AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Advanced Black-Box Tuning of Large Language Models with Limited API Calls

Zhikang Xie, Weilin Wan, Peizhu Gong, Weizhong Zhang, Cheng Jin

AAAI 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 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
train proxy with GP-filter surrogate instead of offline Proxy-TuneQwen3-8B proxy / Qwen3-14B · 11-dataset averageaccuracy (%): 84.03 -> 92.10Improved
GP-filter instead of CPT (about 1.58% of API calls)Qwen3-8B proxy / Qwen3-14B · 11-dataset averageaccuracy (%): 91.76 -> 92.10Improved
GP-random instead of CPT (6.94% of API calls) (ablation)Qwen3-8B proxy / Qwen3-14B · 11-dataset averageaccuracy (%): 91.76 -> 91.28No clear effect
diversity-filtered LogitMap Pairs instead of random sampling for GP training (ablation)GP surrogate · 11-dataset averageaccuracy (%): 91.28 -> 92.10Improved
GP-filter vs pretrained modelLlama2-13B · 11-dataset averageaccuracy (%): 55.92 -> 86.85Improved
GP-filter vs LoRA-TuneLlama2-13B · 11-dataset averageaccuracy (%): 85.79 -> 86.85Improved
GP-filter vs CPTLlama2-13B · 11-dataset averageaccuracy (%): 86.41 -> 86.85Improved
GP-filter vs full fine-tuneLlama2-13B · 11-dataset averageaccuracy (%): 88.58 -> 86.85Hurt
train GP with only 2 API calls (CoLA) (ablation)GP-filter · CoLAaccuracy (%)No clear effect
GP-filter vs pretrainedMistral-7B, Qwen3, DeepSeek-R1-Distill · 11-dataset averageaccuracy (%): +9.94, +7.24, +13.27 pointsImproved

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