Research map / Learning theory and algorithms / Prediction and decision losses
DARC: Disagreement-Aware Alignment via Risk-Constrained Decoding
mingxi Zou, Jiaxiang Chen, Junfan Li, Langzhang Liang, Qifan Wang, Xu Yinghui, Zenglin Xu
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 DARC-ε instead of mean Best-of-K selection | Llama-3.1-8B-Instruct · MT-Bench human eval (high-disagreement top 20%) | human score: 7.62 -> 8.34 | Improved |
| add epsilon tie-breaking (least-controversial among near-optimal) to entropic decoding (ablation) | DARC · MT-Bench human eval (overall) | human score: 7.84 -> 8.08 | Improved |
| apply DARC-ε on top of a cDPO policy | cDPO · MT-Bench human eval (high-disagreement top 20%) | disagreement risk: 0.89 -> 0.63 | Improved |
| apply DARC-ε on top of rDPO policy | rDPO · MT-Bench human eval (overall) | human score: 8.17 -> 8.15 | No clear effect |
| DeAL lookahead decoding vs Best-of-K | Llama-3.1-8B-Instruct · MT-Bench human eval (high-disagreement top 20%) | disagreement risk: 1.00 -> 1.13 | Hurt |
| increase candidate pool size K (ablation) | DARC-ε · MT-Bench | ΔTradeoff and ΔCVaR | Improved |
| increase risk sensitivity β (ablation) | DARC-ε · MT-Bench | ΔTradeoff and ΔCVaR | Improved |
| enlarge near-optimal set via ε (ablation) | DARC-ε · MT-Bench | robustness | Mixed |
| increase perturbation budget Naug (ablation) | DARC-ε · MT-Bench | robustness: +1.54% / +1.98% / +3.50% latency for Naug 4/8/16 | No clear effect |
| multi-scorer robust aggregation with DARC-ε | DARC-ε · Multi-scorer, high-σ subset | W/T/L vs Base: 44 / 54 / 2 | Improved |
| use DARC-ε instead of Best-of-K on low-disagreement prompts | DARC-ε · MT-Bench human bucket Q1 | human score delta: 0.125; W/T/L 11 / 80 / 9 | No clear effect |
| pairwise scalarization instead of absolute ratings (ablation) | DARC-ε · MT-Bench human eval, high-disagreement | mean ∆: +1.037 (S0) vs +0.321 (S1) | No clear effect |
Closest papers by meaning
- Pairwise Calibrated Rewards for Pluralistic Alignment (NeurIPS 2025 · similarity 0.89)
- Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time (ICML 2025 · similarity 0.88)
- Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment (NeurIPS 2026 · similarity 0.88)
- On-the-fly Preference Alignment via Principle-Guided Decoding (ICLR 2025 · similarity 0.88)
- Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization (ICLR 2025 · similarity 0.88)
- Leveraging robust optimization for llm alignment under distribution shifts (NeurIPS 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.
