What does this research mean for the field?

Implementing Universal Compaction and Grammar-Directed Generation in language models eliminates computational waste on structurally determined tokens, achieving 75-93% data compression and reducing forward passes by 40-80%. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To enhance the efficiency of language models by eliminating unnecessary computations during token prediction.

May 18, 2026Open Access

Grammar-Directed Compaction and Generation: Structural Intelligence for Exact-Arithmetic Language Models

Key Points

To enhance the efficiency of language models by eliminating unnecessary computations during token prediction.
Developed Universal Compaction for compressing structured data into tables with high fidelity.
Implemented Grammar-Directed Generation using Prolog grammars to optimize token output.
Validated the system with a Python implementation that passed 178 tests.
Achieved 75-93% compression of structured source material without losing critical information.
Reduced the number of forward passes by 40-80% depending on the type of output generated.
Validated the new systems with a working Python model demonstrating fidelity and grammar inheritance.

Abstract

Language models waste most of their computation predicting tokens that are structurally determined. When generating a Python function, the tokens `def`, `(`, `)`, `:`, and the indentation are not creative decisions — they are grammatical facts. When presenting data in a table, the column separators, row boundaries, and alignment characters are format requirements, not content. Current language models spend a full forward pass on every one of these tokens, running attention over the entire context and softmax over the full vocabulary to predict a closing parenthesis that was inevitable the moment the opening parenthesis appeared. This paper specifies two systems that eliminate this waste. The first is Universal Compaction — a formal system for compressing any structured source material into pipe-delimited tables with typed columns, ID-based cross-references, and self-describing grammars, achieving 75-93% compression while preserving every named concept, relationship, and constraint. The second is Grammar-Directed Generation — a system where Prolog grammars provide the structural tokens of output (brackets, punctuation, formatting, boilerplate) while the language model provides only the content tokens (names, values, creative text), reducing the number of forward passes by 40-80% depending on output type. Both systems are built on the VDR-LLM-Prolog architecture: an exact-arithmetic language model where every number is an exact fraction with zero drift (VDR-1 through VDR-4), knowledge is stored in scoped Knowledge Bases with logical provenance (VDR-5), computation is performed by 448 deterministic primitives invoked through command tokens (VDR-6, VDR-8, VDR-10), and structured reasoning is conducted through an orchestrated inference loop (VDR-9). The grammars live on the Knowledge Base struct as a persistent field, inheriting through the KB tree like constraints, and the language model can create new grammars at any time by asserting facts — making the system self-extending. A working Python implementation with 178 passing tests validates the compaction system's roundtrip fidelity, grammar generation, cross-KB usage grammar creation, and grammar inheritance with override shadowing.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper