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

BBox-Adapter: Lightweight Adapting for Black-Box Large Language Models

ICML 2024 · Large language models › Instruction and fine-tuning · approach: Low-rank adaptation variants · 0 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.

Proposes BBOX-ADAPTER, a 0.1B-0.3B energy-based adapter trained with a ranking-based NCE loss and online positive/negative sampling to adapt black-box LLMs (GPT-3.5-turbo, Mixtral-8x7B) without access to parameters or output probabilities.

Problem

Adapting state-of-the-art black-box LLMs such as GPT-3.5 and GPT-4 to specific tasks is hard because their parameters, embeddings, and output token probabilities are inaccessible, and fine-tuning APIs raise transparency, privacy, and cost concerns.

Approach

Treat black-box LLM adaptation as sampling from an energy-based model in which a small adapter reweights LLM outputs; train it with a ranking-based NCE loss (target data as positives, LLM generations as negatives) in an online loop where positives can come from ground truth, AI feedback, or both, and use the adapter to score sentence-level beam search.

Key results

• BBOX-ADAPTER (Combined) with gpt-3.5-turbo improves accuracy over the base model by up to 6.77% (GSM8K 67.51 to 74.28) and by 6.39% on average across datasets.
• Full-step BBOX-ADAPTER reduces training cost 31.30x and inference cost 1.84x relative to Azure SFT on StrategyQA and GSM8K.
• Plug-and-play adapter trained on gpt-3.5-turbo improves davinci-002 by 6.85% on average and Mixtral-8x7B by 4.50% on average across three datasets.

Measured effects reported in the paper

Surprising findings

• On Mixtral-8x7B, BBOX-ADAPTER (66.08%) underperforms LoRA SFT (73.80%) on StrategyQA, though it still beats the base model (59.91%).

Limitations

• Gains are smaller than direct SFT on several tasks (e.g., Azure-SFT reaches 76.86% on StrategyQA vs 71.62% for the best adapter).
• Full-step inference requires many black-box API calls due to beam search, increasing inference cost relative to single-step.
• Privacy and misuse risks are discussed but not mitigated; the method relies on gradients/logit bias that could be exploited.

Experimental setup

Datasets: GSM8K StrategyQA TruthfulQA ScienceQA

Benchmarks: GSM8K StrategyQA TruthfulQA ScienceQA

Metrics: accuracy True+Info rate training cost ($) inference cost ($) VRAM (GiB)

Baselines: Chain-of-Thought Azure-SFT LoRA MLM loss gpt-3.5-turbo

Builds on: noise contrastive estimation supervised fine-tuning chain-of-thought LoRA proxy-tuning

Closest papers by meaning

• Neutral residues: revisiting adapters for model extension (ICML 2025 · similarity 0.87)
• BLoB: Bayesian Low-Rank Adaptation by Backpropagation for Large Language Models (NeurIPS 2024 · similarity 0.87)
• Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.87)
• Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 2025 · similarity 0.86)
• QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models (ICLR 2024 · similarity 0.86)
• Training-Free Bayesianization for Low-Rank Adapters of Large Language Models (NeurIPS 2025 · similarity 0.86)