Grammar-Constrained Decoding (GCD), a technique for ensuring syntactically correct code, opens a new jailbreak method for attackers with a success rate over 30 percentage points higher than previous approaches.
The security filter in Claude 3.5 Sonnet blocks legitimate security requests, limiting its usability for CTOs performing security audits and vulnerability assessments.
Trust in AI does not emerge automatically but must be systematically built through explainability measures depending on the application context and regulatory requirements.
Anthropic calls for an aviation-like regulatory authority or commissioned private auditors to examine AI models for critical risks before their release.
InternVideo3 enables foundation models to analyze longer video sequences with iterative reasoning and tool use while avoiding efficiency problems in KV cache management.
Arbor enables AI-driven research through systematic hypothesis management and achieved an average of 2.5x higher improvements than existing code models on six test tasks.
Arbor coordinates autonomous AI agents via persistent hypothesis trees and achieved 2.5× better results than Codex and Claude Code on six research tasks.
Bebop uses rejection sampling and TV loss optimization to maintain stable MTP acceptance rates during RL training and accelerates rollouts by up to 1.8x.
