Latent Reasoning

Definition: Latent reasoning is when a language model works through a problem inside its hidden states instead of writing each step out as words. A chain of thought shows its steps as text you can read. Latent reasoning passes vectors from step to step, which is faster and uses fewer tokens, but is much harder for a person to check.

TL;DR: Latent reasoning moves a model's thinking from words into its latent space. The October 2026 Hybrid Reasoning Agents paper reports 35.5% to 54.7% fewer tokens and 1.8 to 2.6 times faster runs by thinking in vectors and acting in text. The price is oversight: there is no readable chain of thought to audit.

Think about how you add up a restaurant bill. You can write every line on a napkin, carry the ones, and show your work. Or you can do it in your head. The head is faster and needs no napkin, but nobody at the table can see where you made a mistake. A chain of thought is the napkin. Latent reasoning is the head.

Why Latent Reasoning Matters in 2026

Latent reasoning matters in 2026 because it reached agents, and agents are where token costs add up. On October 13, 2026, Jiaru Zou, Yunzhe Qi, Sirui Chen, Federico Bianchi, Shang Zhu, Pan Lu, Jingrui He, and James Zou posted Hybrid Reasoning Agents (arXiv 2610.18861, 45 pages, cs.CL). Their argument starts from a gap. Today's agents reason and act in natural language. Latent reasoning is more efficient, but it has no direct way to call a tool and it is hard to inspect.

The paper closes that gap by letting one model switch between the two modes inside a single generation. It thinks in latent space while it works out what to do, and it writes text when it needs to call a tool or answer. The authors tested three model scales on 11 benchmarks across math, science, medicine, search, and agentic reasoning. They report:

These are the authors' own measurements in a first version of the paper. Nobody had replicated them when this article was written. If they hold, they matter for anyone who pays for agents, because an agent loop spends most of its budget on tokens. The inference cost of a long task falls when the model needs fewer written steps to reach the same answer.

Chain of Thought vs Latent Reasoning

The two approaches solve the same problem, giving a model room to think before it answers, in opposite ways.

The readable column is the one that splits researchers. A written chain of thought is not a perfect window. The chain-of-thought faithfulness literature shows that models sometimes write a tidy explanation that does not match what drove the answer. Still, a written trace gives a monitor something to read. A latent trace gives it a list of numbers.

How Latent Reasoning Works

The clearest example is Coconut, short for Chain of Continuous Thought. Shibo Hao, Sainbayar Sukhbaatar, Jason Weston, Yuandong Tian, and colleagues described it in Training Large Language Models to Reason in a Continuous Latent Space (December 9, 2024). In a normal model, the last hidden state at each step is turned into a word, and that word is fed back in. Coconut skips the word. It feeds the last hidden state straight back in as the next input. The authors call that state a continuous thought.

A word commits the model to one next step. A continuous thought does not have to. The Coconut authors argue that one continuous thought can hold several possible next steps at once, so the model can explore options in parallel, a kind of breadth-first search, instead of locking into one path early. They report that Coconut beats chain of thought on logical reasoning tasks that need a lot of search during planning.

Chain of thought (each step becomes a word)

  question -> [model] -> "First, ..." -> [model] -> "So ..." -> [model] -> answer
                           readable          readable

Latent reasoning (each step stays a vector)

  question -> [model] -> h1 -> [model] -> h2 -> [model] -> h3 -> [model] -> answer
                        vector         vector         vector
                     (can hold several candidate next steps at once)

Hybrid Reasoning Agents add a switch to this picture. A small module called HybridLink handles the move from latent steps to text and back. A continuous working memory keeps the hidden context across the turns of an agent run. Training runs in two stages: HybridSFT teaches the model to use HybridLink, and HybridRL uses reinforcement learning to teach it when to think in latent space and when to act through text.

Other routes to latent reasoning exist. Yuntian Deng and colleagues showed in Implicit Chain of Thought Reasoning via Knowledge Distillation (November 2023) that a student model can learn to do multi-digit multiplication through its hidden layers, guided through distillation by a teacher that wrote its steps out, at about the speed of a model that does not reason at all. Jonas Geiping and colleagues built a recurrent-depth model (February 2025) that loops one block of layers more times when a problem is harder, scaled to 3.5 billion parameters and 800 billion tokens.

