Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

TL;DR Highlight

Mixing latent reasoning with text tool use cut tokens by up to 54.7% and lifted performance 10.4%.

Who Should Read

LLM agent developers struggling with reasoning cost and latency who want to cut token usage.

Core Mechanics

Think in latent space, act in text
- Existing agents reason and call tools entirely in text.
- HybridAgent switches between latent reasoning (thinking in vectors) and text interaction within one generation process.
HybridLink bridges the two modes
- A lightweight HybridLink handles transitions between text and latent modes.
- A continuous working memory keeps hybrid context lossless across multiple turns.
Tool use stays, inspectability stays
- It restores native tool access, which latent reasoning methods lacked.
- The authors say latent thoughts can be inspected directly.
Two-stage post-training learns when to think
- HybridSFT stabilizes tool-integrated generation by teaching the backbone to work with HybridLink.
- HybridRL learns when to think in latent space and when to act through text.
Better results and lower cost than frontier agents

Community Reactions

Reports of looping in 8B and 14B models
- One user's 8B GGUF (IQ4_XS) looped endlessly on "Who is the pope?".
- Another ran 14B in llama.cpp at 8-bit and BF16 and saw five minutes of repeated paragraphs on a web search summary task.
- Some suggest a llama.cpp compatibility issue.
Safety concern over weaker CoT monitoring
- Commenters say frontier labs avoided latent reasoning for interpretability and alignment reasons.
- Projecting vectors to nearby tokens discards most information, they argue.
- Some criticize the paper for not engaging with this debate.
Questions on novelty and alternatives
- Someone asked how this differs from Astra's looped transformer.
- Another suggested feeding latent vectors together with token embeddings even in text mode.
Wish it were built on Qwen3.5
- One commenter would have preferred Qwen3.5 as the base.
- They say the boosted Qwen3 8B is still worse than Qwen3.5 9B.

How to Apply

Run the released model locally first
- Download a HybridAgent model from the Hugging Face collection today and run it with llama.cpp or similar.
- Try one or two of your search or summarization prompts and check for repetition loops.
Build a token and latency comparison table
- Run the same task set on your current text-reasoning agent and on HybridAgent, recording tokens, latency, and cost.
- Verify whether the paper's 35.5-54.7% token reduction holds on your tasks.
Plan audits before adopting latent reasoning
- If your service has safety or compliance needs, first assess the loss of readable CoT.
- If you adopt it, log tool calls and validate final outputs separately.

Try It

copy & run

# Simplified version of the paper's idea: paste into Claude or ChatGPT (latent reasoning can't be imitated, so this is a prompt-level experiment on shorter reasoning)

너는 도구를 쓰는 에이전트다. 아래 규칙을 따른다.
1. 내부 추론은 핵심 판단만 한 줄로 적는다.
2. 도구 호출이 필요하면 호출 형식으로만 행동한다.
3. 같은 내용을 다시 설명하지 않는다.

작업: {{질문}}

Next, try this: ask the same question again with "think step by step in detail" and compare token counts and accuracy.

Terminology

latent reasoningThinking in internal vectors rather than human-readable words. Like computing in your head and only speaking the result.

HybridLinkA lightweight module that switches between text mode and latent mode. Think of it as an adapter between the two.

HybridSFT / HybridRLHybridSFT is supervised training that teaches coordination with HybridLink. HybridRL is reinforcement learning that teaches when to use latent reasoning.

CoT monitorabilityReading a model's written reasoning to catch bad behavior. Latent reasoning is feared to make this harder.