Skip to main content Home Creators adu2021 skillxiv t3d-trajectory-distillation
t3d-trajectory-distillation Accelerate diffusion language models from many to few generation steps using trajectory self-distillation. Collect clean-noisy state pairs along teacher trajectories, apply reverse-KL mode-seeking distillation, and weight losses by decoding order to prioritize early predictions where cascading errors compound.
Jump to install Skills Marketplace Discover and explore AI skills built by the community.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Show prompt details A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/ADu2021/skillXiv --skill t3d-trajectory-distillationThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Download Zip Downloading... More from this repository meaningful-kebab-case-name 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.
action-quantization-behavior-cloning 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.
adaptive-lora-personalized-ranks 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.
Related occupations SOC
Based on SOC occupation classification
name t3d-trajectory-distillation title T3D: Few-Step Diffusion LMs via Trajectory Self-Distillation version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2602.12262 keywords ["Diffusion Language Models","Distillation","Few-Step Generation","Self-Supervision","Mode-Seeking"] description Accelerate diffusion language models from many to few generation steps using trajectory self-distillation. Collect clean-noisy state pairs along teacher trajectories, apply reverse-KL mode-seeking distillation, and weight losses by decoding order to prioritize early predictions where cascading errors compound.
T3D: Few-Step Diffusion LMs via Trajectory Self-Distillation
Problem Context
Diffusion language models require many sampling steps (25-50+) for high-quality generation. Few-step variants are needed for deployment, but naive training on random corruption schedules causes train-test mismatch: inference uses confidence-based schedules while training sees uniform random masking. T3D solves this by collecting trajectory pairs from teacher's actual inference schedule and using mode-seeking distillation.
Core Concept
T3D operates in two phases: (1) collect (clean, intermediate) state pairs along teacher trajectories during inference, (2) train student via reverse-KL objective that promotes mode-seeking (concentrating on high-probability outputs) with path-consistency weighting. This addresses the multimodal posterior problem where forward-KL spreads mass across modes.
Architecture Overview
Trajectory collection : Run full teacher steps, save clean (x₀) and intermediate (xₜ) statesReverse-KL distillation : Use discriminative KL divergence promoting mode concentrationPath-consistency weighting : Weight token losses by decoding order (early tokens > later tokens)Mode-seeking loss : Prevent mode-covering behavior in highly multimodal posteriorsFew-step inference : Student generates quality outputs in 2-4 steps
Implementation
Step 1: Collect trajectory pairs from teacher
import torch
import torch.nn.functional as F
from typing import List , Tuple , Dict
class TrajectoryCollector :
"""Collect (clean, corrupted) state pairs from teacher inference."""
def __init__ (
self,
teacher_model,
num_collection_steps: int = 50 ,
corruption_schedule: str = 'linear'
):
self .teacher_model = teacher_model
self .num_collection_steps = num_collection_steps
self
def
get_corruption_schedule
self, num_steps: int
"""
Get corruption level schedule for diffusion.
Args:
num_steps: Number of denoising steps
Returns:
sigma_t: Noise level at each step [num_steps]
"""
if
self
'linear'
return
1.0
0.0
elif
self
'cosine'
0
1
return
2
else
raise
f"Unknown schedule: {self.corruption_schedule} "
def
corrupt_tokens
self,
clean_tokens: torch.Tensor,
sigma: float ,
mask_token_id: int
"""
Add noise by masking tokens.
Args:
clean_tokens: Original tokens
sigma: Noise level (0-1, where 1=all masked)
mask_token_id: Mask token ID
Returns:
corrupted: Noisy version with sigma fraction masked
"""
int
len
len
return
def
collect_trajectories
self,
prompt: str ,
num_trajectories: int = 4
List
Dict
str
"""
Generate trajectories and collect (clean, intermediate) pairs.
