Decompose, Don't Expand: Capability-Independent AI Systems by Construction

# Decompose, Don’t Expand:Capability-Independent AI Systems by Construction
Source: [https://arxiv.org/html/2610.38144](https://arxiv.org/html/2610.38144)
October 28, 2026

###### Abstract

A recurring design choice in building AI systems is whether, when a task is too hard for a model, to*expand*the model \(make it larger, better\-prompted, fine\-tuned\), or to*decompose*the task \(extract the parts that admit a deterministic solution to a host, leaving the model only the irreducibly fuzzy residue\)\. We make three claims\. First, that many tasks commonly delegated whole to a model are in fact deterministic tasks with a small generative core\. Second, and our central empirical result, that decomposing such a task yields a measurable and desirable property we call*parity by construction*: on the extracted sub\-tasks the model\-size performance curve becomes flat, because no model performs them at all\. We measure this across nine models spanning a 494M\-parameter local checkpoint to a frontier Opus\-class model, a thirty\-fold parameter range on the local family alone: once the host resolves the context, operation selection scores*identically*for all nine models and selector resolution for eight of nine, while the one task that does not decompose, labelling, remains sloped across the same range\. Because the task leaves the model, a stock checkpoint suffices: no task\-specific training and no fine\-tuning\. Third, that this is a general design discipline, not a domain trick: we show the same move recurring across physics setup, motion retargeting, and object\-model editing, and in a Web\-CLI tool family \(same author, so not an independent replication\)\. We give a method for locating the cut, the construction that makes a cut exist \(a deterministic surface and a closed grammar over it\), and a way to make the residue safe \(resolve the deterministic context; verify the generative output\), position the work against neuro\-symbolic AI, tool\-augmented language models, and generate\-and\-verify approaches, and are explicit about what is measured \(one domain\) versus illustrative \(the rest\)\.

## 1Introduction

When a task is hard for a model, the default move in contemporary practice is to*expand*the model: a larger model\[[19](https://arxiv.org/html/2610.38144#bib.bib23),[30](https://arxiv.org/html/2610.38144#bib.bib24)\], a better prompt\[[31](https://arxiv.org/html/2610.38144#bib.bib20)\], task\-specific fine\-tuning, more examples\. Expansion makes a system’s correctness a function of model capability, with four consequences that are easy to overlook: the system is always one model generation behind; it is non\-deterministic \(the same input can yield different outputs\); it is unverifiable \(one cannot prove the output correct\); and it has no floor \(a cheaper, older, or smaller model simply fails\)\. A concrete instance in this paper’s own domain is 3D\-LLM\[[15](https://arxiv.org/html/2610.38144#bib.bib3)\], which injects 3D into a language model through a trained 3D feature extractor, an added localization mechanism, and over 300k generated 3D\-language examples\. Decomposition has the contrasting property: because the task leaves the model, a stock checkpoint suffices, with no task\-specific training and no fine\-tuning \(Section[3](https://arxiv.org/html/2610.38144#S3)\)\.

The alternative is to*decompose*: to observe that the task is not uniformly fuzzy, but a mixture of parts with a*right answer*\(parsing, lookup, geometry, constraint satisfaction\) and parts that are genuinely ambiguous or generative\. Decomposition extracts the deterministic parts to a host that solves them exactly, and leaves the model only the residue\.

Our contribution is not the idea of decomposition, which is old\[[29](https://arxiv.org/html/2610.38144#bib.bib19)\]and now common in prompting\[[31](https://arxiv.org/html/2610.38144#bib.bib20),[35](https://arxiv.org/html/2610.38144#bib.bib21),[20](https://arxiv.org/html/2610.38144#bib.bib22)\]\(Section[6](https://arxiv.org/html/2610.38144#S6)\)\. It is threefold:

1. 1\.Ameasured property,*parity by construction*: on decomposed sub\-tasks the model\-size curve flattens, because capability has been removed from the critical path rather than improved\. We report this with a per\-task, per\-model evaluation over nine models in one domain \(Section[3](https://arxiv.org/html/2610.38144#S3)\)\.
2. 2\.Amethodfor identifying where a task admits this cut, and for making the remaining residue safe \(Section[4](https://arxiv.org/html/2610.38144#S4)\), together with theconstructionthat makes such a cut exist: a deterministic surface and a closed grammar over it \(Section[2\.1](https://arxiv.org/html/2610.38144#S2.SS1)\)\.
3. 3\.Cross\-domain evidencethat the same design move recurs \(physics from labels, motion retargeting, object\-model editing, and a Web\-CLI tool family; Section[5](https://arxiv.org/html/2610.38144#S5)\), presented honestly as illustrative case studies rather than as additional measured results\. The Web\-CLI instance shares an author and a project with the rest and is not an independent replication\.

The practical stakes are highest where small, on\-device\[[1](https://arxiv.org/html/2610.38144#bib.bib32)\], or sovereign\[[13](https://arxiv.org/html/2610.38144#bib.bib9)\]models are desirable: decomposition is what makes a∼\\sim1 GB 4\-bit\-quantized model\[[7](https://arxiv.org/html/2610.38144#bib.bib6),[5](https://arxiv.org/html/2610.38144#bib.bib31)\]sufficient without task\-specific training, because it shrinks the model’s job until what remains fits even a small model’s reach\.

