yolay/RAIF-DeepSeek-Qwen-1.5B

Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models

This model is the official implementation of the paper Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models.

You can find the official code and more details on the GitHub repository.

Abstract

Existing large language models (LLMs) face challenges of following complex instructions, especially when multiple constraints are present and organized in paralleling, chaining, and branching structures. One intuitive solution, namely chain-of-thought (CoT), is expected to universally improve capabilities of LLMs. However, we find that the vanilla CoT exerts a negative impact on performance due to its superficial reasoning pattern of simply paraphrasing the instructions. It fails to peel back the compositions of constraints for identifying their relationship across hierarchies of types and dimensions. To this end, we propose RAIF, a systematic method to boost LLMs in dealing with complex instructions via incentivizing reasoning for test-time compute scaling. First, we stem from the decomposition of complex instructions under existing taxonomies and propose a reproducible data acquisition method. Second, we exploit reinforcement learning (RL) with verifiable rule-centric reward signals to cultivate reasoning specifically for instruction following. We address the shallow, non-essential nature of reasoning under complex instructions via sample-wise contrast for superior CoT enforcement. We also exploit behavior cloning of experts to facilitate steady distribution shift from fast-thinking LLMs to skillful reasoners. Extensive evaluations on seven comprehensive benchmarks confirm the validity of the proposed method, where a 1.5B LLM achieves 11.74% gains with performance comparable to a 8B LLM. Evaluation on OOD constraints also confirms the generalizability of our RAIF.

Usage

This model can be loaded and used directly with the Hugging Face transformers library.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "yolay/RAIF-DeepSeek-Qwen-1.5B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Example for instruction following
messages = [
    {"role": "user", "content": "Explain the concept of quantum entanglement in simple terms, using an analogy from everyday life."}
]

# Apply chat template for proper input formatting
input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

# Generate response using parameters from generation_config.json
outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.6,
    top_p=0.9
)

generated_text = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(generated_text)

Model Details

The model DeepSeek-Qwen2.5-1.5B is our optimized model for its advanced instruction-following capabilities under complex instructions. It corresponds to the DeepSeek-Qwen1.5B (Ours) in the Table 1.

Table 1 Performance on seven instruction benchmarks. Best/2nd best are marked bold/underlined.

