菜单
Augment verifiable reward RL (RLVR) with contrastive learning to generate dense auxiliary rewards. Enforce proximity among correct reasoning trajectories in embedding space while suppressing errors, amplifying invariant reasoning patterns.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill clipo-contrastive-policy-optimization命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
正在显示 SKILL.md
| name | clipo-contrastive-policy-optimization |
| title | CLIPO: Contrastive Learning in Policy Optimization Generalizes RLVR |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.10101 |
| keywords | ["RLVR","Contrastive Learning","Policy Optimization","Reasoning","Dense Rewards"] |
| description | Augment verifiable reward RL (RLVR) with contrastive learning to generate dense auxiliary rewards. Enforce proximity among correct reasoning trajectories in embedding space while suppressing errors, amplifying invariant reasoning patterns. |
Sparse verifiable rewards (binary success/failure) provide limited training signal for complex reasoning tasks. CLIPO adds contrastive learning as an auxiliary objective: it embeds reasoning trajectories in latent space and applies InfoNCE loss to cluster correct responses together while repelling errors.
The insight is that successful reasoning paths share consistent underlying logic structures. By enforcing this structure in embedding space, contrastive learning acts as a denoising mechanism, amplifying invariant reasoning patterns while suppressing spurious shortcuts and hallucinations.
CLIPO extends RLVR policy optimization algorithms (GRPO, GSPO, DAPO, GMPO) by introducing a lightweight contrastive head and auxiliary reward:
This dual signal prevents optimization collapse on narrow heuristics while maintaining grounding in task-specific verifiable rewards.
Create embeddings for reasoning trajectories by processing token hidden states.
import torch
import torch.nn as nn
class TrajectoryContrastiveHead(nn.Module):
def __init__(self, hidden_dim, embedding_dim=256, projection_dim=128):
super().__init__()
.projection = nn.Linear(hidden_dim, projection_dim)
.contrastive_head = nn.Sequential(
nn.Linear(projection_dim, embedding_dim),
nn.ReLU(),
nn.Linear(embedding_dim, projection_dim)
)
():
trajectory_repr = hidden_states.mean(dim=)
projected = .projection(trajectory_repr)
embedding = .contrastive_head(projected)
embedding
Within each batch, group by correctness and compute contrastive pairs.
def infonce_loss(embeddings, labels, temperature=0.1):
"""
InfoNCE loss for trajectory embeddings.
embeddings: (batch_size, embedding_dim)
labels: (batch_size,) binary correctness labels
temperature: temperature parameter for softmax
"""
batch_size = embeddings.shape[0]
# Normalize embeddings
embeddings = torch.nn.functional.normalize(embeddings, dim=1)
# Compute similarity matrix
similarity_matrix = torch.mm(embeddings, embeddings.t()) / temperature
# Create positive and negative masks
labels_expanded = labels.unsqueeze(1)
positive_mask = (labels_expanded == labels_expanded.t()).float()
# Set diagonal (self-similarity) to 0 for positive mask
positive_mask.fill_diagonal_(0)
# Negative mask is complement
negative_mask = 1.0 - positive_mask
negative_mask.fill_diagonal_(0)
# InfoNCE: log(exp(sim_pos) / sum(exp(sim_neg)))
exp_sim = torch.exp(similarity_matrix)
# Sum of positive similarities per row
pos_sum = (exp_sim * positive_mask).sum(dim=1, keepdim=True)
# Sum of negative similarities per row
neg_sum = (exp_sim * negative_mask).sum(dim=1, keepdim=True)
# InfoNCE loss (avoiding division by zero)
infonce = -torch.log(
(pos_sum + 1e-8) / (pos_sum + neg_sum + 1e-8)
)
return infonce.mean()
Transform contrastive objective into a dense reward signal that complements verifiable reward.
def contrastive_reward(
embeddings,
labels,
verifiable_rewards,
contrastive_weight=0.5,
temperature=0.1
):
"""
Compute combined reward: verifiable + contrastive auxiliary.
Returns dense reward per trajectory.
"""
batch_size = embeddings.shape[0]
# Normalize embeddings
embeddings_norm = torch.nn.functional.normalize(embeddings, dim=1)
# Compute similarity matrix
similarity_matrix = torch.mm(embeddings_norm, embeddings_norm.t()) / temperature
# Create positive mask (other correct trajectories)
labels_expanded = labels.unsqueeze(1)
positive_mask = (labels_expanded == labels_expanded.t()).float()
positive_mask.fill_diagonal_(0)
# Average positive similarity per trajectory
pos_similarities = (similarity_matrix * positive_mask).sum(dim=1) / (
positive_mask.sum(dim=1) + 1e-8
)
# Average negative similarity per trajectory
negative_mask = 1.0 - positive_mask
negative_mask.fill_diagonal_(0)
neg_similarities = (similarity_matrix * negative_mask).sum(dim=1) / (
negative_mask.sum(dim=1) + 1e-8
)
# Contrastive reward: pull positives, push negatives
contrastive_reward_signal = pos_similarities - neg_similarities
# Combine with verifiable reward
total_reward = (
verifiable_rewards +
contrastive_weight * contrastive_reward_signal
)
return total_reward, contrastive_reward_signal
Extend baseline policy optimization with contrastive auxiliary objective.
def train_step_with_clipo(
model,
input_ids,
verifiable_rewards,
contrastive_head,
policy_optimizer,
contrastive_weight=0.5,
temperature=0.1
):
"""
Single training step combining RLVR with contrastive learning.
"""
# Forward pass through model
outputs = model(
input_ids,
output_hidden_states=True
)
# Extract trajectory embeddings
hidden_states = outputs.hidden_states[-1]
trajectory_embeddings = contrastive_head(hidden_states)
# Compute correctness labels (binarize verifiable rewards)
correctness_labels = (verifiable_rewards > 0).long()
# Compute combined rewards
total_rewards, contrastive_signal = contrastive_reward(
trajectory_embeddings,
correctness_labels,
verifiable_rewards,
contrastive_weight=contrastive_weight,
temperature=temperature
)
# Policy gradient: GRPO style (placeholder)
log_probs = outputs.logits.log_softmax(dim=-1)
baseline = total_rewards.mean()
advantages = total_rewards - baseline
policy_loss = -(log_probs * advantages.unsqueeze(-1)).mean()
# Backward pass
policy_loss.backward()
policy_optimizer.step()
return {
'policy_loss': policy_loss.item(),
'contrastive_signal_mean': contrastive_signal.mean().item(),
'total_reward_mean': total_rewards.mean().item()
}
When to Use:
When NOT to Use:
Hyperparameter Tuning:
Common Pitfalls: