Research map / Large language models / Instruction and fine-tuning
Interactive Multi-fidelity Learning for Cost-effective Adaptation of Language Model with Sparse Human Supervision
Jiaxin Zhang, Zhuohang Li, Kamalika Das, Sricharan Kumar
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 |
|---|---|---|---|
| combine 200 human annotations with 800 GPT-3.5 annotations using IMFL instead of using 600 human annotations | dolly-v2-3b · FPB, Headline, PubMedQA, and MedQA | task score: FPB: 43.16 -> 47.88; Headline: 74.13 -> 81.09; PubMedQA: 70.15 -> 73.76; MedQA: 58.31 -> 67.98 | Improved |
| use exploration-exploitation query rather than random query (ablation) | IMFL · FPB, Headline, PubMedQA, and MedQA | task score: FPB: 41.94 -> 47.88; Headline: 74.32 -> 81.09; PubMedQA: 66.03 -> 73.76; MedQA: 63.77 -> 67.98 | Improved |
| retrieve in-context examples by similarity rather than randomly (ablation) | IMFL with five mini-batches and variable batch size · Headline | Average F-1: 73.98 -> 81.09 | Improved |
| retrieve in-context examples by similarity rather than randomly (ablation) | IMFL with five mini-batches and variable batch size · PubMedQA | Accuracy: 67.44 -> 73.76 | Improved |
| use variable rather than equal batch sizes (ablation) | IMFL with similarity-based retrieval and five mini-batches · Headline | Average F-1: 80.28 -> 81.09 | Improved |
| use variable rather than equal batch sizes (ablation) | IMFL with similarity-based retrieval and five mini-batches · PubMedQA | Accuracy: 72.05 -> 73.76 | Improved |
| use five interactive mini-batches rather than one full batch (ablation) | IMFL with similarity-based retrieval · Headline | Average F-1: 75.48 -> 81.09 | Improved |
| use five interactive mini-batches rather than one full batch (ablation) | IMFL with similarity-based retrieval · PubMedQA | Accuracy: 68.90 -> 73.76 | Improved |
| increase human annotations from 200 to 400 while reducing GPT-3.5 annotations from 800 to 600 (ablation) | IMFL · FPB | F-1: 47.88 ± 0.98 -> 40.23 ± 1.99 | Hurt |
| reduce human annotations from 200 to 100 (ablation) | IMFL · FPB | F-1: 47.88 ± 0.98 -> 43.66 ± 1.42 | Hurt |
| reduce human annotations from 200 to 50 (ablation) | IMFL · FPB | F-1: 47.88 ± 0.98 -> 40.76 ± 1.48 | Hurt |
| use GPT-3.5 retrieval prompting rather than GPT-3.5 zero-shot annotation (ablation) | GPT-3.5 annotator · MedQA | annotation accuracy: 50.82 -> 59.45 | Improved |
Closest papers by meaning
- Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.88)
- DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.87)
- DataMan: Data Manager for Pre-training Large Language Models (ICLR 2025 · similarity 0.87)
- Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models (ICML 2024 · similarity 0.87)
- Cross-model Control: Improving Multiple Large Language Models in One-time Training (NeurIPS 2024 · similarity 0.87)
- ACT as Human: Multimodal Large Language Model Data Annotation with Critical Thinking (NeurIPS 2025 · similarity 0.87)
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.
