Research map / Learning theory and algorithms / Prediction and decision losses
LEC: Linear Expectation Constraints for Selection-Conditioned Risk Control in Selective Prediction and Routing Systems
Zhiyuan Wang, Aniri, Tianlong Chen, Yue Zhang, Heng Tao Shen, Xiaoshuang Shi, Kaidi 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 |
|---|---|---|---|
| add LEC calibration instead of UCB-CLP | single-model selective prediction · TriviaQA, Qwen2.5-14B, alpha=0.05 | power: 0.6240 -> 0.7193 | Improved |
| add LEC calibration instead of UCB-HFD | single-model selective prediction · TriviaQA, Qwen2.5-14B, alpha=0.05 | power: no feasible threshold -> 0.7193 | Improved |
| use LEC instead of UCB-CLP | single-model selective prediction · TriviaQA, Qwen2.5-3B, alpha=0.1 | power: 0.5219 -> 0.5998 | Improved |
| use LEC instead of UCB-CLP | single-model selective prediction · TriviaQA, Qwen2.5-7B, alpha=0.05 | power: 0.3905 -> 0.4889 | Improved |
| use LEC instead of UCB-CLP | single-model selective prediction · TriviaQA, Vicuna-7B-V1.5, alpha=0.1 | power: 0.5686 -> 0.6508 | Improved |
| use LEC instead of UCB-CLP at relaxed risk level | single-model selective prediction · TriviaQA, LLaMA-3.1-70B, alpha=0.1 | power: 1.0 -> 1.0 | No clear effect |
| use LEC instead of UCB-CLP | single-model selective prediction · TriviaQA, Qwen2.5-3B | empirical accepted error rate: 0.0878 -> 0.0987 | Improved |
| add routing to LLaMA-3.1-8B | LEC with Qwen2.5-3B primary · CommonsenseQA, alpha=0.05 | accepted sample fraction (%): 20.3% -> 33.9% | Improved |
| add routing to LLaMA-3.1-8B | LEC with Qwen2.5-7B primary · CommonsenseQA, alpha=0.05 | accepted sample fraction (%): 50.7% -> 51.0% | No clear effect |
| add two-model routing | LEC with Qwen2.5-3B primary · CommonsenseQA, alpha=0.05 | accepted correct samples: 965 -> 1610 | Improved |
| add two-model routing | LEC with Qwen2.5-7B primary · CommonsenseQA, alpha=0.05 | accepted correct samples: 2392 -> 2413 | Improved |
| add joint threshold calibration (ablation) | LEC-Routing · CommonsenseQA two-model routing | system-level error control validity | Improved |
Closest papers by meaning
- CP-Router: An Uncertainty-Aware Router Between LLM and LRM (AAAI 2026 · similarity 0.85)
- Flexible Routing via Uncertainty Decomposition (NeurIPS 2026 · similarity 0.84)
- RACER: Risk-Aware Calibrated Efficient Routing for Large Language Models (ICML 2026 · similarity 0.84)
- Prompt Risk Control: A Rigorous Framework for Responsible Deployment of Large Language Models (ICLR 2024 · similarity 0.84)
- Aligning Language Models with Selective Prediction (NeurIPS 2026 · similarity 0.83)
- Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge (ICML 2026 · similarity 0.83)
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.
