Latent Thought Reasoning

• Latent Thought Reasoning is a framework where internal reasoning is executed in continuous latent spaces rather than as explicit chain-of-thought tokens.
• It decouples planning and verbalization, enabling iterative computation and parallel processing, which can improve inference speed and depth.
• It employs innovative methods like latent-variable models, looped transformers, and diffusion techniques to enhance performance on complex reasoning tasks.

Latent thought reasoning denotes a class of methods in which intermediate reasoning is carried out in continuous latent space or hidden states rather than being verbalized as explicit chain-of-thought tokens. Across the recent literature, the aim is to move deliberation beyond the discrete vocabulary space, reduce the linguistic space bottleneck and token-level inference overhead, and permit forms of iterative computation, planning, exploration, or compression that are difficult under strictly autoregressive textual CoT. The term now covers latent-variable models for reasoning skills, looped and recurrent architectures that implicitly generate latent thoughts, planner–decoder systems that decouple reasoning from verbalization, diffusion and GFlowNet formulations over latent trajectories, and decode-time optimization procedures that refine latent reasoning paths directly (Xu et al., 2023, Saunshi et al., 24 Feb 2025, Wang et al., 29 Jan 2026).

1. Conceptual and formal foundations

An early latent-variable formulation appears in "LaRS: Latent Reasoning Skills for Chain-of-Thought Reasoning" (Xu et al., 2023). There, the latent variable is not a hidden reasoning trace in the usual sense, but a continuous "reasoning skill" used to organize question–rationale pairs for in-context example selection. The model writes rationale generation as

$$P(R \mid Q) = \int_{\mathcal{Z}} P(R \mid z, Q)\, P(z \mid Q)\, dz,$$

with $z$ a continuous latent reasoning skill, $P(z \mid Q)$ a reasoning policy, and $P(R \mid z, Q)$ a rationale model. This established a recurring theme in later work: latent variables are introduced not merely as compressed activations, but as structured carriers of reasoning capability.

"Reasoning with Latent Thoughts: On the Power of Looped Transformers" (Saunshi et al., 24 Feb 2025) made the stronger architectural claim that many reasoning problems require a large depth but not necessarily many parameters. In that account, repeated application of a small transformer block produces latent thoughts implicitly through iterative hidden-state updates. The paper argues that a $k$-layer transformer looped $L$ times nearly matches a $kL$-layer non-looped model on addition, $p$-hop induction, and math problems, and further proves that looped models can simulate $T$ steps of CoT with $T$ loops. This connects latent thought reasoning to effective depth rather than to textual rationale generation.

"A Formal Comparison Between Chain of Thought and Latent Thought" (Xu et al., 25 Sep 2025) supplied the explicit complexity-theoretic separation. In that analysis, latent thought reasoning operates directly in the continuous latent space, enabling computation beyond discrete linguistic representations, and can exploit parallel computation more efficiently than inherently sequential CoT. The summary result is written as

$z$0

whereas

$z$1

The same analysis also argues that CoT retains a distinct advantage for approximate counting and sampling through stochastic decoding. The resulting picture is not that latent thought simply dominates explicit CoT, but that the two paradigms have different computational biases: depth-driven recursion and parallel evaluation on one side, randomized approximation and sampling on the other (Xu et al., 25 Sep 2025).

2. Construction of latent thought states

A central design question is how a latent thought is represented and injected into the model’s computation. "Latent Thoughts Tuning: Bridging Context and Reasoning with Fused Information in Latent Tokens" (Liu et al., 10 Feb 2026) argues that raw hidden-state recycling causes distribution mismatch and instability, while pure soft prediction loses contextual information. Its Context-Prediction-Fusion mechanism therefore combines contextual hidden states with predictive semantic guidance:

$z$2

$z$3

This formulation is explicitly aimed at mitigating feature collapse and enabling dynamically switching between latent and explicit thinking modes.

