A Pilot Mechanistic Study of Cataphoric Pronoun Resolution in GPT-2

Cataphora, a pronoun that precedes its antecedent (as in "Before she left, Maria locked the door"), requires maintaining an unresolved dependency rather than retrieving an already-established referent. Human sentence-processing research shows that comprehenders actively and persistently search for cataphoric antecedents, but to our knowledge no prior work examines this mechanistically in language models. We present a pilot study in GPT-2-small that combines behavioral measurement, activation patching, logit-lens analysis, linear probing and attention analysis. We find a large behavioral effect that is predominantly, though not entirely, attributable to syntactic structure; evidence that gender information is carried in the residual stream between the pronoun and its antecedent; heads and layers whose ordering suggests separate search and read-out stages; and an unexpected gender asymmetry in the model's default resolution behavior. We report these findings with attention to their limits.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24283924
Primary Topic
Natural Language Processing Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Pilot Mechanistic Study of Cataphoric Pronoun Resolution in GPT-2

NEHAA SHRI MS
1 links
Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
preprint

A Pilot Mechanistic Study of Cataphoric Pronoun Resolution in GPT-2

NEHAA SHRI MS
preprint en

Abstract

Cataphora, a pronoun that precedes its antecedent (as in "Before she left, Maria locked the door"), requires maintaining an unresolved dependency rather than retrieving an already-established referent. Human sentence-processing research shows that comprehenders actively and persistently search for cataphoric antecedents, but to our knowledge no prior work examines this mechanistically in language models. We present a pilot study in GPT-2-small that combines behavioral measurement, activation patching, logit-lens analysis, linear probing and attention analysis. We find a large behavioral effect that is predominantly, though not entirely, attributable to syntactic structure; evidence that gender information is carried in the residual stream between the pronoun and its antecedent; heads and layers whose ordering suggests separate search and read-out stages; and an unexpected gender asymmetry in the model's default resolution behavior. We report these findings with attention to their limits.

Zenodo (CERN European Organization for Nuclear Research)
Vellore Institute of Technology University (IN)
Natural Language Processing Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon