Research map / Large language models / Alignment and preferences
Robust Multi-Objective Controlled Decoding of Large Language Models
Seongho Son, William Bankes, Sangwoong Yoon, Shyam Sundhar Ramesh, Xiaohang Tang, Ilija Bogunovic
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 |
|---|---|---|---|
| use RMOD (maximin worst-case weights) instead of uniform-weight controlled decoding | gemma-2-2b-it · HH (gpt-4o judge) | worst-case win rate (%): 57.6% -> 59.1% | Improved |
| distill RMOD outputs into the policy via SFT | gemma-2-2b-it · HH (gpt-4o judge) | worst-case win rate (%): 57.9% vs CD-Uniform 57.6% | Improved |
| increase lambda (ablation) | RMOD · HH | weight sparsity / worst-case focus | Improved |
| increase block size B above 16 (ablation) | RMOD · HH | worst-case win rate | Hurt |
| increase number of objectives up to 10 (ablation) | RMOD · ValuePrism | worst-case reward | Hurt |
| increase K to 16 in best-of-K on harmlessness reward directly (ablation) | BoK-Harmless · HH (gpt-4o judge) | worst-case win rate (%): 52.3% -> 51.1% (K=2 to K=16) | Hurt |
| use worst-case weights computation | RMOD · HH | latency: less than 4.5% higher latency | No clear effect |
Closest papers by meaning
- Decoding-Time Language Model Alignment with Multiple Objectives (NeurIPS 2024 · similarity 0.90)
- Collab: Controlled Decoding using Mixture of Agents for LLM Alignment (ICLR 2025 · similarity 0.88)
- Fast Best-of-N Decoding via Speculative Rejection (NeurIPS 2024 · similarity 0.87)
- Approximately Aligned Decoding (NeurIPS 2025 · similarity 0.86)
- MATO: Multi-objective Personalized Alignment with Test-time Optimization for Large Language Models (NeurIPS 2026 · similarity 0.86)
- Optimized Multi-Token Joint Decoding With Auxiliary Model for LLM Inference (ICLR 2025 · similarity 0.86)
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.
