A developer building Tabibu, a health-information assistant with a safety layer over a language model, ran OpenAI's HealthBench on the full pipeline and found it scored 0.362 versus 0.520 for the bare…
A new arXiv paper (2609.35797v1) finds that collapsing large language model judge rewards to pass/fail {0, 1} hides proportional changes in how well a response meets each criterion, and recommends kee…
The UK AI Security Institute (AISI) is publishing verified evaluation results through the EvalEval Coalition's Evaluation Cards platform, covering five benchmarks — HealthBench, FrontierMath, Humanity…
OpenAI's GPT-6 Astra shows only incremental gains over GPT-5.6 Sol on health benchmarks, with HealthBench Professional rising from 60.5 to 63.4 and HealthBench Hard from 33.1 to 36.3, while Consensus …
Researchers released HealthBench-Psych and HealthBench-Psych-Hard, a mental-health subset of OpenAI's HealthBench, comprising 610 clinician-validated conversations (12.2% of the original 5,000-physici…
OpenAI released GPT-5.6 Terra via its API on July 9, 2026, with availability confirmed by an administrator. The model scored 57.7 on HealthBench Professional, 57 on HealthBench, 32.7 on HealthBench Ha…
Researchers introduced Fully Open Meditron, the first fully open pipeline for building clinical large language models (LLMs), comprising a clinician-audited training corpus, reproducible framework, an…
A study evaluating AI tools on real clinical queries from physicians found that a specialized clinical tool (OpenEvidence) outperformed frontier general-purpose models (Claude Opus 4.8, Gemini 3.1 Pro…
China's medical AI achieved two milestones: Shanghai MicroPort MedBot's Toumai Remote surgical robot received EU CE mark approval, and Baichuan-M4 clinical AI topped OpenAI's HealthBench benchmark. Th…
OpenAI made GPT-5.5 Instant the default ChatGPT model on May 5, achieving a 52.5% reduction in hallucinated medical claims while matching frontier reasoning models on health benchmarks. The model scor…
OpenAI researcher Karan Singhal is leading the company's push to make ChatGPT a trusted healthcare tool, with over 230 million weekly users seeking health advice. Singhal, formerly of Google, is build…
Researchers developed a variance-aware reward framework using Group Relative Policy Optimization (GRPO) to improve heart-focused medical question answering in large language models. The approach, whic…
Researchers at Duke University introduced Reward-Variance Policy Optimization (RVPO), a risk-sensitive alignment method that penalizes inter-reward variance to prevent language models from neglecting …