The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/norgrai/syntax_review
Area: LLM Reasoning
Keywords: Syntactic Knowledge, Transformer Language Models, Interpretability, Systematic Review, Syntax-Semantics Interface
TL;DR¶
This paper presents the first large-scale systematic review of interpretability research evaluating the syntactic capabilities of Transformer-based language models, synthesizing 337 papers, 3,074 datapoints, and 113 languages into an open database, while mixed-effects regression reveals strong mastery of formal syntax contrasting with persistent fragility at the syntax-semantics interface, as well as an extreme concentration on English and BERT-family architectures.
Background & Motivation¶
Since Gulordava et al. (2018) demonstrated that language models trained solely on self-supervised next-token prediction acquire substantial hierarchical linguistic structure, interpretability research interrogating syntactic representations has exploded. The research community has deployed a heterogeneous toolkit spanning behavioral minimal-pair evaluations, representational probing classifiers, and interventional mechanistic circuit analyses. However, this vast literature has developed along fragmented trajectories, characterized by disparate operational definitions of "syntactic knowledge" and isolated empirical findings, leaving the broader NLP community without a unified, reproducible empirical map of what models actually know.
Prior syntheses in this domain have predominantly been narrative reviews that synthesize findings qualitatively. While these qualitative accounts identified valuable consensus patterns—such as the rapid acquisition of syntax early in pretraining, localization within middle layers, and functional specialization in select self-attention heads—they lack the transparent, quantitative evidential backing necessary to disentangle confounding factors. In particular, narrative accounts cannot rigorously determine how much variance stems from architectural scaling, pretraining corpora volume, cross-linguistic digital support, or inherent task difficulty across linguistic phenomena. Furthermore, probing success has frequently been conflated with behavioral competence or causal circuit implementation, obscuring deep methodological divergences.
To resolve these tensions and overcome the limitations of narrative syntheses, computational linguistics urgently requires the systematic review methodology established in medicine and cognitive science. This necessitates transparent search protocols, explicit eligibility boundaries, comprehensive multi-variable annotation schemes, and formal statistical modeling across distributed literature. By synthesizing 337 empirical studies and 3,074 datapoints spanning 113 languages, this work bridges qualitative linguistic intuition and quantitative meta-science. Core idea: Establish an open, PRISMA-compliant systematic review database of 337 studies and 3,074 datapoints, quantitatively decoupling behavioral competence, probing localization, and mechanistic causality to systematically characterize syntactic knowledge encoding and structural blind spots in Transformer language models.
Method¶
Overall Architecture¶
The systematic review adheres to the PRISMA 2020 guidelines, executing a rigorous four-stage pipeline comprising candidate identification, eligibility assessment, multi-dimensional annotation, and quantitative mixed-effects statistical synthesis. First, literature retrieval combines keyword search on Google Scholar with multi-tier backward and forward citation snowballing seeded by six prominent prior reviews, yielding 636 initial candidate papers. Second, four explicit eligibility criteria filter out non-Transformer architectures, non-empirical articles, and non-peer-reviewed preprints, isolating 337 empirical studies. Third, a 32-variable codebook spanning metadata, model attributes, experimental designs, and interpretability methodologies is populated via a hybrid human-in-the-loop workflow combining NotebookLM targeted extraction, ChatGPT summarization, and 100% expert manual verification. Finally, linear mixed-effects modeling (LMM) and rank-correlation analysis are applied to behavioral benchmarks, cross-linguistic evaluations, and representational studies.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Candidate Literature Identification<br/>Keyword Search + Citation Snowballing (636 papers)"] --> B["Eligibility Assessment & Exclusion<br/>Strict Filtering of Non-TLM / Non-Empirical (337 papers)"]
B --> C["Four-Category Systematic Codebook<br/>Meta-data / Models / Experiments / Interpretability (32 vars)"]
C --> D["Semi-Automated Annotation & Verification<br/>NotebookLM Retrieval + ChatGPT + Expert Manual Audit"]
D --> E["Tripartite Quantitative Synthesis<br/>Behavioral LMM / Probing Depth / Mechanistic Causality"]
Key Designs¶
1. Eligibility Assessment & Exclusion Criteria: Grounding Empirical Boundaries and Eliminating Architectural Confounders
To establish a coherent empirical baseline, the review enforces four explicit exclusion boundaries: (1) Architectures are strictly restricted to Transformer-based language models (TLMs), eliminating recurrent neural networks (RNNs/LSTMs) and encoder-decoder machine translation models to prevent architectural confounding; (2) Studies must empirically evaluate syntactic knowledge rather than downstream task performance where syntax is merely incidental; (3) Purely theoretical, position, or qualitative review papers lacking original quantitative datasets are excluded; (4) Publications are limited to peer-reviewed journal articles, conference proceedings, or established archival preprints, filtering out non-archival blog posts and theses. This rigorous screening reduced 636 candidates to 337 high-fidelity core studies covering 54 unique TLM families and 609 distinct model instances.
2. Four-Category Systematic Codebook: Harmonizing Heterogeneous Syntactic Paradigms
To map disparate experimental formulations onto a standardized evaluative schema, the authors formalized 32 annotation variables organized across four dimensions: (1) Meta-data: capturing publication venue, timeline, and inclusion/snowballing lineage; (2) Model-related Information: categorizing pretraining paradigms (59% bidirectional encoders, 28% causal decoders, 13% direct comparisons), language coverage (66% monolingual, 20% multilingual, 14% mixed), and parameter scales; (3) Experimental Design: coding materials (controlled minimal pairs vs. naturalistic corpora) and categorizing syntactic phenomena into 11 bottom-up clusters (agreement & feature-checking, dependency parsing & constituency, POS & lexical properties, hierarchical structure, word order, argument structure & constructions, binding & coreference, negation & NPIs, ellipsis & filler-gap dependencies, benchmarks, and typology); (4) Interpretability Techniques: disentangling behavioral evaluation (e.g., minimal-pair surprisal comparisons), probing classifiers (linear/structural probes), and mechanistic methods (further categorized by causal nature, training phase, and representational locality).
3. Semi-Automated Annotation with Dual Human Validation: High-Precision Large-Scale Evidence Extraction
Given the immense overhead of reading 337 full papers, the authors implemented an AI-assisted extraction workflow for four high-complexity unstructured variables: model type, syntactic phenomena, experimental materials, and interpretability method names. Each paper was processed using NotebookLM via targeted query prompts, followed by ChatGPT harmonization into coherent factual summaries. To guarantee absolute fidelity, the workflow was validated through an independent audit of 50 randomly sampled papers evaluated on a 0/0.5/1 accuracy metric, yielding 97% precision for experimental materials and 93% for interpretability methods (with zero false-positive errors). Furthermore, all 11 syntactic categories, model architectures, and the remaining 28 variables underwent complete manual re-annotation and verification by expert linguists to purge contextual errors stemming from cited related work.
4. Hierarchical Linear Mixed-Effects Modeling: Decoupling Confounding Covariates across Tasks and Languages
To account for cross-study heterogeneity, varied sample sizes, and non-independent observations across 1,577 non-benchmark datapoints spanning 113 languages, the authors fitted a linear mixed-effects regression model:
Fixed effects \(X_k\) encompass syntactic phenomena, model architecture (causal autoregressive vs. bidirectional masked), and language digital resource availability (operationalized via Joshi et al.'s 0–5 support scale). Random intercepts \(u_{\text{model}}\) and \(u_{\text{study}}\) absorb variance attributable to specific model checkpoints and author laboratory methodologies. This statistical formulation isolates genuine linguistic performance disparities from digital resource privilege and dataset idiosyncrasies.
Benchmark & Evaluation Protocol¶
The empirical synthesis evaluates models across three distinct investigative setups: 1. BLiMP English Controlled Benchmark: Synthesizing 25 papers (11 reporting full 12-subphenomenon scores, 14 reporting composite averages across 134 model instances), measuring targeted minimal-pair choice accuracy where random chance is 50%. 2. Non-Benchmark Cross-Linguistic Evaluation: Encompassing 1,577 empirical accuracy datapoints from 76 studies testing 113 natural languages, stratified across Joshi et al.'s (2020) digital support tiers. 3. Probing & Mechanistic Corpora: Analyzing layer-wise localization reports from 56 probing studies, outcome efficacy across 147 probing investigations, and evidence strength across 140 mechanistic papers.
