Research map / Image generation / Fast diffusion sampling
Large Language Models to Diffusion Finetuning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes L2D, a LoRA-based finetuning that turns pre-trained autoregressive LMs into diffusion-style models over token embeddings with cross-entropy loss, enabling test-time scaling via more diffusion steps without changing the original weights.
Problem
Autoregressive LMs lack the ability to scale test-time compute adaptively per token; prior diffusion LMs trained from scratch lag far behind autoregressive models.
Approach
Casts a pre-trained LM as a single-step diffusion model by adding a parallel diffusion path (initialized from the frozen LM's MLP and query weights, trained with LoRA) that operates on noisy token embeddings, trained with cross-entropy loss; inference uses ODE solvers over multiple steps while the original single-step output is preserved at t=0 via zero-initialized weighting, plus classifier-free
Key results
- On Llama 3.2 1B Instruct, L2D raises the 6-task average from 28.54 to 35.50 with 73M trainable parameters, versus 29.97 for LoRA finetuning.
- On Qwen 2.5 7B Instruct, L2D reaches an average of 67.58 versus 63.34 for LoRA finetuning (MATH 43.62 vs 33.82).
- Adaptive RK ODE solver on Llama 3.2 1B averages 118.33 steps and reaches 36.34 overall, exceeding the fixed 15-step midpoint result.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add L2D diffusion finetuning | Llama 3.2 1B Instruct · Overall average score | average score: 28.54 -> 35.50 | Helped |
| Add L2D diffusion finetuning | Qwen 2.5 1.5B Instruct · Overall average score | average score: 40.05 -> 51.01 | Helped |
| Add L2D diffusion finetuning | Llama 3.1 8B Instruct · Overall average score | average score: 54.09 -> 61.33 | Helped |
| Add L2D diffusion finetuning | Qwen 2.5 7B Instruct · Overall average score | average score: 46.65 -> 67.58 | Helped |
| Increase inference budget from the default to 127 steps (ablation) | L2D on Llama 3.2 1B Instruct · Overall average score | average score: 35.50 -> 36.24 | Helped |
| Use an adaptive ODE solver (ablation) | L2D on Llama 3.2 1B Instruct · Math | score: 28.02 -> 30.26 | Helped |
| Optimize all diffusion-path parameters instead of using LoRA (ablation) | L2D on Llama 3.2 1B Instruct · Coding | score: 49.80 -> 50.52 | Helped |
| Initialize L2D from a LoRA-finetuned checkpoint (ablation) | L2D on Llama 3.2 1B Instruct · Overall average score | average score: 35.50 -> 35.51 | No clear effect |
| Initialize L2D from a fully finetuned checkpoint (ablation) | L2D on Llama 3.2 1B Instruct · Overall average score | average score: 35.50 -> 35.84 | Helped |
| Combine L2D with tuned token search (ablation) | L2D on Llama 3.2 1B Instruct · Overall average score | average score: 33.83 -> 38.57 | Helped |
| Use classifier-free guidance with tuned strength (ablation) | L2D on Llama 3.2 1B Instruct · Overall average score | average score: 35.50 -> 36.26 | Helped |
| Use learning rate 1×10⁻⁵ instead of the default (ablation) | L2D on Llama 3.2 1B Instruct · Overall average score | average score: 35.50 -> 33.79 | Hurt |
Surprising findings
- Full weight finetuning of Llama instruct models causes performance drops on coding tasks (e.g., HumanEval 45.26 to 32.08 for Llama 3.2 1B), which L2D does not exhibit.
- Simpler Euler integration with 15 steps slightly outperforms the midpoint solver.
Limitations
- Continuous diffusion LM loses direct access to ground-truth confidence scores and simple logit manipulation.
- Pass@k sampling is constrained; sampling is untempered and token-level logit manipulation is not directly available.
- Full diffusion path finetuning increases parameter cost substantially (992M).
Experimental setup
Datasets: SmolTalk
Benchmarks: GSM8K MATH HumanEval MBPP MMLU MMLU-Pro
Metrics: accuracy pass@10 average score
Baselines: LoRA finetuning full finetuning tuned token search CDCD-style velocity estimate Llama 3.2 1B Instruct (base)
Builds on: DDPM rectified flow CDCD classifier-free guidance LoRA
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.91)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.90)
- Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.90)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.