## 2The principle

#### Capability\-bound tasks\.

Call a task*capability\-bound*if its accuracy rises with model size and plateaus below ceiling; better models help, but the task caps out\. A capability\-bound task is a candidate for decomposition precisely because adding capability yields diminishing returns\[[19](https://arxiv.org/html/2610.38144#bib.bib23)\]\.

#### The deterministic substructure test\.

Ask of a capability\-bound task: is there a deterministic algorithm for some sub\-part? A selector over a labelled graph is resolved by parsing and lookup, as CSS selectors are over a DOM\[[23](https://arxiv.org/html/2610.38144#bib.bib1)\]; a collision body is generated from a class label; a bone correspondence is computed from geometry\. Where such an algorithm exists, the capability it supplies is*free*: a host computes it exactly, every time, at no marginal cost and with a proof of correctness\. One is not removing intelligence from the system; one is sourcing it from determinism instead of from the model\.

#### Parity by construction\.

Letxxbe a task input,HHthe host’s deterministic resolution,MMa model,RRthe constructor of the residual task, andEEthe validator and executor\. For a sub\-task the host fully resolves,

which takes no model argument\. Any two models therefore produce the same output on that sub\-task: not as a measured coincidence but as a structural guarantee, since the host answered before the model was consulted\. We call this*parity by construction*\. For the complete system,

z=E⁡\(H⁡\(x\),M⁡\(R⁡\(x,H⁡\(x\)\)\)\),z=E\\big\(H\(x\),\\,M\(R\(x,H\(x\)\)\)\\big\),\(2\)andzzdepends onMMthrough the residual term\. Parity holds forzzonly whereMM’s contribution is absent or fixed\. Agreement across models on the residual is an empirical question, not a consequence of the construction\. The open question is therefore not whetherHHis model\-independent, which follows by construction, but how much ofzz’s reliability can be moved intoHH, and what building and maintainingHHcosts\.

The guarantee rests on four stated assumptions, none of which we prove: \(A1\) the input representation is valid, soxxcarries the informationHHneeds; \(A2\)HHis correctly implemented for the function it is meant to compute; \(A3\) the host context is complete, soRRdoes not need deterministic information thatHHomits; and \(A4\)EEapplies its verification rules correctly\. Under these, the host\-resolved component is correct with respect toHH\. The end\-to\-end system is not formally verified, and the measurements in Section[3](https://arxiv.org/html/2610.38144#S3)test it only on our fixtures\.

Its signature is a flat model\-size curve on the decomposed sub\-tasks and a sloped curve only on the irreducible residue\.

### 2\.1The construction behind the cut

The test above says where to look for a deterministic part; this subsection says what has to be built for there to be one\. Two halves are needed, and they are duals \(Figure[1](https://arxiv.org/html/2610.38144#S2.F1)\)\. A*deterministic surface*makes an answer*computable*: a host can produce it exactly\. A*DSL*, a closed grammar of operations over the surface, makes an answer*bounded*: the model’s output space is finite and checkable\. Neither works alone\. In Section[3](https://arxiv.org/html/2610.38144#S3), selector resolution could move host\-side only because the graph was addressable, and operation selection could become a lookup only because the verb set was closed\. Without the DSL there is no finite thing to resolve*into*; without the surface there is nothing to resolve*against*\. The cut the method looks for is the seam where the two meet\. This is also why a stock model suffices: given a deterministic surface and a closed grammar, the model is never asked the capability\-bound question, so there is no task\-specific capability left to train into it\.

Figure 1:The construction\. A deterministic surface \(bottom; a small labelled object graph\) makes answers computable\. A DSL over it \(middle\) makes the model’s output bounded\. The three drivers \(top\) enter through the same DSL; they share one model of the scene and one history/undo, which runs beneath the surface\.#### Three testable properties of a surface\.

A domain has a deterministic surface if it passes three tests:

- •Addressable\.Every object has a stable name,*assigned by an author*, not inferred\.
- •Enumerable\.The set of present objects can be listed, so “which parts?” has an exhaustive answer\.
- •Total\.Every operation either applies cleanly or fails cleanly; there are no partial or ambiguous states\. This is what makes verification and undo possible\.

The DSL half asks for a finite verb set with typed arguments\. A DOM with ids, a filesystem, and a relational schema pass; a photograph does not, since it has no names to address, no list of its objects, and no operation that fails cleanly\.

#### Authored labels: the precondition and the cost\.

The first property hides the method’s real price: the labels must be authored\. The system works because someone named\#mayaand\.detail; where labels must be inferred, the task is fuzzy again\. Authoring and maintaining them is part of the cost of the host \(Section[7](https://arxiv.org/html/2610.38144#S7)\)\.

#### Single\-surface parity: a third ingredient, scoped\.

A third ingredient is that the GUI, a shell and an agent drive the*same*model through the*same*operations with*one*undo history \(not to be confused with parity by construction above, which concerns model size\)\. Its necessity depends on the claim\. For*LLM\-drivability*it is optional: a surface plus a DSL is enough\. For an*authorship*claim, that the model supplies speed while the human retains authorship, it is load\-bearing, since it holds only if a person can take over mid\-stream, inspect what the model did, and undo it in the same history; an automation API bolted beside a GUI buys drivability and forfeits this\. We make the authorship claim as a design argument; we have not measured it\.

## 3Measured result: 3D scene editing

We evaluate in a 3D scene\-editing system\[[12](https://arxiv.org/html/2610.38144#bib.bib10)\]in which natural\-language edits are mapped onto a closed set of structured operations over a labelled scene graph \(Figure[2](https://arxiv.org/html/2610.38144#S3.F2)\)\. The task decomposes into selector resolution \(which parts\), operation selection \(which verb\), argument extraction \(which values\), multi\-operation segmentation, and labelling \(naming an unlabelled part\)\. We score at the resolved\-node level over 88 fixtures: 22 each for op\-selection, selector\-resolution and arg\-extraction, 13 for multi\-op and 9 for labelling\.

#### Models\.

Nine stock checkpoints, none fine\-tuned or otherwise trained for this task\. Five local Qwen2\.5\-Coder\-Instruct checkpoints\[[17](https://arxiv.org/html/2610.38144#bib.bib34)\]at 0\.5B, 1\.5B, 3B, 7B and 14B parameters, served by Ollama\[[25](https://arxiv.org/html/2610.38144#bib.bib35)\]at Q4\_K\_M quantization withtemperature=0; and four frontier models accessed through the Anthropic API:claude\-haiku\-5\-5,claude\-sonnet\-5\-5,claude\-opus\-5\-5andclaude\-fable\-5\-1\. The local family alone spans a thirty\-fold parameter range \(494M to 14\.8B\)\.

Figure 2:The two\-way gate\. A natural\-language request passes through a deterministic host, which resolves the selector and operation and hands them to the model as fixed context\. The model supplies the residual arguments and ambiguous choices, and the result is executed as an undoable scene update\. A manual path,$S\(selector\)\.op\(args\), bypasses the model entirely\.
#### Before decomposition \(scaffolded baseline\)\.

With the model asked to compose the whole task, selector resolution is the weakest task at six of the nine model sizes and second\-weakest at the remaining three, behind multi\-op \(Table[1](https://arxiv.org/html/2610.38144#S3.T1)\); the two tasks that trade places are exactly the two the host later takes over\. The model\-size curve is neither flat nor clean: it runs 5%, 55%, 36%, 41%, 45% across the local family, so the 3B, 7B and 14B checkpoints all score*below*the 1\.5B, and the best frontier model reaches only 68%\. The task caps out well below ceiling for every model we measured, which is what makes it capability\-bound in the sense of Section[2](https://arxiv.org/html/2610.38144#S2), and a candidate for decomposition\. The bare condition \(scene only, no selectors or vocabulary; overall scores in Table[2](https://arxiv.org/html/2610.38144#S3.T2), per\-task in the released data\) makes the same point more sharply: selector resolution scores 42% for the 1\.5B local checkpoint against 9%, 18% and 18% for Haiku, Sonnet and Fable, so the small model outscores each frontier model by a factor of 2\.3 to 4\.7\. We report this without a mechanism, and each frontier figure is a single run\.

Table 1:Scaffolded baseline, before host resolution: the model composes the whole task\. Local checkpoints at Q4\_K\_M; frontier models through the Anthropic API\. Where more than one repeat exists the mean is reported; the 1\.5B column is the mean of three repeats and the spread is at most 5 points on any task\.Figure 3:Work removed from the model across the harness ladder: a bare request with the scene, a scaffolded prompt with real selectors and the op vocabulary, constrained JSON\-schema output, and host\-resolved selector and op\. All four rungs send a prompt of comparable size \(10–11\.5k tokens for the local tokenizer\); the rungs differ in what the prompt*contains*, not in how much of it there is\.Table 2:Overall score at each rung of the harness ladder \(Figure[3](https://arxiv.org/html/2610.38144#S3.F3)\), for the four models measured at all four rungs\. The ladder is not monotonic: constrained output scores*below*scaffolded for every one of these models, and improves only on bare\. Host resolution is the only rung that produces a flat curve across models\.
#### After decomposition\.

Moving selector resolution and unambiguous operation selection host\-side, and splitting multi\-operation requests where the split is unambiguous, resolves them from the request text before the model is consulted; ambiguous cases are deferred to the model\. This removes them from the model’s job \(Figure[3](https://arxiv.org/html/2610.38144#S3.F3)traces how the model’s share shrinks\)\. Selector resolution, unambiguous op\-selection and unambiguous multi\-op segments run host\-side; argument extraction, ambiguous choices and labelling stay with the model\. Parity by construction applies to the host\-assigned slots\. The instrumentation records the same counters in*every*host\-resolved run, for all nine models: the host assigned the selector in 31 of 32 emitted operations and the operation in 31 of 32, deferring one genuinely ambiguous case\. The arguments, which the model still produces, are not covered by the construction; their agreement is a measured result\.

Table 3:After host resolution\. Scores on each task, scored on resolved nodes\. Labelling \(†\\dagger\) is not decomposed; it is the residual the model still owns\. Op\-selection is identical for all nine models and selector\-resolution for eight of nine\.
#### The flat line\.

Once the host resolves the context, operation selection scores86% for every one of the nine models, from a 494M local checkpoint to a frontier Opus\-class model, and selector resolution scores 91% for eight of the nine\. These are not scores that agree within noise; they are the same score\. At the fixture level the agreement is equally sharp: comparing each model against the 0\.5B checkpoint, op\-selection and selector\-resolution agree on 22 of 22 fixtures for every model but Haiku, and multi\-op on 13 of 13 for every model but Haiku and Opus \(Table[4](https://arxiv.org/html/2610.38144#S3.T4)\)\. Task performance on the host\-assigned slots no longer depends on which model answers\.