Key Experimental Results¶
Main Results¶
The linear mixed-effects regression reveals sharp, statistically significant performance divides between formal morphosyntactic relations and phenomena situated at the syntax-semantics interface.
| Covariate / Predictor | Coefficient \(\beta\) | Standard Error (SE) | \(p\)-value | Empirical Finding |
|---|---|---|---|---|
| Intercept (Grand Mean Baseline) | 0.400 | 0.030 | \(<0.001^{***}\) | Expected global accuracy anchor across baseline conditions |
| Language Digital Support Category | +0.065 | 0.003 | \(<0.001^{***}\) | Each incremental resource level yields a ~6.5–7.0% accuracy gain |
| Model Type: Causal Decoder (vs Bidir) | +0.018 | 0.030 | \(>0.05\) (n.s.) | No significant main effect between causal and bidirectional on open test sets |
| POS & Other Lexical Properties | +0.120 | 0.030 | \(<0.001^{***}\) | Substantially higher accuracy than mean; surface lexical features easiest |
| Agreement & Feature-Checking | +0.110 | 0.020 | \(<0.001^{***}\) | Highly robust encoding; formal morphosyntactic matching exceeds grand mean |
| Binding & Coreference | +0.030 | 0.040 | \(>0.05\) (n.s.) | Intermediate difficulty; performance aligns closely with global average |
| Ellipsis & Filler-Gap Dependencies | -0.002 | 0.040 | \(>0.05\) (n.s.) | Broad variance across sub-constructions; near overall average |
| Dependency Parsing & Constituency | -0.040 | 0.030 | \(>0.05\) (n.s.) | Full-tree structural decoding exhibits moderate complexity |
| Hierarchical Structure | -0.060 | 0.030 | \(<0.05^{*}\) | Statistically significant deficit in handling nested structural depth |
| Argument Structure & Constructions | -0.090 | 0.030 | \(<0.01^{**}\) | Severe deficit in verb valence and constructional argument mapping |
| Negation & NPIs | -0.090 | 0.030 | \(<0.01^{**}\) | Substantial drop; semantic polarity licensing represents a major bottleneck |
Mixed-Effects Model & Correlation Analysis¶
Regression on the 134 English BLiMP observations quantifies the impact of pretraining data volume, architecture type, and the cross-phenomenon rank consistency of models.
| Predictor (BLiMP Composite Accuracy) | Coefficient \(\beta\) | Standard Error (SE) | Significance | Mechanistic Interpretation |
|---|---|---|---|---|
| Intercept (Bidirectional, 0 Tokens) | 77.06 | 1.91 | \(p < 0.001^{***}\) | Bidirectional pretraining establishes a strong baseline syntactic prior |
| Model Type: Causal Autoregressive | -8.90 | 2.16 | \(p < 0.001^{***}\) | Unidirectional masking incurs an ~8.9% penalty on full-sentence comparisons |
| Pretraining Corpus Size (per 1B Tokens) | +0.002 | 0.000 | \(p < 0.001^{***}\) | Log-linear scaling holds; additional pretraining data steadily boosts syntax |
| Sample Size / Log-Likelihood | \(N = 134\) / -492 | — | AIC: 995, BIC: 1012 | Model ranking variance driven by unified general syntactic competence |
| Strongest Rank Correlation Pair (Kendall's \(\tau\)) | \(\tau = 0.82\) | — | Argument Structure \(\leftrightarrow\) Control/Raising | Valence mastery correlates strongly with subject control resolution |
| Secondary Rank Correlation Pair (Kendall's \(\tau\)) | \(\tau = 0.75\) | — | Filler-Gap \(\leftrightarrow\) Island Effects | Long-distance dependency resolution correlates with constraint enforcement |
| Outlier Phenomenon (Kendall's \(\tau\)) | \(\tau = 0.17 \sim 0.37\) | — | Quantifier Scope | Quantifier interaction exhibits isolated, weak alignment with other syntax |
Key Findings¶
- Bifurcation Between Formal Syntax and the Syntax-Semantics Interface: Across both English BLiMP evaluations and 113 natural languages, models excel on formal morphosyntactic relations (all agreement categories exhibit median accuracies \(>85\%\), with top models matching or surpassing human baselines). Conversely, phenomena located at the syntax-semantics interface (NPI licensing, argument structure, control/raising, quantifiers, island constraints) collapse to medians \(<75\%\).
