See your reviewer panel
at the venue you are targeting
before the deadline does.
Drop your paper as a zip, a PDF or the .tex files. PanelSim reads the venue and the field from the sources, works out who would review a paper like yours, and simulates those reviewers and an area chair at that venue: the distribution of scores and decisions, the meta-review, and a ranked plan priced in hours. Then it executes the plan: passages written from your project, figures regenerated from your data, experiments run, citations verified.
Uploads are deleted after 14 days and never train a model.
Used by nearly all top US universities
Built for every science, not only computer science.
A trial is read by clinicians and a biostatistician against CONSORT; a physics paper by an experimentalist and a theorist; a new material by synthetic chemists who open the supporting information first. Each family of venues carries its own rubric, scale and panel size, from a journal recommendation on a 1 to 5 scale to a conference score out of 10, and every reviewer scores every criterion the venue uses.
One upload. Ten reviewers. A plan you can act on tonight.
Drop the paper, in any form
The zip Overleaf exports, the .tex and .bib files, a folder, or just the PDF. PanelSim resolves every included file, expands your macros, links each figure to the script that makes it, reads the bibliography, extracts the claims and runs mechanical checks. It reads the venue off the LaTeX style and the "under review at" lines, and the field and topic off the paper itself.
Simulate the panel it would draw
Ten reviewer types chosen for this paper, each with an expertise and a stance, blended with behavioural personas and with profiles from your citation neighbourhood, write reviews against the venue's rubric. An area chair runs the discussion and writes the meta-review. Thousands of panels are drawn the way the venue draws them and calibrated to its threshold.
Fix what moves the decision
Concerns are merged across reviewers and ranked by expected gain per hour. Toggle edits to see the predicted outcome move, fit the plan to the time you have, then hand the selection back and get the revised project with a patch.
Scores you can read, a plan you can budget, a patch you can apply.
The three panels below replay the case study's own numbers: the ten reviewer scores before and after, the what-if arithmetic behind the plan, and the part of the patch where the revision run added an experiment it ran itself.
Ten reviewers, scored
first reportA plan priced in hours
what-ifThe revision, as a patch
revision run--- a/sections/experiments.tex+++ b/sections/experiments.tex@@ -79,3 +79,35 @@ \paragraph{Cost.} Table~\ref{tab:latency} reports inference latency. Post-hoc filtering runs t… \paragraph{Qualitative behaviour.} Inspecting the summaries, the baseline most often fails by …++\paragraph{Seeds and repeated runs.}+Every reported number is a mean over five independent fine-tuning runs with seeds $\{0, 1, 2, …++\paragraph{Reproducibility.}+The summarizer is BART-large fine-tuned as a fusion-in-decoder model with $k = 5$ passages of …++\begin{table}[t]+\centering+\caption{Main results with the dense retriever over five seeds, mean with one standard deviati…+\label{tab:variance}+\begin{tabular}{llccc}+\toprule+Dataset & Method & ROUGE-L & \faith{} (\%) & Clean (\%) \\+\midrule+\input{panelsim_table_variance}+\bottomrule+\end{tabular}+\end{table}+
Ten reviewers, in their own words.
The panel for the sample paper was chosen for it: natural language processing, faithfulness of retrieval-augmented summarization, so a faithfulness evaluation specialist, a benchmark methodologist, a contrastive learning theorist and seven more, each writing against the venue's rubric. These are their summaries of the first draft, scores included, before the area chair wrote it up as reject.
The idea is clean and the main table moves in the right direction, but every number in the paper is a single point estimate and no experiment attributes the gain to the contrastive term rather than to the extra training signal. As it stands the ordering of methods in Table 1 is not established.
The evaluation has a structural problem: the metric that carries the main claim is computed by the same entailment model the strongest baseline uses to filter its output. Beyond that, the central assumption behind the negatives, that the lowest-ranked passage is unrelated to the source, is asserted rather than measured.
The formal content is light and mostly correct, but one claim in Section 3.2 is stated in the language of a derivation and never derived. The paper does not overreach on theory, and the objective is stated precisely enough to reproduce.
The paper is honest about inference cost and Table 3 makes the case well. It is silent about training cost, which is the cost this method actually adds: a span tagger pass and a second forward pass per example.
Sequence-level contrastive objectives for summarization exist, and the paper cites one of them, so the contribution is the choice of negatives rather than the objective. That is a legitimate contribution, but it is asserted in the introduction and never isolated by an experiment.
The bibliography is accurate as far as it goes, but it is short for the topic and two of its entries are never discussed. The related work names the right neighbours without saying what each one does that this paper does not.
The abstract claims the method outperforms all existing decoding-time interventions; Table 1 shows post-hoc filtering with higher faithfulness on both datasets. The abstract claims robustness to the choice of retriever from two retrievers on one dataset. The paper establishes a modest and interesting result and describes a larger one.
A competent reader could not reproduce these numbers from the paper. The span tagger, the retrieval index, the entailment model, the training schedule and the seeds are all unspecified, and no artifact is promised.
The method reduces a real harm, unsupported claims in summaries that look grounded, and the paper should say what it does not fix: a summary can be faithful to a retrieved passage that is itself wrong. There is no limitations or impact discussion at all.
