Mechanistic evaluation of Transformers and state space models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Arora, Aryaman, Rathi, Neil, Selvam, Nikil Roashan, Csordás, Róbert, Jurafsky, Dan, Potts, Christopher
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912860432498688
author Arora, Aryaman
Rathi, Neil
Selvam, Nikil Roashan
Csordás, Róbert
Jurafsky, Dan
Potts, Christopher
author_facet Arora, Aryaman
Rathi, Neil
Selvam, Nikil Roashan
Csordás, Róbert
Jurafsky, Dan
Potts, Christopher
contents State space models (SSMs) for language modelling promise an efficient and performant alternative to quadratic-attention Transformers, yet show variable performance on recalling basic information from the context. While performance on synthetic tasks like Associative Recall (AR) can point to this deficiency, behavioural metrics provide little information as to \textit{why} -- on a mechanistic level -- certain architectures fail and others succeed. To address this, we conduct experiments on AR, and find that only Transformers and Based SSM models fully succeed at AR, with Mamba and DeltaNet close behind, while the other SSMs (H3, Hyena) fail. We then use causal interventions to explain why. We find that Transformers and Based learn to store key-value associations in-context using induction. By contrast, the SSMs seem to compute these associations only at the last state using a single layer. We further investigate the mechanism underlying the success of Mamba, and find novel evidence that Mamba \textit{does} implement induction: not via the SSM, but instead via short convolutions. Further experiments on a new hierarchical retrieval task, Associative Treecall (ATR), show that all architectures learn the same mechanism as they did for AR. Furthermore, we show that Mamba can learn Attention-like induction on ATR when short convolutions are removed. These results reveal that architectures with similar accuracy may still have substantive differences, motivating the adoption of mechanistic evaluations.
format Preprint
id arxiv_https___arxiv_org_abs_2505_15105
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Mechanistic evaluation of Transformers and state space models
Arora, Aryaman
Rathi, Neil
Selvam, Nikil Roashan
Csordás, Róbert
Jurafsky, Dan
Potts, Christopher
Computation and Language
Artificial Intelligence
I.2.7
State space models (SSMs) for language modelling promise an efficient and performant alternative to quadratic-attention Transformers, yet show variable performance on recalling basic information from the context. While performance on synthetic tasks like Associative Recall (AR) can point to this deficiency, behavioural metrics provide little information as to \textit{why} -- on a mechanistic level -- certain architectures fail and others succeed. To address this, we conduct experiments on AR, and find that only Transformers and Based SSM models fully succeed at AR, with Mamba and DeltaNet close behind, while the other SSMs (H3, Hyena) fail. We then use causal interventions to explain why. We find that Transformers and Based learn to store key-value associations in-context using induction. By contrast, the SSMs seem to compute these associations only at the last state using a single layer. We further investigate the mechanism underlying the success of Mamba, and find novel evidence that Mamba \textit{does} implement induction: not via the SSM, but instead via short convolutions. Further experiments on a new hierarchical retrieval task, Associative Treecall (ATR), show that all architectures learn the same mechanism as they did for AR. Furthermore, we show that Mamba can learn Attention-like induction on ATR when short convolutions are removed. These results reveal that architectures with similar accuracy may still have substantive differences, motivating the adoption of mechanistic evaluations.
title Mechanistic evaluation of Transformers and state space models
topic Computation and Language
Artificial Intelligence
I.2.7
url https://arxiv.org/abs/2505.15105