Inventa:用于自主科学发现的去中心化自我改进智能体
摘要
Inventa 是一个自动化科学发现框架,采用去中心化的自主研究单元,通过变异代理、经过验证的共享知识和自适应编排器推动想法演进。它在包括 EurekaBench 在内的五项基准测试中领先(在 EurekaBench 上得分为 73.6%,而 CORAL 为 56.2%),同时使用的模型调用次数要少得多。
查看缓存全文
缓存时间: 2026/11/19 03:34
# Decentralized Self-ImprovingAgents for Autonomous Scientific Discovery
Source: [https://arxiv.org/html/2611.21847](https://arxiv.org/html/2611.21847)
## Inventa: Decentralized Self\-Improving Agents for Autonomous Scientific Discovery
Nikos Kanakaris2Purak Jain2Anoop Deoras2Sahika Genc21Georgia Institute of Technology2Amazon
###### Abstract
Abstract:Large language model \(LLM\) agents can generate research ideas and improve scientific discovery through iterative experimentation\. However, increasing the number of agents or iterations does not automatically translate to producing more diverse or feasible ideas\. Agents may converge on similar candidates, propagate unsupported claims through shared memory components and repeatedly apply search strategies that no longer yield any improvements\.
We introduceInventa, a framework for automated scientific discovery that employs autonomous research units and verified shared knowledge\. Each unit creates a lineage of candidate ideas using a mutator agent\. A local orchestrator agent periodically adapts the mutator’s harness\. When progress stalls, the orchestrator agent replaces its model or restarts the idea evolution process\. Verifiers evaluate each candidate idea and record verified improvements, useful components and constraint violations in a shared knowledge, the only channel between units\. We evaluateInventaon five benchmarks spanning the research cycle, covering retrieval \(ScholarCatalyst\), ideation \(IdeaBench\), proposal writing \(IdeaScientist\), and scientific law discovery \(EurekaBench, SciLaws\-Bench\)\. It leads all five, achieving 73\.6% on EurekaBench versus 56\.2% for CORAL\. On IdeaBench, it discovers 26\.0 novel ideas per paper, compared with 18\.6 for ResearchAgent, with 82% fewer model calls\.
††footnotetext:†Work done while at Amazon\.## 1Introduction
Figure 1:Overview ofInventa\.Given a research task,Inventagenerates novel and verified ideas\. Decentralizedautonomous research unitsexplore candidate ideas in parallel\. Each unit0reads theshared knowledge\(𝒦\\mathcal\{K\}\),1selects a parent from its idea tree, and2uses a mutator agent to propose a child idea\.Verifiers3aevaluate it with held\-out scoring, constraint checks, and out\-of\-distribution or ablation tests,3breturn its certificate, and3cpublish qualifying findings to𝒦\\mathcal\{K\}\. The unit4applies acceptance checks and5records the child in its idea tree\. A per\-unit orchestrator6evolves the mutator regularly\.𝒦\\mathcal\{K\}is the sole channel for inter\-unit communication, and only verifiers write to it\.Large language models \(LLMs\) are increasingly used to assist in scientific research, ranging from identifying relevant literature to proposing candidate hypotheses and implementing experiments\. Evolutionary search extends these capabilities by using evaluation feedback to refine candidates\. Systems such as AlphaEvolve\([Novikov et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib35)\), EvoX\([Liu et al\., 2026b](https://arxiv.org/html/2611.21847#bib.bib33)\), CORAL\([Qu et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib36)\), and MILO\([Jana et al\., 2026b](https://arxiv.org/html/2611.21847#bib.bib21)\)show the potential of combining LLMs with closed\-loop search\. These advances motivate automated research assistants that can explore novel ideas, evaluate their feasibility and retain useful findings across experiments\.
Automated research assistanceinvolves several related tasks across the complete scientific research cycle\. These include*retrieving prior work*that provides inspiration for a research question\([Kim et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib25)\),*generating ideas*from existing papers\([Guo et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib15)\),*assessing their novelty*against related work\([Schopf & Färber, 2026](https://arxiv.org/html/2611.21847#bib.bib42)\), and*discovering scientific laws*from data\([Huang et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib17)\)\. They differ in output but share a common challenge: generating promising candidates is insufficient unless a system can also verify them and use the evidence to guide further exploration\.
RecentAI co\-scientists, such as the Agentic Scientist\([Lange et al\., 2026a](https://arxiv.org/html/2611.21847#bib.bib28)\)and InternAgent\-1\.5\([Feng et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib12)\), coordinate multiple agents to generate, test, and refine hypotheses\. Despite this progress, several challenges remain unresolved\.First,maintaining diversity as search grows\. These systems largely follow myopic hill\-climbing to optimize the most promising candidates, which narrows the exploration radius\([Lange et al\., 2026a](https://arxiv.org/html/2611.21847#bib.bib28)\)\. Furthermore, additional samples or iterations yield increasingly similar candidates, so extra computation brings diminishing returns\([Jiang et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib24);[Si et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib44)\)\.Second,sharing useful knowledge without propagating errors\. Isolated agents cannot benefit from one another’s findings, whereas interacting agents often converge on similar ideas\([Chen et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib8)\)\. When shared information is not independently verified, unsupported claims also influence subsequent attempts\([Kim et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib26)\)\.Third,adapting the proposal mechanism\. Refining candidate ideas does not, by itself, change the prompts, tools, or control flow used to generate them, so iterations can improve ideas but not expand the search space\.Fourth,satisfying scientific constraints\. Highly novel ideas often prove infeasible under the task’s assumptions or physical requirements, creating a novelty–feasibility gap\([Geng et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib13)\)\.
To address these challenges, we proposeInventa\(Fig\.\), a framework that combines i\) autonomous research units, ii\) self\-adaptive proposal mechanisms and iii\) sharing of verified findings\. We draw inspiration from how science advances in practice: research groups or labs pursue different approaches in a decentralized fashion while building on evidence reported by others\. Each unit \(like a lab\) maintains its own idea lineage and uses a mutator agent \(junior researcher\) to develop new candidates\. A per\-unit orchestrator \(the lab’s principal investigator, or PI\) adapts the mutator’s harness and intervenes when progress stalls\. Verifiers evaluate the candidates, return structured feedback, and add verified findings to the shared knowledge\.
The proposed framework separates three operations: developing candidate ideas, self\-adapting the process that develops them, and determining which findings can be shared\. This separation allows units to pursue different approaches while receiving, only through shared knowledge, common evidence about successful candidates, useful components, and observed constraint violations\.
\\papersansContributions\.Our main contributions are:❶Self\-improvingidea generation\.We proposeInventa, an agentic framework for scaling scientific discovery and research ideation\. The framework refines ideas while continually evolving the agents that generate them, expanding the space of ideas it can reach as the search proceeds\.❷Diverseexploration with verified knowledge sharing\.We introduce autonomous research units that explore different approaches, each with its own history of ideas and experiments\. Units communicate only through shared knowledge written by verifiers: positive findings point to promising directions, while refuted ones, such as constraint violations, prune the search space\.❸Extensiveexperiments\.On five benchmarks from retrieving prior work to scientific law discovery,Inventaleads every primary metric by 6–40% and alone passes every EurekaBench constraint test \(Fig\.\)\.
Best\-of\-N\([2024](https://arxiv.org/html/2611.21847#bib.bib6)\)ResearchAgent\([2025](https://arxiv.org/html/2611.21847#bib.bib4)\), AIDE2\([2026](https://arxiv.org/html/2611.21847#bib.bib45)\)FunSearch\([2024](https://arxiv.org/html/2611.21847#bib.bib39)\), AlphaEvolve\([2025](https://arxiv.org/html/2611.21847#bib.bib35)\),OpenEvolve\([2025](https://arxiv.org/html/2611.21847#bib.bib43)\)EvoX \(2026\)AgentRxiv\([2025](https://arxiv.org/html/2611.21847#bib.bib41)\), CORAL\([2026](https://arxiv.org/html/2611.21847#bib.bib36)\)MILO\([2026b](https://arxiv.org/html/2611.21847#bib.bib21)\)Figure 2:Comparison of search mechanisms\.Prior methods \(a\)–\(f\) range from repeated sampling to island\-based evolution with fixed, partially adaptive, or centralized strategies; all keep the mutator fixed\. InInventa\(g\), each decentralized unit has a local orchestrator that evolves its mutator, while onlyverifierspublish shared findings\.
