PAWBench: How Far Are We from Probabilistically Aligned World Modeling?

A researcher trained a small transformer in 1.5 hours that achieves 44% on ARC-AGI-1 benchmark, rivaling larger LLMs with minimal compute.

Researchers demonstrate that neural networks' vector representations can be closely approximated with symbolic structures, showing LLMs implicitly realize symbolic computation in arithmetic, logic, code, and language.

Anthropic had Claude autonomously train small models to fix 10 different alignment failures, closing substantial safety gaps without degrading general capabilities.

AI agents can now generate security exploits from just rumors of bugs, rendering traditional security embargoes ineffective as attacks precede patches.

Researchers introduce SCILAWS-BENCH, a benchmark with 118 problems across six disciplines to evaluate LLMs' ability to discover scientific laws from real data and synthetic hidden laws.