Overview
This research examines the scaling properties of implicit deductive reasoning within depth-bounded Transformers. The investigation focuses on models processing Horn clauses. A primary goal was to understand how these Transformers perform when provability is systematically decorrelated from spurious features and algorithmic alignment is enforced.
Research Context
The study delves into the capabilities of Transformers to perform deductive reasoning implicitly. Deductive reasoning, in this context, is applied to Horn clauses. The methodology employed a systematic decorrelation of provability from spurious observational features and prioritized algorithmic alignment. This framework allowed for an analysis of the models' intrinsic reasoning capacities.
Approach
The researchers utilized depth-bounded Transformers for their investigation. A key aspect of the experimental setup involved the systematic decorrelation of provability from features deemed spurious. Concurrently, efforts were made to enforce algorithmic alignment within the models. The architecture specifically included models with a bidirectional prefix mask. Performance was evaluated across varying graph topologies and problem widths to assess the robustness and generalizability of the implicit reasoning capabilities. The comparison point for implicit reasoning performance was explicit Chain-of-Thought (CoT) performance.
Findings
- In sufficiently deep models equipped with a bidirectional prefix mask, implicit deductive reasoning approached the performance levels observed with explicit Chain-of-Thought (CoT).
- This approximation of CoT performance for implicit reasoning was consistent across different graph topologies.
- The similarity in performance between implicit reasoning and explicit CoT also held true across various problem widths.
- Despite the comparable performance in certain scenarios, explicit CoT was found to remain necessary for depth extrapolation. This indicates a limitation of implicit reasoning in maintaining performance when reasoning depth increases beyond the training distribution or a certain bound.