Research map / Learning theory and algorithms / Prediction and decision losses
PETS: A Principled Framework Towards Optimal Trajectory Allocation for Efficient Test-Time Self-Consistency
Zhangyi Liu, Huaizhi Qu, Xiaowei Yin, He Sun, Yanjun Han, Tianlong Chen, Zhun Deng
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 PETS-Offline allocation instead of uniform | self-consistency sampling · GPQA-Diamond, Qwen3-4B | traces to full self-consistency: 11013 -> 2780 | Improved |
| use PETS-Offline allocation instead of uniform | self-consistency sampling · AIME25, Qwen3-4B | traces to full self-consistency: 470 -> 212 | Improved |
| use PETS-Offline allocation instead of confidence-guided | DeepConf allocation · GPQA-Diamond, Qwen3-4B | traces to full self-consistency: 10652 -> 2780 | Improved |
| use PETS-Online allocation instead of uniform | self-consistency sampling · GPQA-Diamond, Qwen3-4B | traces to full self-consistency: 9520 -> 4662 | Improved |
| use PETS-Online instead of uniform | self-consistency sampling · AIME25, gpt-oss-120b | traces to full self-consistency: 416 -> 454 | Hurt |
| use PETS-Online instead of uniform | self-consistency sampling · AIME25, gpt-oss-20b | traces to full self-consistency: 221 -> 195 | Improved |
| allocate by PETS-Offline instead of uniform at matched traces | self-consistency sampling · GPQA-Diamond, Qwen3-4B | accuracy: 68.9 -> 69.7 | Improved |
| add confidence-weighted voting to PETS-Offline | PETS-Offline · GPQA-Diamond, Qwen3-4B | traces to full self-consistency: 2780 -> 3667 | Hurt |
| add confidence-weighted voting to PETS-Offline | PETS-Offline · GPQA-Diamond, Qwen3-4B | accuracy: 69.7 -> 70.3 | Improved |
| vary number of difficulty bins k from 2 to 5 (ablation) | PETS-Online predictor · GPQA-Diamond, Qwen3-4B | predictor accuracy: 0.6845 -> 0.6786 | No clear effect |
| vary KMeans cluster number k from 2 to 8 (ablation) | PETS-Oracle · GPQA-Diamond, gpt-oss-120b | oracle accuracy: 0.8036 -> 0.8274 | No clear effect |
| increase predictor bins from 2 to 3 (ablation) | PETS-Online predictor · AIME 2024, Qwen3-4B | predictor accuracy: 0.9500 -> 0.9000 | Hurt |
Closest papers by meaning
- What If We Allocate Test-Time Compute Adaptively? (ICML 2026 · similarity 0.89)
- Every Rollout Counts: Optimal Resource Allocation for Efficient Test-Time Scaling (NeurIPS 2025 · similarity 0.87)
- Optimal Self-Consistency for Efficient Reasoning with Large Language Models (ICML 2026 · similarity 0.87)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.86)
- Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification (ICML 2025 · similarity 0.86)
- Evaluating Test-Time Scaling of General LLM Agents (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.