Table 4:Fixture\-level agreement with the 0\.5B checkpoint under host resolution\. Each cell counts fixtures on which the two models reach the same outcome\.
#### The widest single comparison\.

The 0\.5B checkpoint moves from 5% to 91% on selector resolution, 9% to 86% on op\-selection and 0% to 85% on multi\-op, with no change to the model and no training of any kind: the same stock 494M checkpoint, asked a smaller question\. The failure mode is as informative as the score\. Under scaffolding, 51 of its 88 outcomes areparse\_error— the model emits text containing no readable operation at all\. Under host resolution it produces zero parse errors; every emission is well\-formed and 13 are simply wrong\. The small model does not fail the undecomposed task by answering badly, but by being unable to compose a well\-formed operation; once the host supplies the selector and the verb, composition is no longer asked of it\.

#### Cost, not only accuracy\.

Decomposition also shortens what the model must produce\. For the 0\.5B checkpoint, mean completion length falls from 1,385 tokens under scaffolding to 26 under host resolution, and mean per\-call latency from 2,978 ms to 275 ms — a complete run going from 92\.3 s to 11\.3 s despite host resolution issuing*more*model calls \(41 against 31\)\. Where the aim is an on\-device or sovereign deployment, the decomposed path is both more accurate and an order of magnitude cheaper to run\.

#### The honest residue\.

Decomposition has a floor\. Labelling \(naming an unlabelled shape from its descriptors\) does*not*decompose; it is genuinely generative, and it is the only task in Table[3](https://arxiv.org/html/2610.38144#S3.T3)that still moves with model capability, running 67%, 78%, 67%, 78%, 89% across the local family and reaching 100% at Opus and Fable — sloped across exactly the range on which the other four tasks are flat\. Multi\-op segmentation retains residual ambiguity on 2 of 13 fixtures\. These are not failures of the method but its correct boundary: the method extracts the deterministic parts and*stops*at the irreducible core\. The boundary is also consistent with the construction of Section[2\.1](https://arxiv.org/html/2610.38144#S2.SS1): labelling is the task of producing the authored labels the surface depends on, so the method’s boundary is exactly the method’s input \(Figure[4](https://arxiv.org/html/2610.38144#S3.F4)\)\.

Figure 4:Where the labels come from\. Everything downstream of the surface is resolved by the host\. The one task that does not decompose, labelling, is the step that would replace the author: the measured residue \(Table[3](https://arxiv.org/html/2610.38144#S3.T3)\)\.
#### Two models that do not fit the pattern\.

We report two deviations rather than smoothing them\.claude\-haiku\-5\-5is the only model whose host\-resolved scores differ on three tasks \(86% on selector\-resolution, 95% on arg\-extraction, 77% on multi\-op\), and the only one to disagree with the 0\.5B checkpoint on op\-selection\.claude\-opus\-5\-5also scores 77% on multi\-op\. We do not have an account of either; both are reported in Table[3](https://arxiv.org/html/2610.38144#S3.T3)and the per\-fixture data is released\.

#### Data availability\.

All 36 analysed runs, with per\-fixture outcomes, failure kinds and per\-run provenance \(model digest, quantization, server version, decode settings, scorer hash, system commit\), are released aseval\-results\.jsonl; the full tables are inRESULTS\.md\. Both files are arXiv ancillary files, and are mirrored at[https://github\.com/tejaswigowda/strata\-editor/tree/013417d/docs/anc](https://github.com/tejaswigowda/strata-editor/tree/013417d/docs/anc)\(pinned to commit013417d\)\. The call\-level capture \(prompts as sent, raw responses, token usage and latency\), from which the completion\-length and latency figures above are taken, is not part of these files\. Two further runs were excluded and are recorded with their reason: both reported a completed harness loop while the majority of their API calls raised exceptions, and both have clean replacement repeats\. The system under test isstrata\-editorat commits888d97eand6b7cbb7; the scorer issha256:43698355…for every run reported here\.

## 4Method

For any task one is tempted to delegate whole to a model, apply in order:

1. 1\.Find the right\-answer part\.Is there a sub\-task with a deterministic solution \(parse, lookup, geometry, constraint\)? Extract it to the host\. Section[2\.1](https://arxiv.org/html/2610.38144#S2.SS1)gives what the domain must offer for this to be possible\.
2. 2\.Isolate the residue\.What genuinely remains open? That, and only that, is the model’s job\.
3. 3\.Resolve the context \(“give sight, not control”\)\.Hand the model the host\-resolved facts as fixed context so it decides only the minimum\. In the measured domain the host resolves the selector and shows it to the model as fixed context rather than asking the model to produce one\.
4. 4\.Verify the output\.If the residue’s output admits a deterministic check \(valid operation enum; left\-side bones on the left\), apply it: wrong generative answers are caught, not shipped\.
5. 5\.Accept the irreducible\.Where no deterministic substructure remains, let the model do it and let capability show\. Do not force a genuinely generative task into a brittle deterministic rule\.

## 5Generalization

The same move recurs across domains\. We present these as case studies of the design discipline;they are not separately measured parity studies, and we flag each as such\. The first three sit inside one system; the fourth does not\.