What Hacker News Pushed Back On

The Hacker News thread on the paper, titled "Hybrid Reasoning Agents: LLM agents that think in latent space, act through text", passed 300 points and 150 comments within a day. We read it on October 19, 2026. Four lines of argument stood out.

Some commenters also pushed back the other way. One argued that written reasoning is not a mere veneer: if it had nothing to do with the result, it would not improve performance, and it does. Another said that text reasoning was never a fully faithful window, and that monitoring a model's activations is the better long-term path.

The Oversight Trade-off

The oversight question is the reason latent reasoning is news outside research labs. In July 2025, 41 researchers, with Tomek Korbak and Mikita Balesni as lead authors, published Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety. Their argument: AI systems that think in human language offer a rare safety opportunity, because people can read those thoughts for signs of intent to misbehave. The monitoring is imperfect, but it works well enough to keep. They also warned that it is fragile, and asked model developers to weigh how their design choices affect it.

Latent reasoning is one of those design choices. A model that thinks in vectors writes no ordinary text trace during the latent stretch. Some systems offer an approximate readout instead. The HybridAgent repository ships an Interpretable Latent Mode inspector that maps each latent step to its nearest token embedding and shows the result as pseudo-text beside the tool calls. That readout is a lossy view, not a faithful account of what the model computed. The other options are the ones from mechanistic interpretability: probes, sparse autoencoders, and other tools that read hidden states directly. Those tools work on narrow questions today, not as a general monitor.

A counter-argument is gaining ground. In April 2026, Wenshuo Wang argued in LLM Reasoning Is Latent, Not the Chain of Thought that reasoning already happens mainly in hidden-state trajectories, and that the written chain of thought is an unreliable proxy for it. On that view, latent reasoning does not remove a window that was truly open. It makes visible a gap that was always there. Both views agree on one practical point: judge an agent by what it does, not only by what it says it thought.

A Short History of Latent Reasoning

Each row links to a source that was checked for this article.

The "dot by dot" result is the quiet turning point. Jacob Pfau, William Merrill, and Samuel Bowman showed that a transformer can use strings of meaningless dots in place of a chain of thought to solve two hard algorithmic tasks it could not solve without intermediate tokens. Part of what a chain of thought buys is extra computation, not the words themselves. They also found that learning to use filler tokens is hard and needs dense supervision.

Limits and Open Problems

• Training is unstable. Most latent methods need careful curricula or distillation from a model that writes its steps. Reinforcement learning on latent steps is newer and harder to tune than RL on text.
• Evaluation is thin. Papers report gains on math and search benchmarks. Few evals test long, messy agent tasks, and few have been replicated by an outside team.
• Decoding is lossy. Turning a latent step back into words, for a log or a monitor, gives an approximation. The readable version can miss what the model computed.
• Monitoring tools lag. Reading hidden states with probes or sparse autoencoders works on narrow questions. No tool yet reads a latent reasoning trace as easily as a person reads text.
• Runtimes lag. The looping reports on Hacker News point at a practical gap: community runtimes are built for text-only models, and hybrid models need support for the switch.

How Latent Reasoning Shows Up in Taskade

Latent reasoning is not a Taskade feature. Taskade agents run on frontier models from top AI labs, and Taskade does not use or expose the hidden states of those models. What you see and review are actions: the reply, the project an agent edits, and the automation step it runs. You judge the work, not the thoughts.

An agent review log is the plain use. An automation runs your agent on each new request, and a second step writes the request, the reply, and the record it changed into a review project. A reviewer marks each row as approved or needs work, and the notes go back into the instructions the agent reads. Taskade Genesis builds the first version of that app from one prompt, and Taskade EVE edits it in chat. To start from a working example, clone the Research Chat Bot kit, an app kit with projects, an AI agent, and automations you can clone. The Free plan covers 3 apps and 1 agent and starts with 0 AI credits.

