AI Paper Atlas

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

HBO: Hierarchical Balancing Optimization for Fine-Tuning Large Language Models

Weixuan Wang, Minghao Wu, Barry Haddow, Alexandra Birch

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems

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
add both global and local actors to proportional sampling (ablation)Llama-3.1-8B · Multilingual (5 tasks avg)accuracy: 46.25 -> 48.07Improved
add local actors on top of the global actor (ablation)HBO Llama-3.1-8B · Multilingual (5 tasks avg)accuracy: 47.22 -> 48.07Improved
add global actor on top of the local actors (ablation)HBO Llama-3.1-8B · Multilingual (5 tasks avg)accuracy: 47.46 -> 48.07Improved
use 4 difficulty groups instead of 1, 2, 8 or 16 (ablation)HBO Llama-3.1-8B · Multilingual (5 tasks avg)accuracy: 47.22 (1 group) -> 48.07 (4 groups); 47.45 (8), 47.87 (16)Improved
discard the easiest training examples while holding compute fixed (ablation)HBO Llama-3.1-8B · Multilingual (5 tasks avg)accuracy: 48.07 -> 47.41 (25%), 47.08 (50%), 46.55 (75%)Hurt
use PPL ratio local reward with L2-norm global reward instead of PPL or loss rewards (ablation)HBO · Multilingual (5 tasks avg)accuracy: 46.94 (PPL) / 47.22 (Loss) -> 48.07Improved
use CosSim instead of L2 norm for the global reward (ablation)HBO · Multilingual (5 tasks avg)accuracy: 48.07 -> 47.69Hurt
use PPL or loss instead of SuperFiltering to split difficulty groups (ablation)HBO · Multilingual (5 tasks avg)accuracy: 48.07 -> 47.42 (PPL), 47.23 (Loss)Hurt
scale model size from 1B to 3B to 8B (ablation)HBO vs proportional sampling · Multilingual (5 tasks avg)accuracy gain: +0.40 (1B), +1.25 (3B), +1.83 (8B)Improved
use HBO instead of DMT or EVIC data-mixing baselinesLlama-3.1-8B · Multilingual (5 tasks avg)accuracy: 45.43 (DMT), 46.10 (EVIC) -> 48.07Improved
use HBO with only 20% of training dataLlama-3.1-8B · Multilingualaccuracy: +3.50 over Prop.Improved
update global and local actors every 200 steps rather than more frequently (ablation)HBO · Multilingualperformance and runtime: about 15% additional training overheadImproved

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