Researchers led by Zhaoyang Wei introduced CALR, a way to link hidden model states with answer supervision. The method compresses derivations into latent blocks that stay tied to the original problem. Tests on five mathematical reasoning benchmarks showed CALR outperformed a similar continuous latent approach by 26 percentage points when using the same compute budget. The paper was posted on arXiv on Oct 7 2026.
Why it matters
The improvement suggests AI systems can solve math problems more reliably without extra compute, which could benefit educational tools.
