Semantic Partial Grounding via LLMs

2025-08-07 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

SPG-LLM is a novel pre-grounding approach that uses Large Language Models (LLMs) to semantically prune Planning Domain Definition Language (PDDL) domain and problem files. This method heuristically identifies and removes irrelevant objects, actions, and predicates before the grounding phase, aiming to reduce computational bottlenecks in classical planning. Evaluated across seven hard-to-ground benchmarks, including Agricola, Blocksworld, and Zenotravel, SPG-LLM consistently achieved faster grounding times, often by orders of magnitude, and reduced the number of grounded operators compared to Full Grounding (FG) and Planning with Learned Object Importance (PLOI). While SPG-LLM, using GPT-5-2025-08-07, produced valid plans for 139 out of 175 tasks (compared to FG's 161/175 and PLOI's 90/100), it sometimes yielded comparable or better plan costs and solving times, particularly in domains like Blocksworld and Zenotravel.

Key takeaway

For AI Scientists and Research Scientists working with classical planning systems, SPG-LLM offers a promising method to mitigate the grounding bottleneck. You should consider integrating LLM-based pre-grounding to significantly accelerate task processing, especially for large PDDL problems. Be aware of potential trade-offs in plan coverage and validate LLM-generated task reductions rigorously to ensure plan soundness for the original task. This approach can enable faster exploration of planning solutions.

Key insights

LLMs can semantically prune PDDL planning tasks to accelerate grounding and improve plan quality.

Principles

• Leverage LLMs for heuristic pruning of planning components.
• Validate pruned tasks syntactically, computationally, and semantically.
• Task-level pruning complements operator-level partial grounding.

Method

SPG-LLM formats PDDL domain/problem files into a prompt for an LLM (e.g., GPT-5-2025-08-07) to identify and prune irrelevant objects, predicates, and actions. The pruned output undergoes multi-level validation, with retries if errors occur.

In practice

• Use LLMs to pre-process PDDL files for large tasks.
• Implement validation checks for LLM-generated task reductions.
• Consider object-level pruning for computational efficiency.

Topics

• Large Language Models
• Classical Planning
• PDDL
• Partial Grounding
• Task Pruning
• Planning Benchmarks

Best for: AI Scientist, Research Scientist

Related on AIssential

• Semantic Partial Grounding via LLMs — LLMs can significantly accelerate classical planning by semantically pre-filtering irrelevant PDDL elements before grounding.
• Hierarchical Task Network Planning with LLM-Generated Heuristics — LLMs can generate effective search heuristics for HTN planning, improving efficiency.
• Google Goes to EXTREMES: In-Context Symbolic AI — LLMs can perform symbolic reasoning with high accuracy when rigorously constrained by in-context PDDL definitions.
• What Makes an LLM a Good Optimizer? A Trajectory Analysis of LLM-Guided Evolutionary Search — Strong LLM optimizers act as local refiners, making incremental improvements within localized semantic search spaces.
• GroundedPlanBench: Spatially grounded long-horizon task planning for robot manipulation — Jointly planning actions and spatial grounding improves robot task success and action accuracy for VLMs.
• Boeing demonstrates large language model for space-grade hardware — Running LLMs on space-grade edge hardware enhances satellite autonomy and reduces data processing latency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.