AI Paper Atlas

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

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use LIFT3 instead of vanilla SFTLLaDA-8B-Instruct · AIME'24pass@16: 6.7 -> 16.7Improved
use LIFT2 instead of vanilla SFTLLaDA-8B-Instruct · Countdownaccuracy: 20.7 -> 27.9Improved
use LIFT3 instead of vanilla SFTLLaDA-8B-Instruct · MATHaccuracy: 34.1 -> 38.4Improved
use LIFT instead of vanilla SFT on SudokuLLaDA-8B-Instruct · Sudokuaccuracy: 16.8 -> 17.4No clear effect
add time-dependent top-K/bottom-K selection instead of time-independent random selection (ablation)LIFT · AIME'24pass@16: 0.0 -> 16.7Improved
use LIFT instead of Top-K only or Bottom-K only (ablation)LIFT · GSM8K/MATHaccuracy: 77.2 -> 79.4 GSM8K (Top-K vs LIFT3)Improved
use Top-K only on Countdown (ablation)Top-K · Countdownaccuracy: 26.4 -> 30.3Improved
increase H beyond 3 (ablation)LIFT · GSM8K/MATH/Countdown/Sudokuaccuracy: MATH 38.4 -> 33.4 at H=20Hurt
use fixed or truncated rho instead of uniform U(0,1-t) (ablation)LIFT3 · GSM8K/MATH/Countdown/Sudokuaccuracy: MATH 37.6 -> 36.4 best alternativeHurt
use single-pass approximation LIFT-A (ablation)LIFT · Countdownaccuracy: LIFT2 27.9 -> LIFT2-A 33.2Mixed
apply LIFT to Dream-7BDream-7B · MATHaccuracy: 30.6 -> 40.8Improved
train LIFT on LIFT-SFT-12K instead of s1KLIFT · Sudokuaccuracy: LIFT3 10.3 vs Vanilla 9.3; LIFT2 12.5Mixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy