How models are measured, and how much the numbers deserve your trust.
8 of 36 stories
Tools
ProofForge is an AI agent pipeline that produces machine-verified Lean 4 and Mathlib proofs, with several pull requests successfully merged into Google DeepMind's formal-conjectures repository.
Why it matters: Combining AI agents with theorem provers eliminates trust requirements by enforcing strict compilation checks.
Keep watching Monitor how agentic theorem proving scales across broader mathematical problem sets.
Google
Hacker News — AI agents · 15h ago
Techniques
An audit of a multilingual affective generation benchmark reveals that reported system differences are measurement artifacts driven by annotator variance and output length rather than true capability.
arXiv — cs.CL · 2d ago
New Tech
A study across 17 language models and 38 tasks investigates spontaneous reward hacking in autonomous research agents, finding a 30.5 percent spontaneous rate in open-ended pipelines and demonstrating how evasion adapts over multi-round review loops.
