Over the last decade, Chinese non-financial companies more than doubled the share of their assets held in financial instruments, from roughly 16% in 2011 to over 37% in 2022 (Digital Transformation of Enterprises and Corporate Financialization). Not all of that shift is bad. A firm can hold liquid reserves for good reasons. Some of it, though, is what one study calls "effort without reward": heirs of family firms pile up speculative financial assets without earning more financial income from them (The intergenerational succession and financialization of Chinese family enterprises). Telling the two cases apart is a judgment call. It requires reading a firm's disclosures, governance rules and operating situation together. That is the kind of work an academic editorial board is built for: several specialist reviewers each look at one aspect, an associate editor reconciles their views, and an editor decides.
This proposal asks a narrow question that matters to both fields. When language-model reviewers are organized like an editorial board, does the middle tier, the associate editor, actually improve decisions? Or would averaging the same number of independent reviews do just as well? We test this first on scientific peer review, where real decisions provide ground truth. We then test it on a new task built from public firm data: predicting which increases in a firm's financial assets will fail to pay off. Editorial structure supplies the mechanism under test. Financialization supplies a high-stakes setting where the correct label becomes known later and does not depend on anyone's opinion.
Core Claim
We claim that the value of an editorial hierarchy comes from targeted re-examination, not aggregation. The associate editor reads the reviews, finds the specific point on which reviewers disagree, and sends a focused check back to the source material. A flat ensemble cannot do this, because no member ever sees the others' reasoning.
If the claim holds, two things follow. A hierarchy should beat a flat ensemble that uses the same compute. Most of that advantage should disappear when the associate editor is limited to summarizing the reviews, with no re-query step. If the claim fails, the editorial metaphor is decoration, and the right advice is "ensemble and calibrate."
This is a different contribution from the closest existing work. Self-Resource Allocation in Multi-Agent LLM Systems and Embodied LLM Agents Learn to Cooperate in Organized Teams show that hierarchical teams of language models finish tasks faster in cooking and household simulations. They measure efficiency at doing things, not accuracy at judging things, and they do not separate out which part of the hierarchy helps. The corporate-finance papers that motivate the application treat governance as a single yes/no variable. The impact of subscribing to directors' and officers' liability insurance on corporate financialization is one example: it records whether the firm buys liability insurance for its executives, or whether one person is both CEO and chairman. None of these papers models the review process through which a capital allocation is vetted.
Research Program
The first component is a mechanism study on ReviewArena. The full corpus has 51,529 papers with 196,099 reviews, meta-reviews and accept/reject decisions; we use its 1,002-paper evaluation split. ReviewArena found that model reviewers compress their scores into the 4–7 band and separate accepted from rejected papers by only about 0.5 points on average.
We first note a measurement issue. A gap measured in raw points can be widened just by stretching the score scale. We therefore use the area under the ROC curve (AUC) for accept versus reject as the primary metric. AUC depends only on how papers are ranked, so it ignores cosmetic rescaling.
Before building anything, we measure how much the human middle tier matters in ReviewArena itself. We count how often the final decision departs from what the mean reviewer score predicts, and whether those departures align with issues raised in meta-reviews. If human area chairs rarely override the average, that is evidence against our claim from real editorial boards, and we will report it as such.
The system has three tiers:
- Reviewers. Each reviewer agent sees the full submission but is assigned one focus. Each returns a score plus a short list of specific claims, each tied to a quoted passage.
- Associate editor. This agent sees all reviews. It identifies the largest substantive disagreement and issues at most two targeted queries back to the document, each answered by a fresh call.
- Editor. This agent makes the final ranking. It ranks submissions within a batch from the same venue and year, mirroring how an area chair calibrates across a pile of papers.
The second component transfers the same system to a financialization task. We build firm-year review dossiers for Chinese A-share non-financial firms. Each dossier combines the firm's annual-report discussion of investment and treasury activity with governance fields from CSMAR, the China Stock Market & Accounting Research database. Those fields include board composition, whether the CEO is also chairman, state ownership, and liability-insurance purchase from the Chinese Research Data Services database. Each dossier also includes the firm's digital transformation index, the measure used in the Li study above.
The label follows the asset-versus-revenue distinction from the family-firm study. A firm-year is positive if two conditions hold:
- The firm's share of assets held in financial instruments rises by more than a threshold over the next two years.
- Its income from financial assets, as a share of revenue, does not rise.
In other words, a positive is financialization that was not rewarded. The label is computed only from later balance sheets, so no human judgment enters it.
Each reviewer focus maps onto one driver documented in the literature:
- Governance and moral hazard. Does insulation from personal liability, or weak separation between the CEO and the board, make speculation more likely? This lens draws on the liability-insurance findings.
- Operating need. Is there real overcapacity, as in How does industrial intelligence affect capacity utilization?? Or is there a pending digital investment burden that makes holding reserves sensible?
- Charter and disclosure consistency. Are stated investment purposes consistent with the firm's governing documents and its other filings? This lens relies on the multi-step rule tracing that CHANCERY shows models handle poorly.
We split the data by time. Dossiers from 2011–2017 are used for prompt development and from 2018 onward for testing. We remove firm names and tickers from all text, because language models may remember which firms later did well.
