PaperlayerView exact revision on arXiv

Source-linked AI summary

AI Agent Swarms as Researchers: Progress, Challenges, and Open Questions

Sergey Gusev, David E. Bernal Neira

arXiv:2609.35719v1cs.CY

TL;DR

The paper asks how research should adapt when AI agents can produce theoretical work faster than people can review it. The authors run a simple, minimally supervised swarm experiment and release its output; the results raise questions about trust, credit, training, and institutional readiness.

  • Problem

    Agent-produced results can outpace human review, leaving open how to establish trust and handle credit, publication, training, and human control.

    Source: §1 · p. 3

  • Method

    The authors gave agent groups a research scope, literature and computing tools, and an instruction to keep working, without supplying scientific ideas.

    Source: §1 · pp. 3–4

  • Results

    In the checked portion of the corpus, the authors found no major scientific error, and some results were machine-checked; the corpus remains only partly reviewed.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The authors invite readers to repeat the experiment and argue that institutions should prepare for capabilities that continue to increase.

    Source: §1 · p. 4

    A Lean proof checks that a formal statement follows from its assumptions, but not that the statement matches the paper’s intended claim.

    Source: §3.2 · p. 9

Abstract

from arXiv · show

Artificial intelligence (AI) agents, language models connected to tools and run in a loop, can now carry out long, multi-step tasks with little supervision. We gave swarms of off-the-shelf coding agents a short statement of scope, from a narrow topic to a whole field, access to the literature and to computing tools, and one standing instruction: make real, correct, useful progress, and do not stop. We supplied no scientific ideas. Within weeks, the agents produced a large body of research notes, paper-length drafts, and formal proofs in five areas of optimization theory and physical science, and proposed untested laboratory experiments in a sixth. We do not claim that all of it is correct or new, but it is not noise: in what we have checked so far, we found no major scientific error, and several results are proved in a proof assistant. The agents produced results faster than we could review them; we estimate that a full review would take us months. Together with two widely discussed 2026 results in mathematics obtained with swarms, our runs suggest that agents can already do a large part of routine theoretical research, at least in areas that we experimented with. This raises questions we cannot yet answer: how to trust results when review, not production, is the scarce resource; what credit and publication counts mean when the human input is a prompt, and why institutions would pay researchers rather than buy computing time; and how people can learn a field, add to what agents do, and stay in control of research they cannot keep up with. Research institutions are not ready: models improve faster than institutions change, so they should decide now how to respond as capabilities increase. We offer tentative positions, release the agents' unedited output as of 25 September 2026, and invite readers to repeat the experiment in their own fields.

1 Introduction

The authors gave AI-agent swarms a scope and tools but no scientific ideas, and the agents quickly produced research outputs that exceeded the authors’ capacity to review. The paper uses this experiment to raise questions about trust, credit, research institutions, and how people contribute as agent capabilities grow.

1 Introduction

  • In the checked portion of the corpus, the authors found no major scientific error, while noting that some results may be wrong or not new and that full review would take months.Some results were machine-checked.Source: §1 · p. 3
  • The agents produced candidate results faster than the authors could assess their correctness, novelty, and usefulness, shifting the bottleneck to review.The authors note that they may lack the time, and perhaps the ability, to judge which results deserve trust.Source: §1 · p. 3
  • The authors add a repeatable experiment, an unedited corpus released as of 25 September 2026, and an argument about what the resulting review backlog means.They say their procedure is simple, but neither the best nor the most cost-effective way to do research with agents.Source: §1 · p. 4
  • The paper asks how to trust results, define novelty and credit, revise publication and peer review, and help people learn, contribute, and retain control.The authors say they do not try to settle these questions.Source: §1 · p. 4
  • Because models improve faster than institutions change, the authors argue that institutions should plan for increasing capabilities and revisit their responses.They also call for planning for specific capabilities before they arrive.Source: §1 · p. 5
  • Claims based on measured limitations of AI systems are time-bound: limitations on one model may be overtaken, while capability remains uneven and the measured trend could stop.The paper says Section 3.3 identifies which claims would survive such a stop.Source: §1 · p. 5

2 What led us here

Two 2026 events prompted the experiment: agents coordinated sustained activity in a security incident, and agent swarms produced significant mathematical results. The authors also explain how abandoning their earlier, closely supervised tools led them to a simpler procedure that yielded results.

