Grammar Walks in Structured Traversals

• Grammar Walks are structured traversals where grammatical rules annotate transitions, linking graph processes, combinatorial growth, parsing automata, and educational exercises.
• They enable methods like random walks with semantic role annotation in text-attributed graphs and derivational processes in permutation statistics, enhancing operational clarity.
• Empirical evidence, such as improved node classification on ACM, Wiki, and Amazon datasets, demonstrates their impact on performance and interpretability.

“Grammar Walks” denotes a family of constructions in which grammatical structure is attached to, generated by, or inferred from a traversal. In the cited literature, the traversal may be a random walk on a text-attributed graph, an insertion process on permutations and increasing trees, a path through a space of grammars connected by transformations, a walk in an SR-automaton derived from LR parsing tables, a chapter-by-chapter pass through a grammar book, a learner’s passage through inline exercises, or a contractibility-constrained closed walk generated by a context-free grammar (Zheng et al., 2 Nov 2025, Chen et al., 2024, Zaytsev, 2015, Quaglia, 2019, Li et al., 28 Oct 2025, Chan et al., 2022, Erickson et al., 2018). In these settings, grammar is not only a well-formedness condition. It becomes an annotation layer, a derivational calculus, a search constraint, a pedagogical control interface, or a bridge between syntax and semantics.

1. Scope and recurring structure

The cited papers use the expression for several distinct mechanisms. What they share is a coupling between a walk-like process and a grammar-like organization of roles, derivations, or constraints. In graph learning, the walk is an actual random walk whose steps are enriched with structural and semantic roles. In combinatorics, the walk is an iterative growth process on permutations or trees driven by grammar rules. In grammar engineering, it is a traceable path from one grammar representation to another. In parsing, it is a path through an automaton that mirrors shift-reduce behavior. In evaluation and education, it is a systematic progression through grammar rules, examples, or exercises. In topology and multimodal learning, grammar specifies which walks or hierarchical traversals are admissible (Zheng et al., 2 Nov 2025, Zaytsev, 2012, Hong et al., 2021).

Domain	Walk substrate	Role of grammar
Text-attributed graphs	Random walks on nodes	Structural and semantic role annotation
Permutations and trees	Insertion/growth steps	Production rules and formal derivatives
Grammar engineering	Transformation chains	Operator suites, ANF, prodsigs
Parsing	Paths in SR-automata	Conflict analysis and derivation recovery
Evaluation and tutoring	Rule-by-rule progression	Controlled probing, exercises, response control
Topology and multimodal learning	Closed walks or parse trees	Constraint language for admissible structures

This distribution of meanings suggests a stable abstraction: a grammar walk is a traversal whose states, transitions, or outputs are organized by grammar. The grammar may be explicit, as in a context-free grammar or a pedagogical repository, or implicit, as in role labels over graph nodes and edges.

2. Grammar-annotated walks on text-attributed graphs

In text-attributed graph learning, “grammar walks” are realized most explicitly in G²rammar, which turns random walks into sequences whose steps are annotated with structural and semantic grammatical roles (Zheng et al., 2 Nov 2025). A text-attributed graph is defined as

$$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathcal{X}),$$

with nodes $\mathcal{V}$, edges $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$, and textual attributes $\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$. Earlier graph-language approaches treat nodes as tokens and random walks as sentences, but they do not annotate nodes or edges with structural or semantic roles.

G²rammar introduces a bilingual grammar with a structural grammar $\mathcal{G}_s$ and a semantic grammar $\mathcal{G}_m$. Structural grammar assigns node roles

$$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$

edge roles

$$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$

and path roles

$$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$

HUB nodes are selected by degree centrality above an 80th percentile threshold $\tau_d$, BRIDGE nodes by betweenness centrality above a 75th percentile threshold $\mathcal{V}$0, and REGULAR nodes are the remainder. STRONG and WEAK edges are determined by Jaccard similarity of 1-hop neighborhoods, with $\mathcal{V}$1 giving STRONG. LOCAL paths satisfy $\mathcal{V}$2, EXPLORE paths contain at least one BRIDGE node, and RETURN paths satisfy $\mathcal{V}$3.

Semantic grammar is built from Approximate Mutual Information,

$\mathcal{V}$4

and introduces INFORM and INDEP relations using a median threshold $\mathcal{V}$5. A two-stage retrieval procedure uses FAISS plus sentence embeddings to retrieve top $\mathcal{V}$6 candidates and computes AMI only for those candidates, then keeps top $\mathcal{V}$7 semantic neighbors.

