An Exact Generate - Transform Decomposition of Small-LLM Team Scaling Across Orchestration Architectures

URL SCAN: An Exact Generate - Transform Decomposition of Small-LLM Team Scaling Across Orchestration Architectures
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

This paper is an instrumentation exercise for AI production. It replaces the lazy slogan “more agents are better” with a measurable split: additional calls expand proposal coverage, while a downstream transform determines whether those candidates become correct answers.

Its result is conditional scaling. Arithmetic gains sharply from added calls; multiple-choice tasks mostly saturate or fail to convert extra coverage; code gains track coverage because generative recovery is weak. Proposer-Critic wins at high budgets on arithmetic but is weak elsewhere. The paper is mapping the production curve of model collectives, not defending human labor.

The Core Fallacy

The technical decomposition is useful. The systemic limitation is treating benchmark accuracy and call budgets as if they were the decisive economic units. They are not. The abstract does not test labor substitution, deployment reliability, integration, latency, energy, oversight, long-horizon work, or whether humans remain economically necessary.

“No architecture wins across tasks” is not a human moat. The Discontinuity Thesis does not require one universal architecture. It requires sufficiently cheap machine performance across enough economically important cognitive work. Task routing, portfolio orchestration, larger models, tools, and future transforms can absorb architecture-specific weaknesses.

The word “exact” describes the accounting decomposition, not a causal proof that the identified components are the sole drivers of real-world performance.

Hidden Assumptions

• Benchmark accuracy is a useful proxy for productive economic value.
• Five short-answer benchmarks and one code benchmark represent cognitive work broadly enough to support scaling conclusions.
• Seven-to-nine-billion-parameter models and a thirty-call ceiling approximate the relevant frontier.
• Call budgets are meaningful cost comparisons despite token costs varying by 2.1x.
• Architectures can be selected per task without material routing, switching, or integration costs.
• Static, bounded tasks generalize to messy, persistent, tool-mediated production environments.
• Human verification and operational supervision are either absent or economically negligible.
• Gains in proposal coverage and transformation will not improve substantially with better models, tools, memory, or domain-specific systems.

Social Function

Classification: partial truth, transition management, and prestige signaling.

The paper punctures crude scaling hype, which is scientifically valuable. But it also relocates the debate into an engineering sandbox: architectures, budgets, benchmarks, and token efficiency. That framing makes AI appear to be an optimization problem rather than an ownership and displacement mechanism. It turns the approaching labor shock into a leaderboard problem.

The Verdict

At the narrow technical level, this is a useful result: team scaling is task- and architecture-dependent, not a uniform lever. Under Discontinuity Thesis logic, it neither proves nor disproves P1–P3. The abstract is too limited to establish durable cognitive dominance or productive-participation collapse.

It does, however, expose a real mechanism compatible with P1: cheap repeated inference creates candidate coverage, and specialized transforms convert that coverage into usable output. The failure of one architecture on one benchmark is engineering friction, not evidence that humans retain a permanent economic role.

This paper is not evidence that the worker survives. It is a small map of the machinery being refined to replace the worker.