Research
Papers, benchmarks, and the science moving the field forward. ·
DeepMind’s Gemini 3 Ultra Wins Gold at the International Math Olympiad
For the first time, a general model solved all six problems under full competition conditions.
ResearchDeepMind’s Gemini 3 Ultra Wins Gold at the International Math Olympiad
For the first time, a general model solved all six problems under full competition conditions.
ResearchThe Quiet Rise of Small Language Models That Run Entirely on Your Phone
Sub-3B models are closing the gap fast, and they never send your data anywhere.
ResearchThe Quiet Rise of Small Language Models That Run Entirely on Your Phone
Sub-3B models are closing the gap fast, and they never send your data anywhere.
ResearchNew Paper Shows Chain-of-Thought Can Be Faked by Smaller Models
Reasoning traces don't always reflect what the model actually computed.
ResearchSparse Attention Returns: A 100x Context Window at Constant Cost
A revived idea makes million-token prompts economically boring.
ResearchDeepMind Details How It Cut Hallucinations With Verifier Models
A second network fact-checks the first before it answers.
ResearchBenchmark Contamination Is Worse Than Anyone Admitted
A new audit finds test data leaking into most popular suites.
ResearchStudy: Fine-Tuning on Synthetic Data Beats Scraping the Web
Curated generations outperform noisy human text.
ResearchInterpretability Team Maps the ‘Deception’ Circuit in a Frontier Model
Researchers isolate where a model decides to withhold.