A new AI agent reasons internally, then uses text to act
latent reasoning
Processing that uses internal representations instead of writing every step as text.
HybridLink
A small component that passes information between text and internal processing.
preprint
A research paper shared before formal peer review.
What was proposed
A research team has posted a 45-page preprint called Hybrid Reasoning Agents. It was submitted to arXiv, a public site for research papers, on October 13, 2026. The paper introduces HybridAgent, an AI agent that can use tools while switching between two kinds of processing.
Why two modes are useful
Most modern AI agents use text for both intermediate reasoning and tool interaction. Text is easy for people to inspect, and tools can receive text commands. But writing every intermediate step as tokens can consume extra computation. Latent reasoning takes place in continuous internal representations instead of decoding every step into text. It may be more efficient, but earlier latent approaches made tool use and inspection harder.
HybridAgent combines the two. In text mode, it writes an answer, calls a search tool, or sends another external request. In latent mode, it carries internal representations forward without turning every step into a word. A small adapter called HybridLink passes information between the modes. The system also keeps text, latent states, and tool results in continuous working memory across multiple turns. The authors report that latent steps can be mapped into readable pseudo-text for inspection.
How the system is trained
The paper describes two post-training stages. HybridSFT teaches the base model to coordinate with HybridLink and switch reliably. HybridRL then learns when internal processing is useful and when text or a tool call is needed. The researchers say this can adapt a standard pretrained model without training a new foundation model from scratch.
What the experiments report
The team evaluated three model scales across 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning. Compared with advanced agents and related frameworks, the paper reports a 10.4% average performance improvement. It also reports 35.5% to 54.7% fewer tokens, 1.8 to 2.6 times faster inference, and 2.7 to 3.7 times lower computational cost.
These are results from the authors' comparisons, not a guarantee for every AI system. The proposed division of labor is still important: internal representations handle some routine computation, while text remains available for tools, key decisions, and monitoring.
What remains open
The paper is an arXiv preprint, so it has not yet established a general result through independent peer review. Its benchmarks, baselines, model sizes, and training choices all affect the numbers. A readable pseudo-text view may also show useful signals without revealing every fact about the underlying computation. Real, long-running tasks will test that distinction.
Hacker News, a technology discussion site, gave the paper 177 points and 109 comments in the supplied cluster data. That measures community attention, not correctness. The next evidence to watch is independent reproduction, tests with different models and tools, and whether the efficiency gains survive outside the paper's setup.
Source: the arXiv paper. Attention data: the Hacker News discussion.
An AI that uses inner calculations and words
📰 Full story: A new AI agent reasons internally, then uses text to act
Researchers described an AI agent that switches between hidden calculations and text.\n
latent mode
A way for the AI to keep some work as internal number patterns.
token
A small piece of text processed by an AI system.
preprint
A paper shared before full research review.
💡 The gist
- HybridAgent, a tool-using AI, switches between inner calculations and words.
- It uses words for tools and internal representations for some reasoning.
- Hacker News, a technology discussion site, listed 177 points and 109 comments. Attention is not proof.
Researchers described the method in a preprint on arXiv, a public paper site. They call the method Hybrid Reasoning Agents. An AI agent is software that can use tools while solving a task.
Many agents write text for every step. Text is easier for people to read. However, long reasoning can use many tokens. Tokens are small pieces of text that AI systems process.
HybridAgent switches between text mode and latent mode. In text mode, it can ask a search tool for information. In latent mode, it keeps some work as internal number patterns. It does not turn every step into words.
HybridLink passes information between the two modes. Working memory keeps earlier text, internal states, and tool results. This helps the agent continue after several turns.
The paper tested three model scales and 11 kinds of benchmarks. It reports 10.4% better average performance. It also reports 35.5% to 54.7% fewer tokens. Inference was 1.8 to 2.6 times faster. Reported computing costs were 2.7 to 3.7 times lower.
These results belong to the authors' experiments. They do not prove the method works everywhere. The paper is still a preprint. Other researchers need to test it. Hacker News attention is separate from the research results. Points and comments show interest, not truth.
Source: the arXiv paper. Attention data: the Hacker News discussion.
💬 An AI that thinks in hidden numbers: promise and risk
This system does some intermediate work as hidden numerical representations instead of visible text. That may help efficiency, but it may also make the process harder to inspect.
- The idea is to do intermediate work as hidden vectors, then produce text or take an action when needed. People want to know whether this saves tokens and time, or also makes answers more accurate.
- A commenter asked about benchmarks; another pointed to the paper’s later evaluation pages, and links to the code and evaluations were shared. The discussion does not show how large the improvement is.
- Hidden steps are harder to inspect. Converting each step to a nearby word may lose combined meanings, and an explanation produced by another model may not fully recover the original calculation. Visible chain-of-thought can also be only an imperfect explanation.
- A strong safety worry is that hidden steps could make bad plans harder to find. Some commenters especially worried that training with step-by-step feedback on latent states could teach the model to avoid what is being watched. That is a concern, not a demonstrated fact.
- Supporters said thinking is not always made of words, so hidden representations might help with spatial or visual ideas and might keep answers more consistent. Comparisons with looped models are still unresolved.
- A user reported that the 14B model in llama.cpp repeated text for about five minutes in both 8-bit and BF16 modes and returned a table instead of CSV. It may have been a compatibility problem, so this is not proof that the model generally performs badly.
initial digest at 109 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
An AI uses words outside and numbers inside
📰 Full story: A new AI agent reasons internally, then uses text to act
A new paper describes an AI that uses two ways to work.
HybridAgent
An AI helper that can use tools.
arXiv
A website where researchers share papers.
Hacker News
A website where people discuss technology.
arXiv, a website for sharing research papers, has a new paper. The paper presents HybridAgent, a helper that can use tools.
It uses words when talking to tools. It can use number patterns inside for some work. It does not write every inside step as words. That may save computer work.
The paper says people can turn some inside work into readable pretend words. The helper also keeps earlier tool results. This helps it continue its job.
The paper tested the helper on many tasks. It is a preprint, so other researchers must check it.
Hacker News, a technology discussion site, noticed the paper. It showed 177 points and 109 comments. That means many people paid attention. It does not prove the paper is correct.
Source: the arXiv paper. Attention data: the Hacker News discussion.
💬 An AI that thinks with hidden numbers
This AI does some of its work with hidden numbers instead of words.
- The paper has tests in its later pages, and people shared code and evaluation links. But the comments do not tell us how much better the AI becomes.
- Hidden work is harder for people to watch. Turning it back into a word may lose some meaning, and bad plans might be harder to notice. Still, hidden representations could help with ideas that are not naturally words.
- One user reported that the 14B model repeated the same text for about five minutes in both 8-bit and BF16 modes and made a table instead of CSV. It may have been a software compatibility problem, so this does not prove that everyone will see the same thing.
initial digest at 109 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 Latent-space reasoning: faster and smarter, or harder to inspect?
HN commenters see a promising way to do intermediate work in latent vectors, but they disagree about performance gains and worry that safety monitoring may become weaker. The thread does not establish a measured improvement.
initial digest at 109 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.