"LTA-thinker: Latent Thought-Augmented Training Framework for LLMs on Complex Reasoning" (Wang et al., 16 Sep 2025) adopts a different construction. Rather than using a small pretrained LLM or auxiliary module, it replaces the latent-thought generator with a lightweight, randomly-initialized Transformer Block trained as a learnable prior. Given instruction $z$4 and question $z$5, it generates latent thought vectors $z$6 according to

$z$7

The motivation is distributional: increasing the variance of generated latent thoughts is argued to better approximate the golden truth distribution, while a learnable prior is expected to raise the performance ceiling without diverging (Wang et al., 16 Sep 2025).

"Latent Chain-of-Thought as Planning: Decoupling Reasoning from Verbalization" (Wang et al., 29 Jan 2026) pushes the separation further by introducing PLaT, where reasoning is modeled as a deterministic trajectory of latent planning states and text generation is delegated to a separate Decoder. The architecture explicitly decouples the Planner, which evolves latent thought trajectories in continuous space, from the Decoder, which grounds those thoughts into natural language when needed. This permits dynamic determination of when to terminate reasoning rather than relying on a fixed number of latent steps, and it allows intermediate inspection by decoding latent states as text. The same decoupling reappears, with domain-specific changes, in later multimodal systems.

3. Learning objectives and control of latent trajectories

The training objectives used for latent thought reasoning vary widely, but most fall into three families: variational latent-variable learning, reinforcement-style trajectory optimization, and reward-weighted sampling of latent paths. In LaRS, the latent skill space is learned with a CVAE whose loss combines reconstruction and KL regularization:

$z$8

CTRLS, by contrast, formulates CoT reasoning as an MDP over latent semantic states, learns a transition prior and state-dependent generator through an ELBO, and then refines latent transitions with on-policy RL, epsilon-greedy exploration, and entropy regularization (Xu et al., 2023, Wu et al., 10 Jul 2025).

LTA-thinker combines standard supervised finetuning with two auxiliary losses intended to shape latent-thought distributions:

$z$9

Here, Semantic Alignment Loss uses KL divergence to keep latent thoughts semantically relevant to the question, while Reasoning Focus Loss uses a contrastive mechanism to bias the model toward critical reasoning steps. The stated purpose is to avoid uninformative high variance while preserving useful, semantically anchored, structurally focused variance (Wang et al., 16 Sep 2025).

A different problem arises in LoopLMs, where standard objectives such as GRPO only assign credit to the final latent state. "Prioritize the Process, Not Just the Outcome: Rewarding Latent Thought Trajectories Improves Reasoning in Looped LLMs" (Jonathan et al., 11 Feb 2026) identifies this as a credit-assignment bottleneck and introduces RLTT, which distributes reward across the full latent reasoning trajectory. Under identical training and inference conditions on Ouro-2.6B-Thinking, RLTT improves accuracy by $P(z \mid Q)$0 on MATH-500, $P(z \mid Q)$1 on AIME24, and $P(z \mid Q)$2 on BeyondAIME. This is a direct claim that latent thought trajectories, rather than only terminal latent states, should be the target of supervision (Jonathan et al., 11 Feb 2026).

Two later approaches move from policy optimization toward latent reward modeling and posterior matching. "Latent Thinking Optimization: Your Latent Reasoning LLM Secretly Encodes Reward Signals in Its Latent Thoughts" (Du et al., 30 Sep 2025) trains a latent classifier to predict answer correctness directly from latent trajectories and uses it as a Latent Reward Model to reweight samples. "Latent Thought Flow: Efficient Latent Reasoning in LLMs" (Zou et al., 15 Jun 2026) instead defines a reward over complete trajectories,

$P(z \mid Q)$3

and trains a continuous GFlowNet sampler to match the corresponding reward-induced posterior, with an Entropy-Weighted Subtrajectory Balance objective and a reference-prior regularizer. This suggests a broader shift from single best latent path learning toward explicit distributions over latent reasoning paths (Du et al., 30 Sep 2025, Zou et al., 15 Jun 2026).