Args:
prompt: Input prompt
num_trajectories: Number of trajectories to collect
Returns:
trajectories: List of {clean: Tensor, intermediates: List[Tensor]}
"""
self
self
for
in
range
self
50
for
in
enumerate
self
self
'clean'
'intermediates'
'schedule'
return
Step 2: Implement reverse-KL distillation objective class ReverseKLDistillation :
"""Reverse-KL (mode-seeking) distillation for few-step training."""
def __init__ (
self,
teacher_model,
student_model,
temperature: float = 1.0
):
self .teacher_model = teacher_model
self .student_model = student_model
self .temperature = temperature
def compute_teacher_logits (
self,
corrupted_tokens: torch.Tensor,
clean_tokens: torch.Tensor,
sigma: float
) -> torch.Tensor:
"""Get teacher's predicted distribution for clean tokens."""
with torch.no_grad():
teacher_logits = self .teacher_model.forward(
corrupted_tokens,
step=sigma,
return_logits=True
)
return teacher_logits
def compute_reverse_kl_loss (
self,
student_logits: torch.Tensor,
teacher_logits: torch.Tensor,
clean_tokens: torch.Tensor
) -> torch.Tensor:
"""
Compute reverse-KL loss: E_q[log q - log p].
Promotes mode concentration on teacher's high-probability outputs.
q = student (learner), p = teacher (reference)
"""
student_log_probs = F.log_softmax(student_logits / self .temperature, dim=-1 )
teacher_log_probs = F.log_softmax(teacher_logits / self .temperature, dim=-1 )
student_probs = F.softmax(student_logits / self .temperature, dim=-1 )
student_samples = torch.multinomial(
student_probs.view(-1 , student_probs.shape[-1 ]),
num_samples=1
).squeeze(-1 )
kl_loss = (
student_log_probs[range (len (clean_tokens)), student_samples] -
teacher_log_probs[range (len (clean_tokens)), student_samples]
).mean()
return kl_loss
def compute_mode_seeking_loss (
self,
student_logits: torch.Tensor,
teacher_logits: torch.Tensor,
clean_tokens: torch.Tensor
) -> torch.Tensor:
"""
Mode-seeking objective: penalize student spreading mass.
Uses reverse-KL which avoids mode-covering behavior.
"""
student_probs = F.softmax(student_logits, dim=-1 )
teacher_probs = F.softmax(teacher_logits, dim=-1 )
student_log_probs = F.log_softmax(student_logits, dim=-1 )
kl_per_position = (
student_probs *
(student_log_probs - F.log_softmax(teacher_logits, dim=-1 ))
).sum (dim=-1 )
return kl_per_position.mean()
Step 3: Path-consistency weighting class PathConsistencyWeighting :
"""Weight losses by token decoding order."""
@staticmethod
def compute_order_weights (
seq_len: int ,
schedule: torch.Tensor,
order: str = 'confidence'
) -> torch.Tensor:
"""
Compute per-token loss weights based on decoding order.
Early tokens (where cascading errors start) get higher weight.
Args:
seq_len: Sequence length
schedule: Corruption schedule
order: 'confidence' (teacher confidence order) or 'sequential'
Returns:
weights: [seq_len] weight for each position
"""
if order == 'sequential' :
weights = torch.linspace(1.0 , 0.3 , seq_len)
elif order == 'confidence' :
confidence_ranks = torch.arange(seq_len, dtype=torch.float32)
weights = (seq_len - confidence_ranks) / seq_len
else :
weights = torch.ones(seq_len)
return weights
@staticmethod
def apply_path_consistency_loss (
loss_per_position: torch.Tensor,
weights: torch.Tensor
) -> torch.Tensor:
"""Weight loss by position importance."""
weighted_loss = (loss_per_position * weights).mean()
return weighted_loss
Step 4: Full training step with distillation class T3DTrainer :
"""Trajectory-based few-step distillation trainer."""
def __init__ (
self,
teacher_model,
student_model,
optimizer,
num_student_steps: int = 4
):
self .teacher_model = teacher_model
self .student_model = student_model
self .optimizer = optimizer
self .num_student_steps = num_student_steps
self .collector = TrajectoryCollector(teacher_model)
self .kl_distiller = ReverseKLDistillation(teacher_model, student_model)
self .path_weighting = PathConsistencyWeighting()
def training_step (
self,
prompt: str ,
num_trajectories: int = 4
) -> Dict [str , float ]:
"""
Single training step on trajectories.
