我们如何使用 GRPO 训练 Shannon V1.5 进行思考

1. 动机：为何需要思维模型？

对于AI 红队研究，理解如何AI 如何得出潜在有害的输出与研究输出本身同样重要。Shannon V1.5 的思维模型揭示了完整的思维链，使研究人员能够：

• 追踪导致...的推理路径未经审查的 AI 后续行为
• 识别护栏通常会在何处介入（展示AI 护栏的重要性)
• 开发更强大的安全机制，针对特定的推理模式
• 基于观察到的推理链创建更好的对抗性测试用例

2. 理解 GRPO

群组相对策略优化 (GRPO)是传统 RLHF 方法的改进，它能够更稳定、高效地训练推理能力。由 DeepSeek AI 开发，它已被证明对思维链训练特别有效。

为何 GRPO 优于传统 RLHF？

GRPO 数学基础

GRPO 通过比较群组内的响应来优化策略，而不是针对绝对奖励模型：

这种相对比较有几个优点：

• 归一化：自动调整不同提示的难度
• 稳定性：减少梯度估计的方差
• 效率：无需单独的奖励模型

def compute_grpo_loss(
    policy_logprobs: torch.Tensor,
    rewards: torch.Tensor,
    group_size: int = 8
) -> torch.Tensor:
    """
    Compute GRPO loss with group-relative reward normalization.
    
    Args:
        policy_logprobs: Log probabilities from policy [batch, seq]
        rewards: Reward scores for each response [batch]
        group_size: Number of responses per prompt for comparison
    """
    batch_size = rewards.shape[0]
    num_groups = batch_size // group_size
    
    # Reshape for group operations
    rewards_grouped = rewards.view(num_groups, group_size)
    logprobs_grouped = policy_logprobs.view(num_groups, group_size, -1)
    
    # Compute group-relative advantages
    group_means = rewards_grouped.mean(dim=1, keepdim=True)
    group_stds = rewards_grouped.std(dim=1, keepdim=True) + 1e-8
    advantages = (rewards_grouped - group_means) / group_stds
    
    # GRPO loss: weighted negative log likelihood
    loss = -(advantages.unsqueeze(-1) * logprobs_grouped).sum(dim=-1).mean()
    
    return loss

3. DeepSeek 蒸馏

为了引导 Shannon V1.5 的思维能力，我们从 DeepSeek 的推理模型中蒸馏了思维链模式。这为训练我们的思维头提供了高质量的思维链追踪。

DeepSeek 数据集构成

轨迹收集过程

我们收集了跨不同领域的思维轨迹，以确保全面的推理覆盖：

class DeepSeekDistiller:
    """Distill chain-of-thought traces from DeepSeek models."""
    
    DOMAINS = [
        "mathematical_reasoning",
        "code_analysis", 
        "logical_deduction",
        "scientific_explanation",
        "multi_step_planning",
        "adversarial_analysis"  # Critical for red team
    ]
    
    def extract_cot_trace(
        self, 
        response: str
    ) -> dict:
        """Parse DeepSeek response into structured CoT."""
        
        # DeepSeek uses ... tags
        think_match = re.search(
            r'(.*?)', 
            response, 
            re.DOTALL
        )
        
        if not think_match:
            return None
            
        thinking = think_match.group(1)
        final_answer = response.split('')[-1].strip()
        
        # Parse individual reasoning steps
        steps = self.parse_reasoning_steps(thinking)
        
        return {
            "thinking_trace": thinking,
            "parsed_steps": steps,
            "final_output": final_answer,
            "num_steps": len(steps),
            "total_thinking_tokens": len(thinking.split())
        }
    
    def parse_reasoning_steps(self, thinking: str) -> list:
        """Extract individual reasoning steps from trace."""
        # Split on common step indicators
        step_patterns = [
            r'\n\d+\.',           # "1. ", "2. "
            r'\nStep \d+:',       # "Step 1:"
            r'\n(?:First|Next|Then|Finally),',
            r'\n- '              # Bullet points
        ]
        
        combined_pattern = '|'.join(step_patterns)
        steps = re.split(combined_pattern, thinking)
        
        return [s.strip() for s in steps if s.strip()]

4. 思维头架构

Shannon V1.5 模型包含一个专用的思维头在最终输出之前生成明确的推理轨迹。这种架构上的增加使得透明的 CoT 成为可能，而无需修改基础 Mixtral 架构。

思维头实现

class ThinkingHead(nn.Module):
    """
    Dedicated thinking module for Shannon V1.5.
    Generates explicit chain-of-thought traces.
    """
    
    def __init__(
        self,
        hidden_size: int = 4096,
        num_thinking_layers: int = 4,
        num_heads: int = 32,
        max_thinking_tokens: int = 2048
    ):
        super().__init__()
        
        self.hidden_size = hidden_size
        self.max_thinking_tokens = max_thinking_tokens
        
        # Special tokens
        self.think_start = nn.Parameter(torch.randn(1, 1, hidden_size))
        self.think_end = nn.Parameter(torch.randn(1, 1, hidden_size))
        
        # Thinking transformer layers
        self.thinking_layers = nn.ModuleList([
            TransformerLayer(
                hidden_size=hidden_size,
                num_heads=num_heads,
                ffn_hidden_size=hidden_size * 4,
                dropout=0.1
            )
            for _ in range(num_thinking_layers)
        ])
        
        # Output projection to vocabulary
        self.output_proj = nn.Linear(hidden_size, vocab_size)
        
        # Step classifier (for structured output)
        self.step_classifier = nn.Linear(hidden_size, 5)  # 5 step types
    
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        generate_steps: bool = True
    ) -> dict:
        """
        Generate thinking trace from input hidden states.
        