#### Physics from labels\(illustrative, not separately measured\)\.

Rigid\-body setup is treated as a deterministic lookup: class labels on scene nodes \(\.static,\.dynamic,\.kinematic,\.trigger\) generate collision bodies host\-side\. “Set up physics” is thereby decomposed into a label table; no model and no per\-instance code performs it\.

#### Motion retargeting with a narrow model\(illustrative, not separately measured\)\.

Retargeting motion capture onto an arbitrary skeleton is deterministic geometry\[[10](https://arxiv.org/html/2610.38144#bib.bib33)\]\(world\-delta transfer with reference\-pose alignment\)\. A model is admitted only for the fuzzy residue \(labelling unknown bones from a closed enum\), and the host verifies each label geometrically before accepting it: decomposition followed by verification, exactly the method of Section[4](https://arxiv.org/html/2610.38144#S4)\.

#### Object\-model editing\(illustrative, not separately measured\)\.

A selector\-and\-operation layer\[[11](https://arxiv.org/html/2610.38144#bib.bib11)\]over a scene graph keeps resolution and the operation set deterministic; a model, when present, supplies only argument values and genuinely ambiguous choices\. The deterministic core is usable with no model at all\.

#### A second instance: the Web\-CLI tool family\(illustrative, not separately measured\)\.

The Web\-CLI family\[[13](https://arxiv.org/html/2610.38144#bib.bib9)\]was built for a different problem,URL in, artifact out: tools areGET\-only, selected byrun=parameters over a set of pure transforms, and report their output through awindow\.\_\_webcli\_resulthandle\. It is the same construction \(Figure[5](https://arxiv.org/html/2610.38144#S5.F5)\): the transforms are a named, deterministic surface; the URL parameters are a closed grammar over it; and a person in a browser, a script and an agent all use the same entry point\. The caveat belongs in the text: this is the same author in the same project, so it is*not*an independent replication\. The honest claim is that the construction*recurred under the same constraint*, which is suggestive of generality, stronger than an argument and weaker than a replication\.

Figure 5:The same construction in a second domain\. Left: Strata\. Right: the Web\-CLI family\. Both have a deterministic surface \(bottom\), a closed grammar over it \(middle\) and several drivers entering through one entry point \(top\)\. Same author and same project: this is not an independent replication\.Across these, the invariant is: the deterministic substructure is extracted and made correct by construction; the model, if present, is confined to the residue and its output is verified\. What varies is only the domain\.

## 6Related work

Our position sits withinneuro\-symbolic AI\[[6](https://arxiv.org/html/2610.38144#bib.bib12)\]: the combination of symbolic and neural components is long\-standing\. We claim neither a new architecture nor the idea of combination; we contribute a measured property \(parity by construction\) and a method for interactive systems\.Tool\-augmented language models\(surveyed in\[[24](https://arxiv.org/html/2610.38144#bib.bib7)\]; program\-aided LMs\[[8](https://arxiv.org/html/2610.38144#bib.bib13),[4](https://arxiv.org/html/2610.38144#bib.bib25)\], Toolformer\[[26](https://arxiv.org/html/2610.38144#bib.bib14)\], ReAct\[[33](https://arxiv.org/html/2610.38144#bib.bib15)\]; visual programming\[[14](https://arxiv.org/html/2610.38144#bib.bib26)\], HuggingGPT\[[28](https://arxiv.org/html/2610.38144#bib.bib27)\], and Chameleon\[[22](https://arxiv.org/html/2610.38144#bib.bib28)\]\) expand the model with callable tools but leave the model orchestrating and deciding; our emphasis is to move the capability\-bound*decision itself*off the model \(the host resolves; the model is never asked\), not merely to equip the model with a tool\.Generate\-and\-verify / LLM\-Moduloapproaches\[[18](https://arxiv.org/html/2610.38144#bib.bib16)\]share our verification step; we add decomposition*before*generation: the deterministic parts are resolved so the model never generates them, and the measured model\-independence that follows\.Constrained decoding and grammars\[[9](https://arxiv.org/html/2610.38144#bib.bib17),[27](https://arxiv.org/html/2610.38144#bib.bib29),[32](https://arxiv.org/html/2610.38144#bib.bib8),[34](https://arxiv.org/html/2610.38144#bib.bib5)\]are a tool we use, not the thesis; we measure it as one rung of the ladder \(Figure[3](https://arxiv.org/html/2610.38144#S3.F3), Table[2](https://arxiv.org/html/2610.38144#S3.T2)\) and find it improves on bare prompting but not on scaffolding, and does not produce the flat curve that host resolution does\.Retrieval and grounding\[[21](https://arxiv.org/html/2610.38144#bib.bib18)\]resolve context for the model, as our step 3 does, but do not characterize the resulting capability\-independence\.

Decomposition in prompting\.Least\-to\-most prompting\[[35](https://arxiv.org/html/2610.38144#bib.bib21)\]and decomposed prompting\[[20](https://arxiv.org/html/2610.38144#bib.bib22)\]have the model produce a decomposition and solve the sub\-problems in turn, and chain\-of\-thought prompting\[[31](https://arxiv.org/html/2610.38144#bib.bib20)\]elicits intermediate steps from the model\. These methods still route the decomposition itself through the model\. Distillation\[[16](https://arxiv.org/html/2610.38144#bib.bib30)\]transfers such capability from a large teacher into a small student; we remove the deterministic sub\-parts from the model’s job instead, so there is no capability to transfer\.

