“How can we build an agent that thinks beyond human language, yet still acts through it in the external environment?” That question opens the Hybrid Reasoning Agents paper (arXiv:2610.18861), a joint effort from Stanford and Together AI. The answer is HybridAgent — a system that pushes most of an agent’s chain of thought out of text and into hidden vector space, while still calling tools through ordinary language.
The numbers reported across coverage are striking. An average gain of +10.4% across 11 benchmarks, paired with 35.5–54.7% fewer tokens. Cheaper reasoning that does not cost accuracy.
TL;DR: HybridAgent, a Stanford and Together AI research project, lets LLM agents reason inside a hidden vector space instead of writing every intermediate thought as text. A lightweight module called HybridLink connects latent states to text-based tool use, trained through HybridSFT and HybridRL. Across 11 benchmarks it scored +10.4% higher on average while cutting token use by 35.5–54.7%, showing cheaper reasoning need not cost accuracy.
What Is a Hybrid Reasoning Agent?
A hybrid reasoning agent alternates between two modes of computation: latent reasoning inside the model’s hidden vector space, and text-based actions in the external environment. Modern LLM agents still reason and act almost entirely through text, which means every intermediate step of thinking consumes generated tokens. HybridAgent breaks that pattern.
The architecture rests on two components described in the paper: a lightweight HybridLink and a continuous working memory. Instead of emitting a full textual chain of thought before each tool call, the agent can think in vectors — continuous representations that never become tokens — and surface into text only when it needs to invoke a tool or produce an answer.
The practical consequence is measurable. Together AI researchers report that hybrid agents cut decoding compute by 62.5% to 72.9% and still scored higher than text-only baselines. Agents run faster and score higher when most thinking happens in hidden vectors, not written text.
The models used were Qwen3-Base at 4B, 8B, and 14B parameter sizes. Notably, coverage of the project reports that the 8B agent, thinking without writing tokens, beats GPT-5.6 Luna on GAIA — though not on HLE. Scale is not everything. Reasoning placement matters.
Why Do Token-Based Agents Waste So Much Compute?
Every word an agent writes during reasoning is a token that must be generated, one at a time, by the decoder. The cost compounds quickly. An agent that plans, self-corrects, and reflects in text can produce thousands of tokens before it performs a single useful action.
The reported results quantify the waste. HybridAgent uses 35.5–54.7% fewer tokens across its benchmark suite, with some sources reporting savings of up to 54.7% and inference up to 2.6 times faster. The explainx.ai writeup characterizes the resulting agents as roughly 3x cheaper. These cost estimates exclude prompt and tool costs, so the comparison focuses narrowly on what the agent itself generates.
Where does the saving come from? Intermediate reasoning steps — the planning, the deliberation, the working-through of a problem — move into the latent space, where they do not require token-by-token decoding. Only the actionable surface, the tool calls and final answers, passes through text.
Why does this matter beyond cost? Because the benchmarks show the latent reasoning did not merely save tokens. It produced +10.4% higher average performance. Thinking in vectors, according to the paper’s framing, lets the model reason beyond the constraints of human language.
How Does HybridLink Connect Latent Thinking to Text Actions?
The bridge between the two worlds is HybridLink, described in the paper as a lightweight component that lets HybridAgents alternate between latent reasoning and text-based tool use. The agent reasons internally in continuous vectors, then transitions into generated text precisely when interaction with the environment is required.
Supporting this alternation is a continuous working memory. Rather than maintaining state as a growing text context, the agent can hold intermediate results as vectors, keeping its deliberation compact. Text appears at the boundary: tool calls go out, observations come back, and the agent folds them into its latent state.
This design targets three goals at once, according to coverage of the paper: better agent performance, fewer tokens, and faster inference with lower cost. The benchmark results suggest the design delivers — improvements across 11 benchmarks, with token reductions of 35.5–54.7%.
The elegance of the approach is that it does not abandon text. Tools, APIs, and environments speak language. HybridAgent keeps that interface intact and changes only what happens inside the agent’s head between actions. Latent for thought. Text for action.
How Were These Agents Trained?
Training a model to interleave vector-space reasoning with text actions requires explicit coordination, and the paper describes two complementary methods: HybridSFT and HybridRL. Together, they teach the agent when to think latently and when to surface into text.
HybridSFT — supervised fine-tuning — establishes the basic alternation pattern between latent reasoning steps and text-based tool use. HybridRL then optimizes this coordination further with reinforcement learning, improving performance on the full benchmark suite.
The training pipeline was applied to Qwen3-Base models at 4B, 8B, and 14B sizes. Coverage from explainx.ai notes that code was released alongside the paper (arXiv 2610.18861), and the project drew attention on Hacker News, where sources describe it as gaining traction quickly.
The results validate the training approach. Across 11 benchmarks, the hybrid agents improved by +10.4% on average while reducing token consumption by 35.5–54.7%. Decoding compute fell by 62.5% to 72.9%.
In my opinion, the SFT-plus-RL combination is the least surprising and most transferable part of the work — it suggests existing agent training pipelines could adopt latent reasoning without reinventing their methodology.
The remaining question is how these numbers break down benchmark by benchmark, how the 8B model’s GAIA result compares against its HLE showing, and where latent reasoning still falls short. Part 2 digs into the results.
What Do the Benchmarks Actually Show?
HybridAgent was evaluated across 11 benchmarks, and the results reported by the Stanford and Together AI team point in one consistent direction: fewer generated tokens, higher scores. The agents cut decoding compute by 62.5% to 72.9% while still scoring higher than their text-only counterparts. That combination is unusual, because cheaper reasoning is usually assumed to cost accuracy.
The headline numbers include a 10.4% average improvement, 35.5% to 54.7% fewer tokens, and inference that is up to 2.6 times faster. The gains come from training the coordination between latent and textual modes rather than from a single clever prompt.
Not every result is a win, though. On GAIA the 8B agent outperformed GPT-5.6 Luna, but on HLE it did not. Coverage of the paper is explicit about that split, and honest readers should be too.
A word of caution on the cost side: the paper’s cost estimates exclude prompt and tool costs. That matters when comparing against commercial APIs, because an agent’s real bill includes every token entering the context window, not just the ones it writes out.
Can an 8B Model Really Beat GPT-5.6 Luna on GAIA?
Yes, on GAIA — and the sources are careful to qualify it. The DEV Community write-up on the paper states directly that an 8B agent which thinks without writing tokens beats GPT-5.6 Luna on GAIA, and just as explicitly notes it does not on HLE. The qualification is part of the claim, not a footnote to it.
How does a small model manage this? The mechanism is HybridLink, the component that lets the agent alternate between latent reasoning and text-based tool use. Instead of spending output tokens narrating every intermediate step, the model reasons in continuous vectors and only commits to text when it needs to act in the environment.
The practical takeaway is that raw parameter count is not the whole story when the reasoning loop itself is redesigned. A 4B, 8B, or 14B Qwen3-Base model with hybrid training can concentrate its capacity where it counts. But the HLE result is a reminder that frontier-level general knowledge is a different target than tool-driven task completion.
How Much Money and Time Do Hybrid Agents Save?
The savings come from a simple observation: decoding is expensive, and HybridAgent decodes far less. Stanford and Together AI researchers measured a 62.5% to 72.9% reduction in decoding compute. Inference runs up to 2.6 times faster, and total token use drops by as much as 54.7%.
Separate coverage of the release frames the economics even more aggressively, describing latent reasoning agents as up to 3x cheaper. Sources also note the project’s traction beyond the paper itself, with the work described as circulating widely on Hacker News.
A representative picture of the reported gains:
- 10.4% average performance improvement across 11 benchmarks
- 35.5% to 54.7% fewer generated tokens
- Up to 2.6x faster inference
- 62.5% to 72.9% less decoding compute
- Up to 3x cheaper per the explainx.ai summary of the release
One caveat belongs in every procurement conversation: the cost estimates in the paper exclude prompt and tool costs. If your workload involves large contexts or expensive tool calls, the real-world savings may look different from the headline figures.