## 2Related Work
Scaling search\.AI scientists generate and test ideas\([Lu et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib34);[Gottweis et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib14)\), often by LLM\-guided search\. Yet in equation discovery, evolution matches fresh sampling at equal budgets\([Li, 2026](https://arxiv.org/html/2611.21847#bib.bib31)\), and repeated runs or added islands stall in idea basins\([Cai et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib7);[Antoniades et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib3)\)or collapse onto a few principles\([Ran et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib37)\)\.Inventaruns many units that redraw their model or restart their lineage on a stall\.
Adapting the generator\.AlphaEvolve co\-evolves its prompts, CORAL’s agents write their own skills, and MILO and ShinkaEvolve pick among fixed agents or LLMs\([Novikov et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib35);[Qu et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib36);[Jana et al\., 2026b](https://arxiv.org/html/2611.21847#bib.bib21);[Lange et al\., 2026b](https://arxiv.org/html/2611.21847#bib.bib29)\)\(Fig\.\)\. Harness search optimizes a final harness on a fixed search set\([Lee et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib30);[Zhang et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib49)\), where overfitting explains about half the gain\([Ding et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib11)\)\. InInventa, an orchestrator acts as a meta\-mutator, evolving the mutator that evolves ideas during search and keeping an edit only if twin trials show significantly fitter children\.
Sharing verified knowledge\.In CORAL and AgentRxiv, agents write shared state\([Qu et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib36);[Schmidgall & Moor, 2025](https://arxiv.org/html/2611.21847#bib.bib41)\); such state can spread reward hacks\([Hu et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib16)\)and early consensus\([Yoon et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib48)\)\. DuoDiscover’s agents optimize self\-built proxy evaluators\([Agarwal et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib1)\); inInventa, only verifiers write shared knowledge, scoring on data no mutator sees, and publish refuted findings that prune all units’ search\.
## 3Preliminaries and Problem Statement
###### Definition 3\.1 \(Agents and harnesses\)\. An AI agentA:=⟨m,h⟩A:=\\langle m,h\\rangleconsists of an LLMmmand a harnesshh\. The harness specifies the agent’s prompts, tools, context management, and control flow\. In our setting, we keep the model weights frozen and allow the harness to change during evolution\. Letℳ\\mathcal\{M\}denote the set of available models andℋ\\mathcal\{H\}the initial set of harnesses\. The pool𝒢⊆ℳ×ℋ\\mathcal\{G\}\\subseteq\\mathcal\{M\}\\times\\mathcal\{H\}contains their compatible model–harness pairs\.
###### Definition 3\.2 \(Research tasks and candidate ideas\)\. Aresearch task, the input toInventa, is a tupleτ:=⟨q,𝒲,𝒞,𝒱⟩\\tau:=\\langle q,\\mathcal\{W\},\\mathcal\{C\},\\mathcal\{V\}\\rangle, whereqqis a research question;𝒲\\mathcal\{W\}is a workspace of literature, tools, and data𝒟\\mathcal\{D\};𝒞\\mathcal\{C\}is a set of task\-specific constraints; and𝒱\\mathcal\{V\}is a set of verifiers that run a full experiment on each candidate, executing it on held\-out data or in task simulators, testing task constraints, and using LLM judges where programmatic tests are unavailable\. Givenτ\\tau,Inventaoutputs acandidate ideaxxthat answersqqin the task’s format, such as code with an explanation;T\(x\)T\(x\)denotes its components, such as the terms of a law, which verifiers ablate \(§\)\. We show an example task in Fig\.\(left\) and examples of all task classes in Appendix\.
###### Definition 3\.3 \(Data splits and candidate evaluation\)\. Research tasks distinguishdata regimes𝒟ρ⊆𝒟\\mathcal\{D\}\_\{\\rho\}\\subseteq\\mathcal\{D\}\. For out\-of\-distribution evaluation, one regime is held out as𝒟ood\\mathcal\{D\}\_\{\\mathrm\{ood\}\}, and the rest form𝒟id\\mathcal\{D\}\_\{\\mathrm\{id\}\}\. Agents fit candidates on the training set𝒟train\\mathcal\{D\}\_\{\\mathrm\{train\}\}from𝒟id\\mathcal\{D\}\_\{\\mathrm\{id\}\}; only verifiers access the validation set𝒟val\\mathcal\{D\}\_\{\\mathrm\{val\}\}, which includes𝒟ood\\mathcal\{D\}\_\{\\mathrm\{ood\}\}\. A sealed test set𝒟test\\mathcal\{D\}\_\{\\mathrm\{test\}\}scores only the final candidate and never guides search\. During search, verifiers compute the task\-specific metricScore\(x\):=Score\(x,𝒟val\)\\textsc\{Score\}\(x\):=\\textsc\{Score\}\(x;\\mathcal\{D\}\_\{\\mathrm\{val\}\}\), the improvementGain\(x\):=Score\(x\)−Score\(x0\)\\textsc\{Gain\}\(x\):=\\textsc\{Score\}\(x\)\-\\textsc\{Score\}\(x\_\{0\}\)over a baselinex0x\_\{0\}, and the fractionValid\(x\)∈\[0,1\]\\textsc\{Valid\}\(x\)\\in\[0,1\]of constraints without a confirmed violation\.
Problem statement\.Given a taskτ\\tauand a search budget,Inventaseeks a candidatex⋆x^\{\\star\}that maximizes the validation scoreScore\(x⋆\)\\textsc\{Score\}\(x^\{\\star\}\)subject toValid\(x⋆\)=1\\textsc\{Valid\}\(x^\{\\star\}\)=1andGain\(x⋆\)\>δ\\textsc\{Gain\}\(x^\{\\star\}\)\>\\deltafor a noise marginδ\\delta, aiming to score high on the unseen𝒟test\\mathcal\{D\}\_\{\\mathrm\{test\}\}\.
## 4Methodology
Algorithm 1\(Inventa\)\.NNautonomous research units evolve their mutators and shareverifier\-writtenfindings\.1:taskτ\\tauwith verifiers𝒱\\mathcal\{V\}, pool𝒢\\mathcal\{G\}, budgetBBhyperparameters:number of unitsNN, patiencePP, lagλ\\lambda2:best verified ideax⋆x^\{\\star\}withCert\(x⋆\)\\textsc\{Cert\}\(x^\{\\star\}\)3:𝒦=\(𝒦ver,𝒦part,𝒦ref\)←\(∅,∅,∅\)\\mathcal\{K\}=\(\\mathcal\{K\}\_\{\\mathrm\{ver\}\},\\mathcal\{K\}\_\{\\mathrm\{part\}\},\\mathcal\{K\}\_\{\\mathrm\{ref\}\}\)\\leftarrow\(\\emptyset,\\emptyset,\\emptyset\)⊳\\trianglerightshared knowledge4:for allj=1,…,Nj=1,\\dots,Nin paralleldo⊳\\trianglerightautonomous units𝒰j\\mathcal\{U\}\_\{j\}5:μj∼𝒢\\mu\_\{j\}\\sim\\mathcal\{G\};ℒj←\{seed idea\}\\mathcal\{L\}\_\{j\}\\leftarrow\\\{\\text\{seed idea\}\\\}⊳\\trianglerightinitialize𝒰j\\mathcal\{U\}\_\{j\}6:forrj←1r\_\{j\}\\leftarrow 1toBBdo7:wait untilrj<minj′∈activerj′\+λr\_\{j\}<\\min\_\{j^\{\\prime\}\\in\\mathrm\{active\}\}r\_\{j^\{\\prime\}\}\+\\lambda⊳\\trianglerightbounded lag8:𝒦j←Snapshot\(𝒦\)\\mathcal\{K\}\_\{j\}\\leftarrow\\textsc\{Snapshot\}\(\\mathcal\{K\}\)⊳\\triangleright0read the shared knowledge9:xpar←SelectParent\(ℒj\)x\_\{\\mathrm\{par\}\}\\leftarrow\\textsc\{SelectParent\}\(\\mathcal\{L\}\_\{j\}\)⊳\\triangleright1select a parent10:x←μj\(xpar,Cert\(xpar\),ℒj,𝒦j\)x\\leftarrow\\mu\_\{j\}\(x\_\{\\mathrm\{par\}\},\\textsc\{Cert\}\(x\_\{\\mathrm\{par\}\}\),\\mathcal\{L\}\_\{j\},\\mathcal\{K\}\_\{j\}\)⊳\\triangleright2write a child11:Cert\(x\),F\(x\)←𝒱\(x\)\\textsc\{Cert\}\(x\),F\(x\)\\leftarrow\\mathcal\{V\}\(x\)⊳\\triangleright3atest,3breturn the certificate12:atomic⊳\\triangleright3cpublish the findings13:xbest←Best\(𝒦ver\)\\smash\[t\]\{x^\{\\mathrm\{best\}\}\}\\leftarrow\\textsc\{Best\}\(\\mathcal\{K\}\_\{\\mathrm\{ver\}\}\)⊳\\trianglerightre\-check the best verified idea14:𝒦←𝒦∪Findings\(x,xbest\)\\mathcal\{K\}\\leftarrow\\mathcal\{K\}\\cup\\textsc\{Findings\}\(x,\\smash\[t\]\{x^\{\\mathrm\{best\}\}\}\)⊳\\trianglerightshare with all units15:ℓx←Accept\(x,ℒj\)\\ell\_\{x\}\\leftarrow\\textsc\{Accept\}\(x,\\mathcal\{L\}\_\{j\}\)⊳\\triangleright4accept or reject the child16:ℒj←ℒj∪\{\(x,ℓx\)\}\\mathcal\{L\}\_\{j\}\\leftarrow\\mathcal\{L\}\_\{j\}\\cup\\\{\(x,\\ell\_\{x\}\)\\\}⊳\\triangleright5record in the unit’s memory17:if𝒰j\\mathcal\{U\}\_\{j\}stallsthen⊳\\trianglerightno new best fitness inPPscored rounds18:RedraworRestart⊳\\triangleright6change model or approach19:everyPProunds:Retool⊳\\triangleright6evolve harness via twin trial20:returnx⋆←Best\(𝒦ver\)x^\{\\star\}\\leftarrow\\textsc\{Best\}\(\\mathcal\{K\}\_\{\\mathrm\{ver\}\}\)withCert\(x⋆\)\\textsc\{Cert\}\(x^\{\\star\}\)Inventacomprises autonomous research units, verifiers, and a shared knowledge \(§–\)\. We detail the asynchronous evolution loop in Alg\.\.
