A new technique lets researchers steer how AI models reason by targeting a small set of stable geometric directions buried inside their hidden states.
Researchers studying "latent reasoning" models — systems that run multi-step inference through hidden-state space rather than generating visible chain-of-thought text — found that the updates separating strong reasoning from weak reasoning cluster into a surprisingly compact, low-rank structure. Everything outside that structure turned out to be noisy: sensitive to how a question was phrased, which model checkpoint was used, or minor perturbations in the reasoning trajectory. They built a training-free framework called Trajectory-Invariant Latent Refinement (TILR) that first maps this stable subspace from contrastive trajectory pairs, then constrains interventions to it while an adaptive gate suppresses misaligned updates. Tested across six reasoning benchmarks, TILR improved answer consistency under paraphrase by roughly 10 percent and cut latent trajectory variance by up to 50 percent without degrading accuracy.
The result matters because latent reasoning models are increasingly attractive for inference efficiency — they skip the token-by-token scratchpad — but their internals have been opaque enough to make targeted improvement guesswork. If a handful of directions really do explain most of the gap between good and bad reasoning, that is a foothold for interpretability and for surgical fine-tuning that does not require retraining from scratch.
The geometric framing here echoes earlier work on linear representation in language models — the idea that concepts and behaviors live along predictable axes in embedding space. Whether those axes stay stable as models scale, or whether TILR's low-rank picture breaks down at frontier scale, is the next question the authors leave open.