메뉴
Enable step-by-step reasoning in diffusion models through iterative latent state refinement. Condition diffusion on evolving thought states across multiple reasoning steps, grounded with textual supervision to prevent drift.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
SOC 직업 분류 기준
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill endocot-internal-chain-of-thought명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SKILL.md 표시 중
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.
| name | endocot-internal-chain-of-thought |
| title | EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.12252 |
| keywords | ["Diffusion","Reasoning","Chain-of-Thought","Latent Space","Generation"] |
| description | Enable step-by-step reasoning in diffusion models through iterative latent state refinement. Condition diffusion on evolving thought states across multiple reasoning steps, grounded with textual supervision to prevent drift. |
Diffusion models generate images by denoising—but they commit to solutions early in this process. EndoCoT enables intermediate reasoning by allowing the conditioning signal to evolve across steps. Rather than static guidance, the framework iteratively refines hidden states representing thoughts, each conditioned on the previous step's reasoning, then uses these refined states to guide generation.
This "endogenous" approach contrasts with external chain-of-thought: reasoning happens within the diffusion process via latent dynamics, enabling truly integrated visual-linguistic reasoning.
EndoCoT combines three mechanisms:
This enables dynamic, evolving conditioning that guides diffusion through a reasoning trajectory.
Refine latent reasoning states across multiple steps before using for generation.
import torch
import torch.nn as nn
class IterativeThoughtRefiner(nn.Module):
def __init__(self, hidden_dim=768, num_reasoning_steps=4):
super().__init__()
self.hidden_dim = hidden_dim
.num_reasoning_steps = num_reasoning_steps
.refiner_layers = nn.ModuleList([
nn.Sequential(
nn.Linear(hidden_dim, hidden_dim * ),
nn.ReLU(),
nn.Linear(hidden_dim * , hidden_dim)
)
_ (num_reasoning_steps)
])
.attention_layers = nn.ModuleList([
nn.MultiheadAttention(hidden_dim, num_heads=, batch_first=)
_ (num_reasoning_steps)
])
():
thought_trajectory = [initial_thought]
current_thought = initial_thought
step (.num_reasoning_steps):
refined = .refiner_layers[step](current_thought)
attended, _ = .attention_layers[step](
refined.unsqueeze(),
reasoning_context,
reasoning_context
)
current_thought = refined + attended.squeeze()
thought_trajectory.append(current_thought)
torch.stack(thought_trajectory, dim=)
Anchor final reasoning state to explicit textual reference to prevent drift.
class TerminalThoughtGrounder(nn.Module):
def __init__(self, hidden_dim=768, text_encoder=None):
super().__init__()
self.hidden_dim = hidden_dim
self.text_encoder = text_encoder
# Projection for alignment
self.thought_projector = nn.Linear(hidden_dim, hidden_dim)
self.text_projector = nn.Linear(hidden_dim, hidden_dim)
def ground_with_text(self, final_thought, reference_text):
"""
Align final thought with textual reference using semantic loss.
final_thought: (batch, hidden_dim)
reference_text: str or tokenized reference
"""
# Encode reference text
if isinstance(reference_text, str):
reference_embedding = self.text_encoder.encode(reference_text)
else:
reference_embedding = self.text_encoder(reference_text)
# Project both to common space
thought_proj = self.thought_projector(final_thought)
text_proj = self.text_projector(reference_embedding)
# Semantic L2 alignment loss
alignment_loss = torch.nn.functional.mse_loss(thought_proj, text_proj)
return alignment_loss
def forward(self, thought_trajectory, reference_text):
"""
Compute grounding loss for final thought.
thought_trajectory: (batch, steps, hidden_dim)
reference_text: reference for grounding
"""
final_thought = thought_trajectory[:, -1, :]
grounding_loss = self.ground_with_text(final_thought, reference_text)
return grounding_loss
Use thought trajectory to condition diffusion generation at each step.
class DiffusionTransformerWithEvolvedThoughts(nn.Module):
def __init__(self, dit_model, hidden_dim=768):
super().__init__()
self.dit = dit_model
self.hidden_dim = hidden_dim
# Project thoughts to condition dimension
self.thought_to_condition = nn.Linear(hidden_dim, dit_model.cond_dim)
def forward(
self,
noise,
timestep,
thought_trajectory,
instruction_tokens=None
):
"""
Run one diffusion step with evolved thought conditioning.
