HybridAgents: LLM Agents That Think in Latent Space, Act Through Text

TL;DR

Stanford and UC Berkeley researchers have unveiled Hybrid Reasoning Agents (HybridAgents), a novel architecture that lets LLM agents seamlessly alternate between latent reasoning (internal, non-verbal thought) and text-based interaction (tool calls, final answers) within a single generation process. A lightweight module called HybridLink enables this dynamic switching, while a continuous working memory preserves context across turns.

In benchmarks spanning mathematics, science, medicine, search, and agentic reasoning, HybridAgents consistently outperform frontier LLM agents by 10.4%, while delivering 35.5%–54.7% token reduction, 1.8x–2.6x faster inference, and 2.7x–3.7x cost savings. The work, submitted to arXiv on October 13, 2026, introduces a two-phase post-training pipeline—HybridSFT and HybridRL—that teaches models when to think latently and when to act through text.

The Core Problem: Why Hybrid Reasoning?

Modern LLM agents are trapped in a fundamental trade-off. On one hand, text-based reasoning—the standard approach—offers transparency and native tool access, but it's token-hungry and slow. On the other hand, latent reasoning (thinking in continuous vector space) is efficient and fast, but it lacks the ability to invoke tools or produce human-readable outputs.

HybridAgents bridges this divide. As the paper's authors ask: "How can we build an agent that thinks beyond human language, yet still acts through it in the external environment?" Their answer is a unified architecture that dynamically navigates between both modes, choosing the most efficient path for each reasoning step.

Inside the Architecture: HybridLink and Working Memory

At the heart of HybridAgents is HybridLink, a lightweight residual adapter that sits between the model's hidden states and its input embedding space. Here's how it works in practice:

• Text mode: The model behaves exactly like a standard LLM, sampling tokens from the vocabulary. When it generates a special <latent> token, it enters latent mode.
• Latent mode: Instead of decoding a token and re-embedding it, HybridLink transforms the last-layer hidden state into the next step's input representation. This allows multiple latent steps to run without token generation.
• Return to text: When the nearest-token readout of a latent state matches a designated marker, the model surfaces back to text mode.

Critically, the transition is dynamic—span lengths and frequency are generated on the fly, not fixed. The residual design of HybridLink preserves much of the hidden state's semantic content while a learned projection aligns it with the input embedding space, keeping overhead negligible.

The architecture also includes a continuous working memory that maintains lossless hybrid context across multiple turns. This ensures latent thoughts aren't lost when the model switches back to text for tool calls or final answers—a key limitation of previous latent-reasoning approaches.

Training: HybridSFT and HybridRL

The researchers developed a two-stage post-training pipeline, mirroring the SFT-then-RL paradigm popularized by reasoning models like DeepSeek-R1 and OpenAI's o-series—but applied to the hybrid routing decision itself.

HybridSFT (Supervised Fine-Tuning) teaches the backbone model to coordinate with HybridLink, stabilizing tool-integrated hybrid generation. This phase ensures the model can fluidly switch between modes without breaking tool calls or producing incoherent outputs.

HybridRL (Reinforcement Learning) then scales reasoning by learning when to think in latent space versus when to act through text. This is the crucial optimization: the model learns to route simple steps through efficient latent reasoning while reserving text for tool invocations and final answers.

Benchmark Results: Consistent Wins Across Domains

The team evaluated HybridAgents at three model scales across 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning. The results are striking:

• Performance: 10.4% average improvement over frontier LLM agents and existing agentic frameworks (both text-based and latent methods).
• Efficiency: 35.5%–54.7% token reduction, meaning fewer tokens processed per task.
• Speed: 1.8x–2.6x faster inference—latent steps avoid the expensive token-by-token decoding loop.
• Cost: 2.7x–3.7x cost savings, a direct consequence of reduced token usage and faster inference.

These gains hold across diverse domains, suggesting the hybrid approach generalizes well rather than overfitting to specific task types.

Why It Matters

The implications extend beyond raw performance metrics. HybridAgents address a fundamental tension in AI agent design: transparency vs. efficiency.

Latent reasoning is powerful but opaque—it's hard to audit why a model made a decision. Text reasoning is transparent but slow. HybridAgents offer a pragmatic middle ground: efficient internal thinking when speed matters, and inspectable text when accountability matters. The paper notes that latent thoughts remain "directly inspectable", which could be critical for debugging and trust in production systems.

For engineers building agentic systems, the cost and speed improvements are immediately actionable. A 2.6x speedup and 3.7x cost reduction at comparable or better accuracy is the kind of win that moves the needle in production environments.

The Road Ahead

The paper (45 pages, 12 figures) is a preprint, and the code and data are publicly available, allowing the research community to build on these findings. While the results are promising, peer review and replication studies will be necessary to fully validate the approach.

Still, HybridAgents represent a significant step toward more efficient, more capable AI agents. By letting models think beyond language while still acting through it, this research could reshape how we build everything from coding assistants to autonomous research tools.