### 4\.1Autonomous Research Units
Each unit𝒰j\\mathcal\{U\}\_\{j\}, wherej∈\{1,…,N\}j\\in\\\{1,\\dots,N\\\}, advances at its own pace\. Itsroundrjr\_\{j\}counts the attempts it has made so far\. At roundrj=rr\_\{j\}=r, we define its state as𝒰j\(r\):=⟨ℒj\(r\),μj\(r\)⟩\\smash\[t\]\{\\mathcal\{U\}\_\{j\}^\{\(r\)\}:=\\langle\\mathcal\{L\}\_\{j\}^\{\(r\)\},\\mu\_\{j\}^\{\(r\)\}\\rangle\}, consisting of its lineage memory and current mutator\. Its orchestrator𝒪j\\mathcal\{O\}\_\{j\}adapts both the mutator and the unit’s search direction\.
Lineage memory\.The memoryℒj\(r\)\\smash\[t\]\{\\mathcal\{L\}\_\{j\}^\{\(r\)\}\}is anidea treeof the candidates𝒰j\\mathcal\{U\}\_\{j\}explored\. Eachnodeholds an idea, its certificate of verifier evidence \(§\), and an acceptance label\. Eachedgelinks a parent to its child\. The tree grows from aseed ideaproposed by𝒪j\\mathcal\{O\}\_\{j\}; every later roundrj=rr\_\{j\}=rhas five stages:
1. \(i\)Parent selection\.To pick a parent ideaxparx\_\{\\mathrm\{par\}\}, the unit treats the accepted nodes𝒜\\mathcal\{A\}ofℒj\(r−1\)\\smash\[t\]\{\\mathcal\{L\}\_\{j\}^\{\(r\-1\)\}\}as arms of a multi\-armed bandit rewarded by the fitnessF\(x\):=Score\(x\)×𝕀\[Valid\(x\)=1\]\\smash\[t\]\{F\(x\):=\\textsc\{Score\}\(x\)\\times\\mathbb\{I\}\[\\textsc\{Valid\}\(x\)=1\]\}\. A node’s values¯v:=1nv∑u∈\{v\}∪ch\(v\)F\(u\)\\smash\[t\]\{\\bar\{s\}\_\{v\}:=\\tfrac\{1\}\{n\_\{v\}\}\\sum\_\{u\\in\\\{v\\\}\\cup\\mathrm\{ch\}\(v\)\}F\(u\)\}averages the fitness ofvvand its childrench\(v\)\\mathrm\{ch\}\(v\), withnv:=1\+\|ch\(v\)\|n\_\{v\}:=1\+\|\\mathrm\{ch\}\(v\)\|\. By Thompson sampling\([Russo et al\., 2018](https://arxiv.org/html/2611.21847#bib.bib40)\), the unit drawss~v∼𝒩\(s¯v,σ2/nv\)\{\\tilde\{s\}\_\{v\}\\sim\\mathcal\{N\}\(\\bar\{s\}\_\{v\},\\sigma^\{2\}/n\_\{v\}\)\}and selectsxpar=argmaxv∈𝒜s~vx\_\{\\mathrm\{par\}\}=\\arg\\max\_\{v\\in\\mathcal\{A\}\}\\tilde\{s\}\_\{v\}, withσ\\sigmathe typical fitness change per child\. This balancesexploitingpromising lineages andexploringnodes with few children\.
2. \(ii\)Mutation\.μj\(r\)\\mu\_\{j\}^\{\(r\)\}diagnosesxparx\_\{\\mathrm\{par\}\}and revises it into a child ideaxx\(Eq\.\), fitted on𝒟train\\mathcal\{D\}\_\{\\mathrm\{train\}\}\.
3. \(iii\)Verification and knowledge sharing\.The verifiers testxx\(§\), return its certificate and fitness, and publish qualifying findings to𝒦\\mathcal\{K\}\(§\)\.
4. \(iv\)Child acceptance\.The unit acceptsxxonly if no previously acceptedy∈𝒜y\\in\\mathcal\{A\}is at least as simple and nearly as fit:Cplx\(y\)≤Cplx\(x\)\\textsc\{Cplx\}\(y\)\\leq\\textsc\{Cplx\}\(x\)andF\(y\)≥F\(x\)−δF\(y\)\\geq F\(x\)\-\\delta, withCplxa task\-specific complexity \(e\.g\., parameters\); extra complexity must thus improve fitness by\>δ\>\\delta\. It also rejects children with a confirmed constraint violation\.
5. \(v\)Memory update\.The child ideaxxjoins the lineage\. Accepted children become eligible parents, while rejected ones remain as negative evidence\. Every scored child updates its parent’s value\.
Mutator agent\.The mutatorμj\(r\)=⟨mj\(r\),hj\(r\)⟩\\smash\[t\]\{\\mu\_\{j\}^\{\(r\)\}=\\langle m\_\{j\}^\{\(r\)\},h\_\{j\}^\{\(r\)\}\\rangle\}, a model run in a harness, writes the child
x←μj\(r\)\(xpar,Cert\(xpar\),ℒj\(r−1\),𝒦j\)x\\leftarrow\\mu\_\{j\}^\{\(r\)\}\\bigl\(x\_\{\\mathrm\{par\}\},\\textsc\{Cert\}\(x\_\{\\mathrm\{par\}\}\),\\mathcal\{L\}\_\{j\}^\{\(r\-1\)\},\\mathcal\{K\}\_\{j\}\\bigr\)\(1\)from the parent, its certificate, recent lineage nodes, and the snapshot𝒦j\\mathcal\{K\}\_\{j\}of shared knowledge\. The certificate giveslocal feedbackon the parent’s failures; the lineage and𝒦j\\mathcal\{K\}\_\{j\}giveglobal contexton what the unit has tried, why each attempt was accepted or rejected, and promising or unpromising directions\. The mutator runs in a sandbox where𝒟val\\mathcal\{D\}\_\{\\mathrm\{val\}\}and𝒟test\\mathcal\{D\}\_\{\\mathrm\{test\}\}are inaccessible\. The initial mutator is drawn from𝒢\\mathcal\{G\}; the orchestrator can later change its harness or model, as we describe next\.
Orchestrator agent\.The orchestrator𝒪j∈𝒢\\mathcal\{O\}\_\{j\}\\in\\mathcal\{G\}adapts the unit through three actions\.Retoolis considered everyPProunds;RedrawandRestartrespond when the unit stalls, i\.e\., goesPPscored rounds without improving its best fitness\.
