10 results on this page
Researchers introduce an open-source benchmark separating olympiad geometry solving from accurate diagram construction. Across 954 problems, current foundation models achieved only a 36.14% average diagram compilation success rate, revealing a substantial gap between mathematical reasoning and faithful visual construction.
Multiple cybersecurity researchers reported suddenly losing access to OpenAI’s Trusted Access for Cyber (TAC) program. The program provides vetted users with models that have fewer guardrails for authorized cybersecurity research.
Google DeepMind researchers found that training AI agents through debate can reduce reward hacking when an LLM judge evaluates their work. On mathematics tasks, debate achieved higher sustained ground-truth accuracy than directly optimizing for LLM judge rewards.
Incogni researchers evaluated 13 AI platforms to assess the privacy risks associated with each service. The analysis suggests that larger platforms generally present greater privacy risks, with one notable exception.
Oxford researcher Carl Benedikt Frey argues that China’s biggest obstacle to large-scale AI deployment may be inadequate social protection for workers displaced by automation, not computing power or algorithms. The analysis compares China with historical British welfare systems and Denmark’s flexicurity model, while noting that the cited Chinese incidents do not prove a broad AI crackdown.
Researchers propose a Metamorphic Artificial Age Score (AAS) prototype that detects drone propeller health issues from flight-log data. It combines six normalized indicators to prioritize routine monitoring, maintenance review, or mandatory inspection across different fault severities.
A researcher applied the same SFT and GRPO recipe to three from-scratch LLMs ranging from 316M to 672M parameters, but observed sharply different outcomes. GRPO barely affected the smallest model, severely degraded the middle model, and caused modest degradation in the largest, with no GSM8K transfer despite curriculum learning.
A security researcher discovered that more than 9 million face photos were stored in an unsecured ClarityCheck database. The database reportedly required no password to access.
Google is working with the UK government on Operation Blue Skies to address aviation contrails with AI. The initiative will subtly reroute flights to test whether contrail formation can be avoided.
A Reddit discussion thread tracks the expected release of EMNLP 2026 notifications and results. The post invites researchers to share updates while hoping to attend the conference in Budapest.