- Middle-Layer Concentration Without Discrete Pipelines: Among 56 probing studies reporting spatial localization, middle layers are identified most frequently as encoding the strongest syntactic information. However, mechanistic evidence indicates that syntactic encoding is distributed and accumulates progressively across layers rather than adhering to rigid, discrete classical NLP pipelines.
- Structural Imbalance in Literature: 69% of analyzed studies examine English exclusively (91% include English). A single model, BERT, accounts for 30% of all experimental datapoints (58% including BERT variants; 62% for BERT, GPT-2, RoBERTa, and mBERT combined). Furthermore, 81% of mechanistic studies remain purely observational, with only 19% conducting causal interventional experiments.
Highlights & Insights¶
- Pioneering PRISMA Systematic Reviews in Computational Linguistics: Translates the gold-standard review protocol from medicine into NLP, curating an openly accessible database of 337 papers and 3,074 datapoints that establishes a transparent empirical ground truth.
- Verification of a Unified "General Syntactic Capability": High cross-phenomenon Kendall's \(\tau\) rank correlations (\(\text{mean } \tau \approx 0.58\)) prove that models do not specialize in isolated grammatical quirks; superior syntactic models demonstrate generalized, uniform competence across almost all categories.
- Deconstruction of Probing Optimism: While 85% of probing studies report successful extraction of syntax, behavioral evaluations exhibit severe degradation on interface phenomena. This stark contrast quantitatively demonstrates that linear decodability of syntax does not imply its functional execution during generation.
Limitations & Future Work¶
- Restricted Architectural Scope: Analysis is intentionally bounded to Transformer models operating on textual inputs, precluding direct generalization to state-space models (e.g., Mamba), recurrent architectures, or multimodal foundation models.
- Anglocentric Search Bias: The primary literature retrieval keywords were conducted in English, likely under-sampling non-English publications from regional venues.
- Mechanistic Heterogeneity Barrier: The immense diversity of intervention methods (activation patching, circuit pruning, sparse autoencoders) prevented meta-analytic statistical pooling of mechanistic circuits, leaving interventional synthesis as a key challenge for future standardization.
Related Work & Insights¶
- vs. Qualitative Reviews (Linzen & Baroni 2021; Kulmizev & Nivre 2022): Replaces qualitative narrative arguments with rigorous meta-analytic regression models over 3,074 datapoints, providing verifiable effect sizes for scaling, pretraining objectives, and digital support tiers.
- vs. Standalone Benchmarks (Warstadt et al. 2020a; Jumelet et al. 2025): Synthesizes multi-benchmark outcomes across diverse literatures, discovering that controlled minimal pairs (e.g., noun ellipsis) can overestimate model capabilities due to constrained token distributions.
- Implications for LLM Reasoning: The systematic failure on syntax-semantics interface tasks reveals that LLM reasoning failures often stem not from formal grammatical incapacity, but from fragile compositional binding and quantifier scope representations at the boundary of syntax and semantics.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Pioneering application of PRISMA systematic review standards and hierarchical regression meta-analysis to NLP syntax interpretability.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Unprecedented scale synthesizing 337 papers, 3,074 empirical datapoints, 113 languages, and 54 model families.
- Writing Quality: ⭐⭐⭐⭐⭐ Exemplary scientific clarity, rigorous methodological conceptualization, and dense quantitative synthesis.
- Value: ⭐⭐⭐⭐⭐ An indispensable foundational reference and benchmark database for linguistic interpretability, LLM reasoning, and multilinguality.