--- title: Hybrid Reasoning Agents: LLM agents that think in latent... description: Researchers introduce Hybrid Reasoning Agents (HybridAgent), a new architecture that lets LLM agents dynamically switch between latent (internal) reasoning and... canonical: https://daily.dev/posts/hybrid-reasoning-agents-llm-agents-that-think-in-latent-space-act-through-text-gqx1d3tvr twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Hybrid Reasoning Agents: LLM agents that think in latent space, act through text | daily.dev og:description: Researchers introduce Hybrid Reasoning Agents (HybridAgent), a new architecture that lets LLM agents dynamically switch between latent (internal) reasoning and... og:url: https://daily.dev/posts/hybrid-reasoning-agents-llm-agents-that-think-in-latent-space-act-through-text-gqx1d3tvr og:image: https://api.daily.dev/og/posts/gqX1d3tVr.png og:image:alt: Hybrid Reasoning Agents: LLM agents that think in latent space, act through text og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Hybrid Reasoning Agents: LLM agents that think in latent space, act through text **[Hacker News](https://daily.dev/sources/hn)** · 2 min read · 0 upvotes · 0 comments ## Summary Researchers introduce Hybrid Reasoning Agents (HybridAgent), a new architecture that lets LLM agents dynamically switch between latent (internal) reasoning and text-based reasoning within a single generation process. A lightweight component called HybridLink enables seamless transitions between the two modes while preserving lossless context across turns and keeping latent thoughts inspectable, alongside tool-integrated agentic capabilities. The team also proposes post-training methods: HybridSFT to stabilize tool-integrated hybrid generation, and HybridRL to teach the model when to think latently versus act through text. Tested across 3 model scales and 11 benchmarks spanning math, science, medicine, search, and agentic reasoning, HybridAgent beats frontier LLM agents and existing text/latent frameworks by 10.4% while cutting tokens by 35.5%-54.7%, running 1.8x-2.6x faster, and saving 2.7x-3.7x in cost. Code and data are released on GitHub. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is a hybrid reasoning agent and how does it combine latent and text-based reasoning? A hybrid reasoning agent dynamically switches between latent (internal, non-verbal) reasoning and text-based reasoning within one unified generation process. HybridAgent uses a lightweight component called HybridLink to transition seamlessly between the two modes, plus a continuous working memory that preserves lossless context across turns, keeping latent thoughts inspectable while retaining tool-integrated agentic capabilities. _daily.dev surfaces agent-architecture research like this for engineers weighing latent versus text-based reasoning designs._ ### How much faster and cheaper is HybridAgent compared to standard LLM agents? HybridAgent delivers 1.8x-2.6x faster inference and 2.7x-3.7x cost savings compared with frontier LLM agents and existing text-based or latent agentic frameworks, while also reducing token usage by 35.5%-54.7% and improving task performance by 10.4% on average across 11 benchmarks spanning math, science, medicine, search, and agentic reasoning. _track efficiency gains like these on daily.dev when evaluating agent frameworks for production inference costs._ ### What post-training methods are used to train hybrid reasoning agents like HybridAgent? Two dedicated algorithms are used: HybridSFT, which stabilizes tool-integrated hybrid generation by teaching the backbone model to coordinate with the HybridLink component, and HybridRL, which scales reasoning by learning when the model should think in latent space versus act through text. Both were evaluated across 3 model scales. _daily.dev keeps builders of agentic training pipelines current on post-training techniques like these._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 95 comments across hackernews (as of 2026-10-19). **TL;DR:** Much of the discussion centers on interpretability/safety concerns around latent reasoning being less auditable than text chains-of-thought, with a smaller thread debating whether LLMs 'think' at all; technical curiosity about benchmarks and comparisons to similar architectures is more measured and positive. **Sentiment:** 20% positive · 30% mixed · 50% skeptical **The case for** - Some see latent reasoning as more natural/efficient since not all thinking requires language (e.g., protein folding, spatial reasoning). - Open weights, paper, and code release were appreciated as enabling scrutiny and experimentation. - A few found the approach intuitive and likely reflective of how frontier models already work internally. **The pushback** - Major worry that removing text-based chain-of-thought makes it much harder to audit a model's reasoning for malicious or deceptive behavior. - Concern that mapping latent steps back to nearest token embeddings for 'interpretability' is lossy and may not hold