Research map / Large language models / Alignment and preferences
FSPO: Few-Shot Optimization of Synthetic Preferences Effectively Personalizes to Real Users
Anikait Singh, Sheryl Hsu, Kyle Hsu, Eric Mitchell, Stefano Ermon, Tatsunori Hashimoto, Archit Sharma, Chelsea Finn
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 |
|---|---|---|---|
| add user description rationalization (RAT) to FSPO (ablation) | FSPO · Reviews, trained users | winrate vs base (%): 80.4 -> 92.3 | Improved |
| add user description rationalization (RAT) to FSPO (ablation) | FSPO · Roleplay (1500 users) | winrate vs base (%): 82.6 -> 90.3 | Improved |
| meta-learn over users with FSPO instead of few-shot prompting alone | Llama 3.2 3B Instruct · Roleplay (1500 users) | winrate vs base (%): 63.2 -> 82.6 | Improved |
| use IPO instead of DPO as the preference objective in FSPO (ablation) | FSPO · Roleplay (1500 users) | winrate vs base (%): 81.3 -> 82.6 | No clear effect |
| increase number of few-shot examples from 1 to 8 (ablation) | FSPO · Roleplay held-out synthetic users | winrate vs base (%): 65.7 -> 82.6 | Improved |
| increase fraction of preference data per user from 10% to 100% (ablation) | FSPO · Roleplay held-out synthetic users | winrate vs base (%): 70.1 -> 82.6 | Improved |
| increase preference data from 10% to 25% (ablation) | FSPO · Roleplay held-out synthetic users | winrate vs base (%): 70.1 -> 69.5 | No clear effect |
| use persona steering instead of low-temperature sampling for response diversity (ablation) | synthetic data generation · Roleplay response diversity (ALOE) | mean semantic similarity: 0.96 -> 0.81 | Improved |
| use view steering plus ensemble of larger models instead of a single model with view steering (ablation) | synthetic data generation · Roleplay response diversity (ALOE) | mean semantic similarity: 0.78 -> 0.71 | Improved |
| iterative persona refinement of underspecified user descriptions (ablation) | synthetic preference labeling · Roleplay preference labels | binary Shannon entropy (nats): 0.64 -> 0.13 | Improved |
| FSPO versus base model with real human participants | FSPO · Roleplay human study | human winrate (%): 50 -> 68.2 | Improved |
| increase shots from 4 to 8 for few-shot prompting (ablation) | Few-shot prompting baseline · Reviews, interpolated users | winrate vs base (%): 61.9 -> 59.1 | Hurt |
Closest papers by meaning
- PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.92)
- Test-Time Personalization: A Diagnostic Framework and Probabilistic Fix for Scaling Failures (NeurIPS 2026 · similarity 0.88)
- Q-Adapter: Customizing Pre-trained LLMs to New Preferences with Forgetting Mitigation (ICLR 2025 · similarity 0.88)
- HYDRA: Model Factorization Framework for Black-Box LLM Personalization (NeurIPS 2024 · similarity 0.88)
- Amulet: ReAlignment During Test Time for Personalized Preference Adaptation of LLMs (ICLR 2025 · similarity 0.88)
- Limited Preference Data? Learning Better Reward Model with Latent Space Synthesis (NeurIPS 2025 · similarity 0.88)
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.