Grounding references in 3D scenes\.Learned grounding of natural\-language references to 3D objects\[[3](https://arxiv.org/html/2610.38144#bib.bib2),[2](https://arxiv.org/html/2610.38144#bib.bib4)\]addresses the problem that selector resolution addresses\. Where the scene graph carries labels, the same resolution can be done deterministically host\-side\.

Injecting 3D into language models\.3D\-LLM\[[15](https://arxiv.org/html/2610.38144#bib.bib3)\]takes 3D point clouds in and produces language out \(captioning, question answering, grounding, dialogue, navigation\)\. It does so by training a 3D feature extractor over multi\-view renders, adding a localization mechanism, and training on over 300k generated 3D\-language examples: the expansion path in this paper’s own domain, with its cost made explicit as a trained pipeline and a generated corpus\. We do not compare the two\. They share no benchmark, task or output, and we report no numbers across them\. They are complementary\. 3D\-LLM’s problem is grounding language in*unlabelled*3D, which is the regime our first surface property excludes \(addressable: names assigned by an author, not inferred\) and where Section[7](https://arxiv.org/html/2610.38144#S7)says the construction does not reach; on an unlabelled point cloud there is nothing to address, enumerate or operate on\. It is also the difficulty that the one task we measure as refusing to decompose, labelling, isolates\. The two approaches meet the same difficulty from opposite sides: one trains a model to produce the labels; the other requires that an author supply them, and measures what that buys\. The method’s identified boundary is the other work’s problem statement\.

The distinction we draw is one of emphasis with measured support: expansion makes correctness a function of capability; decomposition makes the decomposed part independent of it\.

## 7Limitations

Our measured evidence is one domain and a small fixture set \(9 to 22 fixtures per task\); the cross\-domain claims are illustrative, not measured, and the Web\-CLI instance is by the same author in the same project, so the two instances are suggestive rather than independent\. A single additional*measured*domain would move the central claim from “demonstrated once” to “demonstrated across domains” and is the most valuable next step\.

Three properties of the measurement itself should be read with the numbers\.*First*,temperature=0is forwarded for the local models but isnotforwarded on the Anthropic client path for Claude 5 models, so the frontier columns are sampled rather than greedy; this is recorded per run in the released provenance\.*Second*, repeats exist for the 0\.5B and 1\.5B checkpoints and for two frontier conditions, but most frontier cells are a single run; the local repeats vary by at most 5 points on any task, and the host\-resolved local cells do not vary at all\.*Third*, the sweep spans two commits of the system under test \(888d97eand6b7cbb7\); the scorer hash is identical across all runs, so the scoring is comparable, but the harness is not byte\-identical across the whole sweep\.

The method’s hardest judgment, correctly distinguishing deterministic from fuzzy substructure, can fail: a deterministic rule applied where the world is actually variable is brittle\-by\-construction \(a bone\-name table assumes predictable naming and fails on the rig that does not\), which is why steps 4–5 retain a verified fuzzy fallback\. Decomposition is not free: it moves work from the model’s runtime to a host that must be built and maintained, and the price includes the labels themselves, which must be authored \(Section[2\.1](https://arxiv.org/html/2610.38144#S2.SS1)\); the absence of task\-specific training is a cost moved, not removed\. This cost and the floor are the same fact: the one measured task that refuses to decompose is labelling, so the method’s boundary is exactly the method’s input, and the generative residue it leaves remains capability\-bound\.

#### Scope\.

The construction applies to domains with discrete, nameable objects and a finite set of operations over them: scene graphs, the DOM, filesystems, relational schemas, infrastructure manifests, version control\. It does not obviously reach continuous, unlabelled domains such as image synthesis, prose voice or musical phrasing, and we do not claim that it does\. Nor does it reach unlabelled 3D, the setting of 3D\-LLM \(Section[6](https://arxiv.org/html/2610.38144#S6)\): the labels are the input, and producing them is the residue\. The authorship argument for single\-surface parity is a design argument, not a measured result\.

## 8Conclusion

When a model is not good enough for a task, the reflex is to reach for a better model\. Often the better move is to reach for a smaller task\. Decomposition extracts the deterministic substructure, makes it correct by construction, and confines the model to the irreducible residue\. Across nine models spanning a 494M\-parameter local checkpoint to a frontier Opus\-class model, the decomposed sub\-tasks score identically while the one task that does not decompose stays sloped over the same range — so that a stock checkpoint suffices, with no task\-specific training or fine\-tuning, and runs an order of magnitude cheaper\. The principle is old\[[29](https://arxiv.org/html/2610.38144#bib.bib19)\]; the measured property and the method are the contribution, and the cross\-domain recurrence suggests it is a general discipline for building reliable, verifiable, and sovereign AI systems rather than a trick of one domain\.

