APRES: Agentic Paper Revision System

• APRES is an LLM-driven framework for agentically revising manuscripts using citation-predictive rubrics.
• It employs a two-stage pipeline with rubric discovery and targeted revision, using negative binomial regression to forecast citations.
• Evaluated on over 26,000 papers, APRES improves revision quality and citation prediction while strictly preserving core scientific content.

APRES (Agentic Paper Revision and Evaluation System) is a LLM-driven scientific writing framework that autonomously discovers citation-predictive rubrics and revises research manuscripts to maximize future impact, as measured by citation forecasts, under strict preservation of scientific content (Zhao et al., 3 Mar 2026). Its architecture, methodologies, and empirical performance position APRES as a paradigm case of agentic, data-driven augmentation of scholarly communication and peer review. The system was developed and evaluated using a dataset of over 26,000 papers from ICLR and NeurIPS, incorporating public reviews and downstream influential citation counts to calibrate and test rubric and revision procedures.

1. System Architecture and Workflow

APRES operates as a two-stage agentic pipeline:

• Rubric Discovery: Employing a “Rubric Proposer” LLM to generate candidate sets of review criteria, the system uses a “Reviewer” LLM to score a corpus of papers on these criteria, producing k-dimensional rubric vectors. A negative-binomial regression then fits actual citation counts as functions of these LLM-generated rubric vectors. An agentic search loop (MultiAIDE) iteratively refines the rubric set, selecting modifications that decrease the regression’s mean absolute error (MAE) on predicting citations.
• Revision and Evaluation: With the rubric fixed, the “Reviewer” LLM scores each draft on the rubric, while a “Rewriter” LLM proposes diff-based, localized edits (constrained to presentation and exposition). Candidate revisions are scored, the best is selected, and further iterations proceed until no increase in the surrogate “impact score” is available, ensuring edits are optimal under the learned rubric and that core scientific results remain unchanged.

This process is guided by an agentic search that branches revision trajectories, evaluates each by the rubric, and traverses a search tree up to a preset depth, enabling effective exploration of revision space while imposing rigorous edit constraints.

2. Automatic Rubric Discovery: Formulation and Optimization

APRES organizes rubric discovery as a supervised learning problem to identify criteria most predictive of future impact. For each paper $p$, citation count $y_p$ is modeled as negative binomial, with mean determined by a linear function of reviewer score vector $x_p$:

$$y_p \sim \text{NegBinomial}(\mu_p, \alpha), \qquad \log \mu_p = w^\top x_p,$$

where $w \in \mathbb{R}^k$ are regression weights and $\alpha > 0$ is the dispersion parameter. Model fitting and rubric refinement are performed to maximize log-likelihood, with predictive MAE on held-out papers as the objective:

$$\text{MAE}(\mathcal{R}) = \frac{1}{|P|} \sum_p |y_p - \mu_p|.$$

An LLM-driven agentic search proposes and scores rubric modifications, continually updating rubric composition to minimize the MAE until convergence, typically achieved within $\sim$200 iterations. This procedure yields rubrics that outperform human reviewer-derived or ad hoc hand-crafted checklists in citation predictivity.

3. Mathematical Formulation of the Revision Algorithm

Given the discovered rubric $\mathcal{R}^*$, a surrogate paper impact score is computed as the mean across rubric items:

$S(T) = \frac{1}{k} \sum_{i=1}^k s_i(T),$

with each $y_p$0 the LLM-derived score on criterion $y_p$1. The main optimization is:

$y_p$2

where $y_p$3 is the original draft and $y_p$4 is the set reachable via allowed local, non-core edits. Iterative revision proceeds as follows:

1. Compute initial impact score $y_p$5.
2. Generate $y_p$6 candidate diffs $y_p$7.
3. Apply each diff to produce candidate $y_p$8 and compute $y_p$9.
4. Select $x_p$0.
5. If $x_p$1, update; else, terminate.

Revisions are strictly localized and forbidden from modifying results, tables, or scientific arguments, thus ensuring content integrity.

4. Evaluation Design and Benchmarking

APRES performance was evaluated on a corpus of 26,707 arXiv-style papers annotated with public reviews and “influential citation” counts. The experimental protocol entailed:

• Citation Prediction: Baseline methods included mean predictor, negative-binomial on human scores, MLP and PCA+NB on SPECTER embeddings, and prompt-breeder search.
• Paper Revision: Papers partitioned by pre-existing review scores (“Clear Accept,” “Borderline,” “Clear Reject”). Metrics included $x_p$2 (change in impact score) and preference in a blinded human expert study (364 pairs, 3 annotators each, binomial significance testing).

Model robustness was estimated using multiple seeds and confidence intervals. APRES was tested using both OpenAI o1/o3 and Gemini 2.5 (Flash/Pro) LLM families.

5. Quantitative Results

Citation Prediction:

APRES with MultiAIDE search realized a 19.6% relative MAE reduction over the next-best baseline for future citation prediction (MAE $x_p$3 1.92 vs. 2.65 for PCA+NB baseline).

Paper Revision:

Mean $x_p$4 (revision-induced rubric score gain) for “Borderline” and “Clear Reject” papers was $x_p$53.0–3.3 units (o3), with lower gain for “Clear Accept” (saturation at 1.67 units). In blinded side-by-side expert assessment, 79% of revised manuscripts were preferred over originals (95% CI [70.1%, 79.0%], $x_p$6).

Summary Table of Core Benchmarks

Task	Best Baseline MAE	APRES MAE	Human Preference for Revised (%)
Citation Prediction	2.65	1.92	—
Revision, All Papers	—	—	79
Revision, Borderline Only	—	—	—

All observed improvements were statistically significant with non-overlapping confidence intervals for MAE and binomial test $x_p$7-values for human preference.

6. Implications, Limitations, and Prospects

APRES establishes that LLMs, equipped with agentic search and data-driven rubric formation, can discover nontrivial, citation-predictive evaluation metrics and instantiate objective-optimized revision workflows. This framework allows manuscripts to be “stress-tested” prior to peer review, effectively raising quantitative proxies for clarity, rigor, novelty, and impact.

Several caveats are noted:

• Citation counts, while measurable and externally validated, are imperfect proxies for “impact” and are subject to Goodhart’s Law when directly optimized.
• Edits are presentation-constrained and do not address non-textual artifacts (figures, code, reproducibility).
• Maintaining security (e.g., avoiding adversarial prompt-injection) remains an open concern for deployment in real reviewing pipelines.
• Human expert discernment remains central—the system is a tool for augmentation rather than automated replacement.

A plausible implication is that as LLM-driven revision technologies mature, human-in-the-loop workflows could integrate APRES as a pre-submission optimization step, bolstering the reliability and communicability of scientific dissemination while preserving the primacy of expert peer judgment (Zhao et al., 3 Mar 2026).