AI reasoning models think out loud too much — and a new paper argues existing fixes are making the wrong cut.
Researchers have identified two separate sources of bloat in Chain-of-Thought reasoning traces. The first is internal redundancy: the model spins its wheels before landing on a correct answer. The second is external redundancy: it keeps talking after it already got there. Most current compression approaches treat these as one problem and apply a single length penalty across the board. The new paper proposes a dual-penalty reinforcement learning framework that targets each type independently — a sliding-window similarity metric flags low-progress segments mid-reasoning, while a separate metric discourages post-answer rambling.
The results are hard to dismiss. On 1.5B and 7B parameter models tested across GSM8K, MATH500, and AIME24 benchmarks, average reasoning length dropped by roughly 41% and 40%, respectively, with accuracy holding up. The compression behavior also transferred to out-of-domain tasks including GPQA and LiveCodeBench — a sign the model learned something general rather than just memorizing shorter outputs.
The more interesting finding is the asymmetry: external redundancy can be stripped out with almost no accuracy cost, but trimming internal redundancy involves a real trade-off. In other words, getting a model to stop talking after it finishes is easy; getting it to stop meandering before it finishes is genuinely hard. That distinction matters as labs race to cut inference costs on reasoning-heavy models — it suggests a clear low-hanging-fruit target before anyone has to touch the harder problem.