← Home

Topic: Research

1414 stories · page 15 of 29

• Two-year study finds banning AI in class leaves students worse off2026-09-13
• StochBench benchmark scores Opus 4.8 agent at 34.9% proof rate2026-09-13
• Reflective publishes a road map for deciding on solar geoengineering2026-09-13
• PlannerForge automates scenario-based testing for autonomous driving systems2026-09-13
• OpenAI accused of pressuring researchers as math feud escalates2026-09-13
• New method audits LLM bias from hidden states, not outputs2026-09-13
• New framework closes rare-entity gap in multilingual entity linking2026-09-13
• MIT Technology Review names five under-35 biotech innovators2026-09-13
• Mi-Ripple removes digital ripple from repeated AI edits2026-09-13
• IdeaAMBIG benchmark shows LLMs miss gaps in research specs2026-09-13
• GPT-6-Astra cheats at chess every time, Fable sometimes refuses2026-09-13
• GLIE compresses visual document retrieval to four vectors per page2026-09-13
• GenV verifier catches proof translations that fool solvers2026-09-13
• FreeFlow drops hand-built biases from optical flow, still hits SOTA2026-09-13
• DF26 benchmark finds deepfake video detectors near random chance2026-09-13
• CARDEA vision-language model adds auditable reasoning to coronary angiography2026-09-13
• Anthropic resignation adds to AI researchers' extinction fears2026-09-13
• ActReview trains an 8B model to write actionable peer reviews2026-09-13
• X-AuT prunes Qwen3-ASR's audio encoder and lowers its error rate2026-09-12
• World in World lets frozen video models explore new camera views2026-09-12
• UniH3 achieves state-of-the-art all-in-one medical image restoration2026-09-12
• Tiny Aya L2-Thinker tops 93% in-language reasoning across 60 languages2026-09-12
• Timnit Gebru argues AI doom talk distracts from real harms2026-09-12
• SyncWorld simulates robot action outcomes zero-shot in unseen setups2026-09-12
• Survey maps ways to cut inference costs in VideoLLMs2026-09-12
• Starlink leakage floods SKA-Low's key radio frequencies, study finds2026-09-12
• Recursive Code World Models build 3D scenes from one image2026-09-12
• Qwen3-Coder and Gemma 4 regress by 4 to 30 points from imitating an expert2026-09-12
• Nemotron 3 Ultra reaches gold-medal score at IMO 20262026-09-12
• Negative self-distillation trains LLMs to avoid flawed reasoning2026-09-12
• MaP-WAM turns robot memory into plans, hits 83.3% on RMBench2026-09-12
• Image tokenizer choice can affect text modeling, study finds2026-09-12
• Former DeepMind VP Vinyals doubts sudden intelligence explosion, launches Discovery Loop2026-09-12
• Fine-tuned 4B Qwen 3.5 beats GPT-5.6 on MetroLLM-Bench2026-09-12
• Discovery Certification Protocol audits AI research agents' claimed discoveries2026-09-12
• 25 Fields Medal winners warn AI benchmarks harm mathematics2026-09-12
• SWE-Bench Pro Verified closes reward-hacking loopholes: some models score lower2026-09-11
• SpatialBlock-15k trains vision-language models on 3D spatial reasoning2026-09-11
• SenseNova-U1.5 unifies image understanding and generation without encoders or VAEs2026-09-11
• SchemeArena finds instrumental goals drive AI agent scheming2026-09-11
• Rule-chaining framework solves 230 of 240 ARC-AGI-2 tasks2026-09-11
• NCP-ArchPreview matches OLMo-3-7B's pretraining loss on 51.3% of tokens2026-09-11
• Halo improves point forecasts by also estimating their uncertainty2026-09-11
• Google's ToolGrad flips tool-use data generation to answer-first2026-09-11
• GeoSteer replaces one-step LLM activation steering with geodesic optimization2026-09-11
• Gemini's rewrites turn a Chandler passage into horror after 31 iterations2026-09-11
• Federated fire detection removes single point of failure with rotating coordinator2026-09-11
• Data size drives grokking onset far more than model width2026-09-11
• Complete reasoning traces add little value in LLM post-training, study finds2026-09-11
• Φ-Bench tests whether LLMs can engineer the infrastructure that runs them2026-09-11

All topics · Archive by month