Research map / Large language models / Agents, code and math
Improving Large Language Model Planning with Action Sequence Similarity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GRASE-DC, a two-stage in-context exemplar selection pipeline that ranks and clusters exemplars by action sequence similarity (AS) of model-generated plans to improve LLM planning accuracy.
Problem
LLM planning via in-context learning depends heavily on which exemplars are chosen; standard problem/task-description similarity can pick exemplars with drastically different correct plans, misleading the model.
Approach
Score exemplar candidates by action sequence similarity (longest common action sequence normalized by lengths) between the model's initial generated plan and each candidate plan, then apply dynamic agglomerative clustering on AS distances to balance relevance and diversity, optionally iterating and gating with a plan validator (VAL).
Key results
- On Blocksworld with Gemini 1.5 Pro, GRASE raises planning accuracy from 43 (Random) to 58 points.
- GRASE-DC achieves roughly 11-40 point absolute accuracy gains over Random across PDDL tasks while using 27.3% fewer exemplars on average.
- On out-of-distribution Blocksworld (8-20 blocks), GRASE-DC improves accuracy by 23.3 points on average over Random exemplars.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select exemplars by generated-plan action sequence similarity instead of random sampling | GRASE on Blocksworld · Blocksworld | planning accuracy: 43 to 58 | Helped |
| select exemplars by generated-plan action sequence similarity instead of random sampling | GRASE on Minigrid · Minigrid | planning accuracy: 58 to 64.3 | Helped |
| select exemplars by generated-plan action sequence similarity instead of random sampling | GRASE on Tetris · Tetris | planning accuracy: 5.7 to 16 | Helped |
| select exemplars by generated-plan action sequence similarity instead of random sampling | GRASE on Logistics · Logistics | planning accuracy: 20.7 to 26 | Helped |
| add dynamic clustering to GRASE to filter redundancy and balance relevance and diversity | GRASE-DC on PDDL planning tasks · PDDL planning tasks | planning accuracy: ~11-40 point absolute | Helped |
| apply GRASE-DC to out-of-distribution rather than in-distribution Blocksworld examples | GRASE-DC · Blocksworld | planning accuracy improvement over Random: 23.3 ... versus 16.3 | Mixed |
| add validator-based rejection sampling to GRASE (ablation) | GRASE on Blocksworld · Blocksworld | planning accuracy: 51 to 64 | Helped |
| add validator-based rejection sampling to Random (ablation) | Random on Blocksworld · Blocksworld | planning accuracy: 36 to 52 | Helped |
| iterate validator-based rejection sampling with GRASE (ablation) | GRASE-RS on Blocksworld · Blocksworld | planning accuracy: 64 to 70 | Helped |
| replace GRASE's generated-plan similarity selection with embedding-based task similarity (ablation) | Blocksworld exemplar selection · Blocksworld | planning accuracy: 51 to 35 | Hurt |
| represent plans with object-centric action sequences (OAS) rather than action sequences (ablation) | Oracle-plan similarity selection · Blocksworld | planning accuracy | Hurt |
| represent plans with execution-based action sequences (ES) rather than action sequences (ablation) | Oracle-plan similarity selection · Blocksworld | planning accuracy | Helped |
Surprising findings
- Task-description similarity is sometimes worse than random exemplar selection (e.g., Blocksworld and Minigrid), since semantically similar task descriptions can have very different plans.
- Using model-generated plans for exemplar ranking (GRASE) outperforms ranking with oracle test plans (Baseline_AS) in several settings.
Limitations
- Relies on closed-source API LLMs whose details are undisclosed, so cross-model comparisons may be unfair.
- Small open-source models (under 70B) show under 3% planning accuracy on Blocksworld, so gains are limited for weaker models.
- Action sequence representation ignores objects and state, so it is not an equivalent representation of plans.
Experimental setup
Datasets: Blocksworld Minigrid Logistics Tetris Natural Plan Trip Planning Natural Plan Calendar Planning
Benchmarks: Blocksworld Minigrid Logistics Tetris Trip Planning Calendar Planning
Metrics: planning accuracy
Baselines: Random exemplar selection Task description similarity (QA-F1) Gecko embedding retrieval Baseline_AS (oracle plan similarity)
Builds on: Auto-CoT Chain-of-thought prompting VAL plan validator Natural Plan
Closest papers by meaning
- Large Language Models as Planning Domain Generators (Student Abstract) (AAAI 2024 · similarity 0.90)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.89)
- Planning in the Dark: LLM-Symbolic Planning Pipeline Without Experts (AAAI 2025 · similarity 0.89)
- System 1.x: Learning to Balance Fast and Slow Planning with Language Models (ICLR 2025 · similarity 0.88)
- Tree-Planner: Efficient Close-loop Task Planning with Large Language Models (ICLR 2024 · similarity 0.87)
- Language Control Diffusion: Efficiently Scaling through Space, Time, and Tasks (ICLR 2024 · similarity 0.87)
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.