Research map / Large language models / Agents, code and math
POLCA: Stochastic Generative Optimization with LLM
Xuanfei Ren, Allen Nie, Tengyang Xie, Ching-An Cheng
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 epsilon-Net semantic filter to vanilla POLCA (ablation) | POLCA · tau-bench | test score | Improved |
| add LLM Summarizer global context to vanilla POLCA (ablation) | POLCA · tau-bench | test score | Improved |
| use epsilon = 0 (no discretization) instead of positive epsilon (ablation) | POLCA · tau-bench and VeriBench | score | Hurt |
| use large epsilon (ablation) | POLCA · tau-bench and VeriBench | score | Mixed |
| use ensemble logistic-regression reward predictions over embeddings instead of empirical mean selection (ablation) | POLCA · tau-bench | test score | Hurt |
| add UCB bonus to mean priority (ablation) | POLCA · benchmarks | score | No clear effect |
| continuously update empirical-mean statistics in a priority queue versus evaluating each program once | POLCA vs GEPA/OpenEvolve · tau-bench (all 115 tasks) | pass@1: 0.429 -> 0.439 | Improved |
| POLCA vs DSPy on deterministic compilation | POLCA · VeriBench compilation | pass rate: 0.888 -> 0.952 | Improved |
| POLCA vs GEPA on held-out tau-bench tasks | POLCA · tau-bench (last 105 tasks) | pass@1: 0.417 -> 0.425 | Improved |
Closest papers by meaning
- Understanding the Challenges in Iterative Generative Optimization with LLMs (NeurIPS 2026 · similarity 0.88)
- Revolve: Optimizing AI Systems by Tracking Response Evolution in Textual Optimization (ICML 2025 · similarity 0.87)
- MGUP: A Momentum-Gradient Alignment Update Policy for Stochastic Optimization (NeurIPS 2025 · similarity 0.86)
- Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs (NeurIPS 2024 · similarity 0.86)
- Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening (ICML 2026 · similarity 0.86)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · 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.
