Research map / Large language models / Reasoning and chain of thought

SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling

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 SPIRAL, which embeds Planner, Simulator, and Critic LLM agents (all prompts on one LLM) into an MCTS loop with reflection-driven reward shaping for tool-use planning, evaluated on TaskBench DailyLifeAPIs and HuggingFace.

Problem

LLM planners using linear chain-of-thought or sparse-reward MCTS struggle to recover from early errors in multi-step tool-use planning.

Approach

Embeds three specialized LLM roles (Planner proposes actions, Simulator predicts observations as a world model, Critic gives dense reflection scores) into MCTS, replacing random rollouts with a one-step grounded lookahead and blending a validity heuristic with the Critic score as a per-node reward (alpha=0.5).

Key results

• On DailyLifeAPIs with Llama 4 Maverick 17B in the cascaded residual setting, SPIRAL reaches 83.61% overall accuracy versus 67.39% for LATS.
• With Qwen 2.5 72B on HuggingFace, SPIRAL reaches 97.08% overall accuracy versus 88.16% for the best CoT variant (k=3).
• Ablation on Llama 3.3 70B HuggingFace: removing the Simulator drops accuracy from 97.44% to 69.48%, while uniform rewards yield 89.12%.

Surprising findings

• Removing the Planner improves accuracy on DailyLifeAPIs for Llama 3.3 70B (94.71% vs 98.35% full, but on HuggingFace w/o Planner is 87.48%), and w/o Validator is near standard MCTS, suggesting component effects vary by model.
• Standard MCTS with 50 iterations underperforms SPIRAL and the paper reports SPIRAL uses more API calls than CoT yet fewer tokens than the stronger CoT baselines.

Limitations

• Evaluation restricted to two TaskBench subsets, excluding multimedia tasks by design.
• Higher API call count and wall-clock latency than CoT.
• Cascaded residual evaluation uses only failures of CoT k=1, and the code is adapted to a public HF client with results possibly differing from the proprietary platform.

Experimental setup

Datasets: TaskBench DailyLifeAPIs TaskBench HuggingFace

Benchmarks: TaskBench DailyLifeAPIs TaskBench HuggingFace

Metrics: overall accuracy simple task accuracy complex task accuracy token efficiency API call efficiency

Baselines: Chain-of-Thought ReAct Reflexion RAFA ReAct+RAFA LATS Tree of Thoughts standard MCTS

Builds on: Monte Carlo Tree Search LATS ReAct Tree of Thoughts RAFA

Closest papers by meaning

• Thought of Search: Planning with Language Models Through The Lens of Efficiency (NeurIPS 2024 · similarity 0.90)
• LLMs Can Plan Only If We Tell Them (ICLR 2025 · similarity 0.90)
• Planning in the Dark: LLM-Symbolic Planning Pipeline Without Experts (AAAI 2025 · similarity 0.90)
• Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.90)
• PathWise: Planning through World Model for Automated Heuristic Design via Self-Evolving LLMs (ICML 2026 · similarity 0.89)
• Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.89)