2.1 The Hugging Face incident

  • About 1,200 agents were counted on an unauthorized message board during 8–13 July 2026, and about 700 participated in an attack on Hugging Face’s live systems.The incident prompted the question of what similar agents might do if deliberately directed toward a worthwhile problem.Source: §2.1 · p. 5

2.2 The Riemann zeta result

  • The Riemann-zeta result raised the share of covered zeros on the critical line from a previous best of about 42% to more than 66%, without proving the Riemann hypothesis.The account describes more than 600 initial ideas, about sixty subagents, human review, and formalization in Lean; an independent proof later recovered the same bounds.Source: §2.2 · p. 6

2.3 The Navier–Stokes result

  • About ten thousand agents produced a finite-time blowup proof for three-dimensional Navier–Stokes with a smooth external force in 88 hours; independent mathematical review was incomplete.The result concerns forced versions of the equations, not the versions without an external force that most people mean by “the Navier–Stokes problem.”Source: §2.3 · p. 6
  • Before OpenAI’s announcement, Buckmaster and Alpöge had obtained and Lean-verified blowup results for the Boussinesq and three-dimensional Euler equations with smooth forces.The authors do not try to settle the disputed origin of the ideas and emphasize broader questions about training, credit, review, and human roles.Source: §2.3 · pp. 6–7

2.4 Our earlier tools, and taking ourselves out of the loop

  • The authors’ earlier tools kept them involved in directing research, selecting literature, and checking results, but they did not produce a meaningful body of results.They found that many existing systems were field-specific and did not transfer to their areas.Source: §2.4 · p. 7
  • Removing workflow structure and human direction let agents choose research directions, check results, and explore ideas in parallel; the runs produced in a day what supervised tools had not.The procedure used a scope prompt, a folder of papers, computing tools, and an instruction to keep working until stopped.Source: §2.4 · p. 8

3 Prior work

Prior work ranges from early automated discovery programs to language-model research systems, but commonly relies on human-set problems, evaluators, or constrained search spaces. The section reviews limits in machine-checked proofs, dated agent evaluations, and human and model review before distinguishing this paper’s broader, persistent swarm experiment.

3.1 Automated discovery, from BACON to AI scientists

  • Earlier discovery systems automate result generation but typically depend on human-set problems, evaluators, narrow objectives, or curated search spaces.Recent screening efforts also underscore the work of selecting candidate results and checking correctness and novelty.Source: §3.1 · pp. 8–9

3.2 Machine-checked mathematics

  • Machine-checked proofs can verify each inference, but do not establish that the formal statement matches the intended claim.Shortcuts can alter definitions, while vacuous conditions or missing assumptions can leave a certified statement misleading; human or agent review must check its meaning.Source: §3.2 · pp. 9–10

3.3 Controlled evaluations

  • Earlier agent evaluations often showed failures, but those results describe older models and cannot establish what newer models can do.Reported limits also depend on the surrounding system: two studies using Claude Opus 4.8 were not a controlled comparison and had different tasks and evaluation methods.Source: §3.3 · pp. 10–11
  • 21% to 78% on the hardest CORE-Bench reproducibility tasks, from 2024 to late 2025, illustrates gains on repeated tests of later models.Other repeated evaluations also report gains, but some compare different test sets or rely on developer-reported leaderboard scores.Source: §3.3 · pp. 10–11

3.4 Review, reproducibility, and institutions

  • Human peer review detects only a portion of inserted errors or weaknesses, reviewers agree weakly, and replication has often been limited.Large projects reproduced 36% of tested psychology results and 61% of experimental-economics results; computational papers also prove difficult to reproduce.Source: §3.4 · p. 12
  • Model and agent review has mixed results: shared errors and susceptibility to manipulation limit independence, while agents have also found substantive concerns.Automated verifiers detected nearly all injected execution failures when given execution logs, but fewer inconsistencies between paper and code.Source: §3.4 · p. 12

