Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models

Nora Graichen, Iria de-Dios-Flores, Gemma Boleda

TL;DR

The paper addresses what syntactic knowledge Transformer language models actually acquire from language modeling objectives by performing a systematic review of 337 studies and aggregating 1,015 results across languages, phenomena, and interpretability methods. It builds a public, annotated database to enable transparent, large-scale synthesis and benchmarking. Key findings reveal a strong English and BERT-centric bias, stronger performance on formal syntax than on syntax-semantics interfaces, and substantial methodological heterogeneity across studies and benchmarks. The authors advocate for broader language coverage, standardized reporting, increased use of mechanistic causal methods, and multilingual benchmarks to improve generalizability and mechanistic insight, aiming to advance robust interpretability in LLMs.

Abstract

We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models, reporting on 1,015 model results from a range of syntactic phenomena and interpretability methods. Our analysis shows that the state of the art presents a healthy variety of methods and data, but an over-focus on a single language (English), a single model (BERT), and phenomena that are easy to get at (like part of speech and agreement). Results also suggest that TLMs capture these form-oriented phenomena well, but show more variable and weaker performance on phenomena at the syntax-semantics interface, like binding or filler-gap dependencies. We provide recommendations for future work, in particular reporting complete data, better aligning theoretical constructs and methods across studies, increasing the use of mechanistic methods, and broadening the empirical scope regarding languages and linguistic phenomena.

The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models

TL;DR

The paper addresses what syntactic knowledge Transformer language models actually acquire from language modeling objectives by performing a systematic review of 337 studies and aggregating 1,015 results across languages, phenomena, and interpretability methods. It builds a public, annotated database to enable transparent, large-scale synthesis and benchmarking. Key findings reveal a strong English and BERT-centric bias, stronger performance on formal syntax than on syntax-semantics interfaces, and substantial methodological heterogeneity across studies and benchmarks. The authors advocate for broader language coverage, standardized reporting, increased use of mechanistic causal methods, and multilingual benchmarks to improve generalizability and mechanistic insight, aiming to advance robust interpretability in LLMs.

Abstract

We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models, reporting on 1,015 model results from a range of syntactic phenomena and interpretability methods. Our analysis shows that the state of the art presents a healthy variety of methods and data, but an over-focus on a single language (English), a single model (BERT), and phenomena that are easy to get at (like part of speech and agreement). Results also suggest that TLMs capture these form-oriented phenomena well, but show more variable and weaker performance on phenomena at the syntax-semantics interface, like binding or filler-gap dependencies. We provide recommendations for future work, in particular reporting complete data, better aligning theoretical constructs and methods across studies, increasing the use of mechanistic methods, and broadening the empirical scope regarding languages and linguistic phenomena.
Paper Structure (48 sections, 14 figures, 2 tables)

This paper contains 48 sections, 14 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Paper candidate identification.
  4. Eligibility assessment.
  5. Annotation.
  6. Results
  7. General overview.
  8. RQ1: Linguistic Phenomena
  9. RQ2: Methods
  10. RQ3: Knowledge status - BLiMP
  11. Discussion & Conclusion
  12. References
  13. Annotation dimensions
  14. Meta-Information
  15. Source for the article:
  16. ...and 33 more sections

Figures (14)

  • Figure 1: Flowchart summarizing database creation.
  • Figure 2: Left: Number of studies published each year (2025 until 31/07) categorized by model type. Right: 20 most frequently used TLMs. TLMs are counted once per study and grouped by their family or specific model name when available (e.g., one BERT count for a study that compares BERT-base/large/(un)cased). Overall, the database features 54 different TLM names and 609 instances without duplicates per study.
  • Figure 3: Left: Counts of syntactic phenomena by evaluated language setting (only English tested, English and other languages, and non-English). Right: Counts of syntactic phenomena investigated per study.
  • Figure 4: Top: Percentage of methods over years. Bottom: Counts of experimental materials and model types by interpretability method (unknown experimental material with fewer than 10 counts excluded).
  • Figure 5: Left: BLiMP average scores plotted against the number of training tokens (log scale). Right: BLiMP scores from the 11 studies that provide at least one result per-category for English. Human performance taken from Warstadt:etal:2020. Top $\star$ (Roberta large) and bottom black• (Minibert) TLM among those with complete per-category results.
  • ...and 9 more figures