OpenAI Finds 30% of SWE-Bench Pro Tasks Broken
OpenAI found that about 30% of tasks in SWE-Bench Pro are broken, according to a recent audit of coding benchmarks.
Research
190 articles
In-depth coverage of new AI research — papers, benchmarks, and breakthroughs from leading labs and academia, summarized for fast reading and grounded in the methods that matter.
LATEST NEWS
-
-
AWS Unveils BYOKG and GraphRAG for Intelligent Pharmaceutical Research
AWS introduces BYOKG and GraphRAG to address fragmented data in pharmaceutical research, with a 5 percent success rate in early-stage drug discovery.
-
Anthropic's Claude Fable 5 Leads New Industry Benchmarks
Anthropic's Claude Fable 5 topped all six new industry-specific performance indices from Artificial Analysis, despite being significantly more expensive than alternatives.
-
HuggingFace Introduces Atom2.7m for Arithmetic-Aware Language Models
HuggingFace's Atom2.7m model outperforms GPT-2 XL on ArithMark2.0, scoring 29.92% on one-operation expressions.
-
Anthropic's Jacobian Lens Reveals Claude's Hidden Inner Monologue
Anthropic's Jacobian Lens allows researchers to examine Claude's internal working memory, revealing how it processes concepts without explicitly stating them. The tool shows Claude can modify conclusions based on internal changes.
-
Mass Balance Launches Space Lab for Longevity Research
Mass Balance launched a grapefruit-sized lab into orbit on Tuesday to study disease-causing proteins in microgravity, aiming to improve drug development for age-related conditions.
-
Amazon Introduces Nova for Selective Model Unlearning
Amazon announced Amazon Nova, a customizable content moderation tool that reduces over-deflection by 53.74 percentage points on safety-related requests.
-
Hugging Face Teaches 15M French LLM to Think Deeper
Hugging Face researchers developed a 15M parameter French LLM that improves perplexity through entropy-based halting and looping, with results from a single training run.
-
NVIDIA Open Models Drive AI Research at ICML 2026
NVIDIA’s open models and infrastructure were cited in 145 ICML 2026 papers, highlighting their role in advancing AI research across multiple domains.
-
Tencent Hunyuan and Tsinghua Researchers Find AI Search Agents Fail at Asking Right Questions
A new benchmark from Tencent Hunyuan and Tsinghua University shows AI search agents struggle with ambiguous queries, achieving end-to-end accuracy below 50% in most cases.
-
26,000-Student Study Reveals AI Learning Losses Take Two Years to Surface
A 26,000-student study found AI users saw a 24 percent drop in exam scores, with full learning gaps emerging two years after first use.
-
UK AI Security Institute Finds Benchmarks Underestimate AI Agent Capabilities
A study by the UK's AI Security Institute shows standard benchmarks fail to capture AI agents' full potential, with success rates rising by up to 25% when given more computing time.