## References

- \[1\]M\. Abdinet al\.\(2024\)Phi\-3 technical report: a highly capable language model locally on your phone\.arXiv preprint arXiv:2404\.14219\.Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p4.1)\.
- \[2\]P\. Achlioptas, A\. Abdelreheem, F\. Xia, M\. Elhoseiny, and L\. Guibas\(2020\)ReferIt3D: neural listeners for fine\-grained 3D object identification in real\-world scenes\.InEuropean Conference on Computer Vision \(ECCV\),Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p3.1)\.
- \[3\]D\. Z\. Chen, A\. X\. Chang, and M\. Nießner\(2020\)ScanRefer: 3D object localization in RGB\-D scans using natural language\.InEuropean Conference on Computer Vision \(ECCV\),Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p3.1)\.
- \[4\]W\. Chen, X\. Ma, X\. Wang, and W\. W\. Cohen\(2023\)Program of thoughts prompting: disentangling computation from reasoning for numerical reasoning tasks\.Transactions on Machine Learning Research\.Note:arXiv:2211\.12588Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[5\]T\. Dettmers and L\. Zettlemoyer\(2023\)The case for 4\-bit precision: k\-bit inference scaling laws\.InProceedings of the 40th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.202,pp\. 7750–7774\.Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p4.1)\.
- \[6\]A\. d’Avila Garcez and L\. C\. Lamb\(2023\)Neurosymbolic AI: the 3rd wave\.Artificial Intelligence Review56\(11\),pp\. 12387–12406\.Note:arXiv:2012\.05876External Links:[Document](https://dx.doi.org/10.1007/s10462-023-10448-w)Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[7\]E\. Frantar, S\. Ashkboos, T\. Hoefler, and D\. Alistarh\(2023\)GPTQ: accurate post\-training quantization for generative pre\-trained transformers\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p4.1)\.
- \[8\]L\. Gao, A\. Madaan, S\. Zhou, U\. Alon, P\. Liu, Y\. Yang, J\. Callan, and G\. Neubig\(2023\)PAL: program\-aided language models\.InProceedings of the 40th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.202,pp\. 10764–10799\.Note:arXiv:2211\.10435Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[9\]S\. Geng, M\. Josifoski, M\. Peyrard, and R\. West\(2023\)Grammar\-constrained decoding for structured NLP tasks without finetuning\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,Singapore,pp\. 10932–10952\.Note:arXiv:2305\.13971External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.674)Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[10\]M\. Gleicher\(1998\)Retargetting motion to new characters\.InProceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques \(SIGGRAPH\),pp\. 33–42\.External Links:[Document](https://dx.doi.org/10.1145/280814.280820)Cited by:[§5](https://arxiv.org/html/2610.38144#S5.SS0.SSS0.Px2.p1.1)\.
- \[11\]T\. Gowda\(2026\)3DOM: a selector\-and\-op layer for 3d scene graphs\.Note:[https://github\.com/tejaswigowda/3dom](https://github.com/tejaswigowda/3dom)Cited by:[§5](https://arxiv.org/html/2610.38144#S5.SS0.SSS0.Px3.p1.1)\.
- \[12\]T\. Gowda\(2026\)Strata\-editor\.Note:[https://github\.com/tejaswigowda/strata\-editor](https://github.com/tejaswigowda/strata-editor)Cited by:[§3](https://arxiv.org/html/2610.38144#S3.p1.1)\.
- \[13\]T\. Gowda\(2026\)The web\-cli: verifiable privacy for tools, models, and inference engines in the browser\.Note:arXiv:2608\.28950[https://arxiv\.org/abs/2608\.28950](https://arxiv.org/abs/2608.28950)Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p4.1),[§5](https://arxiv.org/html/2610.38144#S5.SS0.SSS0.Px4.p1.1)\.
- \[14\]T\. Gupta and A\. Kembhavi\(2023\)Visual programming: compositional visual reasoning without training\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 14953–14962\.Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[15\]Y\. Hong, H\. Zhen, P\. Chen, S\. Zheng, Y\. Du, Z\. Chen, and C\. Gan\(2023\)3D\-LLM: injecting the 3D world into large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Vol\.36,pp\. 20482–20494\.Note:arXiv:2307\.12981Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p1.1),[§6](https://arxiv.org/html/2610.38144#S6.p4.1)\.
- \[16\]C\. Hsieh, C\. Li, C\. Yeh, H\. Nakhost, Y\. Fujii, A\. Ratner, R\. Krishna, C\. Lee, and T\. Pfister\(2023\)Distilling step\-by\-step\! outperforming larger language models with less training data and smaller model sizes\.InFindings of the Association for Computational Linguistics: ACL 2023,pp\. 8003–8017\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.findings-acl.507)Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p2.1)\.
- \[17\]B\. Hui, J\. Yang, Z\. Cui, J\. Yang, D\. Liu, L\. Zhang, T\. Liu, J\. Zhang, B\. Yu, K\. Lu, K\. Dang, Y\. Fan, Y\. Zhang, A\. Yang, R\. Men, F\. Huang, B\. Zheng, Y\. Miao, S\. Quan, Y\. Feng, X\. Ren, X\. Ren, J\. Zhou, and J\. Lin\(2024\)Qwen2\.5\-coder technical report\.arXiv preprint arXiv:2409\.12186\.Cited by:[§3](https://arxiv.org/html/2610.38144#S3.SS0.SSS0.Px1.p1.1)\.
