AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Difficulty–Diversity Collaborative Filtering for Data-Efficient LLM Fine-Tuning

Long P. Hoang, Wenxuan Zhang, Wei Lu

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression

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 DDCF selection instead of Random 1KQwen2.5-Math-7B · AIME24pass@32: 38.6 -> 49.0Improved
use DDCF instead of Perplexity selectionQwen2.5-Math-7B · ID avgaccuracy: 69.3 -> 70.2Improved
use DDCF instead of Random selectionQwen3-8B-Base · ID avgaccuracy: 83.6 -> 85.0Improved
use DDCF 1K subset instead of full datasetQwen3-8B-Base · OOD avgaccuracy: 80.4 -> 80.5No clear effect
use DDCF 1K subset instead of full datasetQwen2.5-Math-7B · ID avgaccuracy: 75.8 -> 70.2Hurt
increase difficulty weight lambda from 0 to 0.2 (ablation)DDCF · AMC23pass@32: 81.1 -> 83.3Improved
increase difficulty weight lambda from 0.2 to 1.0 (ablation)DDCF · SVAMPaccuracy: about 4 points declineHurt
fine-tune on 1K DDCF subsetQwen2.5-Math-1.5B · ID avgaccuracy: 54.3 -> 45.4Hurt
increase DDCF data size from 1K to 8K+ (ablation)Qwen2.5-Math-1.5B · ID avgaccuracy: 45.4 -> 56.9 (k=16,000)Improved
use DDCF on MMLU STEM instead of DiSFQwen2.5-7B · MMLU STEMaccuracy: 74.7 -> 81.1Improved
add more models to the correctness predictor (1 to 23) (ablation)correctness predictor · held-out accuracyaccuracy: 81.5 -> 82.7Improved
increase seed questions from 1K to 17.5K (ablation)correctness predictor · held-out accuracyaccuracy: 80.0 -> 82.7Improved

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