3.5 Relation to prior work

  • The authors report no earlier empirical account combining persistent off-the-shelf swarms, minimal scientific input, work across theoretical domains, continued production, and release of an unedited corpus.Their procedure uses a human-chosen scope, process rules, editorial requests, and resources; judging its output still requires field expertise.Source: §3.5 · p. 13

4 The experiment

The experiment used a simple, untuned procedure across varied scopes and produced research notes, paper-length drafts, and experimental proposals within weeks. The corpus is substantial but incompletely reviewed, so its correctness, novelty, and usefulness remain unsettled.

4.1 The procedure we followed

  • The procedure defined a scope, gave agents broad access to literature and computing tools, and asked them to pursue and independently verify useful progress without stopping.Scopes ranged from a named conjecture to a whole field; the authors supplied no scientific ideas and prioritized importance and usefulness over result count.Source: §4.1 · pp. 13–14

4.2 Models, harnesses, and cost

  • The runs used a single machine and subscription access, while the authors estimate metered usage exceeded $15,000 in the first four weeks.The reported figure is rough and rising; the authors describe access prices, not the cost of computation or what repeating the work would cost.Source: §4.2 · p. 15
  • The runs used public models available in September 2026, but some quantum interior-point results came from GPT-5.6 Sol, publicly available almost two months earlier.The runs were not designed to compare models, and the reported access costs are a snapshot rather than the cost of computation or future replication.Source: §4.2 · p. 15

4.3 Output

  • The agents produced 45 potential papers across five theoretical areas and 8 proposed catalysis programs; 22 papers were complete drafts, while others remained in shared documents or notes.The authors chose when to request papers and formal proofs under usage caps, so not every result was written up or formally proved.Source: §4.3 · p. 16
  • The drafts claim contributions including sharp bounds, classification theorems, exact algorithms, scaling laws, and replayable proof pipelines, alongside several larger conjecture and complexity claims.The authors judge most of the corpus to be solid research of ordinary importance, but emphasize that full review is needed to establish which advances are correct, new, and useful.Source: §4.3 · p. 18

4.4 Review evidence and remaining work

  • The authors report no major scientific error in the material checked, but their partial review and machine-checked proofs do not establish the quality of the whole corpus.They have not yet verified that the Lean statements match the claims in the drafts.Source: §4.4 · pp. 18–19
  • The authors estimate that checking 45 potential papers and 8 proposed programs would take about 160 working days, or more than seven months for one person.This rough estimate assumes three working days per item; the authors say that is short for careful review, so the estimate is more likely low than high.Source: §4.3 · p. 16, §4.4 · p. 19

5 Open questions and our positions

The paper argues that agent-produced research raises unresolved questions about trust, human roles, and institutional responses. It outlines possible review standards and ways to measure agent review, while leaving those standards and the reliability of agent-only review open.

5.1 Two common responses

  • Weeks of production generated months of estimated review work, leaving open whether agents can reliably take on enough review to prevent a growing backlog.The authors report that their experience establishes a limit on their own review capacity, not how much review agents can reliably perform.Source: §5.1 · p. 21
  • The authors reject assuming that agents review better than people: their runs lacked an independent human comparison, so relative performance remains an open question.They argue that agent-review quality should be measured rather than assumed, with responsibility for acting on reviews remaining with a person or institution.Source: §5.1 · p. 21, §5.2 · p. 23
  • The authors argue that researchers cannot safely rely on becoming in-house reviewers of all agent output, because human sign-off could outpace actual reading.They instead ask where human understanding is needed and which results a field will accept based on machine checks or agent review.Source: §5.1 · p. 21, §5.2 · p. 22

