# Hybrid Reasoning Agents
**LLM Agents that Think in latent space, Act through text**
[](https://arxiv.org/abs/2610.18861)
[](https://huggingface.co/collections/togethercomputer/hybridagent)
[](https://x.com/Jiaru_Zou/status/2110400256647340476)
[](https://lnkd.in/p/g_3PW9PS)
[Jiaru Zou](https://scholar.google.com/citations?user=GzLTey4AAAAJ)1,2, [Yunzhe Qi](https://scholar.google.com/citations?user=Gt17_A0AAAAJ)3,4, [Sirui Chen](https://csr16.github.io/)3, [Federico Bianchi](https://federicobianchi.io/)2, [Shang Zhu](https://shang-zhu.github.io/)2, [Pan Lu](https://lupantech.github.io/)1, [Jingrui He](https://www.hejingrui.org/)3, [James Zou](https://www.james-zou.com/)1,2
1Stanford University · 2Together AI · 3UIUC · 4Meta
**HybridAgent** is an LLM agent that dynamically navigates between latent reasoning and text-based interaction within a single generation process. A lightweight **HybridLink** maps last-layer hidden states back into the input-embedding space, so the agent can enter **latent mode** for fast, inspectable inner thinking and return to **text mode** for tool calls, final answers and external monitoring. The complete hybrid context of text, latent thoughts and tool observations is carried across turns through the agent's own **working memory**, and every latent thought can be rendered as readable pseudo-text.
This repository contains the method: the hybrid generation engine with Python and search tools, the **HybridSFT** and **HybridRL** post-training stages, batch inference, and the **Interpretable Latent Mode** inspector.
## ⚙️ Installation
```bash
git clone https://github.com/togethercomputer/HybridAgent.git && cd HybridAgent
conda create -n hybridagent python=3.11 -y && conda activate hybridagent
pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
pip install -e .
cp .env.example .env # TAVILY_API_KEY for the search tool
```
## 📁 Repository
| Directory | Contents |
|---|---|
| [`hybrid_agent/`](hybrid_agent/) | HybridLink, the model wrapper and loader, the hybrid generation engine (text/latent modes, working memory, concurrent tool calls) and the tools |
| [`hybrid_sft/`](hybrid_sft/) | HybridSFT: data format, objective and trainer |
| [`hybrid_rl/`](hybrid_rl/) | HybridRL: hybrid policy rollout, reward, exact rollout replay, policy update |
| [`hybrid_inference/`](hybrid_inference/) | Batch hybrid inference over a JSONL of questions |
| [`interpretable_latent_mode/`](interpretable_latent_mode/) | Latent thoughts as pseudo-text: HTML reports and terminal view |
| [`scripts/`](scripts/) | Launchers for the three stages |
## 🤗 Released models and data
Checkpoints and data load from a local directory or directly from the Hugging Face Hub. See the [HybridAgent collection](https://huggingface.co/collections/togethercomputer/hybridagent).
| Artifact | Description | Hugging Face |
|---|---|---|
| HybridAgent-8B-Agentic | Qwen3-8B-Base → HybridSFT → HybridRL on agentic reasoning tasks | [`togethercomputer/HybridAgent-8B-Agentic`](https://huggingface.co/togethercomputer/HybridAgent-8B-Agentic) |
| HybridAgent-14B-Agentic | Qwen3-14B-Base → HybridSFT → HybridRL on agentic reasoning tasks | [`togethercomputer/HybridAgent-14B-Agentic`](https://huggingface.co/togethercomputer/HybridAgent-14B-Agentic) |
| HybridAgent-HybridSFT-v1 / -v2 | HybridSFT training sets: hybrid trajectories (`train.jsonl`), v1 and the upgraded v2 | [`togethercomputer/HybridAgent-HybridSFT-v1`](https://huggingface.co/datasets/togethercomputer/HybridAgent-HybridSFT-v1), [`togethercomputer/HybridAgent-HybridSFT-v2`](https://huggingface.co/datasets/togethercomputer/HybridAgent-HybridSFT-v2) |
| HybridAgent-HybridRL | HybridRL training sets: deep reasoning (`math.parquet`) and agentic search (`search.parquet`) | [`togethercomputer/HybridAgent-HybridRL`](https://huggingface.co/datasets/togethercomputer/HybridAgent-HybridRL) |
*Note:* Each model directory holds the backbone in Hugging Face format together with the HybridLink weights (`adapter.pt`, `adapter_config.json`).
