The Continuous Thought Machine: A Stability‑Governed Architecture for Coherent Latent Reasoning
A stability‑governed reasoning architecture that evaluates latent trajectories in real time.
The Continuous Thought Machine (CTM) introduces Clarus, a training‑free, architecture‑agnostic stability layer that evaluates the geometric stability of a model’s hidden‑state trajectory during inference. Clarus computes three activation‑space quantities—curvature, directional drift, and manifold membership—to define a stability manifold and a basin of coherent reasoning. As the paper states, Clarus “measures hidden-state evolution in real time” and detects departures from stability 3–12 steps before text‑level errors appear .
Coupling Clarus to any autoregressive model yields the Continuous Thought Machine, a two‑pole reasoning architecture where the generative pole proposes latent transitions and the stability pole evaluates their coherence. This separation enables early drift detection, regime classification, and stability‑regulated decoding “without modifying model parameters and without interpreting content” .
Across GPT, LLAMA, and Mistral models (7B–70B), Clarus outperforms entropy and token‑loss signals, reduces drift events, and improves long‑horizon reasoning accuracy by 5–15% via soft, non‑intervening decoding. The paper shows that all major LLM failure modes reduce to four geometric instabilities—curvature escalation, drift instability, manifold exit, and failed recovery—providing the first unified, semantics‑free theory of reasoning failure.
Download paper as pdf
Download paper as pdf