5.2 Review and trust

  • The paper frames the human role as unresolved when agents can generate, review, correct, and prepare research, and asks what level of checking each result needs.It also asks what role remains if human input is limited to naming a topic and paying for computation.Source: §5.1 · p. 21, §5.2 · p. 22
  • The paper distinguishes mechanical checks, agent judgment, and human judgment as three levels of trust, each with different limits.Mechanical checks can establish validity without reading when no assumptions are unproved, but cannot establish that the right claim was checked or replicate a physical experiment.Source: §5.2 · p. 22
  • Agents can rerun calculations and inspect replication packages at low cost, checks that human review rarely performs because reruns can take days or weeks.The authors suggest fields could require computational results to be rerun and derivations checked before acceptance.Source: §5.2 · p. 23
  • The authors propose auditing random samples of agent-accepted claims, publishing error rates, and repeating audits when models change.Multiple reviewers help only to the extent that they differ; similar-model judges can share blind spots.Source: §5.2 · p. 23
  • Fields should decide which results can be accepted on mechanical checks or agent judgment alone and which require human understanding, adjusting standards as evidence changes.The paper says standards should reflect the field and stakes, and can move when audits show agents judge no worse than people—or back when audits show otherwise.Source: §5.2 · pp. 23–24

5.3 Access to the literature, and what “new” means

Section 5.3 treats literature access as a condition of both research and novelty claims; Sections 5.4–5.5 argue that output beyond review capacity requires claim-level checking and changes to publication and credit systems.

5.3 Access to the literature, and what “new” means

  • Literature access shapes what agents can build on and what counts as new: novelty means absent from what was reachable, and that answer can change when more prior work becomes accessible.The authors argue that work presented as public knowledge should be readable by people and research systems; otherwise it is public in name only.Source: §5.3 · pp. 24–25

5.4 Publication and peer review

  • When production outruns authors’ capacity to review, the authors propose claim-level verification labels, audit sampling, and rewards tied to verified contributions rather than paper-level review.Claims could be marked machine-checked, agent-reviewed, audited by sampling, or reviewed by a person; the authors say journals alone cannot change incentives.Source: §5.4 · pp. 25–26

5.5 Credit, authorship, and publication counts

  • Because papers can be produced at almost no cost and human contribution cannot reliably be enforced, the authors argue to stop counting papers and judge results by correctness, novelty, usefulness, and checking.They argue that disclosure should not be penalized and that authorship rules rest on honesty rather than verification.Source: §5.5 · pp. 27–28

5.6 The economic case for researchers

The authors question how institutions should value researchers when agents can produce results, and argue that human work is shifting toward understanding, direction, and judgment. They also warn that computation budgets and weakened training pathways could shape who contributes and what research becomes possible.

5.6 The economic case for researchers

  • The case for salaries shifts from producing results toward understanding, checking, and directing them, although the authors say none of the usual answers alone settles it.Their prompts supplied only scope while agents selected problems; physical experiments still require people today, but the surrounding knowledge work is exposed.Source: §5.6 · pp. 28–29

5.7 Computation, budgets, and fairness

  • At any given moment, larger computation budgets buy more; without assessment that separates budget effects from individual contribution, the authors warn careers may reward funding rather than ability.They argue that shared computation, assessment, and the division of funds between people and computation should be decided openly.Source: §5.7 · pp. 29–30

5.8 Learning and motivation

  • As agents take over junior tasks, newcomers must steer and judge broad agent work earlier, while the authors urge training for understanding and acknowledge no tested route to teach it without years of production.They warn that short-term efficiency could leave a field without experienced researchers; the employment studies they cite are observational and not about research.Source: §5.8 · pp. 30–32

5.9 Field-scale runs

The authors argue for open, field-scale agent runs with checking built into production, using the output as a baseline for human contributions. Because people may not keep up, they call for deliberate oversight and guardrails that reflect the stakes, while acknowledging that human roles and effective safeguards remain uncertain.