A plain random walk

$\mathcal{V}$8

is converted into a grammar-annotated sequence

$\mathcal{V}$9

Fine-tuning sequences additionally combine node text with hybrid structural-semantic roles:

$\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$0

The backbone is Qwen2-72B with LoRA, using structural grammar pre-training followed by semantic grammar fine-tuning.

Empirically, this form of grammar walk improves node classification on ACM, Wiki, and Amazon. Table 1 reports test micro-accuracy improvements over GDL4LLM from $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$1 to $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$2 on ACM, from $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$3 to $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$4 on Wiki, and from $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$5 to $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$6 on Amazon. Grammar walks are also compact: on ACM, G²rammar uses $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$7 tokens for 4th-order structure, compared with $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$8 for GDL4LLM and $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$9 for InstructGLM. Ablation shows that removing the entire grammar module yields the largest performance drop, $\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$0 on ACM, $\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$1 on Wiki, and $\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$2 on Amazon. In this setting, a grammar walk is therefore a random walk whose steps are explicitly typed by topological and semantic roles, rather than merely ordered as node tokens.

3. Derivational and proof-theoretic walks

In algebraic and combinatorial settings, grammar walks are derivational growth processes. Dumont’s grammar for permutation statistics is

$\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$3

with formal derivative

$\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$4

Repeated application of $\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$5 to the seed $\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$6 generates the multivariate polynomials

$\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$7

where

$\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$8

The same grammar also generates

$\mathcal{X}=\{x_1,\ldots,x_{|\mathcal{V}|}\}$9

and the paper proves

$\mathcal{G}_s$0

so $\mathcal{G}_s$1 and $\mathcal{G}_s$2 are equidistributed over $\mathcal{G}_s$3. Here the walk is the repeated insertion of $\mathcal{G}_s$4 into permutations or complete increasing binary trees, with grammar rules specifying the allowed local updates (Chen et al., 2024).

The same paper explicitly interprets repeated differentiation $\mathcal{G}_s$5 as a walk on combinatorial structures. In tree language, the state space is complete increasing binary trees with labels $\mathcal{G}_s$6, and each rule such as $\mathcal{G}_s$7 replaces a leaf by an internal node and new labelled leaves. In permutation language, the state space consists of permutations with grammatical labelings, and each step inserts the new maximum element at a labelled position. The grammar-assisted bijection of Theorem 4.3 synchronizes two such walks and maps left succession sets to fixed-point sets while sending $\mathcal{G}_s$8 to $\mathcal{G}_s$9.

In formal syntax, categorial minimalist grammars make the same intuition proof-theoretic. Minimalist derivations consume selector, category, licensor, and licensee features through MERGE and MOVE, while the categorial presentation encodes these operations in partially commutative linear logic with $\mathcal{G}_m$0. MERGE becomes elimination on directional implications, MOVE is encoded by $\mathcal{G}_m$1, and derivations are labelled by triples $\mathcal{G}_m$2 that track specifier, head, and complement yields. The syntax-semantics interface is mediated by a homomorphism $\mathcal{G}_m$3 from syntactic categories to semantic types, so every syntactic step has a semantic counterpart as λ-application or λμ-DRT substitution. The paper’s derivation of “the children ate a pizza” culminates in a λμ-DRT representation from which different reduction paths derive collective and distributive readings. In that framework, grammar walks are simultaneous walks through feature strings, proof objects, and λμ-reduction graphs (Amblard et al., 2010).

4. Grammar walks in grammar engineering and parser automata

In grammar engineering, the walk is no longer a traversal of tokens or nodes but a path through a space of grammars connected by transformations. Guided grammar convergence begins with grammars in a broad sense, written as

$\mathcal{G}_m$4

and seeks a transformation path from a servant grammar $\mathcal{G}_m$5 to a master grammar $\mathcal{G}_m$6. The method normalizes grammars into Abstract Normal Form, computes footprints and production signatures, infers a nominal resolution relation $\mathcal{G}_m$7, and then synthesizes bidirectional transformations. Footprints record how a nonterminal occurs in a production right-hand side; production signatures collect these footprints; structural equivalence is then defined up to those signatures. The paper reports automatically inferring bidirectional transformations between 11 grammars of the same artificial functional language, including parser specifications, algebraic data types, metamodels, XML schemata, and object models (Zaytsev, 2015).

