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

Can Models Learn Skill Composition from Examples?

NeurIPS 2024 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 1 citations (OpenAlex)

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.

Fine-tunes LLaMA-2-13B and Mistral-7B-Instruct-v0.2 on GPT-4-generated skill-composition texts (k≤3) to show that compositional skill-writing ability transfers to unseen k (4-5) and held-out skills.

Problem

Whether LLMs' ability to compose k language skills, which larger pretrained models show only weakly, can be induced by fine-tuning on suitable examples, and whether this generalizes to unseen k and unseen skills.

Approach

Extend the SKILL-MIX evaluation into a training pipeline: generate skill-combination texts with GPT-4 (filtered for full marks) using only training skills and k≤3, fine-tune small models, and test on held-out skills/topics and larger k, with skill-rich data shown to be more data-efficient.

Key results

• Fine-tuning LLaMA-2-13B on k≤3 data raised Ratio of Full Marks on held-out skills with k=3 from 4% to 37%, and Mistral-7B from 8% to 49%.
• Fine-tuned LLaMA-2-13B on k=1,2,3 data improved SKILL-MIX k=4 Ratio of Full Marks on training skills from 0% to 8%, despite no k=4 training data.
• Using a sub-sampled 8000-example k≤3 set with more skill-rich data outperformed k≤2 data at k=4 (about 11% vs 1% on all skills), indicating skill-richer data is more data-efficient.

Measured effects reported in the paper

Surprising findings

• Models fine-tuned only on k≤3 combinations improved on k=4 and k=5 compositions never seen in training.
• Performance on held-out skill categories was comparable to or higher than on training skill categories.

Limitations

• High computational and financial costs limited hyperparameter sweeps and repeated runs with different seeds.
• Study restricted to the SKILL-MIX language-skill setting; generalization to other compositional settings is untested.
• Held-out skills are from different categories but still near relatives of training skills, so out-of-distribution generalization is limited.

Experimental setup

Datasets: SKILL-MIX skills and topics (101 skills, 100 topics) SKILL-MIX k=1,2,3 GPT-4 generations filtered to full marks RedPajama-style pretraining mixture (common crawl and code) as regularizer

Benchmarks: SKILL-MIX (k=1 to 5) on training skills SKILL-MIX on held-out skills SKILL-MIX on all skills

Metrics: Ratio of Full Marks Ratio of All Skills Skills Fraction

Baselines: LLaMA-2-13B-Chat Mistral-7B-Instruct-v0.2 GPT-4 GPT-3.5-turbo

Builds on: SKILL-MIX

Closest papers by meaning

• SKILL-MIX: a Flexible and Expandable Family of Evaluations for AI Models (ICLR 2024 · similarity 0.90)
• From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones (ICLR 2026 · similarity 0.88)
• Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.87)
• How does RL Post-training Induce Skill Composition? A Case Study on Countdown (NeurIPS 2026 · similarity 0.87)
• Limits of Transformer Language Models on Learning to Compose Algorithms (NeurIPS 2024 · similarity 0.87)
• Training on the Test Task Confounds Evaluation and Emergence (ICLR 2025 · similarity 0.87)