AGENTIC-SYSTEMSJuly 14, 20261 min read

Interpreting Latent CoT Reasoning as Dynamical Systems

Researchers have introduced a breakthrough framework to decode the "black box" of latent Chain-of-Thought (CoT) reasoning in AI. By applying dynamical systems analysis, the study reveals how models like CODI and COCONUT evolve their reasoning trajectories within the representation space.

Computer Science > Artificial Intelligence

Title:Interpreting Latent CoT Reasoning as Dynamical Systems

View PDF HTML (experimental)Abstract:Recent latent reasoning methods, such as CODI and COCONUT, face a fundamental interpretability problem: they maintain multiple superimposed candidate traces in the hidden space at each step, unlike explicit- CoT, which follows a single transparent reasoning trace. Existing mechanistic methods show compression, shortcuts, and superposition without explaining how reasoning evolves across latent steps. To address this gap, we model latent token sequences as trajectories in representation space and apply dynamical systems analysis to characterize the evolution of reasoning. Using quantitative measures, such as step-to-step change, direction consistency, and Lyapunov sensitivity, alongside qualitative projections, such as UMAP and DMD/PHATE, we show that latent CoT exhibits structured, non-random dynamics with two distinct stability classes. CODI behaves as a stable attractor, while COCONUT behaves as an unstable expanding system, and SIM-CoT supervision tightens both behaviors without changing the underlying dynamics. This framework advances the interpretability of latent CoT reasoning dynamics and provides actionable insights for improving latent reasoning performance. Code1 and Project page2 available online.

Current browse context:

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

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.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

Faithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent

A new study reveals that message formats in multi-hop LLM agent relays significantly impact data fidelity, especially for smaller models. While structured formats like JSON prevent information drift, they act as faithful, error-localizing channels rather than error-correcting codes.

agentic-systems

GES-TSP: Graph Edge Sparsification for TSP

Researchers have proposed GES-TSP, a learning-based graph edge sparsification method that accelerates solving large-scale Traveling Salesman Problems (TSP). This approach can prune up to 99% of unnecessary edges while maintaining near-optimal accuracy.

agentic-systems

The Verifier is the Curriculum: Execution-Gated Self-Distillation for Cross-Family Game Generation

The study introduces execution-gated self-distillation using a strict-launch filter to boost out-of-family game generation from 8.8% to 42.2% on GameCraft-Bench. It demonstrates that a precise, ungameable verifier acts as an effective curriculum for training code generators.

agentic-systems

Boltzmann MapReduce: A Partition-Function Reduce for Forkable Sandboxes

A new research paper introduces 'Boltzmann MapReduce', bridging statistical physics and distributed computing by modeling data chunks as thermodynamic systems to optimize the Reduce step.

NOW LET US Related – YUKTI: From Natural-Language Situations to Robust, Verifiable Decisions An Uncertainty-Typed Proposition IR, Assumption-Robust Pareto Frontiers, and a Regret Certificate

agentic-systems

YUKTI: From Natural-Language Situations to Robust, Verifiable Decisions An Uncertainty-Typed Proposition IR, Assumption-Robust Pareto Frontiers, and a Regret Certificate

Current LLM-based optimization pipelines are fragile due to their reliance on single-point assumptions. YUKTI addresses this by introducing an uncertainty-typed proposition graph and Assumption-Robust Pareto Frontiers, reducing decision regret by over 90% compared to naive approaches.

agentic-systems

From ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation

Researchers have proposed a new approach based on the Toulmin model of argumentation to structure and interpret AI-based image diagnoses. By combining specialized models like MedGemma and MedSigLip, the framework allows medical experts to critically assess ML-generated diagnoses rather than accepting them at face value.