Model	Method	IFEval	CELLO	CF Bench	Complex Bench	FB Bench	Follow Bench	Info Bench	Avg.
Qwen2.5-1.5B-Instruct	I/O	45.28	71.00	36.00	50.97	39.81	40.00	71.24	50.61
Qwen2.5-1.5B-Instruct	CoT	28.65	59.30	22.00	32.94	37.31	29.28	62.22	38.81 (-11.79%)
Qwen2.5-1.5B-Instruct	SDC	41.95	66.10	30.00	41.70	36.52	37.39	67.55	45.89 (-4.71%)
Qwen2.5-1.5B-Instruct	SFT	65.61	71.20	48.00	57.46	42.75	56.47	76.22	59.67 (+9.06%)
Qwen2.5-1.5B-Instruct	Ours	44.91	73.50	53.66	63.92	58.67	59.82	81.95	62.35 (+11.74%)
DeepSeek-Qwen1.5B	I/O†	36.04	62.50	27.99	39.89	34.51	20.29	52.00	39.03
DeepSeek-Qwen1.5B	SFT	45.29	63.20	25.33	35.53	37.59	22.18	51.96	40.15 (+1.12%)
DeepSeek-Qwen1.5B	Ours	57.67	69.00	40.00	44.38	37.78	37.79	60.48	49.58 (+10.54%)
DeepScaleR-1.5B	I/O†	41.77	65.00	30.00	40.70	40.24	26.01	60.31	43.43
DeepScaleR-1.5B	SFT	48.24	62.90	28.00	36.68	35.72	26.50	54.22	41.75 (-1.67%)
DeepScaleR-1.5B	Ours	55.63	67.30	39.33	43.23	37.81	36.80	60.08	48.60 (+5.17%)
Qwen2.5-7B-Instruct	I/O	72.82	76.50	64.33	74.47	59.29	75.03	85.60	72.58
Qwen2.5-7B-Instruct	CoT	69.50	75.20	61.66	72.00	42.65	74.86	82.13	68.28 (-4.29%)
Qwen2.5-7B-Instruct	SDC	60.44	72.60	65.66	76.53	60.07	76.09	86.88	71.18 (-1.39%)
Qwen2.5-7B-Instruct	SFT	72.45	77.50	63.33	74.23	58.76	75.92	84.31	72.36 (-0.21%)
Qwen2.5-7B-Instruct	Ours	70.06	79.20	65.00	77.40	64.45	75.32	82.67	73.44 (+0.85%)
Ministral-8B-Instruct	I/O	59.51	76.20	62.33	70.03	54.54	73.49	84.00	68.58
Ministral-8B-Instruct	CoT	48.79	61.90	49.66	61.31	39.17	61.75	79.73	57.47 (-11.11%)
Ministral-8B-Instruct	SDC	58.59	63.60	56.99	68.32	48.06	69.37	84.08	64.14 (-4.43%)
Ministral-8B-Instruct	SFT	68.57	66.30	48.66	67.20	37.26	54.37	76.62	59.85 (-8.72%)
Ministral-8B-Instruct	Ours	72.64	72.6	59.33	70.45	54.35	76.08	75.33	68.68 (+0.10%)
DeepSeek-Qwen7B	I/O†	60.81	72.39	57.99	66.86	59.59	62.80	79.64	65.73
DeepSeek-Qwen7B	SFT	67.09	69.10	58.66	58.42	55.60	65.96	79.15	64.85 (-0.88%)
DeepSeek-Qwen7B	Ours	71.35	71.40	58.67	62.04	59.65	59.38	82.00	66.35 (+0.62%)

---

Table 2 Performance on ComplexBench (Qwen2.5-7B-Instruct). Best/2nd best are marked bold/underlined. OD, SC, CNFR, FC, and SR stand for Oracle Decomposition, Self-Consistency, Conifer, FollowComplex, and Self-Refine.

Category	ND	I/O	OD	SC	CNFR	FC	SR	Ours
And	1	85.85	84.27	84.03	75.10	84.77	85.66	86.57
Chain
	1	72.18	74.68	73.54	60.95	66.27	75.25	73.96
	2	70.56	72.70	69.63	64.43	70.66	73.07	76.88
Avg.	-	70.96	73.18	70.57	63.59	69.60	73.59	76.18
Selection
	1	77.25	76.61	72.08	60.52	71.67	69.61	73.39
	2	65.61	71.83	68.23	53.25	61.96	64.34	72.92
	3	63.39	68.45	56.13	46.04	51.70	58.67	60.75
Avg.	-	65.67	70.49	65.83	51.92	60.92	62.69	69.16
Selection & Chain
	2	65.64	65.94	60.81	47.33	61.07	52.01	61.06
	3	59.70	65.77	64.08	48.53	57.65	60.41	65.00
Avg.	-	62.68	65.85	62.44	47.93	59.36	56.20	63.03
Overall	-	74.47	76.26	73.76	63.51	71.97	74.00	77.40

---

Acknowledgement🫡

In this project, we follow the SimpleRL and the OpenRLHF framework to prepare the codebase. We acknowledge their great work for open-sourcing the implementations of reinforcement learning algorithms.

• [SimpleRL]
• [OpenRLHF]

We also would like to express gratitude to the research community that organize the existing benchmarks for validating the LLMs of solving complex instructions.

License🪪

Please refer to License_RAIF for the license of this project.

Citation🎓

If you find this work useful, please consider the following citation:

@article{qin2025incentivizingreasoningadvancedinstructionfollowing,
      title={Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models},
      author={Yulei Qin and Gang Li and Zongyi Li and Zihan Xu and Yuchen Shi and Zhekai Lin and Xiao Cui and Ke Li and Xing Sun},
      year={2025},
      eprint={2506.01413},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2506.01413}
}