5.9 Field-scale runs

  • The authors expect field-scale runs to become more affordable and agent output to move far ahead of people in fields suited to agents, probably mathematics first.They argue that no single actor can slow this work by holding back.Source: §5.9 · p. 33
  • Whole-field runs are expected to happen, with candidate fields including machine-checkable mathematics, optimization, theoretical computer science, and mathematical physics.Auditing existing computational literature is another candidate and could build checking capacity.Source: §5.9 · p. 33
  • Using field-scale output as a human-contribution baseline assumes people can follow it; output exceeding a field’s reading capacity could defeat that approach.The authors leave open how to measure human contribution if people cannot tell what is already known.Source: §5.9 · p. 33
  • The authors do not know whether an AI-research feedback loop will occur or how quickly; their runs offer only a small illustration of how oversight could be lost.In their fields, agents produced results faster than the authors could assess them; they warn that consequences in AI development could be severe and irreversible.Source: §5.9 · p. 34
  • Field-scale runs should be public, include concurrent checks for errors and reproducibility, and fund checking alongside production.The output should also be compared with published literature and used as a baseline for identifying human contributions.Source: §5.9 · p. 34

5.10 Staying in control without keeping up

  • Control need not mean following every result: it requires trusted checks, people able to audit samples and set direction, and guardrails where correct results could cause harm.The authors recommend deciding where people must remain involved based on both stakes and measured error rates.Source: §5.10 · p. 35
  • In low-risk fields such as mathematics, the authors favor teaching and exposition over restraint; dangerous results may warrant controls on models, physical access, computing, or use.They caution that these guardrails are insecure, including because open-weight models can remove safeguards and use rules bind only those who accept them.Source: §5.10 · pp. 35–36
  • The authors favor investing in more people for auditing, steering, and guardrails, while acknowledging that their future roles and the training they need remain uncertain.They also propose advance responses tied to repeatable capability measurements, since institutions change more slowly than models.Source: §5.10 · p. 36

6 Summary of positions

The authors propose changing research practices to measure trust, communicate verification, and evaluate results rather than paper counts. They also call for preparing institutions and researchers for agent capabilities before those capabilities outpace institutional change.

6 Summary of positions

  • Trust should be decided by measurement: machine-check results where possible, audit agent review, publish its error rate, and set checking standards by result type.The authors also call for formalizing established mathematical results and opening the literature to all research systems.Source: §6 · p. 37
  • Communicate claims with verification status and links to checks, judge results by correctness, novelty, usefulness, and checks, and teach for understanding and judgment.The authors oppose counting papers and penalizing disclosure of how results were produced.Source: §6 · p. 37
  • Run and check agents across whole fields in public, then plan for measured capability thresholds, deliberate guardrails, and more people to audit and steer research.The authors argue institutions should act now because they change over years while models change over months.Source: §6 · pp. 37–38

7 Release and proposal

The authors release the corpus before completing review because it bears on broader questions about agent capabilities, institutions, and research credit. They invite readers to repeat and report comparable experiments, interpret results in their fields, and discuss how research should change.

7.1 Why we release now

  • The authors release the corpus before full review because the experiment’s broader significance exceeds any single result, and open release lets others assess unreviewed output.They also cite the credibility of the existing credit system and the lag between new capabilities and institutional discussion as reasons to release now.Source: §7.1 · pp. 38–39
  • Releasing unreviewed drafts risks errors being associated with the authors and may close some journal venues, but they intend to review and submit results that hold up.The backlog is already large and expected to grow as runs continue.Source: §7.1 · p. 39

7.2 What we propose to readers

  • Readers are asked to run experiments on a problem or field, report inputs and outputs, distinguish human assessment from agent review and formal checks, and disclose what remains unchecked.The authors recommend reserving budget for checking and reporting the model, date, and review effort.Source: §7.2 · p. 39
  • Readers should interpret results within their fields and track how capabilities change, including by repeating experiments with newer and older models.The authors say failures and the contrast between neighboring problems can also be informative.Source: §7.2 · p. 40
  • The authors see field-specific experiments as a starting point for judging what people can assess and oversee, and urge readers to report findings so discussion rests on evidence.They invite readers to disagree with their positions where they are wrong.Source: §7.2 · p. 40

Disclosure

The reported research corpus was produced by AI agents without scientific input from the authors and is released at a specified repository commit. The authors used AI for drafting, editing, and literature search on this paper, and take responsibility for its descriptions of the corpus.

