Learning to Search with Subgoals

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F68407700%3A21730%2F25%3A00388665" target="_blank" >RIV/68407700:21730/25:00388665 - isvavai.cz</a>

Výsledek na webu

<a href="https://doi.org/10.1007/978-3-032-02725-2_24" target="_blank" >https://doi.org/10.1007/978-3-032-02725-2_24</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-3-032-02725-2_24" target="_blank" >10.1007/978-3-032-02725-2_24</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Learning to Search with Subgoals

Popis výsledku v původním jazyce

This work investigates whether a Transformer-based language model can learn to imitate a problem-solving process that decomposes tasks into subgoals, akin to human cognitive strategies. We train the model to replicate a solver that employs a greedy approach, switching to subproblems upon encountering obstacles. Using two synthetic tasks—the Countdown arithmetic puzzle and a Reachability with Obstacles pathfinding task—we demonstrate successful imitation of a simple solver, with generalization to unseen input samples and solution path lengths. We evaluate several variants of the Pythia model, finding that even a compact model (310k parameters) performs competitively, though larger models converge faster. Our results suggest that even small language models can internalize structured, hierarchical problem-solving, highlighting their potential for understanding how human-like subgoal decomposition can be implemented with neural networks.

Název v anglickém jazyce

Learning to Search with Subgoals

Popis výsledku anglicky

This work investigates whether a Transformer-based language model can learn to imitate a problem-solving process that decomposes tasks into subgoals, akin to human cognitive strategies. We train the model to replicate a solver that employs a greedy approach, switching to subproblems upon encountering obstacles. Using two synthetic tasks—the Countdown arithmetic puzzle and a Reachability with Obstacles pathfinding task—we demonstrate successful imitation of a simple solver, with generalization to unseen input samples and solution path lengths. We evaluate several variants of the Pythia model, finding that even a compact model (310k parameters) performs competitively, though larger models converge faster. Our results suggest that even small language models can internalize structured, hierarchical problem-solving, highlighting their potential for understanding how human-like subgoal decomposition can be implemented with neural networks.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

—

Rok uplatnění

2025

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Advances in Computational Intelligence 18th International Work-Conference on Artificial Neural Networks, IWANN 2025, A Coruña, Spain, June 16–18, 2025, Proceedings, Part I

ISBN

978-3-032-02724-5

ISSN

0302-9743

e-ISSN

1611-3349

Počet stran výsledku

12

Strana od-do

310-321

Název nakladatele

Springer

Místo vydání

Cham

Místo konání akce

A Coruña

Datum konání akce

16. 6. 2025

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—