Research map / Bandits and online learning / Multi-armed bandits
Cutting LLM Evaluation Costs with SySRs: A Bandit Algorithm That Provably Exploits Model Similarity
Zifan Lyu, Chahine Nejma, Tobias Wegel, Fanny Yang, Florian E. Dorner
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 synchronization (same queries for all models) to Successive Rejects (ablation) | Successive Rejects · 15 benchmarks avg, 12% budget | identification accuracy (%): 89.7 -> 95.1 | Improved |
| add synchronization to Successive Rejects (ablation) | Successive Rejects · 15 benchmarks, 100% confidence budget (max) | confidence budget (%): 43 -> 35 | Improved |
| add synchronization to uniform sampling (ablation) | Uniform Sampling · 15 benchmarks avg, 12% budget | identification accuracy (%): 36.7 -> 46.8 | Improved |
| add synchronization to UCB-E (a=1.0) (ablation) | UCB-E · 15 benchmarks avg, 12% budget | identification accuracy (%): 86.8 -> 89.4 | Improved |
| add synchronization to UCB-E (a=0.1) (ablation) | UCB-E · 15 benchmarks avg, 12% budget | identification accuracy (%): 66.9 -> 74.0 | Improved |
| add synchronization to UCB-E (a=10.0) (ablation) | UCB-E · 15 benchmarks avg, 12% budget | identification accuracy (%): 63.3 -> 73.3 | Improved |
| add synchronization to uniform sampling (ablation) | Uniform Sampling · 15 benchmarks, 90% confidence budget (mean) | confidence budget (%): 73 -> 59 | Improved |
| add synchronization to Successive Rejects (ablation) | Successive Rejects · 15 benchmarks, 90% confidence budget (mean) | confidence budget (%): 11 -> 8 | Improved |
| lower UCB-E exploration parameter a from 10.0 to 1.0 | UCB-E · 15 benchmarks, 90% confidence budget (mean) | confidence budget (%): 19 -> 10 | Improved |
| raise UCB-E exploration parameter a from 1.0 to 10.0 | UCB-E · 15 benchmarks, 100% confidence budget (mean) | confidence budget (%): 68 -> 22 | Improved |
| remove the 5% warm-up from UCB-E-LRF (ablation) | UCB-E-LRF · 15 benchmarks avg, 12% budget | identification accuracy (%): 60.8 -> 46.7 | Hurt |
Closest papers by meaning
- On Speeding Up Language Model Evaluation (ICLR 2025 · similarity 0.91)
- Active Evaluation Acquisition for Efficient LLM Benchmarking (ICML 2025 · similarity 0.88)
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.88)
- Adaptive Testing for LLM Evaluation: A Psychometric Alternative to Static Benchmarks (ICML 2026 · similarity 0.87)
- RocketEval: Efficient automated LLM evaluation via grading checklist (ICLR 2025 · similarity 0.87)
- Tuning LLM Judge Design Decisions for 1/1000 of the Cost (ICML 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.