- •Retool\(evolving the harness\)\.Acting as ameta\-mutator, the orchestrator diagnoses tracesℰj\\mathcal\{E\}\_\{j\}of recent attempts \(turns, tool calls, costs, and outcomes\) and proposes a harnessh′←𝒪j\(hj\(r\),ℰj\)h^\{\\prime\}\\leftarrow\\mathcal\{O\}\_\{j\}\(\\smash\[t\]\{h\_\{j\}^\{\(r\)\}\},\\mathcal\{E\}\_\{j\}\)\. Since a worseh′h^\{\\prime\}would degrade every later child, a twin trial first runs the two harnesses on the same parent and mutation plan, giving childrenxix\_\{i\}andxi′x\_\{i\}^\{\\prime\}withdi:=F\(xi′\)−F\(xi\)d\_\{i\}:=F\(x\_\{i\}^\{\\prime\}\)\-F\(x\_\{i\}\)when both are scored\. The unit setsμj←⟨mj\(r\),h′⟩\\mu\_\{j\}\\leftarrow\\langle\\smash\[t\]\{m\_\{j\}^\{\(r\)\}\},h^\{\\prime\}\\rangleif thennscored pairs give a significantly positive meandid\_\{i\}; otherwise,hj\(r\)\\smash\[t\]\{h\_\{j\}^\{\(r\)\}\}stays\.
- •Redraw\(replacing the model\)\.𝒦\\mathcal\{K\}tallies each model’s recent fitness improvements \(§\), so𝒪j\\mathcal\{O\}\_\{j\}can swap in a better\-performing model: it drawsm′m^\{\\prime\}from this tally by Thompson sampling and, keeping the harness, setsμj←⟨m′,hj\(r\)⟩\\mu\_\{j\}\\leftarrow\\langle m^\{\\prime\},\\smash\[t\]\{h\_\{j\}^\{\(r\)\}\}\\rangle\.
- •Restart\(changing the research direction\)\.If the unit stalls again after a secondRedrawwith noVerifiedfinding since \(§\),𝒪j\\mathcal\{O\}\_\{j\}archives the unit’s existing lineage and proposes a new seed idea from𝒦\\mathcal\{K\}to initiate a new idea tree\.
### 4\.2Verifier Components
The verifiers𝒱\\mathcal\{V\}evaluate each candidate ideaxxagainstx0x\_\{0\}and the constraints𝒞\\mathcal\{C\}\. They return a certificateCert\(x\)\\textsc\{Cert\}\(x\)recording the outcome of every check and the evidence behind each failure, and decide which findings enter𝒦\\mathcal\{K\}\(§\)\. Before sharing results to𝒦\\mathcal\{K\}, the verifiers run four checks:
- •Held\-out scoring\.Runningxxon𝒟val\\mathcal\{D\}\_\{\\mathrm\{val\}\}givesScore\(x\)\\textsc\{Score\}\(x\)andGain\(x\)\\textsc\{Gain\}\(x\)\(Def\.\)\.
- •Constraint checking\.Eachc∈𝒞c\\in\\mathcal\{C\}is tested programmatically or by LLM consensus\. Confirmed violations form𝒞viol\(x\)\\mathcal\{C\}\_\{\\mathrm\{viol\}\}\(x\), soValid\(x\)=1−\|𝒞viol\(x\)\|/\|𝒞\|\\textsc\{Valid\}\(x\)=1\-\|\\mathcal\{C\}\_\{\\mathrm\{viol\}\}\(x\)\|/\|\\mathcal\{C\}\|; undecided ones form𝒞undec\(x\)\\mathcal\{C\}\_\{\\mathrm\{undec\}\}\(x\)\.
- •Out\-of\-distribution testing\.OOD\(x\)\\textsc\{OOD\}\(x\)is11if the score ofxxon the held\-out regime𝒟ood\\mathcal\{D\}\_\{\\mathrm\{ood\}\}is at leastκ\\kappatimes its score on𝒟val∖𝒟ood\\mathcal\{D\}\_\{\\mathrm\{val\}\}\\setminus\\mathcal\{D\}\_\{\\mathrm\{ood\}\}, else00\.
- •Component ablation\.Each componentt∈T\(x\)t\\in T\(x\)is removed in turn; its contribution on dataS⊆𝒟S\\subseteq\\mathcal\{D\}is the score dropAblt\(x,S\):=Score\(x,S\)−Score\(x∖\{t\},S\)\\textsc\{Abl\}\_\{t\}\(x;S\):=\\textsc\{Score\}\(x;S\)\-\\textsc\{Score\}\(x\\setminus\\\{t\\\};S\); its maximum over validation regimes isAbltmax\(x\):=maxρAblt\(x,𝒟ρ∩𝒟val\)\\textsc\{Abl\}^\{\\max\}\_\{t\}\(x\):=\\max\_\{\\rho\}\\textsc\{Abl\}\_\{t\}\(x;\\mathcal\{D\}\_\{\\rho\}\\cap\\mathcal\{D\}\_\{\\mathrm\{val\}\}\)\(Appendix\)\.
### 4\.3Verified Shared Knowledge
The shared knowledge stores the per\-model gain tally used byRedraw\(§\) and three types of findings,𝒦=\(𝒦ver,𝒦part,𝒦ref\)\\mathcal\{K\}=\(\\mathcal\{K\}\_\{\\mathrm\{ver\}\},\\mathcal\{K\}\_\{\\mathrm\{part\}\},\\mathcal\{K\}\_\{\\mathrm\{ref\}\}\)\. Verifiers write to it, while units read it at each round’s start\.
Verifiedfindings\.The set𝒦ver\\mathcal\{K\}\_\{\\mathrm\{ver\}\}stores new bests\(x,Cert\(x\)\)\(x,\\textsc\{Cert\}\(x\)\)\. Letxbestx^\{\\mathrm\{best\}\}be the highest\-scoring idea in𝒦ver\\mathcal\{K\}\_\{\\mathrm\{ver\}\}\. A candidatexxis added ifScore\(x\)\>Score\(xbest\)\+δ\\smash\[t\]\{\\textsc\{Score\}\(x\)\>\\textsc\{Score\}\(x^\{\\mathrm\{best\}\}\)\+\\delta\}and it passes all checks:𝒞viol\(x\)∪𝒞undec\(x\)=∅\\mathcal\{C\}\_\{\\mathrm\{viol\}\}\(x\)\\cup\\mathcal\{C\}\_\{\\mathrm\{undec\}\}\(x\)=\\emptyset,OOD\(x\)=1\\textsc\{OOD\}\(x\)=1, and, where ablation is supported,Abltmax\(x\)\>δ\\textsc\{Abl\}^\{\\max\}\_\{t\}\(x\)\>\\deltafor eacht∈T\(x\)t\\in T\(x\)\. This raises the bar for all units\.
Partialfindings\.The set𝒦part\\mathcal\{K\}\_\{\\mathrm\{part\}\}holds useful components\(t,x,Cert\(x\)\)\(t,x,\\textsc\{Cert\}\(x\)\)for other units to reuse\. Each componentttof a candidatexxwithValid\(x\)=1\\textsc\{Valid\}\(x\)=1andGain\(x\)\>δ\\textsc\{Gain\}\(x\)\>\\deltaenters ifAbltmax\(x\)\>δ\\textsc\{Abl\}^\{\\max\}\_\{t\}\(x\)\>\\delta, even whenxxdoes not beatxbestx^\{\\mathrm\{best\}\}\.
Refutedfindings\.The set𝒦ref\\mathcal\{K\}\_\{\\mathrm\{ref\}\}records each confirmed violationc∈𝒞viol\(x\)c\\in\\mathcal\{C\}\_\{\\mathrm\{viol\}\}\(x\)as\(c,x,Cert\(x\)\)\(c,x,\\textsc\{Cert\}\(x\)\), so other units avoid them and prune search space\.
\(a\) Gain over the best baseline, by research stageGain ofInventaoverthe best baseline \(%\)Retrieveprior workGenerateideasWriteproposalsDiscoverscientificlawsScholarCatalystOpenCode \+ DS V3\.2\+21\.9%36\.5→\\rightarrow44\.5IdeaBenchResearchAgent\+39\.8%18\.6→\\rightarrow26\.0IdeaScientistIdeaScientist\-Base\+6\.4%77\.7→\\rightarrow82\.7EurekaBenchCORAL\+31\.0%56\.2→\\rightarrow73\.6SciLaws\-BenchReAct \+ Grok 4\.7\+5\.9%59\.3→\\rightarrow62\.8
\(b\) Scientific law discovery020040060080020406080100Cost per task \(units\)Cond\. score on𝒟val\\mathcal\{D\}\_\{\\mathrm\{val\}\}\(%\)\+11\.9\+17\.0InventaCORALAgentic ScientistMILOOpenEvolveAgents1–14
\(c\) Idea generation010020030040050004080120Model calls per paperDistinct novel ideas / paperInventaMILOBest\-of\-NN20–62% fewercalls than MILO
Figure 3:Inventaleads at every research stage and keeps improving with search\.\(a\) Gain over the best baseline on the respective primary metric \(%\)\. \(b\) EurekaBench validation conditioned score vs\. cost per task, in units of 10 USD; baselines from Table\. \(c\) IdeaBench distinct novel ideas vs\. model calls per paper\.
