Beyond the Parameter Monolith: Reconstructive Memories, Executable Skills, and Residual Assembly for Language Models

Beyond the Parameter Monolith. Reconstructive Memories, Executable Skills, and Residual Assembly for Language Models. By A. Bochkov. Published on the second of October. No code repository is linked to it.

The starting position is that a language-model system does not have to keep everything in one shared set of parameters. Contextual computation, persistent storage, and exact execution can be separated, rather than all capabilities being updated through a single parameter system. The paper investigates one such organization, called FEM-ASM, inspired by the finite element method. Independently constructed document states, and deterministic executable skills, each contribute typed proposals to a shared language-model state. Where proposals attach to common interface nodes, an explicit residual operator reconciles them.

The author evaluates this through controlled experiments and openly reported negative results, and is explicit that the work is not claiming a physical finite-element formulation of language. The pieces are tested one by one. An attention-free Multi-Mesh prototype learns causal language modeling, but does not establish competitive general capability. A versioned store holds fifty-two thousand eight hundred and nine reconstructive memory elements, near a budget that runs past a billion floating values.

Reconstruction from that store is incomplete, at roughly seventy-five percent token accuracy. Support-aware lexical indices make the elements addressable, under query construction that controls for provenance. For executable arithmetic, positional result observations substantially improve neural rendering compared with a repeated global result vector, and substituting the output changes which answer the model prefers. A bounded attachment demonstration measures what happens when selected evidence is made available.

The conclusions are deliberately narrow. The results support separating storage, execution and neural coordination as an organization. They do not show that the prototype is competitive in general. The attachment demonstration does not establish the value of loading an entire store of several billion values. And the author names three limitations that remain open: question-only retrieval, unrestricted answer generation, and end-to-end efficiency.

This reads less as a new model than as a careful test of an architectural stance. Memory, skills and coordination are pulled apart and checked separately, with the failures reported alongside the successes, so a listener comes away knowing which parts of the separation held up and which are still unproven.