Decomposition Buys Integrity, Not Yield
cs.MA · cs.AI · cs.DC
Quick verdict
Clean theory with unusually honest measurement (δ=0.34, C=0.571); compelling, but rests on one scaffold and an admittedly mis-specified integrity term.
Summary
The paper asks whether decomposing a task across a tree of LLM agents preserves what leaf agents discover. Modeling an agent handed b items as retaining each with probability r(b), it proves a conservation law (r=1/b makes every tree yield exactly one finding) and a factorization Y=C^k·N^(1−δ): architecture contributes only C≤1 per tier, so flat structures are yield-optimal. On 600 production deep-research traces it measures δ=0.34 [0.30,0.38] via three identifications that do not share a failure mode, and reads C=0.571 directly at b=1 over 16,082 hops; a framework-fixed-effects fit on 1,012 MAST-Data traces gives briefing fidelity μ=0.939, so each tier costs Cμ=0.536. Depth is instead justified on two other axes: root-context integrity (exposure falls from N to N^(1/k)) and token cost (flat sessions bill as N^1.392; two tiers overtake flat at 403 findings). A hazard model over 743,819 tool calls finds delegation ignores a filling context (OR 0.969 per doubling) and is an opening move. The model predicts 0.7%–11.3% of sessions worth delegating against 7.8% observed. Main limitations: δ and C come from one scaffold and one tool vocabulary; C is assumed constant across levels despite measured exponent heterogeneity (0.13 vs 0.34); the integrity term is admitted to be mis-specified; μ is identified essentially within one framework.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
6 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under power-law retention r(b)=Cb^-δ, a depth-k decomposition tree yields exactly Y=C^k·N^(1−δ), so architecture cannot raise yield and flat is optimal whenever C≤1 (Theorems 1-3).
- On 600 production deep-research traces, δ=0.34 [0.30,0.38] by three identifications, and C=0.571 [0.527,0.615] is read directly at b=1 over 16,082 hops; with briefing loss μ=0.939, each tier costs Cμ=0.536 of the findings.
- Depth pays only on two other axes: root-context integrity (exposure falls from N to N^(1/k)) and token cost (flat sessions bill as N^1.392; two tiers overtake flat at 403 findings at equal spend).
- The optimal number of tiers is 2-3 at measured production parameters and rises only seven tiers across fourteen orders of magnitude of task size.
- Production systems delegate in ~7.8% of sessions, inside the model's predicted 0.7-11.3% range, but the allocation ignores load: a hazard model shows delegation odds fall slightly (OR 0.969) as context fills.
Strengths
Falsifiable formal core with exact verification
Theorem 1 (conservation at r=1/b) is verified to 2.4e-15 on 20,000 random irregular trees; Theorem 2 to 8.8e-16 over 400 (C,δ,b,k) combinations; the closed-form optimum (Eq. 13) reproduces the argmax of Eq. 12 in 80/80 parameter combinations.
Genuinely triangulated measurement of δ
Three identifications on 600 traces that the authors argue do not share a failure mode: aggregate log-log fit u=0.45b^0.601 (Eq. 6), instantaneous-choice concentration six times above the uniform null at b≈300 (Table 1), and a per-item hazard over 98,768 URLs with stratum intercepts and cluster bootstrap (δ=0.341, Eq. 7), plus a fourth outcome (cited, δ=0.398) landing nearby.
C is observed, not extrapolated
Unusual epistemic honesty
Full reproducibility stack
Sharp connection to prior theory
Weaknesses
Narrow external validity for the headline constants
Functional-form tension inside the paper's own data
The integrity axis named in the title is empirically the weakest
μ is identified essentially by one framework
Yield proxy validity untested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a narrow, well-posed question: when a task is split across a tree of LLM agents, how many of the atomic findings discovered at the leaves survive to the root? The theoretical core (Sections 2-3) models an agent handed b items as retaining each with probability r(b). From this it derives a conservation law (r=1/b makes every tree yield exactly one finding, Theorem 1, verified to 2.4e-15 on 20,000 irregular trees) and a factorization Y=C^k·N^(1−δ) (Theorem 2), implying flat architectures maximize yield since C≤1 (Theorem 3). The empirical program measures δ=0.34 [0.30,0.38] and C=0.571 [0.527,0.615] on production traces, adds briefing fidelity μ=0.939 from annotated multi-agent data, then argues depth pays on two other axes: root-context integrity (exposure falls N to N^(1/k), Section 5) and token cost (flat bills N^1.392; two tiers overtake flat at 403 findings, Section 6). Section 7 closes with a negative result: deployed systems choose topology in the first rounds and never revise it (delegation OR 0.969 per context doubling).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 6 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of 'Decomposition Buys Integrity, Not Yield'?
That splitting a task across a tree of LLM agents cannot increase how many findings survive to the final answer: each extra tier multiplies yield by about 0.54 given the measured constants. Depth is justified instead by protecting the root's persistent context from contamination and by token cost, with an optimal depth of 2-3 tiers.
What do δ and C mean in this paper?
δ≈0.34 is the exponent at which carried-forward findings fall as an agent's context crowds (measured on 600 production research traces three different ways). C≈0.571 is the single-hop retention at batch size 1, i.e., the fraction of findings one summarization tier keeps; combined with briefing fidelity μ=0.939, each tier costs Cμ=0.536.
How many agent tiers should a multi-agent system use?
The paper's model gives 2-3 tiers at measured production parameters, rising slowly with task size (seven tiers across fourteen orders of magnitude). Its design rule: fan out only as wide as summarization stays near-lossless, and add a tier when the token bill, not information loss, is the binding constraint.
Can the paper's results be reproduced?
According to the paper, yes: every number is generated by scripts in anc/code/ from three public corpora fetched by get_data.sh, runs on a laptop CPU, and audit.py re-derives each numeric assertion with a pass/fail verdict.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as extracted captions and axis labels in the provided text; visual data quality was not independently verifiable.
- Ancillary code (anc/code/) was not inspected; reproducibility claims rest on the paper's own description of the scripts.
- Reference [11] is an unpublished preprint; its two corpora could not be independently confirmed as public.
- The metadata abstract is truncated mid-sentence, but the embedded full text contains the complete abstract, so this did not affect the review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.