Research map / Large language models / Language, safety and interpretability
HYDRA: Model Factorization Framework for Black-Box LLM Personalization
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 HYDRA, a black-box LLM personalization framework that trains a retrieval reranker and a best-of-b response adapter as a shared base model with per-user heads, using gpt-3.5-turbo as a frozen generator.
Problem
Personalizing black-box LLMs whose parameters are inaccessible, where prompt-based methods (RAG, profile augmentation) fail to capture shared knowledge across users and retrieved history may be relevant but not useful.
Approach
Factorize personalization into a shared base model (Longformer-base) plus lightweight user-specific classification heads, used for both a usefulness-based history reranker and a best-of-b response adapter that scores black-box LLM candidates.
Key results
- HYDRA improves over the best prompt-based baseline by an average relative 9.01% across five LaMP tasks (e.g., LaMP-2N accuracy 0.780 vs 0.680).
- On LaMP-3, HYDRA reaches MAE 0.400 vs 0.500 for the best RAG baseline, a 20% relative MAE reduction.
- In the 1000-user scale-up, HYDRA leads all baselines on every task (e.g., LaMP-2N accuracy 0.748 vs 0.656 for RAG k=4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HYDRA instead of the strongest listed baseline | Black-box LLM personalization · LaMP-2N | accuracy: 0.680 -> 0.780 | Helped |
| Use HYDRA instead of the strongest listed baseline | Black-box LLM personalization · LaMP-2M | accuracy: 0.520 -> 0.540 | Helped |
| Use HYDRA instead of HYDRA-Reranker (k=4) | Black-box LLM personalization · LaMP-3 | MAE: 0.480 -> 0.400 | Helped |
| Use HYDRA instead of RAG (k=2) | Black-box LLM personalization · LaMP-4 | ROUGE-1: 0.173 -> 0.178 | Helped |
| Use HYDRA instead of RAG (k=2) | Black-box LLM personalization · LaMP-5 | ROUGE-1: 0.419 -> 0.434 | Helped |
| add personalized training to the adapter (ablation) | HYDRA · LaMP-2N | accuracy: 0.740 -> 0.780 | Helped |
| add personalized training to the reranker (ablation) | HYDRA · LaMP-2N | accuracy: 0.700 -> 0.780 | Helped |
| add personalized training to adapter and reranker (ablation) | HYDRA · LaMP-2N | accuracy: 0.780 -> 0.780 | No clear effect |
| add personalized training to the reranker (ablation) | HYDRA-Reranker (k=4) · LaMP-2N | accuracy: 0.700 -> 0.760 | Helped |
| add personalized training to the adapter (ablation) | HYDRA-Adapter · LaMP-2N | accuracy: 0.673 -> 0.680 | Helped |
Surprising findings
- Updating the full model performs worse than updating only user heads, attributed to the larger parameter count needing more data.
- RAG performance on LaMP-3 is inconsistent with increasing k, attributed to noisy retrieved history.
Limitations
- Experiments restricted to gpt-3.5-turbo(1106) via Azure OpenAI API; generalization to other black-box LLMs untested.
- Adapter is 110M parameters and inference requires many black-box LLM samples (best-of-b), so on-device use is infeasible.
- Potential privacy leakage from retrieved historical records.
Experimental setup
Datasets: LaMP-2N LaMP-2M LaMP-3 LaMP-4 LaMP-5
Benchmarks: LaMP-2N LaMP-2M LaMP-3 LaMP-4 LaMP-5
Metrics: accuracy F-1 MAE RMSE ROUGE-1 ROUGE-L BLEU
Baselines: gpt-3.5-turbo zero-shot ICL-Random RAG PAG
Builds on: LaMP retrieval-augmented generation BM25 Longformer
Closest papers by meaning
- Rethinking Personalization in Large Language Models at the Token Level (ICML 2026 · similarity 0.89)
- PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.89)
- PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026 · similarity 0.88)
- RLPF: Reinforcement Learning from Prediction Feedback for User Summarization with LLMs (AAAI 2025 · similarity 0.87)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.87)
- Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs (ICLR 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.