4. Inference-time scaling, adaptive compute, and selective compression

A major attraction of latent thought reasoning is inference-time scaling without explicit token-level rationales. PLaT reports a trade-off in which greedy accuracy is lower than explicit CoT and existing latent baselines, but reasoning diversity scales better: on GSM8K Pass@128, PLaT-2 scores $P(z \mid Q)$4, exceeding Coconut at $P(z \mid Q)$5 and CODI at $P(z \mid Q)$6, while inference is up to $P(z \mid Q)$7 faster for PLaT-1 than CoT because intermediate token-by-token generation can be skipped (Wang et al., 29 Jan 2026). "LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning" (Kang et al., 6 Oct 2025) reaches adaptive computation through a different route, using a VAE to encode blocks of thought tokens and a latent diffusion model with blockwise bidirectional attention for iterative refinement; more denoising steps monotonically improve accuracy, and increasing the number of denoising steps from 10 to 30 yields a $P(z \mid Q)$8 point improvement (Kang et al., 6 Oct 2025).

Structured sampling in latent space is another line of work. "GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler" (Wang et al., 15 Feb 2026) replaces heuristic perturbations such as dropout or fixed Gaussian noise with a conditional diagonal-Gaussian sampler trained with GRPO-style policy optimization. On COCONUT, GTS reaches $P(z \mid Q)$9, $P(R \mid z, Q)$0 rate $P(R \mid z, Q)$1, and $P(R \mid z, Q)$2; on CODI, it reaches $P(R \mid z, Q)$3, $P(R \mid z, Q)$4 rate $P(R \mid z, Q)$5, and $P(R \mid z, Q)$6. The argument is that latent inference-time scaling requires structured and optimizable exploration, not merely stronger stochasticity (Wang et al., 15 Feb 2026).

Several methods perform direct test-time optimization of latent thoughts. "Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization" (Ye et al., 5 Oct 2025) treats the latent thought vectors themselves as per-instance parameters and updates them by policy gradient using an intrinsic confidence-based reward. On AIME24 and AIME25, LTPO achieves $P(R \mid z, Q)$7 and $P(R \mid z, Q)$8, compared with $P(R \mid z, Q)$9 and $k$0 for Zero-Shot CoT and $k$1 and $k$2 for SoftCoT. "Inference-Time Rethinking with Latent Thought Vectors for Math Reasoning" (Kong et al., 6 Feb 2026) alternates between generating a candidate trace and optimizing the latent thought vector in a Gibbs-style procedure; with 30 rethinking iterations, a $k$3B-parameter model reaches $k$4 on GSM8K versus $k$5 for the best baseline, despite baselines having 10 to 15 times more parameters (Kong et al., 6 Feb 2026).

Not all steps in a reasoning chain need equal compression. "Selective Latent Thinking: Adaptive Compression of LLM Reasoning Chains" (Xie et al., 25 May 2026) argues that existing latent reasoning methods treat reasoning as uniformly compressible and thereby over-compress precision-critical steps. SLT therefore predicts an upcoming span, uses confidence-based gating to decide the longest span that can be reliably compressed, and leaves uncertain or precision-critical reasoning explicit. On the reported benchmarks, it achieves $k$6 higher accuracy than latent reasoning baselines at comparable compression ratios, while reducing reasoning chain length by $k$7 with only $k$8 accuracy degradation compared to explicit CoT (Xie et al., 25 May 2026). A related scaling result appears in LTA-thinker, where with only $k$9 it surpasses baseline methods that require $L$0 or even $L$1, and at $L$2 reaches a GSM8K score of $L$3 while SoftCoT++ needs $L$4 for $L$5 (Wang et al., 16 Sep 2025).

5. Empirical scope and application domains

Although mathematical reasoning is the dominant evaluation regime, latent thought reasoning is not confined to arithmetic or symbolic tasks. LaRS evaluates latent reasoning skills on TabMWP, GSM8K, COGS, and Spider, with backbones including gpt-3.5-turbo, text-davinci-003, Claude-v2, and Falcon-40B-Instruct. It reports that LaRS consistently outperforms SOTA skill-based selection methods, processes example banks four times faster, reduces LLM inferences during the selection stage by half, and shows up to $L$6 absolute gain in answer accuracy over non-oracle baselines (Xu et al., 2023). This is an important reminder that latent reasoning can operate at the level of example retrieval and rationale alignment, not only inside the forward pass of a reasoning model.

