Research Paper
KVRM is a finite-registry neural selection pattern. Instead of asking a model to generate arbitrary outputs, it asks the model to choose from a fixed, audited registry of valid options.
With a fixed registry, the selector can only emit an in-registry key. That is the strongest claim and the part that is architectural rather than empirical.
Picking the right registry key is learned behavior. It depends on the workload family, backend, registry contract, training data, and evaluation protocol.
The deployment claim here is narrower: the interface can be made type-safe and fail-closed, and then semantic performance can be measured honestly.
The strongest open-world result is the contract guard: on pooled larger-bank MPS and CPU runs, it achieved 0.0% supported false rejectsand 100.0% unsupported rejects.
The strongest deployment caveat is registry drift: if runtime registry order changes without stable keys or version guards, structural validity can remain intact while semantic correctness collapses. In the active drift artifact, unguarded reorder fell to 4.7% correctness, while stable-key decoding restored the original behavior.
Within the supported contract, KVRM clearly beats a cold heuristic on node-level boundary decisions, but local transition sensitivity is still weak on Apple MPS and only modest on CPU. The current evidence supports operating-point tradeoffs, not a universal robustness win.
The repo surface is cleaner now too: the active selection evidence, qualification docs, and paper draft sit in the foreground, while older GPU and KV-compression benchmark helpers are explicitly published as historical context rather than current headline support.
XGrammar and KVRM solve different problems. XGrammar is a constrained decoding runtime around a base language model. KVRM is a purpose-built finite selector architecture.
Very fast finite audited selection. On the active measured-oracle MPS artifact with SmolLM, KVRM beat zero-shot XGrammar on the same workloads by 33 wins to 0.
General structured generation, dynamic schemas, and expressive output formats. It is the better tool when the task is JSON or grammar-constrained text generation rather than finite system dispatch.
The honest interpretation is narrow: KVRM clearly beats zero-shot XGrammar on this finite selection benchmark, but stronger non-neural baselines like retrieval remain competitive and are often not statistically separable at the current sample sizes.
KVRM is worth using when you need a finite audited selector with explicit interface safety. It is not a general-purpose claim about universal robustness or “non-hallucinogenic” AI.