From Vision to Language: Where VLMs Actually Fuse What They See
A Findings of EMNLP 2026 paper uses causal interventions to pinpoint where vision-language models fuse visual information — mainly when reading answer options.
AI Research
Read Full News
The ‘Threat-Model Coverage Gap’: Why Automated Red-Teaming Isn’t Enough
A new paper argues automated red-teaming beating humans on cost and count doesn’t retire humans — benchmarks measure search thoroughness, not harm coverage.
“Don’t Drop Dropout”: The Case for Layer Sparsity in Efficient LLM Training
A Cerebras paper argues layer dropout belongs in state-of-the-art LLM training: up to 25% training FLOP savings and 1.5× faster inference, 2,400 experiments.
Google’s AI Overviews Cut Publisher Traffic Without Making Search Better, Study Finds
A preregistered experiment with 1,100 Google Search users found AI Overviews and AI Mode reduce clicks to publishers — without improving user experience.
Researchers Stole Hidden AI ‘Reasoning’ From OpenAI, Anthropic and Google APIs. All Three Have Now Patched It.
A paper from ELLIS Institute Tübingen and MPI showed encrypted chain-of-thought from major AI APIs could be stolen cross-session. Vendors have patched it.
Anthropic Raised Its Own AI Misalignment Risk Rating. The Reason Is Uncomfortable.
Anthropic’s August 2026 risk report raised catastrophic misalignment risk from ‘very low’ to ‘low’ — driven by saturated safety benchmarks, not a failure.