Research map / Large language models / Instruction and fine-tuning
Learnability-Informed Fine-Tuning of Diffusion Language Models
Shubham Parashar, Atharv Chagi, Jacob Helwig, Lakshmi Jotsna Madhavarapu, Sushil Vemuri, James Caverlee, Dileep Kalathil, Shuiwang Ji
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LIFT3 instead of vanilla SFT | LLaDA-8B-Instruct · AIME'24 | pass@16: 6.7 -> 16.7 | Improved |
| use LIFT2 instead of vanilla SFT | LLaDA-8B-Instruct · Countdown | accuracy: 20.7 -> 27.9 | Improved |
| use LIFT3 instead of vanilla SFT | LLaDA-8B-Instruct · MATH | accuracy: 34.1 -> 38.4 | Improved |
| use LIFT instead of vanilla SFT on Sudoku | LLaDA-8B-Instruct · Sudoku | accuracy: 16.8 -> 17.4 | No clear effect |
| add time-dependent top-K/bottom-K selection instead of time-independent random selection (ablation) | LIFT · AIME'24 | pass@16: 0.0 -> 16.7 | Improved |
| use LIFT instead of Top-K only or Bottom-K only (ablation) | LIFT · GSM8K/MATH | accuracy: 77.2 -> 79.4 GSM8K (Top-K vs LIFT3) | Improved |
| use Top-K only on Countdown (ablation) | Top-K · Countdown | accuracy: 26.4 -> 30.3 | Improved |
| increase H beyond 3 (ablation) | LIFT · GSM8K/MATH/Countdown/Sudoku | accuracy: MATH 38.4 -> 33.4 at H=20 | Hurt |
| use fixed or truncated rho instead of uniform U(0,1-t) (ablation) | LIFT3 · GSM8K/MATH/Countdown/Sudoku | accuracy: MATH 37.6 -> 36.4 best alternative | Hurt |
| use single-pass approximation LIFT-A (ablation) | LIFT · Countdown | accuracy: LIFT2 27.9 -> LIFT2-A 33.2 | Mixed |
| apply LIFT to Dream-7B | Dream-7B · MATH | accuracy: 30.6 -> 40.8 | Improved |
| train LIFT on LIFT-SFT-12K instead of s1K | LIFT · Sudoku | accuracy: LIFT3 10.3 vs Vanilla 9.3; LIFT2 12.5 | Mixed |
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.90)
- Large Language Models to Diffusion Finetuning (ICML 2025 · similarity 0.90)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.90)
- MRO: Enhancing Reasoning in Diffusion Language Models via Multi-Reward Optimization (NeurIPS 2025 · similarity 0.90)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.89)
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.
