Artificial Intelligence (AI) has made rapid progress in recent years, with large language models (LLMs) leading the way toward artificial general intelligence (AGI). OpenAI’s o1 has introduced advanced inference-time scaling techniques, significantly improving reasoning capabilities. However, its closed-source nature limits accessibility.

A new breakthrough in AI research comes from DeepSeek, which has unveiled DeepSeek-R1, an open-source model designed to enhance reasoning capabilities through large-scale reinforcement learning. The research paper, “DeepSeek-R1: Incentivizing Reasoning Capability in Large Language Models via Reinforcement Learning,” offers an in-depth roadmap for training LLMs using reinforcement learning techniques. This article explores the key aspects of DeepSeek-R1, its innovative training methodology, and its potential impact on AI-driven reasoning.

Revisiting LLM Training Fundamentals

Before diving into the specifics of DeepSeek-R1, it’s essential to understand the fundamental training process of LLMs. The development of these models generally follows three critical stages:

1. Pre-training

The foundation of any LLM is built during the pre-training phase. At this stage, the model is exposed to massive amounts of text and code, allowing it to learn general-purpose knowledge. The primary objective here is to predict the next token in a sequence. For instance, given the prompt “write a bedtime _,” the model might complete it with “story.” However, despite acquiring extensive knowledge, the model remains ineffective at following human instructions without further refinement.

2. Supervised Fine-Tuning (SFT)

In this phase, the model is fine-tuned using a curated dataset containing instruction-response pairs. These pairs help the model understand how to generate more human-aligned responses. After supervised fine-tuning, the model improves at following instructions and engaging in meaningful conversations.

3. Reinforcement Learning

The final stage involves refining the model’s responses using reinforcement learning. Traditionally, this is done through Reinforcement Learning from Human Feedback (RLHF), where human evaluators rate responses to train the model. However, obtaining large-scale, high-quality human feedback is challenging. An alternative approach, Reinforcement Learning from AI Feedback (RLAIF), utilizes a highly capable AI model to provide feedback instead. This reduces reliance on human labor while still ensuring quality improvements.

DeepSeek-R1-Zero: A Novel Approach to RL-Driven Reasoning

One of the most striking aspects of DeepSeek-R1 is its departure from the conventional supervised fine-tuning phase. Instead of following the standard process, DeepSeek introduced DeepSeek-R1-Zero, which is trained entirely through reinforcement learning. This innovative model is built upon DeepSeek-V3-Base, a pre-trained model with 671 billion parameters.

By omitting supervised fine-tuning, DeepSeek-R1-Zero achieves state-of-the-art reasoning capabilities using an alternative reinforcement learning strategy. Unlike traditional RLHF or RLAIF, DeepSeek employs Rule-Based Reinforcement Learning, a cost-effective and scalable method.

The Power of Rule-Based Reinforcement Learning

DeepSeek-R1-Zero relies on an in-house reinforcement learning approach called Group Relative Policy Optimization (GRPO). This technique enhances the model’s reasoning capabilities by rewarding outputs based on predefined rules instead of relying on human feedback. The process unfolds as follows:

Generating Multiple Outputs: The model is given an input problem and generates multiple possible outputs, each containing a reasoning process and an answer.

Evaluating Outputs with Rule-Based Rewards: Instead of relying on AI-generated or human feedback, predefined rules assess the accuracy and format of each output.

Training the Model for Optimal Performance: The GRPO method trains the model to favor the best outputs, improving its reasoning abilities.

Key Rule-Based Rewards

Accuracy Reward: If a problem has a deterministic correct answer, the model receives a reward for arriving at the correct conclusion. For coding-related tasks, predefined test cases validate the output.

Format Reward: The model is instructed to format its responses correctly. For example, it must structure its reasoning process within tags and present its final answer within tags.

By leveraging these rule-based rewards, DeepSeek-R1-Zero eliminates the need for a neural-based reward model, reducing computational costs and minimizing risks like reward hacking—where a model exploits loopholes to maximize rewards without actually improving its reasoning.

