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

• Decomposition assigns right-answer sub-tasks such as parsing, lookup, geometry, and constraint satisfaction to a host, leaving the model only the ambiguous or generative residue.The authors distinguish this from expanding the model through size, prompting, or fine-tuning.Source: §1 · pp. 1–2
• The authors distinguish their contribution from the older idea of decomposition, which is also common in prompting.Source: §1 · p. 2
• The central measured contribution is parity by construction: model-size performance flattens on decomposed sub-tasks because model capability is removed from their critical path.The evaluation covers nine models in one domain.Source: §1 · p. 2
• The method identifies where a deterministic cut exists and makes the residue safe, using a deterministic surface and a closed grammar over it.Source: §1 · p. 2
• Illustrative examples span physics setup, motion retargeting, object-model editing, and a Web-CLI tool family, but are not additional measured results.The Web-CLI instance shares an author and project with the other examples, so it is not an independent replication.Source: §1 · p. 2
• Decomposition is presented as a way to make a roughly 1 GB 4-bit-quantized model sufficient without task-specific training, especially where small on-device or sovereign models are desirable.Source: §1 · p. 2

2 The principle

• A capability-bound task improves with model size but plateaus below ceiling; it is a decomposition candidate when some sub-part has a deterministic algorithm.For example, parsing and lookup can resolve a selector over a labelled graph without asking a model.Source: §2 · p. 2
• For a host-resolved sub-task, any two models produce the same output because the host answers before the model is consulted; residual agreement remains empirical.The complete system still depends on the model wherever its contribution remains in the residual.Source: §2 · pp. 2–3
• The guarantee assumes valid input, correct host implementation, complete host context, and correct verification; the end-to-end system is not formally verified and was tested only on fixtures.Source: §2 · p. 3

2.1 The construction behind the cut

• A deterministic surface makes answers computable, while a closed DSL makes outputs bounded; the cut is the seam where these two halves meet.Neither half works alone, and this construction means the model is not asked the capability-bound question.Source: §2.1 · p. 3
• A deterministic surface must have stable authored names, an enumerable object set, and operations that apply cleanly or fail cleanly; the DSL needs a finite verb set with typed arguments.DOMs with ids, filesystems, and relational schemas pass these tests, unlike photographs.Source: §2.1 · p. 4
• The surface depends on authored labels, so inferring labels makes the task fuzzy again and authoring and maintaining them adds to host cost.Source: §2.1 · p. 4
• Sharing one model of the scene, operations, and undo history across GUI, shell, and agent is optional for LLM-drivability but load-bearing for the authorship claim.The authorship claim is a design argument and has not been measured.Source: Figure 1 · p. 3, §2.1 · p. 4

Evaluation and decomposition

• Across 88 fixtures, the evaluation scores selector resolution, operation selection, argument extraction, multi-operation segmentation, and labelling; nine stock models include a 494M-to-14.8B local family.Host resolution assigns the selector and operation in 31 of 32 emitted operations, deferring one ambiguous case; arguments and labelling remain model-owned.Source: §3 · pp. 4–5

Before decomposition

• Before decomposition, selector resolution was the weakest scaffolded task at six of nine model sizes and second-weakest at the other three.In the bare condition, the 1.5B checkpoint scores 42% on selector resolution, versus 9% for Haiku, 18% for Sonnet, and 18% for Fable.Source: §3 · pp. 4–5

After decomposition

• 86% operation selection across all nine models and 91% selector resolution across eight show flat host-resolved scores; fixture agreement is 22/22 for both tasks except Haiku.Multi-op agreement is 13/13 except for Haiku and Opus.Source: §3 · p. 6, Table 3 · p. 7, Table 4 · p. 7

The widest single comparison

• The 0.5B checkpoint rises from 5% to 91% on selector resolution, 9% to 86% on op-selection, and 0% to 85% on multi-op after host resolution.Its parse errors fall from 51 of 88 scaffolded outcomes to zero under host resolution.Source: §3 · p. 6

Cost, not only accuracy

• For the 0.5B checkpoint, mean completion length falls from 1,385 to 26 tokens and mean per-call latency from 2,978 to 275 ms after decomposition.A complete run falls from 92.3 to 11.3 seconds despite increasing model calls from 31 to 41.Source: §3 · p. 6

The honest residue

• Labelling is the only task that moves with model capability, scoring 67%, 78%, 67%, 78%, and 89% across the local family, then 100% at Opus and Fable; multi-op segmentation retains ambiguity on 2 of 13 fixtures.Haiku is the only model whose host-resolved scores differ on three tasks (86% on selector-resolution, 95% on arg-extraction, 77% on multi-op), and Opus also scores 77% on multi-op.Source: §3 · pp. 7–8