noise: (batch, channels, height, width)
timestep: int or (batch,)
thought_trajectory: (batch, steps, hidden_dim)
instruction_tokens: optional additional conditioning
"""
batch_size = noise.shape[0]
# Select appropriate thought based on diffusion progress
# (early steps use earlier thoughts, late steps use refined thoughts)
progress = timestep / 1000 # Normalized to [0, 1]
step_idx = min(
int(progress * thought_trajectory.shape[1]),
thought_trajectory.shape[1] - 1
)
current_thought = thought_trajectory[:, step_idx, :] # (batch, hidden_dim)
# Project to condition
conditioning = self.thought_to_condition(current_thought) # (batch, cond_dim)
# Optional: concatenate with instruction tokens
if instruction_tokens is not None:
conditioning = torch.cat([conditioning, instruction_tokens], dim=-1)
# Denoising step
model_output = self.dit(
noise,
timestep,
c=conditioning
)
return model_output
First stage: supervise all intermediate reasoning steps. Second stage: optimize only final output.
def train_endocot(
model,
data_loader,
text_encoder,
num_epochs=10,
stage1_epochs=5
):
"""
Two-stage training: reasoning -> output quality.
"""
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
for epoch in range(num_epochs):
is_stage1 = epoch < stage1_epochs
for batch_idx, (images, prompts, texts) in enumerate(data_loader):
# Forward pass
image_noise = torch.randn_like(images)
# Get initial thought
prompt_embedding = text_encoder.encode(prompts)
# Refine thoughts
thought_refiner = model.thought_refiner
thought_trajectory = thought_refiner(
prompt_embedding,
text_encoder.encode(texts).unsqueeze(1)
)
if is_stage1:
# Stage 1: Supervise all intermediate thoughts
total_loss = 0
for step_idx in range(thought_trajectory.shape[1]):
thought = thought_trajectory[:, step_idx, :]
# Generate with this thought
outputs = model.dit_with_thoughts(
image_noise,
torch.tensor([500]), # Midpoint diffusion step
thought.unsqueeze(1)
)
# Supervise against ground truth
step_loss = torch.nn.functional.mse_loss(outputs, images)
total_loss += step_loss
# Also supervise terminal grounding
grounding_loss = model.grounder.ground_with_text(
thought_trajectory[:, -1, :],
texts
)
total_loss += 0.1 * grounding_loss
else:
# Stage 2: Optimize final output quality only
# Freeze intermediate supervision
thought_trajectory = thought_trajectory.detach()
final_thought = thought_trajectory[:, -1, :]
# Full generation with final thought
outputs = model.dit_with_thoughts(
image_noise,
torch.tensor([0]), # Final step
final_thought.unsqueeze(1)
)
total_loss = torch.nn.functional.mse_loss(outputs, images)
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
if batch_idx % 100 == 0:
stage = "Stage1" if is_stage1 else "Stage2"
print(f"Epoch {epoch} [{stage}]: Loss = {total_loss.item():.4f}")
return model
Demonstrate full workflow with reasoning prompts.
def generate_with_endogenous_reasoning(
model,
text_encoder,
initial_prompt,
reasoning_steps=4,
num_diffusion_steps=50
):
"""
Full generation pipeline with endogenous reasoning.
"""
# Step 1: Generate reasoning trajectory
prompt_embedding = text_encoder.encode(initial_prompt)
thought_refiner = model.thought_refiner
thought_trajectory = thought_refiner(
prompt_embedding,
torch.zeros(1, 1, 768) # Placeholder context
)
# Step 2: Prepare diffusion
noise = torch.randn(1, 4, 64, 64) # Latent space noise
dit_model = model.dit_with_thoughts
# Step 3: Iterative diffusion with evolving thoughts
for diffusion_step in range(num_diffusion_steps):
timestep = (num_diffusion_steps - diffusion_step) / num_diffusion_steps
# Update thought based on diffusion progress
if diffusion_step % (num_diffusion_steps // reasoning_steps) == 0:
# Optionally re-refine thought mid-generation
pass
# Denoising step
noise = dit_model(
noise,
int(timestep * 1000),
thought_trajectory
)
# Step 4: Decode from latent space
image = model.vae.decode(noise)
return image
When to Use:
When NOT to Use:
Hyperparameter Tuning:
Common Pitfalls: