Research map / Probabilistic and scientific modeling / Molecules and proteins
Efficient Test-time Inference for Generative Planning Models with OCL Search
Robert Gieselmann, Mihai Samson, Federico Pecora, Jeremy L Wyatt
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 OCLGen instead of MCTS with full rollouts | GPT-2 planner · Blocksworld | mean plan length: 54.56 -> 43.88 | Improved |
| use OCLGen instead of Best-of-N | GPT-2 planner · Labyrinth | mean plan length: 17.42 -> 12.99 | Improved |
| add depth-partitioned selection (ablation) | OCLGen · Blocksworld | mean plan length: 48.81 -> 43.88 | Improved |
| add depth-partitioned selection (ablation) | OCLGen · Sokoban | instances solved optimally: 353 / 489 (without) vs 377 / 489 (with); length 129.39 vs 128.67 | Mixed |
| add confidence-based adaptive expansion (ablation) | OCLGen · Blocksworld | mean plan length: 44.35 -> 43.88 | Improved |
| add confidence-based adaptive expansion (ablation) | OCLGen · Sokoban | mean plan length: 129.73 -> 128.67 | No clear effect |
| use lower-percentile cost-to-go instead of mode (ablation) | OCLGen · Blocksworld | mean plan length: 44.68 -> 43.88 | Improved |
| use lower-percentile cost-to-go instead of mode (ablation) | heuristic model · Blocksworld heuristic MAE | MAE: 8.68 -> 4.07 | Improved |
| scan vs uniform depth selection (ablation) | OCLGen · 4 planning domains | mean plan length: 43.88 vs 44.10 (Blocksworld) | No clear effect |
| generate self-improvement training plans with OCLGen instead of MCTS | GPT-2 planner · Blocksworld, Best-of-N after 3 rounds | mean plan length: 53.15 -> 41.40 | Improved |
| 3 rounds of self-improvement with OCLGen search | OCLGen test-time search · Blocksworld | instances solved optimally: 528 / 630 -> 630 / 630 | Improved |
| self-improvement fine-tuning on search plans, Best-of-N completion on Sokoban | GPT-2 planner · Sokoban | completion rate: 95.8 -> 93.4 | Hurt |
Closest papers by meaning
- SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search (AAAI 2026 · similarity 0.85)
- GenPlan: Generative Sequence Models as Adaptive Planners (AAAI 2025 · similarity 0.85)
- System 1.x: Learning to Balance Fast and Slow Planning with Language Models (ICLR 2025 · similarity 0.85)
- PathWise: Planning through World Model for Automated Heuristic Design via Self-Evolving LLMs (ICML 2026 · similarity 0.84)
- What If We Allocate Test-Time Compute Adaptively? (ICML 2026 · similarity 0.84)
- Beyond Test-Time Memory: State-Space Optimal Control for LLM Reasoning (ICML 2026 · similarity 0.84)
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.