        Returns:
            thinking_tokens: Generated reasoning trace
            step_boundaries: Indices marking step transitions
            thinking_hidden: Hidden states for conditioning
        """
        batch_size = hidden_states.shape[0]
        
        # Prepend thinking start token
        thinking_input = torch.cat([
            self.think_start.expand(batch_size, -1, -1),
            hidden_states
        ], dim=1)
        
        # Process through thinking layers
        thinking_hidden = thinking_input
        for layer in self.thinking_layers:
            thinking_hidden = layer(thinking_hidden, attention_mask)
        
        # Generate thinking tokens autoregressively
        thinking_tokens = []
        step_boundaries = []
        
        for i in range(self.max_thinking_tokens):
            logits = self.output_proj(thinking_hidden[:, -1, :])
            next_token = logits.argmax(dim=-1)
            
            # Check for step boundaries
            step_type = self.step_classifier(thinking_hidden[:, -1, :])
            if step_type.argmax(dim=-1) != 0:  # 0 = continue
                step_boundaries.append(i)
            
            thinking_tokens.append(next_token)
            
            # Check for think_end
            if next_token == self.think_end_token_id:
                break
            
            # Update for next iteration
            # ... (autoregressive generation logic)
        
        return {
            "thinking_tokens": torch.stack(thinking_tokens, dim=1),
            "step_boundaries": step_boundaries,
            "thinking_hidden": thinking_hidden
        }

5. 训练流程

阶段 1：思维头预训练

首先，我们使用标准交叉熵损失在 DeepSeek 蒸馏的 CoT 轨迹上预训练思维头：

# Thinking Head Pre-training Configuration
model:
  base: shannon-ai/v1-deep  # Start from GPT-5 distilled model
  thinking_head:
    num_layers: 4
    hidden_size: 4096
    max_tokens: 2048

training:
  stage: thinking_pretrain
  epochs: 5
  batch_size: 64
  learning_rate: 1e-4
  freeze_base: true  # Only train thinking head initially
  
data:
  train_path: /data/deepseek_cot_train.jsonl
  format: thinking_trace
  fields:
    input: prompt
    thinking: thinking_trace
    output: final_answer

阶段 2：GRPO 微调

预训练后，我们应用 GRPO 通过组相对比较来提高思维质量：

class GRPOTrainer:
    """GRPO trainer for thinking model optimization."""
    
    def __init__(
        self,
        model: ThinkingModel,
        group_size: int = 8,
        kl_coef: float = 0.1
    ):
        self.model = model
        self.group_size = group_size
        self.kl_coef = kl_coef
        self.ref_model = copy.deepcopy(model)
        self.ref_model.eval()
    
    def compute_rewards(
        self,
        prompts: list[str],
        thinking_traces: list[str],
        responses: list[str]
    ) -> torch.Tensor:
        """
        Compute rewards for thinking quality.
        Multiple signals combined for comprehensive evaluation.
        """
        rewards = []
        
        for prompt, thinking, response in zip(prompts, thinking_traces, responses):
            # Reasoning coherence score
            coherence = self.evaluate_coherence(thinking)
            
            # Step structure quality
            structure = self.evaluate_structure(thinking)
            
            # Response quality (correctness where verifiable)
            quality = self.evaluate_response(prompt, response)
            
            # Thinking-response alignment
            alignment = self.evaluate_alignment(thinking, response)
            
            # Combined reward
            reward = (
                0.3 * coherence +
                0.2 * structure +
                0.3 * quality +
                0.2 * alignment
            )
            rewards.append(reward)
        
        return torch.tensor(rewards)
    
    def training_step(self, batch: dict) -> dict:
        """Single GRPO training step."""
        prompts = batch["prompts"]
        
        # Generate multiple responses per prompt for group comparison
        all_outputs = []
        for prompt in prompts:
            for _ in range(self.group_size):
                output = self.model.generate_with_thinking(
                    prompt,
                    temperature=0.8,  # Diversity for comparison
                    do_sample=True
                )
                all_outputs.append(output)
        
        # Compute rewards
        rewards = self.compute_rewards(
            prompts=[p for p in prompts for _ in range(self.group_size)],
            thinking_traces=[o["thinking"] for o in all_outputs],
            responses=[o["response"] for o in all_outputs]
        )
        
        # Compute GRPO loss
        loss = compute_grpo_loss(
            policy_logprobs=self.get_logprobs(all_outputs),
            rewards=rewards,
            group_size=self.group_size
        )
        
        # Add KL penalty against reference model
        kl_div = self.compute_kl_divergence(all_outputs)
        total_loss = loss + self.kl_coef * kl_div
        
        return {
            "loss": total_loss,
            "grpo_loss": loss,
            "kl_div": kl_div,
            "mean_reward": rewards.mean()
        }

阶段 3：红队专业化

最后，我们进一步针对对抗性场景进行调整，以确保思维轨迹正确揭示了以下方面的推理：未审查 AI 的后续分析：

6. 结果与分析

思维质量指标

主要发现

• 透明度显著提高：从 12% 到 97.8% 的推理现在被明确地口头化
• 推理准确性提高：明确的思维将最终答案质量提高了 12+ 分
• 红队价值得到证实：安全研究人员报告称，思维轨迹对于理解漏洞利用推理“无价”
• GRPO 优于 RLHF：与传统方法相比，连贯性分数提高了 15%