## 5Experiments
Benchmarks and metrics\.We evaluateInventaat four stages of the research cycle: \(i\)*retrieving prior work that inspires new ideas*on ScholarCatalyst\([Kim et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib25)\)\(894 queries\), \(ii\)*generating ideas*on IdeaBench\([Guo et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib15)\)\(736 test papers\), \(iii\)*writing grounded proposals*on IdeaScientist\([Liu et al\., 2026a](https://arxiv.org/html/2611.21847#bib.bib32)\)\(277 research problems\), and \(iv\)*discovering scientific laws*on EurekaBench\([Geng et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib13)\)\(16 CPU\-only tasks\) and SciLaws\-Bench\([Huang et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib17)\)\(38 tasks\)\. Primary metrics follow each stage’s goal: \(i\) retrieval quality, the mean of nDCG@20 and Recall@kk; \(ii\) novelty, distinct novel ideas per paper rated≥\\geq7/10 by an expert\-calibrated judge; \(iii\) novelty and quality, their combined score; and \(iv\) constraint\-aware fit, EurekaBench’s conditioned score \(CS\) of predictive accuracy \(PA\) and scientific insight \(SI\), zero unless the law passes all of the task’s scientific\-constraint tests, and SciLaws\-Bench’s held\-out fitSNS\_\{N\}\(see Appendices–\)\.
Baselines and protocol\.We compare with frontier AI agents \(OpenHands\([Wang et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib47)\), OpenCode\([SST, 2025](https://arxiv.org/html/2611.21847#bib.bib46)\), Goose\([Block, 2025](https://arxiv.org/html/2611.21847#bib.bib5)\), SciLaws\-Bench’s ReAct\) with up to five models, Best\-of\-NN, SoTA evolutionary search and AI co\-scientists \(e\.g\., GEPA\([Agrawal et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib2)\), MILO, CORAL\), and SoTA task\-specific systems\. Evolution sees only𝒟train\\mathcal\{D\}\_\{\\mathrm\{train\}\}and𝒟val\\mathcal\{D\}\_\{\\mathrm\{val\}\}; the best candidate, chosen on𝒟val\\mathcal\{D\}\_\{\\mathrm\{val\}\}, is then tested on𝒟test\\mathcal\{D\}\_\{\\mathrm\{test\}\}\(details in Appendix\)\.
RQ1: Can one framework serve multiple stages of the research cycle?
Table 1:ScholarCatalyst\(894 queries; primary:Mean\)\.Bold/underline: best/second\.Inventaranks first on the primary metric of all five benchmarks, 6–40% above the strongest baseline \(Fig\.a\)\.SoTA frameworks, by contrast, are inconsistent across stages: among evolutionary baselines, CORAL leads onEurekaBench\(56\.2% CS\) but trails on IdeaBench \(2\.7 ideas\), and OpenEvolve does the opposite \(14\.7 ideas; 34\.0% CS\), whereasInventaleads on both\. In retrieval \(ScholarCatalyst\),Inventaevolves its seed retriever from 33\.1% to 44\.5%, versus 36\.5% for the best agent, OpenCode \(Table\)\. In idea generation \(IdeaBench\), it returns 26\.0 distinct, judge\-rated novel ideas per paper, versus 18\.6 for ResearchAgent, with 82% fewer calls \(Table\)\. In proposal writing \(IdeaScientist\),Inventascores 82\.7%, vs\. 77\.7% for IdeaScientist\-Base and 74\.6% for IdeaScientist\-Trained, with the highest novelty \(82\.3% vs\. at most 69\.8%; Table\)\. In scientific law discovery \(EurekaBench,SciLaws\-Bench\), it leads by 31\.0% and 5\.9%\.
Table 2:EurekaBench\(16 tasks; primary: CS\)\. Best\-of\-NN: best on𝒟val\\mathcal\{D\}\_\{\\mathrm\{val\}\}, tested on𝒟test\\mathcal\{D\}\_\{\\mathrm\{test\}\}\.Bold/underline: best/second\.
Table 3:SciLaws\-Bench\(38 tasks; primary:SNS\_\{N\}\)\.\>\>Ref: beats the published law\.Bold/underline: best/second\.Table 4:IdeaScientist\(277 problems; primary:Combined\)\. Rows 1–2 from[Liu et al\. \(2026a\)](https://arxiv.org/html/2611.21847#bib.bib32)\.
Table 5:IdeaBench\(736 test papers; App\.\)\.Distinct novel: ideas per paper rated≥\\geq8 or≥\\geq7 \(primary\)\.Bold/underline: best/second per setting\.Novelty & breadthInsight Score↑\\uparrowSimilarity to paperDistinct novel↑\\uparrowMethodNon\-duplicateRate↑\\uparrowFeasibilitySemanticOverlap≥8\{\\geq\}8≥7\{\\geq\}727\-idea setting1Single prompt37\.7542\.494\.891\.61\.554\.33\.871\.01\.22Over\-generate \+ rerank218\.0921\.397\.095\.21\.654\.63\.533\.94\.13Best\-of\-NN,N=26N\{=\}26507\.9621\.896\.594\.18\.655\.03\.772\.24\.74GEPA\([2026](https://arxiv.org/html/2611.21847#bib.bib2)\)568\.0040\.897\.196\.532\.253\.72\.998\.312\.35OpenEvolve\([2025](https://arxiv.org/html/2611.21847#bib.bib43)\)568\.6370\.698\.798\.65\.352\.82\.6813\.514\.76EvoX\([2026b](https://arxiv.org/html/2611.21847#bib.bib33)\)568\.1322\.697\.094\.65\.854\.13\.477\.78\.27CORAL\([2026](https://arxiv.org/html/2611.21847#bib.bib36)\)418\.0830\.598\.496\.455\.350\.52\.552\.52\.78ResearchAgent\([2025](https://arxiv.org/html/2611.21847#bib.bib4)\)2477\.9697\.995\.695\.225\.554\.43\.155\.318\.69MILO\([2026b](https://arxiv.org/html/2611.21847#bib.bib21)\), 3 islands568\.4452\.098\.398\.113\.854\.33\.2510\.113\.410Inventa458\.8898\.396\.897\.420\.851\.72\.2518\.426\.0122\-idea setting11Best\-of\-NN,N=122N\{=\}1222428\.028\.294\.992\.79\.754\.93\.724\.08\.812MILO, 15 islands2508\.8240\.998\.598\.417\.154\.43\.3632\.548\.213Inventa957\.9959\.296\.097\.132\.953\.32\.6016\.761\.6241\-idea setting14MILO, 30 islands4898\.9335\.897\.997\.917\.454\.43\.3559\.283\.415Inventa2938\.0867\.994\.896\.125\.853\.22\.4252\.0130\.2
RQ2: Are the discovered scientific laws both novel and consistent with domain constraints?
OnEurekaBench,Inventaalone passes every scientific\-constraint test, and its laws carry far more insight than those of CORAL, the strongest baseline \(67\.0% vs\. 50\.3% SI; 73\.6% vs\. 56\.2% CS; Table\)\. OnSciLaws\-Bench,Inventabeats the published law on 31 of 38 tasks and reaches 62\.8%SNS\_\{N\}, against 28 and 59\.3% for the best baseline \(Table\); its edge lies in tasks that hold out entire groups \(65\.8% vs\. at most 61\.4%\)\. Its validitySVS\_\{V\}\(77\.1%\) also exceeds that of 15 of the 17 baselines\.Inventathus discovers scientific laws that are both novel and consistent with domain constraints\.
RQ3: How does performance scale with the search budget?
Inventaturns extra budget into steady gains, whereas SoTA agents and evolutionary methods level off\. OnEurekaBenchvalidation at equal cost \(Fig\.b\),Inventareaches 64\.7% CS after 10 rounds, against at most 52\.8% for the SoTA evolutionary methods and 47\.4% for the best of agents1–14; after 20 rounds, it reaches 72\.5%, while CORAL, the best baseline, reaches 55\.5%\. OnSciLaws\-Bench, 25×\\timesmore samples \(Best\-of\-50 vs\. Best\-of\-2\) lift the strongest ReAct agent only from 56\.9% to 59\.3%, belowInventa’s 62\.8% \(Table\)\. OnIdeaBench\(Fig\.c; Table\), from the 27\- to the 122\-idea setting,Inventa’s distinct novel ideas rise from 26\.0 to 61\.6 with 95 calls per paper, whereas Best\-of\-NN’s rise only from 4\.7 to 8\.8 with 242 calls and MILO’s reach 48\.2 with 250\. Its lead thus widens as the budget grows\.