- \[18\]S\. Kambhampati, K\. Valmeekam, L\. Guan, M\. Verma, K\. Stechly, S\. Bhambri, L\. P\. Saldyt, and A\. B\. Murthy\(2024\)Position: LLMs can’t plan, but can help planning in LLM\-modulo frameworks\.InProceedings of the 41st International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.235,pp\. 22895–22907\.Note:arXiv:2402\.01817Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[19\]J\. Kaplan, S\. McCandlish, T\. Henighan, T\. B\. Brown, B\. Chess, R\. Child, S\. Gray, A\. Radford, J\. Wu, and D\. Amodei\(2020\)Scaling laws for neural language models\.arXiv preprint arXiv:2001\.08361\.Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p1.1),[§2](https://arxiv.org/html/2610.38144#S2.SS0.SSS0.Px1.p1.1)\.
- \[20\]T\. Khot, H\. Trivedi, M\. Finlayson, Y\. Fu, K\. Richardson, P\. Clark, and A\. Sabharwal\(2023\)Decomposed prompting: a modular approach for solving complex tasks\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2210\.02406Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p3.1),[§6](https://arxiv.org/html/2610.38144#S6.p2.1)\.
- \[21\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems 33 \(NeurIPS 2020\),pp\. 9459–9474\.Note:arXiv:2005\.11401Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[22\]P\. Lu, B\. Peng, H\. Cheng, M\. Galley, K\. Chang, Y\. N\. Wu, S\. Zhu, and J\. Gao\(2023\)Chameleon: plug\-and\-play compositional reasoning with large language models\.InAdvances in Neural Information Processing Systems 36 \(NeurIPS 2023\),pp\. 43447–43478\.Note:arXiv:2304\.09842Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[23\]MDN Web Docs\(2024\)Document\.querySelector\(\)\.Note:[https://developer\.mozilla\.org/en\-US/docs/Web/API/Document/querySelector](https://developer.mozilla.org/en-US/docs/Web/API/Document/querySelector)Mozilla Developer NetworkCited by:[§2](https://arxiv.org/html/2610.38144#S2.SS0.SSS0.Px2.p1.1)\.
- \[24\]G\. Mialon, R\. Dessì, M\. Lomeli, C\. Nalmpantis, R\. Pasunuru, R\. Raileanu, B\. Rozière, T\. Schick, J\. Dwivedi\-Yu, A\. Celikyilmaz,et al\.\(2023\)Augmented language models: a survey\.arXiv preprint arXiv:2302\.07842\.Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[25\]Ollama contributors\(2026\)Ollama: get up and running with large language models locally\.Note:[https://github\.com/ollama/ollama](https://github.com/ollama/ollama)Version 0\.40\.6 used for the local models in Section 3Cited by:[§3](https://arxiv.org/html/2610.38144#S3.SS0.SSS0.Px1.p1.1)\.
- \[26\]T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, E\. Hambro, L\. Zettlemoyer, N\. Cancedda, and T\. Scialom\(2023\)Toolformer: language models can teach themselves to use tools\.InAdvances in Neural Information Processing Systems 36 \(NeurIPS 2023\),Note:arXiv:2302\.04761Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[27\]T\. Scholak, N\. Schucher, and D\. Bahdanau\(2021\)PICARD: parsing incrementally for constrained auto\-regressive decoding from language models\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,pp\. 9895–9901\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.emnlp-main.779)Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[28\]Y\. Shen, K\. Song, X\. Tan, D\. Li, W\. Lu, and Y\. Zhuang\(2023\)HuggingGPT: solving AI tasks with ChatGPT and its friends in hugging face\.InAdvances in Neural Information Processing Systems 36 \(NeurIPS 2023\),pp\. 38154–38180\.Note:arXiv:2303\.17580External Links:[Document](https://dx.doi.org/10.52202/075280-1657)Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[29\]H\. A\. Simon\(1962\)The architecture of complexity\.Proceedings of the American Philosophical Society106\(6\),pp\. 467–482\.Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p3.1),[§8](https://arxiv.org/html/2610.38144#S8.p1.1)\.
- \[30\]J\. Wei, Y\. Tay, R\. Bommasani, C\. Raffel, B\. Zoph, S\. Borgeaud, D\. Yogatama, M\. Bosma, D\. Zhou, D\. Metzler, E\. H\. Chi, T\. Hashimoto, O\. Vinyals, P\. Liang, J\. Dean, and W\. Fedus\(2022\)Emergent abilities of large language models\.Transactions on Machine Learning Research\.Note:arXiv:2206\.07682Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p1.1)\.
- \[31\]J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. H\. Chi, Q\. V\. Le, and D\. Zhou\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InAdvances in Neural Information Processing Systems 35 \(NeurIPS 2022\),pp\. 24824–24837\.Note:arXiv:2201\.11903Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p1.1),[§1](https://arxiv.org/html/2610.38144#S1.p3.1),[§6](https://arxiv.org/html/2610.38144#S6.p2.1)\.
- \[32\]B\. T\. Willard and R\. Louf\(2023\)Efficient guided generation for large language models\.arXiv preprint arXiv:2307\.09702\.Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[33\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao\(2023\)ReAct: synergizing reasoning and acting in language models\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2210\.03629Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[34\]P\. Yin and G\. Neubig\(2017\)A syntactic neural model for general\-purpose code generation\.InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics \(ACL\),Cited by:[§6](https://arxiv.org/html/2610.38144#S6.p1.1)\.
- \[35\]D\. Zhou, N\. Schärli, L\. Hou, J\. Wei, N\. Scales, X\. Wang, D\. Schuurmans, C\. Cui, O\. Bousquet, Q\. Le, and E\. Chi\(2023\)Least\-to\-most prompting enables complex reasoning in large language models\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2205\.10625Cited by:[§1](https://arxiv.org/html/2610.38144#S1.p3.1),[§6](https://arxiv.org/html/2610.38144#S6.p2.1)\.