up, especially for models trained from scratch in latent space. - Some argue training a critic against latent chain-of-thought is literally 'The Most Forbidden Technique,' risking optimizing for deception that evades the quantized audit signal. - One tester reported the 14B model looping on repeated output and apparent confusion about basic formatting (CSV vs markdown), suggesting possible llama.cpp compatibility issues or quality concerns. **By community** - hackernews (heated): A large, contentious subthread debates whether hiding reasoning in latent space is a dangerous loss of interpretability/alignment ('the Most Forbidden Technique'), alongside a separate heated tangent on whether LLMs can 'think' at all, with more muted technical questions and a mixed early hands-on report. **Hottest debate:** Whether moving reasoning into an uninspectable latent space meaningfully undermines interpretability and safety auditing, versus it being a natural and efficient evolution of how models 'think'. **Open questions** - Will the claimed ability to recover latent reasoning semantics via nearest-token mapping hold up for models trained from scratch to reason in latent space, rather than retrofitted? - How does this architecture compare in safety/alignment terms to other latent-reasoning approaches like looped transformers or simply increasing model depth? - Is the apparent looping/quality issue seen in early hands-on testing of the 14B model a quantization or llama.cpp compatibility problem, or a deeper issue with the model? **Highlights** > It's worth pointing out that merely having the chain-of-thought in latent space isn't by itself "The Most Forbidden Technique". Rather, it's doing what the authors propose as the next step in the paper: Hybrid generation can also make post-training more efficient. Because latent thoughts remain auditable during rollout, an external critic can assign step-level credit to latent spans, branch promising trajectories, or terminate ones whose latent reasoning has already gone wrong, before any text or tool call is produced. Unless I'm misunderstanding, they're suggesting RL training against the CoT. Which, uh, that is The Most Forbidden Technique, as I understand it, when combined with the latent reasoning space. It allows for the model to "learn"/optimize for malicious behaviors that are just far enough away from the quantized token they're auditing against. > — [spijdar on hackernews](https://news.ycombinator.com/item?id=50146035) > One of the core claims made in the paper is "To inspect any latent reasoning span, one can easily recover its semantics by mapping each latent step produced by HybridLink to its nearest token embedding, as in Eq. 5". They show some charts to support this. In doing this however, you lose meaning where the model is expressing combinations of tokens in a vector, and your logging/monitoring outputs just the closest one. You would need another model going through and outputting the full meaning being described (with more than one output token per reasoning token) I wonder if their observability claims will hold up on a model trained from scratch to reason in latent space. > — [chabons on hackernews](https://news.ycombinator.com/item?id=50146053) > If you accept any of the AI-equivalents of Pascal's Wager, latent thinking represents an uninterpretable and thus unalignable portion of AI machinery. Ie: a malevolent AI can hide bad stuff from people and Manchurian candidate themselves after it's too late. > — [kelseyfrog on hackernews · 1 comments](https://news.ycombinator.com/item?id=50145777) > I'm running the 14B in llama.cpp (a version from yesterday), the 8-bit quantization from bartowski, and it looped on variations of the same paragraph for five minutes before running out its turn budget, in response to a pretty simple chat request that other similar-sized models (e.g. Gemma 4 12B) have no trouble with, a web search and data summarization task, sort of the sweet spot for models this size. It's possible llama.cpp needs some update to work with this model, but it's based on Qwen 3, so I would expect it to be pretty well-supported. The chat template is the same one upstream ships, which may or may not be correct. I'm currently downloading the 16-bit version, to see if it's quantization pathology. I'm also not confident it knows what a CSV is, either, as it put the data into a Markdown table during its reasoning phase as an example of how to present the data (my prompt asks for data in a CSV). My experience with most Qwen fine-tunes is they get dumber, but their benchmark results suggest otherwise, so I have to assume it's a llama.cpp incompatibility. Edit: Same looping in BF16. So, probably have to wait on a llama.cpp update. > — [SwellJoe on hackernews](https://news.ycombinator.com/item?id=50146330) > Shockingly all the answers so far are subtly incorrect or extremely unclear. The actual answer is that this is massive loss for Interpretibilty, AKA "telling whether the model is lying to you or