Research map / Large language models / Agents, code and math
PPDL: LLM-Based Flows as Probabilistic Programs
Louis Mandel, Guillaume Baudart, Mandana Vaziri, Martin Hirzel
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 majority voting with importance sampling over factors (ablation) | PPDL code-generation flow · MBPP | accuracy: 69.7% -> 80.5% (granite4-small) | Improved |
| use IS instead of majority voting (ablation) | PPDL flow with gpt-oss-120b · LiveCodeBench | accuracy: 19.5% -> 36.1% | Improved |
| majority voting on LiveCodeBench for gpt-oss models (ablation) | gpt-oss-120b · LiveCodeBench | accuracy: 30.3% -> 19.5% | Hurt |
| IS on LiveCodeBench for gpt-oss-20b (ablation) | gpt-oss-20b · LiveCodeBench | accuracy: 20.8% -> 8.0% | Hurt |
| add inference scaling (IS, 5 particles) with factors | granite4-small · GSM8k | accuracy: 83.8% -> 93.7% | Improved |
| add inference scaling on strong models | llama4-maverick · GSM8k | accuracy: 95.3% -> 95.4% (IS) | No clear effect |
| add inference scaling (5 particles) on Fever for llama4-maverick | llama4-maverick · Fever | accuracy: 81.6% -> 80.9% (IS) | No clear effect |
| use SMC instead of IS with same particle count (ablation) | Rocq theorem-proving agent · MiniF2F-Rocq | problems solved of 244: 87.7 -> 95.0 (40 particles) | Improved |
| increase particles from 5 to 40 in SMC (ablation) | Rocq theorem-proving agent · MiniF2F-Rocq | problems solved of 244: 84.7 -> 95.0 | Improved |
| SMC vs IS at 5 particles (ablation) | Rocq agent · MiniF2F-Rocq | problems solved of 244: 82.7 -> 84.7 (+2.0) | Improved |
| add parallel particles under constant token budget (ablation) | Rocq agent · MiniF2F-Rocq | proof/repair steps: 74 steps (IS@1) -> 9 steps (SMC@40) | Mixed |
| SMC vs IS on GSM8k granite4-micro (ablation) | granite4-micro · GSM8k | accuracy: 87.8% (IS) -> 84.2% (SMC) | Hurt |
Closest papers by meaning
- Grammars of Formal Uncertainty: When to Trust LLMs in Automated Reasoning Tasks (NeurIPS 2025 · similarity 0.89)
- APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning (NeurIPS 2025 · similarity 0.88)
- Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference (NeurIPS 2025 · similarity 0.87)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.87)
- Does Reasoning Emerge? Examining the Probabilities of Causation in Large Language Models (NeurIPS 2024 · similarity 0.86)
- The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility (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.
