Research map / Probabilistic and scientific modeling / Variational and generative sampling
Cache Coherent Resampling for Efficient Test Time Scaling in LLM Reasoning via Adaptive Sequential Monte Carlo
KE WANG, ZEHAO YU, Luwei Wang, Yongchao Huang
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 |
|---|---|---|---|
| replace sequential MCMC with ASMC at 128x budget | ASMC · MATH500 | p95 latency (s): 1317.5 -> 75.4 | Improved |
| replace sequential MCMC with ASMC at 128x budget | ASMC · MATH500 | accuracy (%): 73.8 -> 80.2 | Improved |
| use KV reorder instead of full cache rebuild for resampling (ablation) | cache-coherent ASMC · single resampling event, Qwen2.5-Math-7B | event latency: 53 to 77x faster | Improved |
| use KV reorder instead of rebuild at L=3072, R=50 (ablation) | cache-coherent ASMC · feasibility, N>=80 | feasibility: rebuild OOM at N>=80; reorder feasible through N=96 | Improved |
| use adaptive-N instead of fixed-N ASMC at 16x budget (ablation) | ASMC · MATH500 | accuracy (%): 77.6 -> 76.4 | Hurt |
| use adaptive-N instead of fixed-N ASMC at 128x budget (ablation) | ASMC · MATH500 | accuracy (%): 80.2 -> 80.6 | Improved |
| switch from multinomial to residual resampling (ablation) | ASMC · MATH500 (200 problems) | accuracy (%): +0.5% | No clear effect |
| vary ESS threshold from 0.3 to 0.7 (ablation) | ASMC-adaptive · MATH500 | accuracy (%): 79.4 -> 79.0 | No clear effect |
| more resampling events (>=16 vs 1-2) (ablation) | ASMC · MATH500 | accuracy (%): 100.0% -> 43.9% | Hurt |
| increase best-of-n from n=4 to larger n (log-prob reranking) (ablation) | Best-of-n · MATH500 | accuracy (%): 79.2 at n=4 | No clear effect |
| use ASMC (N=4) instead of greedy at 8x budget on code | ASMC · HumanEval | pass@1 (%): 55.5 -> 49.4 | Hurt |
| scale ASMC to N=64 at 128x budget on code | ASMC · HumanEval | pass@1 (%): 56.1 -> 64.0 | Improved |
Closest papers by meaning
- Mitigating Premature Exploitation in Particle-based Monte Carlo for Inference-Time Scaling (ICML 2026 · similarity 0.87)
- Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening (ICML 2026 · similarity 0.86)
- Accelerated Speculative Sampling Based on Tree Monte Carlo (ICML 2024 · similarity 0.85)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.85)
- Rollout Roulette: A Probabilistic Inference Approach to Inference-Time Scaling of LLMs using Particle-Based Monte Carlo Methods (NeurIPS 2025 · similarity 0.85)
- Tuning Sequential Monte Carlo Samplers via Greedy Incremental Divergence Minimization (ICML 2025 · similarity 0.85)
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.
