Light-IF: Endowing LLMs with Generalizable Reasoning via Preview and Self-Checking

---

Aug. 13 update: SOTA 14B

🧪 Benchmarks

Model	SuperClue	IFEval	CFBench	IFBench
Qwen3-4B	0.225	0.888	0.787	0.382
Qwen3-8B	0.225	0.888	0.813	0.417
Qwen3-14B	0.227	0.898	0.827	0.422
Qwen3-32B	0.234	0.877	0.823	0.384
Qwen3-235B-A22B	0.244	0.882	0.834	0.423
Qwen3-235B-A22B-Thinking-2507	0.434	0.916	0.843	0.475
DeepSeek-R1-0528	0.436	0.863	0.827	0.415
Doubao-seed-1-6-thinking-250615	0.362	0.832	0.82	0.477
Doubao-seed-1-6-thinking-250715	0.345	0.856	0.84	0.366
ChatGPT-4o-latest	0.260	0.836	0.807	0.365
Deepseek-v3-250324	0.306	0.859	0.833	0.405
Doubao-1.5-pro-32k-250115	0.285	0.889	0.797	0.375
Kimi-K2	0.227	0.921	0.820	0.395
GLM-4.5	0.395	0.893	0.833	0.466
Light-IF-4B (ours) 🤗	0.445	0.916	0.80	0.443
Light-IF-8B (ours) 🤗	0.471	0.932	0.82	0.473
Light-IF-32B (ours) 🤗	0.575	0.938	0.85	0.575
Light-IF-14B (ours) 🤗	0.589	0.962	0.833	0.697

Light-IF is a powerful instruction-following large language model (LLM) series that leverages Preview-Checking reasoning to handle complex instructions with generalizable behavior — all trained with less than $3,000 in compute.

Light-IF-14B 🤗 is the most powerful 14B instruction-following model we have open-sourced, even outperforming Light-IF-32B.

This remarkable performance is largely attributed to our carefully designed curriculum learning strategy.

---

📌 Highlights

• 🔍 Identifies and overcomes lazy reasoning in LLMs.
• 🧩 Integrates Preview + Self-Checking mechanisms.
• 🚀 Combines Entropy-SFT and TEA-RL for robust training.
• 💡 Achieves state-of-the-art results on instruction benchmarks.
• 💰 Trained efficiently on A800 GPUs at very low cost.

---

🔨 Technical Overview

Light-IF addresses the challenge of poor instruction-following due to lazy reasoning. Its pipeline includes:

1. Hardness-aware Prompt Synthesis

• Construct prompts with complex verifiable constraints.
• Filter invalid outputs using LLMs to form high-quality datasets.

2. Zero-RL Training

• Train a base model to reject lazy thinking with length-based and correctness-based rewards.

3. Entropy-Preserving SFT

• Select tokens by balancing NLL and entropy.
• Prevents overfitting and preserves model diversity.

4. TEA-RL (Token-wise Entropy-Adaptive RL)

• Dense rewards for partially satisfying constraints.
• Entropy-regularized policy gradient for stable learning.

The overall framework of the proposed method:

---

💻 Quick Usage

The following contains a code snippet illustrating how to use the model generate content based on given inputs.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "qihoo360/Light-IF-32B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Please help me write a poem with a total of 15 lines and no more than 300 words. The poem should be divided into 4 stanzas, each beginning with a **highlighted subtitle**."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True 
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
print("thinking content:", thinking_content)
print("content:", content)

---

⚙️ Training Cost

Model	GPUs	Hours	Cost (USD)
Light-IF-1.7B	A800×4	10	~$342
Light-IF-32B	A800×88	30	~$2,800

---

📜 License

This repository is licensed under the Apache 2.0 License.

---

Citation

@article{Light-IF,
  title={Light-IF: Endowing LLMs with Generalizable Reasoning via Preview and Self-Checking for Complex Instruction Following},
  author={Light-IF Team},
  journal={arXiv preprint arXiv:2508.03178},
  year={2025}
}

Star History