The article explores latent reasoning as an alternative to chain-of-thought in AI, categorizing five families of approaches and discussing implications for AGI and interpretability.
After following various arXiv papers and researcher discussions on X/bluesky about latent reasoning and continual learning, one idea which resonates strongly is that path forward (towards AGI) may depend less on generating ever-longer chains of thought and more on finding architectures that can reason beyond the token stream. LLMs routinely reach correct answers through flawed or fabricated CoT steps, and produce perfectly logical steps that end in wrong answers (Kambhampati, 2025). The trace doesn't track the computation which clarifies that verbalized CoT is an imitation of reasoning and not the mechanism itself. The alternative mechanism which gets the most attention is latent reasoning: instead of verbalizing every intermediate result, the model repeatedly transforms its continuous hidden state and decodes only the answer. I’m breaking latent reasoning down into at least five distinct families: Continuous thoughts in autoregressive LMs: Coconut (Hao et al., 2024) feeds the model's own final hidden state back in as the next input embedding while Soft Thinking (Zhang et al., 2025) reasons in a continuous concept space. Theory here argues a single continuous state can hold several search frontiers at once and expand them in parallel (Zhu et al., 2025) . Compressed discrete non-linguistic tokens: Abstract-CoT (Ramji et al., 2026) swaps verbal rationales for a short sequence from a learned vocabulary. It is non-linguistic, but still serial and externally decoded, the discrete end of the spectrum . Recurrent depth and looped models: recurrent-depth LMs (Geiping et al., 2025) and looped Transformers (Saunshi et al., 2025; Zhu et al., 2026) reapply a shared block to a latent state. Mostly framed as parameter efficiency and test-time-compute scaling, not as a new reasoning interface . Task-trained recursive solvers: HRM (Wang et al., 2025) and TRM (Jolicoeur-Martineau, 2025) recursively refine latent and candidate-answer states. Their ARC pipelines are transductive: evaluation-task demonstrations get augmented into optimization with learned per-puzzle identities, so an unseen task needs a backward pass before it can be answered . In-context recurrent latent solvers: this is where BDH-CQ (Engdahl et al., 2026) sits, built on the Dragon hatchling architecture (Kosowski et al., 2025). Demonstrations write directly into a recurrent memory at inference time, and new test inputs are then solved by iterative computation in a separate continuous latent space. The authors report a point beyond the previously published cost–accuracy Pareto frontier on public ARC-AGI-1, as well as early pretraining experiments show transformer-like scaling laws upto 600B parameters while preserving the latent reasoning behavior. Two distinctions seem especially important: how a system acquires a new task (through context, memory, or gradient-based optimization or finetuning) and where its intermediate computation happens (through language tokens, abstract tokens, or continuous latent states) Lmk if I have missed any family or papers. More importantly, if latent reasoning wins on efficiency, what happens to the readable traces on which much of industry's interpretability and evaluation work currently depends? Was CoT legibility a temporary consequence of how we scaled LLMs or is it a safety property worth paying an efficiency penalty to keep?
This article argues that Chain of Thought reasoning is a temporary hack and that the future lies in latent reasoning methods like Coconut, HRM, and RecursiveMAS. It discusses the black box problem and proposes outer loop governance with DAGs and verification, while positioning BDH as a model that combines latent computation with stateful memory.
This paper introduces BDH-CQ, a 150M-parameter reasoning model that combines in-context learning with recurrent latent reasoning, achieving 29.5% pass@2 on ARC-AGI-1 at very low inference cost and establishing a new cost-accuracy frontier.
This paper applies dynamical systems analysis to interpret latent chain-of-thought reasoning in models like CODI and COCONUT, revealing structured dynamics with stable and unstable classes.
This survey provides a comprehensive overview of latent reasoning in LLMs, exploring methods that perform multi-step inference in continuous hidden states without explicit token-level supervision.
This paper introduces RIS, a framework for spatial-semantic grounded latent visual reasoning in Multimodal Large Language Models to overcome information bottlenecks. It proposes anchoring latent tokens to spatial and semantic evidence, showing improvements on benchmarks like V* and HRBench.