AI models that reason silently instead of writing out their steps are getting closer to models that reason out loud.
Researchers studied implicit chain-of-thought (iCoT) methods, which try to get large language models to internalize multi-step reasoning without generating the visible scratchpad text that explicit chain-of-thought prompting relies on. iCoT models have historically lagged their explicit counterparts on accuracy. The team found that the hidden-state trajectories models use when reasoning internally follow a low-rank structure, meaning the reasoning process can be compressed without losing much information. Building on that, they built LoRi, a distillation framework that trains a student model to match a teacher's reasoning by aligning both in a shared low-rank subspace using first- and second-order statistics, rather than copying tokens one by one. Tested on LLaMA and Qwen models at multiple sizes on math reasoning benchmarks, LoRi beat earlier iCoT distillation methods and gained the most ground on harder, multi-step problems.
Explicit chain-of-thought prompting works, but it is slow and costly: every reasoning step is extra tokens, extra latency, extra spend. A method that recovers most of that accuracy without the verbose output matters for anyone running reasoning models at scale, especially on math and logic tasks where the chains run long.
Still, approaching explicit CoT accuracy is not matching it, and a model that reasons silently is also one you can no longer audit by reading its scratchpad.