Daily AI Paper

Today’s pick is CritICL, a paper about making language models reason better at test time without paying the usual cost of generating answer after answer or relying on external verifiers. The key idea is surprisingly simple: when smaller models fail, they don’t fail randomly. Their mistakes often follow structured patterns. CritICL turns those failure modes into useful guidance by converting them into critique-style in-context examples. In other words, instead of only showing the model what a good answer looks like, it also shows what kinds of bad reasoning to avoid. The method comes in two versions: one uses a global profile of common failures, and the other predicts likely failures for each input and retrieves matching critiques. Across benchmarks, it improves reasoning over standard prompting and competes with heavier test-time scaling methods, while using far fewer generations and tokens. That makes it especially interesting for anyone who wants stronger inference without much extra latency or cost.