aiqbench analysis
This mechanistic study investigates whether Latent Chain-of-Thought (CoT) models genuinely perform step-by-step reasoning or rely on compressed shortcuts when handling sequential tasks. The research reveals that while simple tasks allow for faithful intermediate state representations, complex regimes cause the model to either compress computation into partial pathways or lose its reasoning signature entirely due to memory constraints imposed by the task's algebraic structure. These findings are critical for understanding the reliability and architectural limits of distillation-based approaches in advanced AI systems.
Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks
arXiv:2602.00449v2 Announce Type: replace-cross Abstract: Latent Chain-of-Thought aims to enable step-by-step computation without emitting long rationales, yet its internal mechanisms remain unclear. We study CODI, a continuous-thought teacher-student distillation model, on strictly sequential polynomial-iteration tasks with known intermediate states. Using logit-lens decoding, linear probes, attention analysis, and activation patching, we localize intermediate-state representations and trace how they are routed to the final readout. In short-horizon, low-hop tasks, CODI forms faithful bridge states across latent-thought positions, while the final input follows a separate near-direct route; predictions arise through late fusion at the answer readout. As task depth and difficulty increase, however, CODI does not reliably sustain a full latent rollout: it either compresses computation into a partial late-intermediate pathway or, in harder regimes, loses the latent reasoning signature altogether. To explain this transition, we show theoretically that the task's algebraic structure controls its effective memory: compressible regimes support late-bottleneck reasoning, while incompressible regimes preserve full-history dependence and destabilize latent rollouts. Overall, our results characterize when CODI-style latent-CoT yields faithful iterative computation versus compressed or shortcut strategies.
Title:Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks
Abstract:Latent Chain-of-Thought aims to enable step-by-step computation without emitting long rationales, yet its internal mechanisms remain unclear. We study CODI, a continuous-thought teacher-student distillation model, on strictly sequential polynomial-iteration tasks with known intermediate states. Using logit-lens decoding, linear probes, attention analysis, and activation patching, we localize intermediate-state representations and trace how they are routed to the final readout. In short-horizon, low-hop tasks, CODI forms faithful bridge states across latent-thought positions, while the final input follows a separate near-direct route; predictions arise through late fusion at the answer readout. As task depth and difficulty increase, however, CODI does not reliably sustain a full latent rollout: it either compresses computation into a partial late-intermediate pathway or, in harder regimes, loses the latent reasoning signature altogether. To explain this transition, we show theoretically that the task's algebraic structure controls its effective memory: compressible regimes support late-bottleneck reasoning, while incompressible regimes preserve full-history dependence and destabilize latent rollouts. Overall, our results characterize when CODI-style latent-CoT yields faithful iterative computation versus compressed or shortcut strategies.
Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2602.00449 [cs.AI]
(or arXiv:2602.00449v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2602.00449 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 277 words
Related coverage
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · Oct 8, 2026
Deep recurrent models are less robustly CoT-monitorable than normal CoT models in a toy setting
LessWrong · Sep 18, 2026
Generalization and infinite width
LessWrong · Aug 1, 2026
[AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment
Latent Space · Oct 16, 2026