## 🚀 Quick start
### Batch Hybrid Inference
```bash
python -m hybrid_inference.run_batch --model togethercomputer/HybridAgent-8B-Agentic \
--input examples/questions.jsonl --output outputs/results.jsonl --num_samples 4
bash scripts/run_batch_inference.sh 8 --model togethercomputer/HybridAgent-8B-Agentic \
--input questions.jsonl --output outputs/results.jsonl # one process per GPU
```
One record per question: the hybrid trajectory, the extracted answer, the pseudo-text readout of every latent span and the tool-call statistics. See [`hybrid_inference/`](hybrid_inference/).
### Interpretable latent mode
```bash
python -m interpretable_latent_mode --model togethercomputer/HybridAgent-14B-Agentic \
--question "How many trailing zeros does 100! have?" --output report.html
```
Renders the latent thoughts as pseudo-text inline with the text reasoning, tool calls and observations. HybridAgent-14B-Agentic on the first question of `examples/questions.jsonl` (terminal view, text segments omitted):
```text
QUESTION At the speed of sound in air at 20 degrees Celsius, how many seconds does sound take
to travel 10 kilometers? Round to the nearest second.
LATENT [span 1 · 42 steps · readability 0.84]
Wait, I me the the speed of speed in air at 20° is about 343 meters per second. But I
should verify that. Let me do a search search to be sure.
SEARCH speed of sound in air at 20 degrees celsius meters per second
RESULT Page 1: The speed of sound in dry air at 20 °C is 343 meters per second. ...
LATENT [span 2 · 22 steps · readability 0.88]
Now, I need to calculate the time it takes for speed to travel 10 km at this speed.
LATENT [span 3 · 21 steps · readability 0.96]
So, time = 10,000 meters / 343 meters per second.
PYTHON ... print(f"Rounded to the nearest second: {rounded_time} seconds")
RESULT Time taken: 29.154518950437318 seconds
Rounded to the nearest second: 29 seconds
LATENT [span 7 · 10 steps · readability 0.92]
So, the answer is 29 seconds.
ANSWER \boxed{29}
```
See [`interpretable_latent_mode/`](interpretable_latent_mode/) for the full trace and the HTML report.
### HybridSFT
```bash
MODEL=Qwen/Qwen3-8B-Base DATA=togethercomputer/HybridAgent-HybridSFT-v2 OUT=outputs/hybrid_sft \
bash scripts/run_hybrid_sft.sh all
```
Train HybridSFT on the curated hybrid trajectories (format activation, HybridLink activation and self-conditioned training). Recommended hardware: 8 or 16 GPUs. See [`hybrid_sft/`](hybrid_sft/).
### HybridRL
```bash
TASK=search MODEL=outputs/hybrid_sft/self_conditioned DATA=hf://togethercomputer/HybridAgent-HybridRL/search.parquet \
NNODES=4 NODE_RANK= MASTER_ADDR= bash scripts/run_hybrid_rl.sh
```
Train HybridRL on the agentic search tasks. Recommended hardware: 32 GPUs (4 nodes × 8), the launcher default. See [`hybrid_rl/`](hybrid_rl/).
## 📚 Citation
```bibtex
@misc{zou2026hybridagent,
title={Hybrid Reasoning Agents},
author={Jiaru Zou and Yunzhe Qi and Sirui Chen and Federico Bianchi and Shang Zhu and Pan Lu and Jingrui He and James Zou},
year={2026},
eprint={2610.18861},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2610.18861},
}
```