Research map / Learning theory and algorithms / Combinatorial and search optimization

System 1.x: Learning to Balance Fast and Slow Planning with Language Models

ICLR 2025 · Learning theory and algorithms › Combinatorial and search optimization · approach: Classical and HTN planning · 1 citations (OpenAlex)

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 System-1.x, a controllable hybrid LLM planner that decomposes problems into sub-goals, routes easy ones to a fast System-1 and hard ones to a search-trained System-2, all fine-tuned from one Mistral-7B with LoRA using search traces.

Problem

LLM planners face a trade-off: direct plan generation (System-1) is cheap but inaccurate, while search-based planning (System-2) is accurate but token-expensive and infeasible for long plans. Users also lack a way to control this compute-accuracy trade-off.

Approach

A controller fine-tuned on one LLM decomposes a plan into sub-goals and labels each as easy (System-1) or hard (System-2), governed by a user-specified hybridization factor x; training data comes from a sliding-window split of gold search traces ranked by a hardness function, and test-time bias on the controller adjusts the System-2 share.

Key results

• On Maze at default budget (13.6 states), System-1.5 reaches 70.4% plan validity vs 48.7% for System-1 and 37.2% for System-2 at matched states.
• Neuro-symbolic System-1.5 with A* reaches 70.5% validity at 11.6 states vs 31.0% for A* at matched budget.
• On OOD Blocksworld, System-1.5 with sub-goals gets 25% validity at 30 states vs 13% for System-2.

Measured effects reported in the paper

Surprising findings

• Sub-goal decomposition can reduce plan optimality (plans sometimes sub-optimal) even while improving validity.
• Without sub-goal decomposition, the controller predicts almost all Blocksworld OOD samples as hard, collapsing to System-2 behavior.

Limitations

• Evaluated only on two synthetic planning domains (5x5 mazes and Blocksworld).
• Sub-goal decomposition may hurt plan optimality.
• Hardness functions are hand-designed per domain.

Experimental setup

Datasets: synthetic 5x5 maze planning problems synthetic Blocksworld problems

Benchmarks: Maze Navigation (5x5) Blocksworld (plan lengths 7-10 OOD)

Metrics: plan validity #states explored

Baselines: System-1 (direct plan generation) System-2 (A*-trace trained LLM) System-1.5 without sub-goal decomposition A* search Searchformer Stream-of-Search

Builds on: Stream-of-Search Searchformer A* search LoRA Dual-Process Theory

Closest papers by meaning

• SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search (AAAI 2026 · similarity 0.89)
• Language Control Diffusion: Efficiently Scaling through Space, Time, and Tasks (ICLR 2024 · similarity 0.88)
• SayCanPay: Heuristic Planning with Large Language Models Using Learnable Domain Knowledge (AAAI 2024 · similarity 0.88)
• Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.88)
• Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.88)
• Non-myopic Generation of Language Models for Reasoning and Planning (ICLR 2025 · similarity 0.88)