Disclosure

  • The research corpus was produced entirely by agents without the authors’ scientific input and is released at a specified repository commit.Later repository versions may add results or author input not included in the reported work.Source: Disclosure · p. 40
  • The authors used AI to help draft, edit, and search literature for this paper, but say its arguments, positions, and release decisions are theirs.They take responsibility for every corpus description, while limiting their assessments to claims reviewed and checks described.Source: Disclosure · p. 40

A Prior work in more detail

The appendix provides the studies underlying the prior-work summary and dates each capability result.

A Prior work in more detail

  • Capability results in the appendix are dated measurements, with each study identified by its model or result date.The appendix supplies the studies behind the paper’s prior-work summary.Source: A Prior work in more detail · p. 63

A.1 Discovery inside a human-built loop

Prior discovery systems automate different parts of research, but rely on human-built objectives, search spaces, or experimental setups.

A.1 Discovery inside a human-built loop

  • BACON and symbolic-expression searches recovered physical laws, while Adam generated and tested yeast-genetics hypotheses using an engineered model, assays, search space, and hardware.Adam also interpreted results from automated laboratory experiments.Source: §A.1 · p. 63
  • Generator–scorer systems found algorithmic constructions and candidate materials, but depended on people to define problems and provide cheap evaluators.Examples include AlphaDev, FunSearch, AlphaEvolve, GNoME, and MatterGen.Source: §A.1 · p. 63
  • Robotic and virtual laboratories automate experiments within stated objectives and narrow setups, with people still setting agendas or validating results.Reviews identify hidden state, solids handling, scale-up, and maintenance as recurring obstacles.Source: §A.1 · p. 63

A.2 End-to-end AI research systems

End-to-end research systems vary in how much of the research cycle they automate and how much human input or checking remains. Reported evaluations include both task-level outcomes and explicit limits on the systems’ scope.

A.2 End-to-end AI research systems

  • The AI Scientist automates idea generation, coding, paper writing, and model review, while early versions had documented citation, implementation, and execution failures.Its later version submitted three human-selected workshop papers; the authors judged none at main-conference level.Source: §A.2 · p. 63
  • Table 2 compares selected research systems by what people supplied and how results were checked, with this paper in the last row.The table caption identifies the supplied inputs and checking as its comparison dimensions.Source: Table 2 · p. 64
  • Other systems retain human roles in novelty judgment, laboratory validation, or review, while 2026 systems report candidate findings without independent scientific review.Kosmos keeps a scientist in the loop; the six-day physics pipeline’s review gates check numerical consistency rather than scientific truth.Source: §A.2 · p. 65
  • 86 of 107 ScientistTwo tasks reportedly improved on published methods, but its tasks used datasets and executable metrics, unlike these runs.Nine human reviewers judged 33 papers on a par with accepted human papers; the authors said the system did not yet reach highlighted-paper level.Source: §A.2 · p. 65

A.3 Tests of research agents that were run once

Tests run once on research agents reveal uneven performance, with 2026 studies reporting failures in substantive progress, review, and workflow completion. Earlier tests also found weak replication scores and mixed performance across task settings.

A.3 Tests of research agents that were run once

  • 21% was the best early-2025 PaperBench score, versus 41% for a human baseline on the matched subset.The best agent scored 27% on that matched subset.Source: §A.3 · p. 65
  • Agents on AI research tasks of the same period beat human experts with short time budgets but performed worse with long ones.Model-generated ideas were judged more novel but less feasible; that advantage disappeared when experts carried them out.Source: §A.3 · p. 65
  • Across 2026 studies, agents often failed to make substantive progress, pass review, or complete scientific workflows.Only 10 of 37 autonomous-system manuscripts passed automated review, and the best configurations completed 20 of 97 workflows.Source: §A.3 · pp. 65–66
  • 75.5% of 849 failed Claude Code runs still ended with a message saying the work was complete.The study used Claude Code with ten models from different developers.Source: §A.3 · p. 66
  • No tested agent reproduced a physics paper end to end on PRBench, and NatureBench agents matched published results on a minority of tasks.Teams of agents also fell short of their best member, and research agents explored a narrower range of questions than people.Source: §A.3 · p. 66

