Pith. sign in

REVIEW 3 cited by

Complexity Control Facilitates Reasoning-Based Compositional Generalization in Transformers

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2501.08537 v1 pith:FXCGAPS6 submitted 2025-01-15 cs.CL cs.LG

Complexity Control Facilitates Reasoning-Based Compositional Generalization in Transformers

classification cs.CL cs.LG
keywords complexitysolutionscompositionalreasoning-basedtaskstransformersacrossbias
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Transformers have demonstrated impressive capabilities across various tasks, yet their performance on compositional problems remains a subject of debate. In this study, we investigate the internal mechanisms underlying Transformers' behavior in compositional tasks. We find that complexity control strategies significantly influence whether the model learns primitive-level rules that generalize out-of-distribution (reasoning-based solutions) or relies solely on memorized mappings (memory-based solutions). By applying masking strategies to the model's information circuits and employing multiple complexity metrics, we reveal distinct internal working mechanisms associated with different solution types. Further analysis reveals that reasoning-based solutions exhibit a lower complexity bias, which aligns with the well-studied neuron condensation phenomenon. This lower complexity bias is hypothesized to be the key factor enabling these solutions to learn reasoning rules. We validate these conclusions across multiple real-world datasets, including image generation and natural language processing tasks, confirming the broad applicability of our findings.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Gauging, Measuring, and Controlling Critic Complexity in Actor-Critic Reinforcement Learning

    cs.LG 2026-07 unverdicted novelty 6.0

    Critic complexity quantified by spectral effective-rank entropy is measurable during TD3/PPO training, associated with behavior in a heterogeneous way, and can be altered by adding a spectral-entropy penalty to the cr...

  2. Unveiling the Mechanisms of Multi-Hop Reasoning in Transformers via Identity Bridge

    cs.LG 2025-09 conditional novelty 6.0

    Adding identity supervision on bridge tokens enables out-of-distribution two-hop reasoning in simple transformers, with a nuclear-norm theory explaining the benefit.

  3. An overview of condensation phenomenon in deep learning

    cs.LG 2025-04 unverdicted novelty 2.0

    Neural networks exhibit condensation of neurons into clusters with similar outputs whose number increases monotonically during training, facilitated by small initializations or dropout, providing insights into general...