Research map / Label-efficient and robust learning / Recognition and long-tail
Structurally Guided Task Decomposition in Spatial Navigation Tasks (Student Abstract)
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.
Extends a computational task-decomposition model of human planning with A* search using a Manhattan-distance heuristic, and tests it on an online spatial maze navigation experiment with 40 participants.
Problem
How people plan efficiently under limited cognitive resources in complex spatial navigation tasks, where existing task decomposition models only explained simple, unstructured planning problems.
Approach
Adds structural (spatial) information to a cost-based task decomposition framework by computing planning costs with A* using a Manhattan-distance heuristic, so the number of visited states approximates planning cost, and predicts which of two maze subtasks people choose.
Key results
- In the last trial, participants chose the model-favored subtask with logistic regression intercept 0.731 (p = 0.030).
- The proportion of the computationally simpler subtask chosen rose from 42.5% in the first trial to 67.5% in the last trial, a significant trial-number effect (slope 0.063, p = 0.021).
- About 32.5% of participants did not choose the subtask with lower planning cost after 12 trials.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Progress through repeated trials | Participants choosing between subtasks with different planning costs · Spatial navigation | proportion choosing favorable subtask: 42.5% -> 67.5% | Helped |
| Prefer the path with lower overall cost | Participants choosing between the two maze subtasks · Spatial navigation, last trial | choice of favorable path: 0.731 (sd = 0.338, p = 0.030) | Helped |
| Increase trial number | Participants choosing the computationally simpler subtask · Spatial navigation | logistic regression trial-number coefficient: 0.063 (sd = 0.027, p = 0.021) | Helped |
Surprising findings
- A learning effect was observed: choices shifted toward the lower-cost subtask over trials, implying a learned bias rather than immediate use of the model's predictions.
Limitations
- Predicts the choice of only the majority of participants; about 32.5% chose the higher-cost subtask.
- Small sample (40 participants) and a fixed, simplified set of two subtasks adjacent to the start state.
- Planning cost estimated only via A* visited-state counts; the choice of planner and heuristic may need improvement.
Experimental setup
Datasets: 48 custom maze navigation tasks (6 base mazes with 8 transformations each)
Metrics: proportion of participants choosing the lower-planning-cost logistic regression intercept and slope
Builds on: Correa et al. 2023 task decomposition framework A* search options framework
Closest papers by meaning
- EvoEmpirBench: Dynamic Spatial Reasoning with Agent-ExpVer (AAAI 2026 · similarity 0.84)
- CitySeeker: How Do VLMs Explore Embodied Urban Navigation with Implicit Human Needs? (ICLR 2026 · similarity 0.84)
- TaskLAMA: Probing the Complex Task Understanding of Language Models (AAAI 2024 · similarity 0.84)
- STRIDER: Navigation via Instruction-Aligned Structural Decision Space Optimization (NeurIPS 2025 · similarity 0.84)
- Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests (NeurIPS 2026 · similarity 0.83)
- Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations (ICLR 2026 · similarity 0.83)
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.