KAT

• We released the technical report of the KAT-V1 model, available at https://arxiv.org/pdf/2507.08297.
• Kwaipilot-AutoThink ranks first among all open-source models on LiveCodeBench Pro, a challenging benchmark explicitly designed to prevent data leakage, and even surpasses strong proprietary systems such as Seed and o3-mini.

Introduction

KAT (Kwaipilot-AutoThink) is an open-source large-language model that mitigates over-thinking by learning when to produce explicit chain-of-thought and when to answer directly.

Its development follows a concise two-stage training pipeline:

    • Think-off queries labeled via a custom tagging system.
    • Think-on queries generated by a multi-agent solver.

Data Format

KAT produces responses in a structured template that makes the reasoning path explicit and machine-parsable. Two modes are supported:

Special Tokens

🔧 Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Kwaipilot/KAT-V1-40B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=65536,
    temperature=0.6,
    top_p=0.95,
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
content = tokenizer.decode(output_ids, skip_special_tokens=True).strip("\n")
print("prompt:\n", prompt)
print("content:\n", content)
"""
prompt:
Give me a short introduction to large language model.
content:
The user's request is to provide a concise factual introduction to large language models, which involves retrieving and summarizing basic information. This task is straightforward as it only requires recalling and presenting well-known details without deeper analysis. No complex reasoning is needed here—just a simple explanation will suffice.

A **Large Language Model (LLM)** is an advanced AI system trained on vast amounts of text data to understand, generate, and process human-like language. Here’s a concise introduction:

### Key Points:
1. **Training**: Trained on diverse text sources (books, websites, etc.) using deep learning.
2. **Capabilities**:
   - Answer questions, generate text, summarize content, translate languages.
   - Understand context, sentiment, and nuances in language.
3. **Architecture**: Often based on **transformer models** (e.g., BERT, GPT, LLaMA).
4. **Scale**: Billions of parameters, requiring massive computational resources.
5. **Applications**: Chatbots, content creation, coding assistance, research, and more.

### Examples:
- **OpenAI’s GPT-4**: Powers ChatGPT.
- **Google’s Gemini**: Used in Bard.
- **Meta’s LLaMA**: Open-source alternative.

### Challenges:
- **Bias**: Can reflect biases in training data.
- **Accuracy**: May hallucinate "facts" not grounded in reality.
- **Ethics**: Raises concerns about misinformation and job displacement.

LLMs represent a leap forward in natural language processing, enabling machines to interact with humans in increasingly sophisticated ways. 🌐🤖
"""

Future Releases

Looking ahead, we will publish a companion paper that fully documents the AutoThink training framework, covering:

• Cold-start initialization procedures
• Reinforcement-learning (Step-SRPO) strategies
• Data curation and reward design details

At the same time, we will open-source:

• Training resources – the curated dual-regime datasets and RL codebase
• Model suite – checkpoints at 1.5B, 7B, and 13B parameters, all trained with AutoThink gating

Citation

@techreport{Zhan2025KATV1,
  title={KAT-V1: Kwai-AutoThink Technical Report},
  author={Zizheng, Zhan and Ken, Deng and Huaixi, Tang and Wen, Xiang and Kun, Wu and others},
  year={2025},
  institution={arXiv preprint arXiv:2507.08297},
  number={arXiv:2507.08297},
  url={https://arxiv.org/abs/2507.08297}
}