Editorial Desk
Papers, methods, and deep-dives from frontier labs.
Research Wire summarises and contextualises the most consequential AI research — new architectures, training advances, safety evaluations, interpretability work — distilled into reads aimed at practitioners. Each piece links to the source paper and surfaces the empirical claims that drove the headline.
24 articles on this page
Fields Medalist Terence Tao published a ChatGPT conversation showing the model assisting with algebraic calculations while he explored a potential counterexample to the Jacobian Conjecture.
Allen Institute for AI unveiled DiScoFormer, a transformer model cutting high-dimensional density error by 37x without requiring retraining per dataset.
OpenAI's review of SWE-Bench Pro found that a large share of model failures stem from flaky tests, environment mismatches, and bad ground truth — not actual coding mistakes.
ServiceNow AI Research evaluates leading ASR models on bilingual enterprise audio, uncovering significant accuracy drops on code-switched speech.
DeepSeek-AI's DSpark paper reports a 2-3x throughput gain over standard decoding using a hybrid draft-and-verify approach, tested on models including DeepSeek-V2 and Llama 3.
Allen AI researchers published findings showing Olmo Hybrid beats pure transformers on content tokens while revealing trade-offs in token prediction accuracy.
Anthropic reveals Claude authors over 80% of its codebase, driving an 8x code output increase and demonstrating early recursive self-improvement.
Dharma AI researchers successfully applied Direct Preference Optimization to OCR models, cutting text degeneration by an average of 59.4%.
Artificial Analysis and IBM launch ITBench-AA, exposing that frontier AI models score under 50% on agentic Kubernetes site reliability engineering tasks.
Google Antigravity 2.0 took top honors in ModelRift's OpenSCAD benchmark, proving the power of text-based parametric models for complex spatial tasks.
A hackathon-built open-source framework, OncoAgent, reports fine-tuning dual-tier clinical models on 266,854 cases in 50 minutes using AMD MI300X hardware, per its own published paper.
OpenAI has introduced a WebSocket-based version of its Responses API that keeps connections open to cut per-call overhead and adds connection-scoped caching for agentic workflows.
NVIDIA and Hugging Face have released Nemotron-Personas-Korea, offering 7 million synthetic personas grounded in South Korean census data for culturally accurate AI.
NVIDIA releases Nemotron OCR v2, utilizing 12.2 million synthetic images to achieve 34.7 pages per second throughput across six different languages.
An independent benchmark reveals Anthropic's Claude Opus 4.7 tokenizer uses up to 47% more tokens on technical files, driving session costs up 20% to 30%.
IBM Research released VAKRA, an execution-centric benchmark with 8,000 APIs across 62 domains to diagnose AI agent failures in complex enterprise tasks.
New research reveals a major weakness in today's most advanced vision AI models: they often can't cite the specific data source for their answers. A new benchmark called ViTaB-A highlights this critical attribution gap. Why does this matter for AI trust?
Researchers have developed CrispEdit, a new algorithm to edit large language models without corrupting their core capabilities. How does this solve the key challenge of 'proxy hacking' in AI?
Andrej Karpathy detailed an LLM Wiki framework that replaces dynamic RAG with persistent, agent-maintained Markdown files that compound knowledge over time.
Researchers introduced EpiScreen, an LLM framework that detects early epilepsy from health records, achieving up to 0.980 AUC and boosting expert diagnostic performance by 10.9%.
A new research paper reveals a GPU acceleration method that boosts transformer inference speed by 64.4x. The technique slashes memory usage by 63%, enabling real-time performance for models like BERT and GPT-2. Can this finally make low-latency AI a reality?
Researchers introduce the R3 framework to solve the conflict in multimodal AI where boosting visual generation weakens model understanding capabilities.
UC Berkeley researchers created a simple 'copy-paste' bot that scored 97 on a major AI benchmark, outperforming GPT-4o. This stunning result exposes critical vulnerabilities in how we measure AI progress. What does this mean for the future of AI testing?
A new arXiv paper proposes GlobeDiff, a diffusion-based model that helps multi-agent AI systems infer a global environment state from each agent's limited local observations.