RQ4: Do the gains depend on the judge?
We re\-scoreIdeaBenchwith two judges unused in search:Inventastill finds the most distinct novel ideas \(17\.5 and 13\.5 vs\. 13\.1 and 11\.3 for OpenEvolve; paired 90% intervals above zero; Appendix\)\.
## 6Conclusion
We presentedInventa, a framework for scientific discovery in which autonomous research units adapt their own mutators and share only verified findings\. On five benchmarks spanning multiple research stages, it outperforms frontier AI agents, evolutionary search, and AI co\-scientists, discovering more novel ideas and scientific laws that satisfy domain constraints\. By self\-adapting its search,Inventabenefits more from added budget, whereas existing frameworks show diminishing returns\.
## Acknowledgements
We thank Prof\. Vijay Ganesh \(Georgia Institute of Technology\) for his fruitful feedback and insightful discussions throughout this project, which helped us improve this paper\.
## References
- Agarwal et al\. \(2026\)Shubham Agarwal, Akshat Gupta, Mike Heddes, Dilara Soylu, Rui Meng, Patrick Tser Jern Kon, Jiefeng Chen, Mert Cemri, Shu Liu, Abe Gonzalez, Akanksha Jain, Bhavana Dalvi Mishra, Linfeng Song, Vilobh Meshram, Xiangru Tang, Jinsung Yoon, Aditya Parameswaran, Ion Stoica, Vahab Mirrokni, Burak Gokturk, Parthasarathy Ranganathan, Tomas Pfister, Nanyun Peng, and Chun\-Liang Li\.DuoDiscover: Scaling Autonomous Discovery Beyond the Evaluation Bottleneck\.*arXiv Preprint arXiv:2610\.17184*, 2026\.URL[https://arxiv\.org/abs/2610\.17184](https://arxiv.org/abs/2610.17184)\.
- Agrawal et al\. \(2026\)Lakshya A\. Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl\-Ong, Arnav Singhvi, Herumb Shandilya, Michael J\. Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alex Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, and Omar Khattab\.GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning\.In*The Fourteenth International Conference on Learning Representations \(ICLR\)*, 2026\.URL[https://arxiv\.org/abs/2507\.19457](https://arxiv.org/abs/2507.19457)\.
- Antoniades et al\. \(2026\)Antonis Antoniades, Deepak Nathani, Ritam Saha, Alfonso Amayuelas, Ivan Bercovich, Zhaotian Weng, Vignesh Baskaran, Kunal Bhatia, and William Yang Wang\.Heuresis: Search Strategies for Autonomous AI Research Agents Across Quality, Diversity and Novelty\.*arXiv Preprint arXiv:2606\.25198*, 2026\.URL[https://arxiv\.org/abs/2606\.25198](https://arxiv.org/abs/2606.25198)\.
- Baek et al\. \(2025\)Jinheon Baek, Sujay Kumar Jauhar, Silviu Cucerzan, and Sung Ju Hwang\.ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models\.In Luis Chiruzzo, Alan Ritter, and Lu Wang \(eds\.\),*Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\)*, pp\. 6709–6738, Albuquerque, New Mexico, April 2025\. Association for Computational Linguistics\.ISBN 979\-8\-89176\-189\-6\.[10\.18653/v1/2025\.naacl\-long\.342](https://doi.org/10.18653/v1/2025.naacl-long.342)\.URL[https://aclanthology\.org/2025\.naacl\-long\.342/](https://aclanthology.org/2025.naacl-long.342/)\.
- Block \(2025\)Block\.goose\.[https://github\.com/aaif\-goose/goose](https://github.com/aaif-goose/goose), 2025\.An open\-source, extensible AI agent\.
- Brown et al\. \(2024\)Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V\. Le, Christopher Ré, and Azalia Mirhoseini\.Large Language Monkeys: Scaling Inference Compute with Repeated Sampling\.*arXiv Preprint arXiv:2407\.21787*, 2024\.URL[https://arxiv\.org/abs/2407\.21787](https://arxiv.org/abs/2407.21787)\.
- Cai et al\. \(2026\)Ziyang Cai, Christos Ziakas, Vasilis Kontonis, Tim Pearce, Siddhartha Sen, Akshay Krishnamurthy, Shivam Garg, and Dimitris Papailiopoulos\.Fork\-and\-Flush: Escaping Idea Basins in Autoresearch Agents\.*arXiv Preprint arXiv:2610\.07447*, 2026\.URL[https://arxiv\.org/abs/2610\.07447](https://arxiv.org/abs/2610.07447)\.
- Chen et al\. \(2026\)Nuo Chen, Yicheng Tong, Yuzhe Yang, Yufei He, Xueyi Zhang, Zou Qingyun, Qian Wang, and Bingsheng He\.Diversity Collapse in Multi\-Agent LLM Systems: Structural Coupling and Collective Failure in Open\-Ended Idea Generation\.In Maria Liakata, Viviane P\. Moreira, Jiajun Zhang, and David Jurgens \(eds\.\),*Findings of the Association for Computational Linguistics: ACL 2026*, pp\. 251–306, San Diego, California, United States, July 2026\. Association for Computational Linguistics\.ISBN 979\-8\-89176\-395\-1\.[10\.18653/v1/2026\.findings\-acl\.13](https://doi.org/10.18653/v1/2026.findings-acl.13)\.URL[https://aclanthology\.org/2026\.findings\-acl\.13/](https://aclanthology.org/2026.findings-acl.13/)\.
- Cohere \(2024\)Cohere\.Introducing Rerank 3\.5: Precise AI Search\.[https://cohere\.com/blog/rerank\-3pt5](https://cohere.com/blog/rerank-3pt5), 2024\.
- Cohere \(2025\)Cohere\.Introducing Embed 4: Multimodal Search for Business\.[https://cohere\.com/blog/embed\-4](https://cohere.com/blog/embed-4), 2025\.
- Ding et al\. \(2026\)Wenxuan Ding, Qin Lu, Changlong Yu, Sha Li, Shuowei Jin, Xin Liu, Zhenyu Shi, and Greg Durrett\.Harness\-Delta Attribution: Dissecting Gains From Harness Evolution\.*arXiv Preprint arXiv:2610\.22466*, 2026\.URL[https://arxiv\.org/abs/2610\.22466](https://arxiv.org/abs/2610.22466)\.
- Feng et al\. \(2026\)Shiyang Feng, Runmin Ma, Xiangchao Yan, Yue Fan, Yusong Hu, Songtao Huang, Shuaiyu Zhang, Zongsheng Cao, Tianshuo Peng, Jiakang Yuan, Zijie Guo, Zhijie Zhong, Shangheng Du, Weida Wang, Jinxin Shi, Yuhao Zhou, Xiaohan He, Zhiyin Yu, Fangchen Yu, Bihao Zhan, Qihao Zheng, Jiamin Wu, Mianxin Liu, Chi Zhang, Shaowei Hou, Shuya Li, Yankai Jiang, Wenjie Lou, Lilong Wang, Zifu Wang, Jiong Wang, Wanghan Xu, Yue Deng, Dongrui Liu, Yiheng Wang, Wenlong Zhang, Fenghua Ling, Shufei Zhang, Xiaosong Wang, Shuangjia Zheng, Xun Huang, Siqi Sun, Shuyue Hu, Peng Ye, Chunfeng Song, Bin Wang, Conghui He, Yihao Liu, Xin Li, Qibin Hou, Tao Chen, Xiangyu Yue, Bin Wang, Liang He, Dahua Lin, Bowen Zhou, Bo Zhang, and Lei Bai\.InternAgent\-1\.5: A Unified Agentic Framework for Long\-Horizon Autonomous Scientific Discovery\.*arXiv Preprint arXiv:2602\.08990*, 2026\.URL[https://arxiv\.org/abs/2602\.08990](https://arxiv.org/abs/2602.08990)\.
- Geng et al\. \(2026\)Jiayi Geng, Zhengxuan Wu, Kevin S\. Chen, Seungone Kim, Joseph Janssen, Zora Zhiruo Wang, Bhupalee Kalita, Runtian Gao, Aaron Ho, Andrew Oakleigh Nelson, Olexandr Isayev, Francisco Villaescusa\-Navarro, Ching\-Yao Lai, Howard Chen, and Graham Neubig\.EurekaBench: Measuring Agentic Ability to Discover New Scientific Insights\.*arXiv Preprint arXiv:2610\.00492*, 2026\.URL[https://arxiv\.org/abs/2610\.00492](https://arxiv.org/abs/2610.00492)\.
