Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Balancing forward and backward reasoning, by studying the Gadget Game

Josephine Lee¹, Jacob Loader², Jonas Bayer², Tracey Mills¹, Junyi Chu³, Samuel J Cheyette¹, Katherine M. Collins², Joshua B. Tenenbaum¹, Timothy Gowers²; ¹Massachusetts Institute of Technology, ²University of Cambridge, ³Stanford University

Presenter: Katherine M. Collins

Many real-world problems involve planning toward long-horizon goals, often requiring a balance of goal-directed backward search and forward exploration. Different domains may yield different tradeoffs for whether backward or forward reasoning is preferrable. Here, we study the utility and limitations of backward reasoning in a domain inspired by theorem proving: the "Gadget Game". We implement and evaluate a progressive suite of computational reasoning models in this puzzle environment. We begin with baseline backward reasoning solvers, systematically augmenting them with heuristics, constraint propagation, and smarter pruning to mitigate compute challenges that often arise in backwards planning. We demonstrate that highly optimized backward search, in this domain, is fundamentally limited when intermediate ``lemmas'' are required. To bridge this gap, we introduce mixed and adaptive models that integrate forward reasoning with backward reasoning, enabling bypassing the limits of pure goal-directed search. By testing these models in simulations across a range of puzzles, we aim to take initial steps to better understand how to strategically trade off cognitive effort and solution quality when navigating complex problem spaces.

Topic Area: Decision-Making, Cognitive Control & Event Cognition