Deterministic Operator Aggregation (UnitBoost)

A defined, non-generative meta-level operator replaces the generative LLM manager in a compound system: a task-given unit map turns each worker's output into slot-value proposals, a constrained argmax assembles the final output order-invariantly, and slots left unfilled or unsupported become an explicit residual that targets the next round of work.

Properties

family: Independence & Ensembling
maturity: frontier
diagram: layers
whyMultiAgent: The operator only has something to assemble because multiple distinct workers each contribute independent slot-value proposals; a single agent produces one output stream with nothing to reconcile, order, or leave as a residual. Distinct from the already-tracked mixture-of-agents and adaptive-mixture-of-agents patterns, where the aggregator is itself a generative LLM call (skipped only when experts already agree): UnitBoost's aggregator is never generative, it is a deterministic, order-free operator with a stated dominance guarantee absent coupling constraints, and it tracks per-slot unit provenance plus an explicit label-free residual signal that redirects the next round at unsupported slots specifically, rather than concentrating selection, allocation, and stopping decisions in one opaque model call.
useCases: [Compound LLM systems where a generative manager concentrates opaque control decisions][Multi-round refinement targeted at label-free residual or unsupported slots rather than full retries][Order-invariant, provenance-tracked aggregation of heterogeneous worker outputs]
source: UnitBoost: Managing Compound LLM Systems with a Merge Operator, Not a Model · arXiv:2609.09815

Provenance

Source: https://arxiv.org/abs/2609.09815

Deterministic Operator Aggregation (UnitBoost)

Properties

family

Independence & Ensembling

maturity

frontier

diagram

layers

whyMultiAgent

The operator only has something to assemble because multiple distinct workers each contribute independent slot-value proposals; a single agent produces one output stream with nothing to reconcile, order, or leave as a residual. Distinct from the already-tracked mixture-of-agents and adaptive-mixture-of-agents patterns, where the aggregator is itself a generative LLM call (skipped only when experts already agree): UnitBoost's aggregator is never generative, it is a deterministic, order-free operator with a stated dominance guarantee absent coupling constraints, and it tracks per-slot unit provenance plus an explicit label-free residual signal that redirects the next round at unsupported slots specifically, rather than concentrating selection, allocation, and stopping decisions in one opaque model call.

useCases

[Compound LLM systems where a generative manager concentrates opaque control decisions][Multi-round refinement targeted at label-free residual or unsupported slots rather than full retries][Order-invariant, provenance-tracked aggregation of heterogeneous worker outputs]

source

UnitBoost: Managing Compound LLM Systems with a Merge Operator, Not a Model · arXiv:2609.09815