analiza aiqbench
This research introduces C-MTP, a direct supervision method for continuous Chain-of-Thought models that compresses verbose reasoning traces into dense latent representations more efficiently than prior indirect approaches. While the technique performs competitively with simplified tasks under 100 tokens, it reveals significant limitations when scaling to complex problems requiring hundreds of reasoning steps, where current methods suffer up to a 65% performance drop. The study highlights critical challenges in maintaining reasoning fidelity during compression as models attempt to balance speed and accuracy in advanced inference scenarios.
Training Continuous Chain of Thought Models: A Tale of Two Regimes
arXiv:2607.16972v1 Announce Type: new Abstract: Continuous Chain-of-Thought methods replace verbose reasoning traces with a short sequence of dense latent representations. Earlier continuous CoT methods indirectly supervise the latent representations such that its final state match that of verbose reasoning traces, requiring autoregressive, slow generation during training. We introduce C-MTP, a simpler, faster direct supervision approach that models each latent as an average of the embeddings in the CoT traces to be compressed. Our approach outperforms a prior direct supervision method that approximates the distribution of compressed tokens, and performs competitively to slower indirect supervision approaches in existing evaluation setup with simplified CoT traces (less than 100 tokens). Lastly, we extend the evaluation of Continuous CoT methods to complex tasks with longer reasoning traces ($\ge$ few hundreds reasoning tokens). We find both direct and indirect supervision training methods perform poorly (roughly 65\% performance drop) in this setting, revealing the limitations of current continuous CoT methods. The code and checkpoints are released at https://github.com/Varun221/cmtp_research
Abstract:Continuous Chain-of-Thought methods replace verbose reasoning traces with a short sequence of dense latent representations. Earlier continuous CoT methods indirectly supervise the latent representations such that its final state match that of verbose reasoning traces, requiring autoregressive, slow generation during training. We introduce C-MTP, a simpler, faster direct supervision approach that models each latent as an average of the embeddings in the CoT traces to be compressed. Our approach outperforms a prior direct supervision method that approximates the distribution of compressed tokens, and performs competitively to slower indirect supervision approaches in existing evaluation setup with simplified CoT traces (less than 100 tokens). Lastly, we extend the evaluation of Continuous CoT methods to complex tasks with longer reasoning traces ($\ge$ few hundreds reasoning tokens). We find both direct and indirect supervision training methods perform poorly (roughly 65\% performance drop) in this setting, revealing the limitations of current continuous CoT methods. The code and checkpoints are released at this https URL
Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2607.16972 [cs.AI]
(or arXiv:2607.16972v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2607.16972 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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.
Text integral sub licența deschisă a publicației · 257 cuvinte
Subiecte conexe
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · 5 oct. 2026
The Illusion of Superposition? A Principled Analysis of Latent Thinking in Language Models
arXiv · 6 aug. 2026
Projection sampling preserves capabilities but diminishes gains from alignment fine-tuning
LessWrong · 16 oct. 2026
Problems of Proliferation
LessWrong · 29 sept. 2026