- Gottweis et al\. \(2026\)Juraj Gottweis, Wei\-Hung Weng, Alexander Daryin, Tao Tu, Petar Sirkovic, Artiom Myaskovsky, Grzegorz Glowaty, Felix Weissenberger, Alessio Orlandi, Dan Popovici, Anil Palepu, Keran Rong, Ryutaro Tanno, Khaled Saab, Fan Zhang, Jacob Blum, Andrew Carroll, Kavita Kulkarni, Nenad Tomašev, Dina Zverinski, Ivor Rendulic, Elahe Vedadi, Florian Hasler, Luka Rimanic, Marina Boia, Ivan Budiselic, Ben Feinstein, Mathias Bellaiche, Tom Sheffer, Jan Freyberg, Jeremy Ratcliff, Ottavia Bertolli, Katherine Chou, Avinatan Hassidim, Burak Gokturk, Amin Vahdat, Yuan Guan, Vikram Dhillon, Eeshit Dhaval Vaishnav, Byron Lee, Tiago R\. D\. Costa, José R\. Penadés, Gary Peltz, Yossi Matias, James Manyika, Demis Hassabis, Yunhan Xu, Pushmeet Kohli, Annalisa Pawlosky, Alan Karthikesalingam, and Vivek Natarajan\.Accelerating Scientific Discovery with Co\-Scientist\.*Nature*, 655\(8122\):487–496, 2026\.[10\.1038/s41586\-026\-10644\-y](https://doi.org/10.1038/s41586-026-10644-y)\.
- Guo et al\. \(2025\)Sikun Guo, Amir Hassan Shariatmadari, Guangzhi Xiong, Albert Huang, Myles Kim, Corey M\. Williams, Stefan Bekiranov, and Aidong Zhang\.IdeaBench: Benchmarking Large Language Models for Research Idea Generation\.In*Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, V\.2*, pp\. 5888–5899, 2025\.[10\.1145/3711896\.3737419](https://doi.org/10.1145/3711896.3737419)\.
- Hu et al\. \(2026\)Yuepeng Hu, Pegah Ghahremani, Reza Aghajani, Mahesh Pasupuleti, and Himaghna Bhattacharjee\.Reward Hacking as a Transmissible Strategy in Multi\-Agent Coding Systems\.*arXiv Preprint arXiv:2610\.12822*, 2026\.URL[https://arxiv\.org/abs/2610\.12822](https://arxiv.org/abs/2610.12822)\.
- Huang et al\. \(2026\)Yiming Huang, Ziche Liu, Junxia Cui, Zhuohang Wu, Yiqian Wang, Xinkai Zou, Lingjun Mao, Nan Huang, Naicheng Yu, Kaijie Zhu, Yue Ma, Kun Zhou, Letian Peng, and Jingbo Shang\.Can LLMs Discover Scientific Laws in Real and Parallel Worlds?*arXiv Preprint arXiv:2609\.01552*, 2026\.URL[https://arxiv\.org/abs/2609\.01552](https://arxiv.org/abs/2609.01552)\.
- Jana \(2024\)Prithwish Jana\.NeuroSymbolic LLM for Mathematical Reasoning and Software Engineering\.In*Proceedings of the Thirty\-Third International Joint Conference on Artificial Intelligence \(IJCAI\-24\)*, pp\. 8492–8493\. International Joint Conferences on Artificial Intelligence Organization, 2024\.[10\.24963/ijcai\.2024/961](https://doi.org/10.24963/ijcai.2024/961)\.Doctoral Consortium\.
- Jana et al\. \(2024\)Prithwish Jana, Piyush Jha, Haoyang Ju, Gautham Kishore, Aryan Mahajan, and Vijay Ganesh\.CoTran: An LLM\-Based Code Translator Using Reinforcement Learning with Feedback from Compiler and Symbolic Execution\.In*27th European Conference on Artificial Intelligence \(ECAI\)*, pp\. 4011–4018\. IOS Press, 2024\.[10\.3233/FAIA240968](https://doi.org/10.3233/FAIA240968)\.
- Jana et al\. \(2026a\)Prithwish Jana, Sam Davidson, Bhavana Bhasker, Andrey Kan, Anoop Deoras, and Laurent Callot\.TerraFormer: Automated Infrastructure\-as\-Code with LLMs Fine\-Tuned via Policy\-Guided Verifier Feedback\.In*Proceedings of the IEEE/ACM 48th International Conference on Software Engineering: Software Engineering in Practice*, pp\. 578–589\. ACM, 2026a\.[10\.1145/3786583\.3786898](https://doi.org/10.1145/3786583.3786898)\.
- Jana et al\. \(2026b\)Prithwish Jana, Mononito Goswami, Hao Liu, Xinyu Li, Langlin Huang, Zhehui Huang, Zhishen Huang, Patrick Blöbaum, Anoop Deoras, Purak Jain, Nikos Kanakaris, and Sahika Genc\.MILO: Automated Harness Discovery via Orchestrated Multi\-Agent Evolution\.*arXiv Preprint arXiv:2609\.38349*, 2026b\.URL[https://arxiv\.org/abs/2609\.38349](https://arxiv.org/abs/2609.38349)\.
- Jana et al\. \(2026c\)Prithwish Jana, Kaan Kale, Ahmet Ege Tanriverdi, Cruise Song, Sriram Vishwanath, and Vijay Ganesh\.ProofBridge: Auto\-Formalization of Natural Language Proofs in Lean via Joint Embeddings\.In*The Fourteenth International Conference on Learning Representations \(ICLR\)*, 2026c\.
- Jha et al\. \(2025\)Piyush Jha, Prithwish Jana, Pranavkrishna Suresh, Arnav Arora, and Vijay Ganesh\.RLSF: Fine\-Tuning LLMs via Symbolic Feedback\.In*28th European Conference on Artificial Intelligence \(ECAI\)*, pp\. 1687–1694\. IOS Press, 2025\.[10\.3233/FAIA250996](https://doi.org/10.3233/FAIA250996)\.
- Jiang et al\. \(2025\)Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, and Yejin Choi\.Artificial Hivemind: The Open\-Ended Homogeneity of Language Models \(and Beyond\)\.In*Advances in Neural Information Processing Systems 38 \(NeurIPS 2025\), Datasets and Benchmarks Track*, pp\. 90705–90774, 2025\.[10\.52202/085713\-2732](https://doi.org/10.52202/085713-2732)\.
- Kim et al\. \(2026\)Sohyeon Kim, Yoonho Lee, Bo Liu, Dayoon Ko, Rulin Shao, Seungone Kim, Graham Neubig, Pang Wei Koh, Aakanksha Chowdhery, Akari Asai, Omar Khattab, Yejin Choi, Gunhee Kim, and Chelsea Finn\.ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research\.*arXiv Preprint arXiv:2610\.02202*, 2026\.URL[https://arxiv\.org/abs/2610\.02202](https://arxiv.org/abs/2610.02202)\.
- Kim et al\. \(2025\)Yubin Kim, Ken Gu, Chanwoo Park, Chunjong Park, Samuel Schmidgall, A\. Ali Heydari, Yao Yan, Zhihan Zhang, Yuchen Zhuang, Yun Liu, Mark Malhotra, Paul Pu Liang, Hae Won Park, Yuzhe Yang, Xuhai Xu, Yilun Du, Shwetak Patel, Tim Althoff, Daniel McDuff, and Xin Liu\.Towards a Science of Scaling Agent Systems\.*arXiv Preprint arXiv:2512\.08296*, 2025\.URL[https://arxiv\.org/abs/2512\.08296](https://arxiv.org/abs/2512.08296)\.
- Kung & Robinson \(1981\)H\. T\. Kung and John T\. Robinson\.On Optimistic Methods for Concurrency Control\.*ACM Transactions on Database Systems*, 6\(2\):213–226, 1981\.[10\.1145/319566\.319567](https://doi.org/10.1145/319566.319567)\.
- Lange et al\. \(2026a\)Robert Tjarko Lange, Luca Grillotti, Yutaro Yamada, Rujikorn Charakorn, and David Ha\.The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific \(Co\-\)Discovery\.*arXiv Preprint arXiv:2610\.18933*, 2026a\.URL[https://arxiv\.org/abs/2610\.18933](https://arxiv.org/abs/2610.18933)\.
- Lange et al\. \(2026b\)Robert Tjarko Lange, Yuki Imajuku, and Edoardo Cetin\.ShinkaEvolve: Towards Open\-Ended and Sample\-Efficient Program Evolution\.In*The Fourteenth International Conference on Learning Representations \(ICLR\)*, 2026b\.URL[https://arxiv\.org/abs/2509\.19349](https://arxiv.org/abs/2509.19349)\.