Describe yours and build it free →

Related Concepts

• Latent Space: the hidden map where latent reasoning happens
• Chain of Thought: the written alternative that latent reasoning replaces
• Chain-of-Thought Faithfulness: whether a written chain of thought matches what the model computed
• Reasoning Models: models trained to spend more compute before they answer
• Reasoning Effort: the setting that trades thinking time against cost
• Mechanistic Interpretability: reading a model's hidden states directly
• Sparse Autoencoders: a tool that turns hidden states into named features
• Next-Token Prediction: the word-by-word process that latent steps skip
• Inference Cost: why fewer tokens means a cheaper agent run
• Large Language Models: the models that latent reasoning changes
• Transformer: the architecture whose hidden states carry latent steps
• Token: the unit that written reasoning spends and latent reasoning saves
• AI Reasoning Models Explained: the longer guide to how reasoning models think
• Representation Learning: how models learn useful features from raw data
• Steering Vectors: directions in latent space that change behavior
• Platonic Representation Hypothesis: whether different models learn the same map
• Embedding Models: how to choose the model that turns text into vectors

Frequently Asked Questions About Latent Reasoning

What is latent reasoning in simple terms?

Latent reasoning means a model thinks in its head instead of on paper. It passes hidden vectors from one reasoning step to the next instead of writing each step as words. That saves tokens and time. The cost is that a person cannot read the steps, so mistakes and intentions are harder to spot.

Is latent reasoning the same as chain of thought?

No. They are opposites. A chain of thought writes each reasoning step as text, which you can read. Latent reasoning keeps the steps as vectors inside the model. Both give the model room to think before answering. Some systems, such as Hybrid Reasoning Agents, mix the two: latent steps for thinking, text for tool calls and answers.

What are hybrid reasoning agents?

Hybrid Reasoning Agents are language model agents from an October 2026 paper by Jiaru Zou, James Zou, and colleagues. They think in latent space and switch to text when they call a tool or answer. A module called HybridLink handles the switch. The authors report a 10.4% average gain and 35.5% to 54.7% fewer tokens across 11 benchmarks.

Is latent reasoning faster?

Usually, yes. Writing each step as text means generating many tokens one at a time. Latent steps skip that. The Hybrid Reasoning Agents authors report 1.8 to 2.6 times faster inference, and the Latent-SFT paper reports reasoning chains 2.7 to 5.5 times shorter on math benchmarks. Both are author-reported results on research benchmarks.

Why do safety researchers worry about latent reasoning?

Because a readable chain of thought is one of the few ways to see what a model is planning. A July 2025 paper by 41 researchers called chain-of-thought monitoring a new and fragile safety opportunity. Latent reasoning removes the text that such a monitor reads, so oversight has to rely on reading hidden states or on checking actions.

Can you read a model's latent thoughts?

Only partly. Researchers can decode a latent step into approximate words, or use probes and sparse autoencoders to detect specific features in hidden states. Each method gives a partial view, and a decoded version can miss what the model computed. No current tool reads latent reasoning as easily as a person reads a written explanation.

Does ChatGPT use latent reasoning?

Every language model computes in hidden states, so some reasoning always happens there. "Latent reasoning" names research methods that replace written steps with hidden ones, such as Coconut or Hybrid Reasoning Agents. Vendors do not publish whether their production models use such methods, so no public evidence shows that any major chat assistant ships them.

What is Coconut in AI?

Coconut, short for Chain of Continuous Thought, is a December 2024 method from Shibo Hao, Jason Weston, Yuandong Tian, and colleagues. It feeds a model's last hidden state straight back in as the next input, instead of turning it into a word first. The authors argue that one such continuous thought can hold several possible next steps at once, which helps on tasks that need search.

Further Reading

• Zou et al., Hybrid Reasoning Agents, arXiv 2610.18861, October 13, 2026
• HybridAgent code, data, and model links, GitHub
• Hacker News discussion, read on October 19, 2026
• Hao et al., Training Large Language Models to Reason in a Continuous Latent Space, December 2024
• Zhu et al., A Survey on Latent Reasoning, July 2025
• Korbak, Balesni et al., Chain of Thought Monitorability, July 2025
• Wang, LLM Reasoning Is Latent, Not the Chain of Thought, April 2026
• Deng et al., LLM Latent Reasoning as Chain of Superposition, October 2025
• Pfau, Merrill, and Bowman, Let's Think Dot by Dot, April 2024