4 Method

• Extract any sub-task with a deterministic solution—such as parsing, lookup, geometry, or constraint solving—to the host.The domain must offer the right-answer part for this cut to be possible.Source: §4 · p. 8
• Keep only what genuinely remains open as the model’s job.Source: §4 · p. 8
• Give the model host-resolved facts as fixed context so it decides only the minimum.In the measured domain, the host resolves the selector rather than asking the model to produce one.Source: §4 · p. 8
• Apply deterministic checks to outputs that admit them, catching invalid generative answers before they are shipped.Examples include checking a valid operation enum or that left-side bones remain on the left.Source: §4 · p. 8
• Leave genuinely generative tasks to the model rather than forcing them into brittle deterministic rules.Source: §4 · p. 8

5 Generalization

• Physics setup uses class labels to generate collision bodies host-side, without a model or per-instance code.This is presented as an illustrative, not separately measured, case study.Source: §5 · p. 8
• Motion retargeting uses deterministic geometry for transfer and alignment, while a model labels unknown bones from a closed enum and the host verifies each label geometrically.This case is illustrative, not separately measured.Source: §5 · pp. 8–9
• Object-model editing keeps selector resolution and the operation set deterministic, leaving argument values and genuinely ambiguous choices to a model when present.The deterministic core can also be used with no model.Source: §5 · p. 9
• The Web-CLI family combines deterministic transforms, a closed URL-parameter grammar, and one entry point for browser, script, and agent drivers.Figure 5 depicts the deterministic surface, grammar, and drivers; the example shares an author and project and is not an independent replication.Source: §5 · p. 9, Figure 5 · p. 9
• Across the cases, the invariant is to extract and make the deterministic substructure correct, while confining any model to a verified residue.Source: §5 · pp. 8–9

6 Related work

• Neuro-symbolic AI has long combined symbolic and neural components; the paper claims neither a new architecture nor that combination as its contribution.It instead claims a measured property and a method for interactive systems.Source: §6 · p. 9
• Tool-augmented language models leave orchestration and decisions to the model, whereas this approach moves capability-bound decisions off-model; generate-and-verify methods verify outputs but do not first remove deterministic generation.The cited approaches include program-aided LMs [8] [4], Toolformer [26], ReAct [33], and LLM-Modulo [18].Source: §6 · p. 10
• Least-to-most prompting [35], decomposed prompting [20], and chain-of-thought prompting [31] still route decomposition through the model, while this approach removes deterministic sub-parts rather than transferring them through distillation [16].Source: §6 · p. 10
• For labelled scene graphs, selector resolution can be deterministic host-side, addressing a problem also studied in learned 3D-reference grounding [3].Source: §6 · p. 10
• The paper does not compare with 3D-LLM [15]: its unlabelled-point-cloud grounding setting falls outside the addressable, author-labelled regime required here.The paper describes the approaches as complementary and reports no cross-work benchmark or numbers.Source: §6 · p. 10
• The distinction is supported as an emphasis with measured support: expansion makes correctness depend on capability, while decomposition makes the decomposed part independent of it.Source: §6 · p. 10

7 Limitations

• The measured evidence covers one domain with 9 to 22 fixtures per task; cross-domain claims are illustrative, and the Web-CLI example is not independent.The authors identify one additional measured domain as the most valuable next step.Source: §7 · p. 10
• Frontier Claude runs are sampled rather than greedy because temperature=0 is not forwarded; most frontier cells are single runs, and the sweep spans two system commits.Local repeats vary by at most 5 points per task, host-resolved local cells do not vary, and the scorer hash is identical although the harness is not byte-identical throughout.Source: §7 · pp. 10–11
• Misclassifying variable structure as deterministic can make rules brittle, so the method retains a verified fuzzy fallback; decomposition also shifts work to a host that must be built and maintained.The costs include authoring labels, and the generative residue remains capability-bound.Source: §7 · p. 11
• The construction is scoped to discrete, nameable objects and finite operations, not claimed for continuous unlabelled domains or unlabelled 3D.The labels are inputs to the method; producing them remains the residue.Source: §7 · p. 11

8 Conclusion

• Across nine models, decomposed subtasks score identically while the task that does not decompose remains sloped over the same range.A stock checkpoint suffices, with no task-specific training or fine-tuning, and runs an order of magnitude cheaper.Source: §8 · p. 11
• Decomposition extracts deterministic substructure, makes it correct by construction, and confines the model to the irreducible residue.Source: §8 · p. 11
• The conclusion presents decomposition as a general discipline for building reliable, verifiable, and sovereign AI systems rather than a trick of one domain.The principle is old; the measured property and method are the contribution.Source: §8 · p. 11