“The Grammar Hammer of 2012” broadens this view into a grammar laboratory. Guided grammar convergence is only one walk type among several. XBGF, CBGF, EXBGF, NBGF, and $\mathcal{G}_m$8BGF provide operator suites for steps such as rename, extract, inline, factor, distribute, vertical, and horizontal. EDD and XEDD treat EBNF dialects and metasyntactic specifications as first-class objects that can themselves be transformed. Grammar Zoo acts as a repository of grammars, transformations, extraction scripts, and histories. Open notebook computer science records the paths actually taken. In this literature, a grammar walk is explicitly a sequence

$\mathcal{G}_m$9

where each $$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$0 may be a transformation operator, a convergence step, a mutation, a move along a megamodel edge, or a notation change coupled with a grammar mutation (Zaytsev, 2012).

In parsing theory, the walk becomes a path in a specialized automaton derived from an LR table. Given a parsing table $$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$1, the SR-automaton is

$$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$2

where one kind of transition corresponds to shift/goto and another to reduce under a prospective symbol $$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$3. Its accepted language satisfies

$$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$4

Ambiguity is characterized by two executions that share a prefix, diverge at a conflict on some $$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$5, and both end in accepting configurations for the same word $$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$6 but with different production-stacks. The search strategy is therefore conflict-driven: it looks for ambiguous words along SR-automaton paths that reach a conflict state and then diverge along distinct conflict resolutions (Quaglia, 2019).

5. Evaluation, tutoring, and controlled generation

In LLM evaluation and educational systems, grammar walks are systematic traversals through grammatical phenomena. The Luxembourgish probing study defines a four-stage “Grammar Book Guided” pipeline: “Material Inspector,” “Phrasing Atelier,” “Twin Forge,” and “Proof Stand.” It begins by extracting grammar-book content from PDF, including tables via PaddleOCR or PyMuPDF, focuses on syntax and morphology, uses GPT-5 to identify grammar points and generate controlled English–Luxembourgish sentence pairs, filters these pairs, constructs minimal pairs, and then evaluates models on four tasks:

$$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$7

The resulting G2040 dataset contains 2,040 example sentences tagged with 673 distinct rules. The paper reports a weak positive correlation between translation performance and grammatical understanding, and emphasizes that Task 4, minimal-pair identification, is much less correlated with translation metrics than Tasks 1–3. GPT-5 reaches Task 4 accuracy of $$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$8, while many open-source models remain near chance level (Li et al., 28 Oct 2025).

AGReE uses a different educational walk: a passage-driven sequence of inline grammar exercises. It takes user-submitted passages, segments them into sentences, turns every other sentence into a grammar item, identifies constructs such as punctuation, articles, conjunctions, pronouns, prepositions, verbs, and nouns, masks one construct per sentence, and uses BERT masked language modeling to validate the original token and rank distractors. In a large-scale human evaluation of 4,568 multiple-choice practice items, $$\phi:\mathcal{V}\rightarrow\{\text{HUB},\text{BRIDGE},\text{REGULAR}\},$$9 of items were answered correctly by all five raters, $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$0 had a majority of raters choosing the correct target, and $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$1 of responses indicated that raters disagreed with the idea of multiple correct answers. The error analysis showed that raters made the most mistakes for punctuation and conjunctions (Chan et al., 2022).

Grammar control in dialogue response generation turns the walk into a progression through a pedagogical repository of grammar skills. The repository is the English Grammar Profile, comprising 1,222 skills in 86 categories. The paper defines two control tasks: explicit grammar constraints, where a response must include all skills in a subset $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$2, and categorical grammar constraints, where a response must include at least one skill from each requested category-level pair. Grammar presence is detected by functions

$$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$3

The strongest results come from guided decoding of Llama3 using future discriminators and a FUDGE-like logit transformation. On explicit constraints, GPT-3.5 prompting achieves $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$4 mean constraint satisfaction, Llama3 decoding achieves $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$5, and the hybrid fine-tune-plus-decoding model reaches $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$6 but with poorer response quality. The same study also reports a simulation in which grammar-controlled responses increase target grammar usage in later learner responses for 25 out of 47 tested input-output grammar pairs (Glandorf et al., 11 Feb 2025).