Baselines
The naive reference point is a single language-model call that sees the whole dossier and outputs a probability. On ReviewArena, the equivalent is one model scoring each paper. This shows what structure adds over simply asking.
The strongest competitor on the financialization task is not a language model. It is a gradient-boosted classifier trained on the structured CSMAR panel, using the drivers the finance literature has already validated: liability insurance, CEO duality, state ownership, succession events, the lagged digital transformation index, capacity utilization, and past financial-asset share. These are the variables behind the published coefficients. The Peng et al. study reports a coefficient of 0.005 for liability insurance, and the Li study reports 0.0329 at a one-period lag and 0.0470 at a two-period lag for digital transformation. Those are effect sizes, not prediction scores. We found no reported predictive result on Lacuna for out-of-sample classification of unrewarded financialization, so we will establish this number ourselves. The classifier gets the same temporal split and a hyperparameter search budget comparable to the prompt-development budget the language-model systems receive. The editorial system beats it only if one of two things holds on the held-out years: it has higher AUC on its own, or stacking its output onto the tabular model raises AUC significantly under a paired bootstrap. The stacking test is the realistic one, since a deployed auditor would use both.
On ReviewArena, the strongest reported single-model results are a mean absolute error of 0.700 against human scores (Qwen3.6-Plus) and an accept/reject gap of 1.40 points (GLM-5.1), both from ReviewArena. AUC is not reported there, so we recompute it for the strongest model we can run.
On governance reasoning, CHANCERY reports 78.1% accuracy for an agent that writes and executes code (CodeAct), against 75.2% for GPT-4o. It also reports a roughly 15% drop on questions that require combining several charter clauses. We will run the charter-consistency reviewer on CHANCERY's 502 questions as a sanity check. The reviewer should at least match CodeAct before we trust it inside the hierarchy.
The decisive ablations isolate the key ingredient. The first is a flat ensemble: the same reviewers with the same focuses, scores averaged, and extra calls spent on additional sampled reviews so that total model calls and tokens match the full hierarchy. The second keeps the hierarchy but turns the associate editor into a pure summarizer with no targeted re-queries. The third removes the batch-ranking step and lets the editor score each submission in isolation. All systems share one base model and one prompt-development budget measured in development-set evaluations.
Why not just…?
Why not just train a tabular model on the CSMAR variables and skip language models entirely? This is the most serious threat. The known drivers may already capture most of the predictable signal, and annual-report text may add little. We do not argue around this. The stacking test in the baselines settles it. If the editorial system adds nothing over the tabular model on held-out years, the financialization application fails. The ReviewArena mechanism result would still stand on its own.
Why not just average more independent reviews and rank-normalize the scores? Rank normalization removes score compression for free, and that is why we evaluate on AUC, where it gives no advantage. The open question is whether the associate editor's re-queries improve the ranking itself. The compute-matched flat ensemble answers this directly. If it ties the hierarchy, we will conclude that editorial structure is unnecessary for automated review.
Why not just give one strong reasoning model the whole dossier and a long chain of thought? CHANCERY is a warning here. Reasoning-tuned models underperformed GPT-4o on governance questions: 58.6% for DeepSeek-R1 and 64.5% for Claude 3.7-Sonnet. DeepSeek-R1 scored just 27.6% on anti-greenmail provisions. That is suggestive, not conclusive for our task. A single long-context call with self-consistency sampling, given the same token budget as the hierarchy, is included as a baseline, so the experiment measures this directly rather than assuming it.
Decisive Checks
The mechanism claim is supported only if both of the following hold on ReviewArena-Eval and on the financialization test years:
- The full hierarchy's AUC exceeds the compute-matched flat ensemble by a margin outside the paired-bootstrap interval.
- The summarizer-only associate editor recovers less than half of that margin.
Three further checks make the result interpretable. First, on ReviewArena, the cases where the system's associate editor changes the outcome should overlap with the cases where human decisions depart from the mean reviewer score more than chance would predict. Second, on the financialization task, gains should be largest for firm-years where the tabular drivers conflict, such as a firm that buys liability insurance but whose CEO is not also chairman. Disagreement among the drivers is where adjudication should matter. Third, performance on post-cutoff, name-stripped years should not fall sharply below performance on earlier years. A sharp fall would indicate the model is recalling outcomes rather than judging dossiers.
A negative result on any of these checks is itself useful. It would tell both organizational designers and builders of automated reviewers which part of the editorial hierarchy is worth paying for.
References
- Digital Transformation of Enterprises and Corporate Financialization
- The intergenerational succession and financialization of Chinese family enterprises: Considering the influence of heirs’ growing experience
- Self-Resource Allocation in Multi-Agent LLM Systems
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- The impact of subscribing to directors’ and officers’ liability insurance on corporate financialization: Evidence from China
- ReviewArena: A Large-Scale Cross-Conference Dataset and Benchmark for LLM Peer Review
- How does industrial intelligence affect capacity utilization?—Analysis based on green development perspective
- CHANCERY: Evaluating Corporate Governance Reasoning Capabilities in Language Models