attack methods, 3 defenses, and 13 VLMs under a black-box threat model. The benchmark…