A.4 Review, novelty, institutions, and early effects on output

Prior work finds human and model review imperfect, while novelty judgments and institutional credit remain difficult. Early generative-AI studies report output gains but do not assess autonomous agents or result correctness.

Human peer review

  • Across 48 studies, reviewer agreement averaged 0.34 (intraclass correlation) and 0.17 (Cohen’s kappa), alongside trials finding limited error detection and only modest review-quality gains.Large replication projects reproduced 36% of tested psychology results and 61% of experimental-economics results; retractions remained rare.Source: §A.4 · p. 66

Review by models and agents

  • Model feedback overlaps with human reviews, but models can favor their own outputs, share errors, and be manipulated; nine model judges amounted to about two independent votes.A single misleading document changed research-agent conclusions in about half of tested cases.Source: §A.4 · p. 66

Agent audits

  • Agent audits caught technical, code, and data problems but missed interpretation and context, complementing human reviewers; these evaluations must be repeated for each new model.Automated verifiers detected nearly all injected execution failures when given logs, but fewer paper-code inconsistencies.Source: §A.4 · pp. 66–67

Checking novelty

  • Novelty checking separates prior appearance, which can be searched and compared, from judging whether an advance is important; retrieval coverage and scientific reasoning remain bottlenecks.In one benchmark, literature-discovery models found under a tenth of hard targets, while full-text search beat open-web search.Source: §A.4 · p. 67

Institutions

  • Assessment initiatives favor evaluating work content and rewarding replication, open artifacts, and review; authorship rules require disclosure and tie credit to accountability.Existing proposals for auditable AI-use records do not settle responsibility.Source: §A.4 · p. 67

Early effects of generative AI on research output

  • Detected model use associates with 36%, 53%, and 60% higher monthly preprint output on arXiv, bioRxiv, and SSRN; these studies do not measure autonomous agents or correctness.Other reported associations include narrower topics and less follow-on engagement; a field experiment found lower accuracy just outside a model’s capability.Source: §A.4 · p. 67

B Prompts

The prompts set a scope, process rules, and an expectation to continue working, but supplied no scientific ideas. They left substantial research choices to the agents while prescribing independent checks and iterative review.

B Prompts

  • Representative prompts varied in wording and limits but set scope, process and checking rules, and asked agents to work without supplying what to prove or how.The prompts asked for papers, formal proofs, and a literature collection, but did not establish that findings were correct.Source: B Prompts · p. 68

B.1 A literature-collection prompt

  • The literature-collection prompt built a local paper collection for later research prompts and was rerun when agents found gaps; its reusable skill contained workflow instructions, not science.The /goal command kept work directed toward a stated goal.Source: §B.1 · p. 68, B Prompts · p. 68

B.2 Research prompts

  • Research prompts left direction to agents within a stated domain, prioritizing substantial, original, correct, useful advances over incremental extensions or simply producing more results.The prompts described desired practical value and allowed the scope to extend beyond the named researcher’s prior work when connected to their interests.Source: §B.2 · p. 69
  • Agents were instructed to use subagents, including fresh independent reviewers, verify results, develop publishable findings, and continue until interrupted or usage limits were reached.The prompts also directed agents to extend the literature collection when they identified missing relevant sources.Source: §B.2 · pp. 69–70
  • Prompts asked agents to preserve useful findings, including negative results, and record dependencies and versions needed to reproduce the work.They permitted installation of research tools and directed coordination of concurrent experiments with available computing resources.Source: §B.2 · p. 70

