Year: 2026
Type: dissertation
Abstract: This thesis studies how input representations and supervised reasoning traces affect out-of-distribution (OOD) generalisation in FrozenLake planning. FrozenLake is a deterministic grid-world task in which a model receives a complete map and outputs a move sequence that reaches the goal while avoiding holes; candidate plans can be checked exactly by simulation. The thesis extends an existing benchmark with coordinate-based inputs and trace formats, solution augmentation, and structural stress tests based on direction and turn complexity. It also tests whether the trace-supervision findings transfer to a different model family. Using simulator-based valid-plan accuracy, the strongest non-augmented coordinate-based input–trace combinations reach 0.83 accuracy on larger maps with unseen start-goal distances and 0.62 on fixed-size distance extrapolation. When coordinate-description supervision is augmented with all optimal paths for each training map, the corresponding OOD averages reach 0.91 and 0.85. These gains do not imply full structural generalisation: performance drops sharply on maps whose shortest solutions require more distinct movement directions or more turns. Targeted fine-tuning improves transfer only once sufficiently many structurally complex examples are added. Overall, the results show that reasoning-trace supervision can induce useful procedure-like planning behaviour, with substantial gains from coordinate-based representations, structured trace formats and broader training coverage of valid solutions and structurally difficult examples. At the same time, size and distance extrapolation alone overestimates structural robustness.
Source: Aaltodoc (Aalto University)
Author: Tanchotikul, Chanoknan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204017244