The same logic extends to interactive speech systems. "The Silent Thought: Modeling Internal Cognition in Full-Duplex Spoken Dialogue Models via Latent Reasoning" (Wu et al., 18 Mar 2026) introduces FLAIR, which performs latent thinking simultaneously with speech perception by recursively feeding the latent embedding output from the previous step into the next step during the user’s speaking phase. The method is designed to be strictly causal and to add zero extra inference latency. Empirically, it achieves competitive results on a range of speech benchmarks and competitive performance on full-duplex interaction metrics, while remaining robust to conversational dynamics such as interruptions and overlapping speech (Wu et al., 18 Mar 2026).

A multimodal extension appears in "CoCoVa: Chain of Continuous Vision-Language Thought for Latent Space Reasoning" (Ma et al., 4 Nov 2025). CoCoVa introduces a Latent Q-Former that iteratively refines a chain of latent thought vectors through cross-modal fusion, together with a token selection mechanism for salient visual regions and a multi-task objective combining contrastive learning and diffusion-based reconstruction. With a $L$7B backbone, it competes with or surpasses larger $L$8B-$L$9B models on almost all benchmarks, and with $kL$0B LLM backbones it remains competitive with state-of-the-art models (Ma et al., 4 Nov 2025). This suggests that latent thought reasoning is best understood as a general computational pattern—continuous internal deliberation before or alongside verbal output—rather than as a technique specific to text-only LLMs.

6. Interpretability, superposition, and causal scrutiny

A persistent claim in the area is that continuous latent thoughts may support superposition, understood as the ability to maintain multiple candidate solutions simultaneously within a single representation. "The Illusion of Superposition? A Principled Analysis of Latent Thinking in LLMs" (Rizvi-Martel et al., 7 Apr 2026) directly tests this claim across training-free, fine-tuned, and from-scratch regimes. Its conclusion is restrictive: only models trained from scratch exhibit signs of using superposition; in the training-free and fine-tuned regimes, the superposition either collapses or is not used at all, with models discovering shortcut solutions instead. The paper further argues that pretraining biases models to commit to a token in the last layers, and that model capacity has a large effect on whether shortcut solutions are favored (Rizvi-Martel et al., 7 Apr 2026). This suggests that the main benefit of latent reasoning, as currently implemented, may be flexibility in hidden-state computation rather than stable superpositional reasoning.

A second line of critique concerns observational interpretability. "Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models" (Aswal et al., 10 Jun 2026) shows that BFS-like frontiers and decodable arithmetic computations also appear in controls lacking the proposed recurrence or curriculum and do not always causally affect behavior. Causal interventions reveal that latent-thought utilization is graded rather than binary, and geometric analysis localizes behavioral influence to low-rank directions whose structure becomes more organized as their causal effect increases. Its summary statement is unusually explicit: latent thoughts should be treated as hidden computation, not hidden explanation; decodability, attention, or static structure alone cannot establish mechanism (Aswal et al., 10 Jun 2026).

This does not imply that latent reasoning is uninterpretable in practice. "Unlocking the Black Box of Latent Reasoning: An Interpretability-Guided Approach to Intervention" (Chang et al., 31 May 2026) reports structural, causal, and geometric probes indicating that latent vectors encode compressed, faithful representations of reasoning steps, with high CKA values, linear recoverability, and early vectors acting as critical causal hubs. The same work turns these findings into training-free, decode-time interventions—semantic, causal, and geometric edits of latent vectors—that consistently improve reasoning accuracy without parameter updates (Chang et al., 31 May 2026). A plausible implication is that interpretability in latent thought reasoning is most productive when it is tied to matched controls, causal tests, and actionable interventions, rather than to surface analogies between latent states and human-readable rationales.