togethercomputer/HybridAgent

togethercomputer/HybridAgent

Language: Python

License: MIT

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-10-09T02:59:46Z

Pushed: 2026-10-11T05:26:12Z

Default branch: main

Fork: no

Archived: no

README:

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

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.

| Artifact | Description | Hugging Face | |---|---|---| | HybridAgent-8B-Agentic | Qwen3-8B-Base → HybridSFT → HybridRL on agentic reasoning tasks | togethercomputer/HybridAgent-8B-Agentic | | HybridAgent-14B-Agentic | Qwen3-14B-Base → HybridSFT → HybridRL on agentic reasoning tasks | togethercomputer/HybridAgent-14B-Agentic | | HybridAgent-HybridSFT-v1 / -v2 | HybridSFT training sets: hybrid trajectories (train.jsonl), v1 and the upgraded v2 | togethercomputer/HybridAgent-HybridSFT-v1, togethercomputer/HybridAgent-HybridSFT-v2 | | HybridAgent-HybridRL | HybridRL training sets: deep reasoning (math.parquet) and agentic search (search.parquet) | 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

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

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):

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

MODEL=Qwen/Qwen3-8B-Base DATA=togethercomputer/HybridAgent-HybridSFT-v2 OUT=outputs/hybrid_sft \
bash scripts/run_hybrid_sft.sh all

Runs the three HybridSFT stages in sequence: format activation, HybridLink activation and self-conditioned HybridSFT. Recommended hardware: 8 or 16 H100 80 GB GPUs; the provided configs are for one node with 8 GPUs, and on 16 GPUs across two nodes --grad_accum 1 keeps the same effective batch of 16. See [hybrid_sft/](hybrid_sft/).

HybridRL

TASK=math MODEL=outputs/hybrid_sft/self_conditioned DATA=hf://togethercomputer/HybridAgent-HybridRL/math.parquet \
NNODES=4 NODE_RANK= MASTER_ADDR= bash scripts/run_hybrid_rl.sh

Trains HybridRL on the deep-reasoning tasks; TASK=search switches to the agentic search setting. Recommended hardware: 32 H100 80 GB GPUs (4 nodes × 8), the launcher default; the rollout batch is split across ranks, so --prompts_per_step must be divisible by the number of GPUs. See [hybrid_rl/](hybrid_rl/).

📚 Citation

@article{zou2026hybrid,
title = {Hybrid Reasoning Agents},
author = {Zou, Jiaru and Qi, Yunzhe and Chen, Sirui and Bianchi, Federico and Zhu, Shang and
Lu, Pan and He, Jingrui and Zou, James},
journal = {arXiv preprint arXiv:XXXX.XXXXX},...