The paper reads well and the method section is clear. The contribution is stated three different ways in the abstract, the introduction and the results, and one figure caption does not say what the figure shows.
Everything a real panel would say, a week before it says it.
A distribution, not a verdict
Predicted panel mean with an 80% band, acceptance probability with an interval, and where each simulated reviewer lands. Most of the spread is the venue sampling reviewers, and the report says so.
Concerns that point into your paper
Every concern names a section, figure, table or claim, carries a severity, and shows how many reviewers raised it. Absent content is reported as absent, never invented.
A plan you can budget
Edits with author hours and compute hours, the concerns each one resolves, and the expected score gain. Fit two hours or eight and watch the estimate update.
The area chair's meta-review
The reviews are weighed, the disagreements resolved, and the decision written up the way an AC writes it: recommendation, decisive factors, what would change the outcome, and a senior area chair's sign-off. An ethics review is flagged when the content calls for one.
The revision, executed
Selected edits become LaTeX inserted or replaced in the right file and section, tables and figures computed from your data by code that runs, and references verified against Crossref and OpenAlex before they are written. You get the revised project, a unified patch and a predicted after score.
A rebuttal preview, yours to share
Draft responses to the top concerns written for you to adapt, a read-only link for co-authors and advisors, an HTML export, and your reports kept under your account for 14 days.
From likely reject to likely accept in one revision.
The first report put the draft at 1 percent at ICLR: no variance, no ablation, no limitations, and a faithfulness metric shared with the strongest baseline. PanelSim executed 11 of the planned edits into the project, the author finished the 6 measurements only they could make, and the revised paper came back at 79 percent with every remaining concern minor.
Where the panels landed
Two stages, priced separately. No subscription.
The first look costs less than an hour of anyone's time. The second is paid only when you want the work done, and one purchase covers 3 runs on the same report so you can change the selection and run again.
- The sample paper, run on any venue
- The example report and the case study
- Read-only share links for your reports
- Any venue: typed, or detected from your sources
- Ten reviewers chosen for your paper, plus the area chair's meta-review
- Outcome distribution and acceptance probability
- Concerns merged and ranked, plan with hours, what-if, rebuttal preview, share link
- Up to 12 edits executed per run, 3 runs included
- Passages written from your project, in place
- Tables and figures computed from your data
- Citations verified before they are written
- Revised project zip, unified patch, predicted after score
Labs and departments: volume pricing and a shared account are available. Card payments are processed by Stripe; your card details never reach PanelSim.
What people ask before their first run.
How accurate is the outcome estimate?
It is a model of the venue, not a forecast of your particular panel. Scores are calibrated to each venue's scale and threshold, the spread reflects reviewer sampling, and the report states an interval rather than a single number. The useful part is the ordering: which concerns cost the most and which edits buy the most.
Does it write my paper?
It writes only what your project supports. Passages are drafted from your own content, numbers are never invented, code runs against your data files and says when the data is insufficient, and a reference that does not resolve in an open index is discarded rather than written.
What happens to my upload?
It stays on the server that ran the report, tied to your account, and is deleted after 14 days. It is never used to train a model and never shown to anyone without your share link.
Which venues and formats?
Any venue in any science. 131 conferences and journals are in the catalog with their scales, thresholds and rubrics: NeurIPS, ICLR and ACL, but also Nature, Science and PNAS, Physical Review Letters, JACS and Advanced Materials, Cell and eLife, NEJM and The Lancet, the Journal of Fluid Mechanics and the IEEE Transactions, JASA, the American Economic Review, and their neighbours. Type any other name and it gets the rubric of its field. Leave the venue empty and it is read from the LaTeX class (revtex, achemso, elsarticle, aastex, jfm, imsart and the rest), the ACM and IEEE macros, or the "under review at" line. Upload a zip or tar.gz of the project, the .tex and .bib files, a folder, or a PDF alone.
How do you know who would review my paper?
The title, abstract, headings and bibliography give the field, the topic and the kind of contribution. From those, PanelSim describes the reviewers a paper like yours typically draws at that venue the way an assignment list reads: researchers named by their area, seniority and how close their own work is to yours, plus the area chair who would handle it. Each one judges your paper on every criterion the venue scores; their area sets what they can check closely and their reviewing style, taken from a library of twelve, sets what weighs most in their overall score. Where your bibliography allows, a profile from your citation neighbourhood sharpens each of them further. The report lists the whole panel.
How do I pay?
By card, on the pay page, once your paper is parsed and the venue and panel are set: you see what was detected before you pay. Payments are processed by Stripe and the card details never reach PanelSim. One payment per report, one per revision purchase, no subscription.
Can co-authors see the report?
Every report has a read-only share link and an HTML export. The revision run and the what-if stay with the account that ran the report.
What if the run stops?
Each run advances in small steps and resumes from where it stopped. Runs continue on the server whether or not a page is open, and Your runs shows every one of them with its progress.
Know before you submit.
One upload, ten reviewers, a plan you can act on tonight, and a revision run that moved the sample paper from 1% to 79%.