Confidence-Aware Self-Consistency with Adaptive Early Stopping
- 주제(키워드) LLM efficiency , test-time scaling
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Hyunsouk Cho
- 발행년도 2026
- 학위수여년월 2026. 8
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000036581
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Self-consistency improves the reasoning accuracy of large language models （LLMs） by majority voting over multiple sampled solutions, but it always exhausts the sampling budget and trusts every solution equally. Early-stopping methods cut token cost substantially via convergence monitoring; however, deciding when to stop cannot change what is voted on, so they save to- kens without meaningfully improving accuracy. This thesis introduces CASE （Confidence-Aware Self-consistency with Adaptive Early stopping）, which improves both accuracy and efficiency by leveraging token-level confidence normalized at two scales: within-problem to filter low-quality solutions and weight survivors, and cross-problem to adapt the stopping threshold to problem difficulty. Across five models and six benchmarks, CASE simultaneously saves 85% of tokens and improves accuracy by +2.24 percentage points over majority voting, outperforming every early-stopping baseline on both accuracy and efficiency. Component and role-swap analyses confirm that filtering and weighting are complementary, and that the two normalization scales are non-interchangeable, each encoding information the other cannot recover. CASE requires no additional training and no modification to the standard serving pipeline, demonstrating how a single token-level confidence signal can serve multiple roles in self-consistency.
more목차
1 Introduction 1
1.1 Contributions 3
2 Related Work 4
2.1 Early Stopping for Self-Consistency 4
2.2 Solution Quality Estimation 4
3 CASE: Confidence-Aware Self-Consistency 5
3.1 A Modular Decomposition of Self-Consistency 5
3.2 Solution Quality Estimation 6
3.3 Candidate Filtering (F) 7
3.4 Weighting (W) 7
3.5 Difficulty-Aware Early Stopping (S) 8
4 Experiments 10
4.1 Experimental Setup 10
4.2 Experimental Results 11
4.3 Component Analysis 13
4.4 Intra-z vs. Inter-z: Role Validation 14
4.5 Budget Sensitivity 15
5 Conclusion 16
5.1 Limitations and Future Work 16
5.2 Concluding Remarks 17
References 18
Appendix 22
A Method Taxonomy 22
B Confidence Signal and Sampling Details 23
B.1 Confidence Signal Design 23
B.2 Sampling Parameters 24
B.3 Confidence Discrimination 24
C Robustness Analyses 26
C.1 Inter-Problem Calibration 26
C.2 Shuffle Variance 27
C.3 Hyperparameter Sensitivity 28
D DeepConf Comparison 31