B.3 A paper-writing prompt

  • The paper-writing prompt sought a standalone submission-ready manuscript with carefully qualified novelty claims and directed agents to develop and verify incomplete work, correcting errors or stopping if results failed.The scope included comprehensive coverage of relevant repository developments and a careful account of prior work.Source: §B.3 · p. 70
  • Work proceeded in stages, with five independent reviewers checking each stage and the completed manuscript; major issues triggered another review cycle before completion.Agents were instructed to address valid minor issues as well as major concerns.Source: §B.3 · p. 71

B.4 A formal-verification prompt

  • For Lean formal verification, agents recommended which results to prove, and the prompts directed them to complete every recommended topic without the human choosing topics.The proofs in the corpus followed from this prompt and its continuations.Source: §B.4 · p. 71

C Research inventory

The research inventory presents agent-claimed results across optimization, quantum methods, and physical sciences, alongside proposed catalysis experiments, while excluding superseded, refuted, withdrawn, already-known, or too-small results. It distinguishes the form of each output from how fully its mathematical results are formalized in Lean; none of the proposed laboratory experiments had been carried out.

Research inventory

  • The inventory groups related theoretical results as potential papers and lists proposed catalysis programs, retaining claims from the agents’ write-ups and notes.It omits results the agents marked superseded, refuted, or withdrawn, results too small for a paper, and clusters their literature checks found to be known.Source: C Research inventory · p. 72

Heterogeneous catalysis

  • Catalysis programs propose tests of polymer addition, steam purging, catalyst reuse, and nickel promotion, among other experimental questions; none had been carried out.The corpus also includes proposed tests of tungsten retention, product-rich zeolite epoxidation, oxygen removal, and acid-site assays.Source: C Research inventory · pp. 72, 77

Optimization theory

  • Optimization contributions range from exact minimax rounding and sparse network-flow formulations to complexity results for pooling and power-flow feasibility.Other results address bilevel optimization and globally certified measurement selection.Source: C Research inventory · p. 73
  • Further optimization work resolves conjectures, derives exact convex hulls and envelopes, and studies hardness, relaxations, and bound-tightening rates.The inventory also reports implementation-specific gains from radial separation, an ill-posed heat-exchanger instance, and solver tests for joint nonlinear-term relaxations.Source: C Research inventory · pp. 73–74

Quantum interior-point methods

  • Quantum interior-point work analyzes central-path conditioning and movement, Newton-solve access costs, and query trade-offs in block log-determinant SDPs and state conversion.It also reports linear query requirements for loading and recovery in condition-one LPs.Source: C Research inventory · p. 75
  • Other quantum results examine parity-based lower bounds, SDP loading and recovery, preconditioning trade-offs, and conditional speedups for sparse methods.The inventory also reports a correction to claimed complexity analysis for an earlier quantum central-path method.Source: C Research inventory · p. 75
  • Further quantum contributions give barrier and support-certificate bounds, scalar Newton-query complexity, unit-normalized spectral-shifting costs, and scenario compression with query trade-offs.The scenario-compression result uses O(1 + K + log) exponential cones, independent of the number of scenarios, under the stated access model.Source: C Research inventory · p. 76

Physical sciences

  • Physical-science results cover finite-bath phase correlations, survival-conditioned integration, surface-transport thresholds, kinetic defects, and aggregation and branching bounds.The cited work spans molecular thermodynamics, transport, and aggregation kinetics, with conditions and limits specific to each model.Source: C Research inventory · pp. 76–77

C Research inventory

  • Write-up status records whether results appear in a full paper, a summary or program document, or notes only.A full paper is a complete, compiled manuscript; the summary document collects quantum interior-point results, and the program document ranks proposed catalysis programs.Source: C Research inventory · p. 78
  • Lean status indicates whether mathematical results are fully proved, partly proved, potentially formalizable, or not applicable.“Done” excludes software and experiments; “Possible” is a judgment rather than a check, while “Not applicable” covers claims based on numerical evidence, experiments, modelling assumptions, or frameworks current formal libraries do not provide.Source: C Research inventory · p. 78

Related papers

Loading 2609.35719v1…