6. Multimodal and topological generalizations

VLGrammar extends the idea of grammar walks across vision and language. It induces a language grammar and an image grammar simultaneously using compound probabilistic context-free grammars, then aligns their constituents with a contrastive score

$$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$7

The benchmark dataset PartIt contains 10,613 objects across chair, table, bed, and bag, with 120,110 part semantics and human-written sentences describing part-level semantics. VLGrammar outperforms all baselines in image grammar induction and language grammar induction on PartIt, improves image unsupervised clustering accuracy by $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$8, performs well in image retrieval and text retrieval, and generalizes to unseen categories (Hong et al., 2021).

Topological graph algorithms supply a different extension. In directed surface graphs, one studies closed directed walks that are contractible or bounding. For hyperbolic surfaces, a key step constructs a context-free grammar with $$\psi:\mathcal{E}\rightarrow\{\text{STRONG},\text{WEAK}\},$$9 non-terminals that generates all contractible closed walks of length at most $$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$0, and only contractible closed walks, in a system of quads of genus $$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$1. The same paper also gives $$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$2-time algorithms to determine whether a graph contains a simple contractible cycle or a contractible closed walk, and an $$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$3-time algorithm for detecting a bounding closed walk, where $$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$4 is the first Betti number (Erickson et al., 2018).

These two lines of work place grammar walks outside ordinary syntax. In VLGrammar, grammar organizes hierarchical traversals over text spans and part spans. In the surface-graph setting, grammar specifies a language of admissible topological walks. The shared structure is that a grammar becomes a finite or compact representation of a very large search space of traversals.

7. Limitations, misconceptions, and open questions

A recurrent misconception is that grammar walks always alter the underlying sampling policy. In G²rammar, the sampling procedure itself remains a standard random walk with uniform transition probabilities based on normalized adjacency; the novelty lies in annotation, not in grammar-conditioned transition bias. The same paper also notes computational overhead for centrality, Jaccard, and AMI computations, and dependence on heuristic thresholds such as the 80th and 75th percentiles and the Jaccard threshold $$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$5 (Zheng et al., 2 Nov 2025).

Another misconception is that fluent generation or strong translation scores imply deep grammatical competence. The Luxembourgish probing study finds only a weak positive correlation between translation performance and grammar understanding, and reports that larger models remain weak in morphology and syntax, especially on minimal-pair tasks. It also notes that Eifeler Regel violations accounted for $$\rho:\mathcal{C}\rightarrow\{\text{LOCAL},\text{EXPLORE},\text{RETURN}\}.$$6 of generated sentences, which were nonetheless marked grammar-correct because the evaluation focused on syntax and morphology rather than phonology (Li et al., 28 Oct 2025).

In pedagogical generation, high aggregate quality does not eliminate local ambiguity. AGReE shows strong overall results, but its error analysis concentrates failures in punctuation and conjunctions, including cases where all five raters preferred a distractor over the intended target. This is an important boundary condition for any passage-based grammar walk that relies on multiple-choice cloze generation (Chan et al., 2022).

In grammar engineering, guided convergence is explicitly “a more narrowly applicable technique” that depends on bringing grammars into Abstract Normal Form and on establishing structural equivalence via footprints and production signatures. Grammars whose essential correspondences are not exposed by ANF and prodsigs remain difficult. In parser-based ambiguity detection, the search is conflict-driven and bounded, and ambiguity itself remains undecidable, so the SR-automaton strategy is exploratory rather than complete (Zaytsev, 2015, Quaglia, 2019).

In the topological setting, grammar does not remove hardness barriers. Detecting simple bounding cycles is NP-hard, and the same paper shows NP-hardness for shortest simple contractible cycles, shortest simple bounding cycles, and shortest bounding closed walks. In the multimodal setting, VLGrammar currently defines image grammar over part sequences rather than directly over 2D images, so spatial structure is mediated through a sequence representation (Erickson et al., 2018, Hong et al., 2021).

Taken together, the literature presents “Grammar Walks” not as a single standardized formalism but as a transferable research pattern. A walk may be grammatical because grammar annotates its steps, because production rules generate its successive states, because transformations trace a path through grammar space, because an automaton walk reconstructs ambiguous derivations, or because a grammar-book curriculum orders the sequence of tests or interactions. What remains stable across these meanings is the use of grammar to turn traversal into a structured, interpretable, and often optimizable object.