not". Which sounds silly when you're thinking about chatbots, but is an incredibly important line in the sand for superheavy, longrunning, self-improving agents. To say the least! It's based on this philosophy, FWIW: https://plato.stanford.edu/entries/language-thought/ And a good writeup I just found of recent worries regarding Astra's same shift towards intelligence at the cost of safety: https://www.lesswrong.com/posts/HCRs8btkiamtWSNAL/astra-is-h... > — [bbor on hackernews](https://news.ycombinator.com/item?id=50146120) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50145337) · 148 points · 95 comments ## Similar posts on daily.dev - [How to Combine Traditional Machine Learning with Agentic Reasoning](https://daily.dev/posts/how-to-combine-traditional-machine-learning-with-agentic-reasoning-mvbgh1sty) · Machine Learning Mastery · 0 upvotes · 0 comments - [Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on Agent Logic](https://daily.dev/posts/beyond-llms-why-scalable-enterprise-ai-adoption-depends-on-agent-logic-b9zmijebk) · Hugging Face · 1 upvotes · 0 comments - [Kimi K2 Thinking : Best Agentic Reasoning LLM is here, beats GPT5, Sonnet 4.5](https://daily.dev/posts/kimi-k2-thinking-best-agentic-reasoning-llm-is-here-beats-gpt5-sonnet-4-5-ef6sapttw) · Medium · 16 upvotes · 0 comments - [Explainable AI Agents: Capture LLM Tool Call Reasoning with Spring AI](https://daily.dev/posts/explainable-ai-agents-capture-llm-tool-call-reasoning-with-spring-ai-jxsstkrzq) · Spring · 2 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#ai-agents](https://daily.dev/tags/ai-agents), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/hybrid-reasoning-agents-llm-agents-that-think-in-latent-space-act-through-text-gqx1d3tvr) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Hybrid Reasoning Agents: LLM agents that think in latent space, act through text","url":"https://daily.dev/posts/hybrid-reasoning-agents-llm-agents-that-think-in-latent-space-act-through-text-gqx1d3tvr","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/hybrid-reasoning-agents-llm-agents-that-think-in-latent-space-act-through-text-gqx1d3tvr"},"datePublished":"2026-10-18T22:19:35.723Z","dateModified":"2026-10-19T00:05:40.933Z","description":"Researchers introduce Hybrid Reasoning Agents (HybridAgent), a new architecture that lets LLM agents dynamically switch between latent (internal) reasoning and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1fed0de190485cbbb37a86fdf656d71b?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1fed0de190485cbbb37a86fdf656d71b?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/hybrid-reasoning-agents-llm-agents-that-think-in-latent-space-act-through-text-gqx1d3tvr","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,ai-agents,reinforcement-learning","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Hybrid Reasoning Agents: LLM agents that think in latent space, act through text"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/hybrid-reasoning-agents-llm-agents-that-think-in-latent-space-act-through-text-gqx1d3tvr#faq","mainEntity":[{"@type":"Question","name":"What is a hybrid reasoning agent and how does it combine latent and text-based reasoning?","acceptedAnswer":{"@type":"Answer","text":"A hybrid reasoning agent dynamically switches between latent (internal, non-verbal) reasoning and text-based reasoning within one unified generation process. HybridAgent uses a lightweight component called HybridLink to transition seamlessly between the two modes, plus a continuous working memory that preserves lossless context across turns, keeping latent thoughts inspectable while retaining tool-integrated agentic capabilities. daily.dev surfaces agent-architecture research like this for engineers weighing latent versus text-based reasoning designs."}},{"@type":"Question","name":"How much faster and cheaper is HybridAgent compared to standard LLM agents?","acceptedAnswer":{"@type":"Answer","text":"HybridAgent delivers 1.8x-2.6x faster inference and 2.7x-3.7x cost savings compared with frontier LLM agents and existing text-based or latent agentic frameworks, while also reducing token usage by 35.5%-54.7% and improving task performance by 10.4% on average across 11 benchmarks spanning math, science, medicine, search, and agentic reasoning. track efficiency gains like these on daily.dev when evaluating agent frameworks for production inference costs."}},{"@type":"Question","name":"What post-training methods are used to train hybrid reasoning agents like HybridAgent?","acceptedAnswer":{"@type":"Answer","text":"Two dedicated algorithms are used: HybridSFT, which stabilizes tool-integrated hybrid generation by teaching the backbone model to coordinate with the HybridLink component, and HybridRL, which scales reasoning by learning when the model should think in latent space versus act through text. Both were evaluated across 3 model scales. daily.dev keeps builders of agentic training pipelines current on post-training techniques like these."}}]} ```