DeepSeek-R1-Zero’s Performance and Benchmarking

The effectiveness of DeepSeek-R1-Zero is evident in its performance benchmarks. When compared to OpenAI’s o1 model, it demonstrates comparable or superior reasoning abilities across various reasoning-intensive tasks.

In particular, results from the AIME dataset showcase an impressive improvement in the model’s performance. The pass@1 score—which measures the accuracy of the model’s first attempt at solving a problem—skyrocketed from 15.6% to 71.0% during training, reaching levels on par with OpenAI’s closed-source model.

Self-Evolution: The AI’s ‘Aha Moment’

One of the most fascinating aspects of DeepSeek-R1-Zero’s training process is its self-evolution. Over time, the model naturally learns to allocate more time to complex reasoning tasks. This means that as training progresses, the model increasingly refines its thought process, much like a human would when tackling a challenging problem.

A particularly intriguing phenomenon observed during training is the “Aha Moment.” This refers to instances where the model reevaluates its reasoning mid-process. For example, when solving a math problem, DeepSeek-R1-Zero may initially take an incorrect approach but later recognize its mistake and self-correct. This capability emerges organically during reinforcement learning, demonstrating the model’s ability to refine its reasoning autonomously.

Why Develop DeepSeek-R1?

Despite the groundbreaking performance of DeepSeek-R1-Zero, it exhibited certain limitations:

Readability Issues: The outputs were often difficult to interpret.

Inconsistent Language Usage: The model frequently mixed multiple languages within a single response, making interactions less coherent.

To address these concerns, DeepSeek introduced DeepSeek-R1, an improved version of the model trained through a four-phase pipeline.

The Training Process of DeepSeek-R1

DeepSeek-R1 refines the reasoning abilities of DeepSeek-R1-Zero while improving readability and consistency. The training follows a structured four-phase process:

1. Cold Start (Phase 1)

The model starts with DeepSeek-V3-Base and undergoes supervised fine-tuning using a high-quality dataset curated from DeepSeek-R1-Zero’s best outputs. This step improves readability while maintaining strong reasoning abilities.

2. Reasoning Reinforcement Learning (Phase 2)

Similar to DeepSeek-R1-Zero, this phase applies large-scale reinforcement learning using rule-based rewards. This enhances the model’s reasoning in areas like coding, mathematics, science, and logic.

3. Rejection Sampling & Supervised Fine-Tuning (Phase 3)

In this phase, the model generates numerous responses, and only accurate and readable outputs are retained using rejection sampling. A secondary model, DeepSeek-V3, helps select the best samples. These responses are then used for additional supervised fine-tuning to further refine the model’s capabilities.

4. Diverse Reinforcement Learning (Phase 4)

The final phase involves reinforcement learning across a wide range of tasks. For math and coding-related challenges, rule-based rewards are used, while for more subjective tasks, AI feedback ensures alignment with human preferences.

DeepSeek-R1: A Worthy Competitor to OpenAI’s o1

The final version of DeepSeek-R1 delivers remarkable results, outperforming OpenAI’s o1 in several benchmarks. Notably, a distilled 32-billion-parameter version of the model also exhibits exceptional reasoning capabilities, making it a smaller yet highly efficient alternative.

Final Thoughts

DeepSeek-R1 marks a significant step forward in AI reasoning capabilities. By leveraging rule-based reinforcement learning, DeepSeek has demonstrated that supervised fine-tuning is not always necessary for training powerful LLMs. Moreover, the introduction of DeepSeek-R1 addresses key readability and consistency challenges while maintaining state-of-the-art reasoning performance.

As the AI research community moves toward open-source models with advanced reasoning capabilities, DeepSeek-R1 stands out as a compelling alternative to proprietary models like OpenAI’s o1. Its release paves the way for further reinforcement learning and large-scale AI training innovation.