What Are the Limitations of Latent Reasoning?
The clearest documented limit is the HLE result. The same 8B agent that beats GPT-5.6 Luna on GAIA does not do so on HLE, and sources state this openly. Latent reasoning helps with agentic, tool-driven tasks; it is not a shortcut to frontier-level performance on every benchmark.
A second limitation is architectural. HybridAgents need a lightweight HybridLink and continuous working memory to coordinate the two reasoning modes, and this coordination has to be trained through HybridSFT and HybridRL. That means you cannot bolt latent reasoning onto an arbitrary model with a prompt; the pipeline involves supervised fine-tuning followed by reinforcement learning.
There is also an interpretability question. When the model reasons in hidden vectors rather than written text, the intermediate steps leave no token trail to inspect. The paper’s framing — thinking beyond human language, yet still acting through it — captures the trade-off: the thoughts are efficient precisely because they are not written down.
A third practical limit: cost analyses that exclude prompt and tool costs will understate total spend for context-heavy workflows.
How Can Developers Try HybridAgent Today?
The code is out. Coverage of the release states plainly that the code was released alongside the paper, which means developers can inspect the HybridLink implementation and the HybridSFT/HybridRL training pipeline rather than reconstructing them from the arXiv text alone.
The paper is available as arXiv 2610.18861, indexed under Hybrid Reasoning Agents on Papers with Code. The models used are Qwen3-Base at 4B, 8B, and 14B scales, so anyone with hardware for those sizes can explore the setup.
A short getting-started checklist based on the published material:
- Read the paper (arXiv 2610.18861) for the HybridLink and continuous working memory design
- Pull the released code and review the training stages
- Start with the 4B Qwen3-Base model if compute is tight
- Compare token counts and latency against your text-only baseline
- Remember that prompt and tool costs sit outside the paper’s estimates
The concept has also attracted independent explainers across DEV Community in multiple languages, which helps newcomers parse the architecture before touching the code.
Frequently Asked Questions
What is HybridAgent?
HybridAgent is a hybrid reasoning agent from Stanford and Together AI researchers, described in arXiv paper 2610.18861. It alternates between latent reasoning (thinking in vectors) and text-based tool use, coordinated through a lightweight component called HybridLink and a continuous working memory.
How many tokens does HybridAgent save?
Sources report 35.5% to 54.7% fewer tokens, with decoding compute cut by 62.5% to 72.9% according to the Stanford and Together AI paper. Inference is up to 2.6 times faster, and one summary of the release describes the approach as up to 3x cheaper overall.
Which base models does HybridAgent use?
The models are Qwen3-Base at 4B, 8B, and 14B scales. Across 11 benchmarks the hybrid approach delivered a 10.4% average improvement, and the 8B variant beat GPT-5.6 Luna on GAIA while not matching it on HLE.
Is the code available?
Yes — coverage of the release states that the code was released alongside the paper. The paper itself is on arXiv as 2610.18861 and indexed on Papers with Code, so both the implementation and the methodology are publicly accessible.
Summary
HybridAgent is one of the more concrete answers yet to the question of whether reasoning has to happen in text. The evidence from 11 benchmarks suggests it largely does not.
- Latent reasoning plus text-based action, coordinated by HybridLink and continuous working memory, raised average performance by 10.4% across 11 benchmarks.
- Efficiency gains are large: 35.5% to 54.7% fewer tokens, up to 2.6x faster inference, and 62.5% to 72.9% less decoding compute per the Stanford and Together AI paper.
- The 8B Qwen3-Base agent beat GPT-5.6 Luna on GAIA but not on HLE — a qualified result the sources state honestly.
- The cost estimates exclude prompt and tool costs, so real-world savings depend on your workload.
- The code and the paper (arXiv 2610.18861) are public, so the fastest way to judge the approach is to run it against your own text-only baseline.
If you work with agentic systems, try the released code on a small Qwen3-Base model and measure token spend on your own tasks. Have you hit a wall with token costs in your agents? The comments are open.