let-late-early-training

name	let-late-early-training
title	Late-to-Early Training: LET LLMs Learn Earlier
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2602.05393
keywords	["LLM Training","Knowledge Distillation","Accelerated Training","Pretraining","Layer Alignment"]
description	Accelerate LLM pretraining by leveraging small pretrained models as teachers, enabling large models to explicitly learn late-layer knowledge in early layers via alignment loss, achieving 1.6x convergence speedup and 5% downstream improvement even with 10x smaller teachers.

LET: Late-to-Early Training for Faster LLM Pretraining

Standard LLM training develops knowledge progressively from early to late layers. LET inverts this by leveraging small pretrained models to teach large models to learn late-layer representations in early layers, dramatically accelerating convergence. The approach uses late-to-early layer alignment loss combined with late-to-early step learning (phasing out teacher influence), enabling models to skip intermediate development stages.

Core Concept

The key insight is that small pretrained models contain useful knowledge about which representations are important. By explicitly teaching large models to develop these representations early, we can bypass inefficient exploration phases. This works even with teachers 10x smaller than the target model, suggesting fundamental acceleration in learning dynamics rather than just parameter transfer.

Architecture Overview

• Late-to-Early Layer Learning: Align early layers of target with late layers of small teacher via cosine similarity loss
• Late-to-Early Step Learning: Teacher influence decays during training (λ linearly annealed from λ₀ to 0)
• Combined Alignment Loss: Standard NLL loss + alignment loss during pretraining
• Architecture-Agnostic: Works across different model sizes and architectures
• Composable: Combines with standard training optimizations

Implementation

Step 1: Prepare Teacher Model

Initialize and freeze small pretrained teacher model.

import torch
import torch.nn.functional as F

def prepare_teacher_model(teacher_config, teacher_checkpoint=None):
    """
    Load or initialize small pretrained teacher model.
    Teacher is frozen during training; only used for guidance.
    """
    from transformers import AutoModelForCausalLM, AutoConfig
    
    if teacher_checkpoint:
        teacher = AutoModelForCausalLM.from_pretrained(teacher_checkpoint)
    else:
        config = AutoConfig.from_pretrained(teacher_config)
        teacher = AutoModelForCausalLM.from_config(config)
    
    # Freeze teacher parameters
    for param  teacher.parameters():
        param.requires_grad = 
    
    teacher.()
     teacher