Args:
prompt: Input prompt
num_trajectories: Trajectories to collect and train on
Returns:
metrics: Loss and auxiliary metrics
"""
trajectories = self .collector.collect_trajectories(
prompt, num_trajectories=num_trajectories
)
total_loss = 0.0
num_samples = 0
for traj in trajectories:
clean_tokens = traj['clean' ]
intermediates = traj['intermediates' ]
schedule = traj['schedule' ]
path_weights = self .path_weighting.compute_order_weights(
len (clean_tokens), schedule, order='confidence'
)
num_train_steps = len (schedule) // max (1 , len (schedule) // self .num_student_steps)
for step_idx in range (0 , len (schedule), num_train_steps):
corrupted = intermediates[step_idx]
sigma = schedule[step_idx]
student_logits = self .student_model.forward(
corrupted, step=sigma.item(), return_logits=True
)
teacher_logits = self .kl_distiller.compute_teacher_logits(
corrupted, clean_tokens, sigma.item()
)
kl_loss = self .kl_distiller.compute_mode_seeking_loss(
student_logits, teacher_logits, clean_tokens
)
weighted_loss = self .path_weighting.apply_path_consistency_loss(
kl_loss.view(-1 ), path_weights
)
self .optimizer.zero_grad()
weighted_loss.backward()
torch.nn.utils.clip_grad_norm_(
self .student_model.parameters(), 1.0
)
self .optimizer.step()
total_loss += weighted_loss.item()
num_samples += 1
avg_loss = total_loss / max (1 , num_samples)
return {'loss' : avg_loss}
Step 5: Few-step inference and evaluation def generate_few_step (
student_model,
prompt: str ,
num_steps: int = 4 ,
schedule: str = 'cosine'
) -> torch.Tensor:
"""
Generate with few-step student model.
Args:
student_model: Distilled few-step model
prompt: Input prompt
num_steps: Number of generation steps (2-4)
schedule: Corruption schedule
Returns:
generated_tokens: Final prediction
"""
if schedule == 'cosine' :
sigmas = torch.linspace(1.0 , 0.0 , num_steps)
else :
sigmas = torch.linspace(1.0 , 0.0 , num_steps)
seq_len = 128
corrupted = torch.full((seq_len,), student_model.mask_token_id)
for step_idx, sigma in enumerate (sigmas):
logits = student_model.forward(corrupted, step=sigma.item(), return_logits=True )
probs = F.softmax(logits, dim=-1 )
corrupted = torch.multinomial(probs, num_samples=1 ).squeeze(-1 )
return corrupted
def evaluate_few_step_quality (
student_model,
teacher_model,
test_prompts: List [str ],
verifier,
num_steps: int = 4
) -> Dict [str , float ]:
"""Benchmark few-step vs full student performance."""
few_step_scores = []
full_step_scores = []
for prompt in test_prompts:
few_tokens = generate_few_step(student_model, prompt, num_steps=num_steps)
few_text = student_model.decode(few_tokens)
few_score = verifier(few_text)
full_tokens = teacher_model.generate(prompt, num_steps=50 )
full_text = teacher_model.decode(full_tokens)
full_score = verifier(full_text)
few_step_scores.append(few_score)
full_step_scores.append(full_score)
return {
'few_step_accuracy' : sum (few_step_scores) / len (few_step_scores),
'full_step_accuracy' : sum (full_step_scores) / len (full_step_scores),
'speedup' : 50 / num_steps
}
Practical Guidance When to use : Diffusion language models needing deployment-ready few-step variants; trading some quality for 10-15× speedup
num_student_steps : 2-4 (speedup-quality tradeoff)temperature : 1.0-1.5 (higher = softer distributions)path_consistency_order : 'confidence' (early predictions matter more)num_collection_trajectories : 4-8 per batch
Addresses train-test mismatch via trajectory collection
Mode-seeking prevents spread across multiple outputs
Path-consistency weighting prioritizes early predictions
Maintains quality under aggressive step reduction
Reverse-KL can underflow on very multimodal tasks
Path weights too extreme → ignores later predictions
Too few trajectories → limited training signal
Not validating that early predictions actually cascade errors
Scaling : Linear in number of trajectories collected.
Reference Paper: https://arxiv.org/abs/2602.12262
Related work: Diffusion models, distillation, few-step inference, mode-seeking losses
Benchmarks: Text generation, code synthesis, reasoning tasks