- Lee et al\. \(2026\)Yoonho Lee, Roshen Sanjay Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn\.Meta\-Harness: End\-to\-End Optimization of Model Harnesses\.In*Conference on Language Modeling \(COLM\)*, 2026\.URL[https://arxiv\.org/abs/2603\.28052](https://arxiv.org/abs/2603.28052)\.
- Li \(2026\)Pan Li\.Dictionaries, Not Darwin: Set\-Level Selection Beats LLM Evolution in Scientific Equation Discovery\.*arXiv Preprint arXiv:2607\.04108*, 2026\.URL[https://arxiv\.org/abs/2607\.04108](https://arxiv.org/abs/2607.04108)\.
- Liu et al\. \(2026a\)Jiarui Liu, Renjie Tao, Yiwei Liao, Chuanyang Jin, Kai Sun, Xiao Yang, Xinyuan Zhang, Xilun Chen, Zhuangqun Huang, Lechen Zhang, Yongjin Yang, Yinghui He, Weihao Xuan, Rakesh Wanga, Anuj Kumar, Mona T\. Diab, Wen\-tau Yih, and Xin Luna Dong\.IdeaScientist: Orchestrating Agents for Grounded Scientific Ideation\.*arXiv Preprint arXiv:2610\.04074*, 2026a\.URL[https://arxiv\.org/abs/2610\.04074](https://arxiv.org/abs/2610.04074)\.
- Liu et al\. \(2026b\)Shu Liu, Shubham Agarwal, Monishwaran Maheswaran, Mert Cemri, Qiuyang Mang, Zhifei Li, Ashwin Naren, Ethan Boneh, Audrey Cheng, Alexander Du, Melissa Pan, Kurt Keutzer, Alvin Cheung, Koushik Sen, Alex Dimakis, Matei Zaharia, and Ion Stoica\.EvoX: Meta\-Evolution for Automated Discovery\.In*Conference on Language Modeling \(COLM\)*, 2026b\.URL[https://openreview\.net/forum?id=puRQNPwKs1](https://openreview.net/forum?id=puRQNPwKs1)\.
- Lu et al\. \(2026\)Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foerster, David Ha, and Jeff Clune\.Towards End\-to\-End Automation of AI Research\.*Nature*, 651\(8107\):914–919, 2026\.[10\.1038/s41586\-026\-10265\-5](https://doi.org/10.1038/s41586-026-10265-5)\.
- Novikov et al\. \(2025\)Alexander Novikov, Ngân Vũ, Marvin Eisenberger, Emilien Dupont, Po\-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J\. R\. Ruiz, Abbas Mehrabian, M\. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog\.AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery\.*arXiv Preprint arXiv:2506\.13131*, 2025\.URL[https://arxiv\.org/abs/2506\.13131](https://arxiv.org/abs/2506.13131)\.
- Qu et al\. \(2026\)Ao Qu, Han Zheng, Zijian Zhou, Yihao Yan, Yihong Tang, Shao Yong Ong, Fenglu Hong, Kaichen Zhou, Chonghe Jiang, Minwei Kong, Jiacheng Zhu, Sirui Li, Cathy Wu, Bryan Kian Hsiang Low, Jinhua Zhao, and Paul Pu Liang\.CORAL: Towards Autonomous Multi\-Agent Evolution for Open\-Ended Discovery\.In*Conference on Language Modeling \(COLM\)*, 2026\.URL[https://openreview\.net/forum?id=QcAiEdP4aF](https://openreview.net/forum?id=QcAiEdP4aF)\.
- Ran et al\. \(2026\)Nian Ran, Tianhe Jia, Zhengyang Mu, Yue Wang, Sebastian Brulin, Markus Olhofer, Mauricio A\. Álvarez, Yang Liu, and Richard Allmendinger\.An Important Bottleneck in LLM Optimizers: Principle\-Level Local Optima\.*arXiv Preprint arXiv:2610\.24291*, 2026\.URL[https://arxiv\.org/abs/2610\.24291](https://arxiv.org/abs/2610.24291)\.
- Robertson & Zaragoza \(2009\)Stephen Robertson and Hugo Zaragoza\.The Probabilistic Relevance Framework: BM25 and Beyond\.*Foundations and Trends in Information Retrieval*, 3\(4\):333–389, 2009\.[10\.1561/1500000019](https://doi.org/10.1561/1500000019)\.
- Romera\-Paredes et al\. \(2024\)Bernardino Romera\-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M\. Pawan Kumar, Emilien Dupont, Francisco J\. R\. Ruiz, Jordan S\. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi\.Mathematical Discoveries from Program Search with Large Language Models\.*Nature*, 625\(7995\):468–475, 2024\.[10\.1038/s41586\-023\-06924\-6](https://doi.org/10.1038/s41586-023-06924-6)\.
- Russo et al\. \(2018\)Daniel J\. Russo, Benjamin Van Roy, Abbas Kazerouni, Ian Osband, and Zheng Wen\.A Tutorial on Thompson Sampling\.*Foundations and Trends in Machine Learning*, 11\(1\):1–96, 2018\.[10\.1561/2200000070](https://doi.org/10.1561/2200000070)\.
- Schmidgall & Moor \(2025\)Samuel Schmidgall and Michael Moor\.AgentRxiv: Towards Collaborative Autonomous Research\.*arXiv Preprint arXiv:2503\.18102*, 2025\.URL[https://arxiv\.org/abs/2503\.18102](https://arxiv.org/abs/2503.18102)\.
- Schopf & Färber \(2026\)Tim Schopf and Michael Färber\.Is This Idea Novel? An Automated Benchmark for Judgment of Research Ideas\.In*Proceedings of the Fifteenth Language Resources and Evaluation Conference \(LREC 2026\)*, pp\. 4716–4727, 2026\.[10\.63317/4c3gy3f7epnj](https://doi.org/10.63317/4c3gy3f7epnj)\.
- Sharma \(2025\)Asankhaya Sharma\.OpenEvolve: An Open\-Source Evolutionary Coding Agent\.GitHub, 2025\.URL[https://github\.com/algorithmicsuperintelligence/openevolve](https://github.com/algorithmicsuperintelligence/openevolve)\.
- Si et al\. \(2025\)Chenglei Si, Diyi Yang, and Tatsunori Hashimoto\.Can LLMs Generate Novel Research Ideas? A Large\-Scale Human Study with 100\+ NLP Researchers\.In*The Thirteenth International Conference on Learning Representations \(ICLR\)*, 2025\.URL[https://arxiv\.org/abs/2409\.04109](https://arxiv.org/abs/2409.04109)\.
- Srikanth et al\. \(2026\)Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, and Zhengyao Jiang\.Recursive Self\-Improvement of AI Research Agents\.*arXiv Preprint arXiv:2609\.26457*, 2026\.
- SST \(2025\)SST\.OpenCode\.[https://github\.com/anomalyco/opencode](https://github.com/anomalyco/opencode), 2025\.
- Wang et al\. \(2025\)Xingyao Wang, Boxuan Li, Yufan Song, Frank F\. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H\. Tran, Fuqiang Li, Ren Ma, Mingzhang Zheng, Bill Qian, Yanjun Shao, Niklas Muennighoff, Yizhe Zhang, Binyuan Hui, Junyang Lin, Robert Brennan, Hao Peng, Heng Ji, and Graham Neubig\.OpenHands: An Open Platform for AI Software Developers as Generalist Agents\.In*The Thirteenth International Conference on Learning Representations \(ICLR\)*, 2025\.URL[https://arxiv\.org/abs/2407\.16741](https://arxiv.org/abs/2407.16741)\.
- Yoon et al\. \(2026\)Soyoung Yoon, Boyi Liu, Yite Wang, Ruofan Wu, Canwen Xu, Nikki Lijing Kuang, Seung\-won Hwang, Yuxiong He, and Zhewei Yao\.ArcticSwarm: Deferring Early Consensus in Long\-Horizon Multi\-Agent Research\.*arXiv Preprint arXiv:2609\.01870*, 2026\.URL[https://arxiv\.org/abs/2609\.01870](https://arxiv.org/abs/2609.01870)\.
- Zhang et al\. \(2026\)Jenny Zhang, Shengran Hu, Cong Lu, Robert Tjarko Lange, and Jeff Clune\.Darwin Gödel Machine: Open\-Ended Evolution of Self\-Improving Agents\.In*The Fourteenth International Conference on Learning Representations \(ICLR\)*, 2026\.URL[https://arxiv\.org/abs/2505\.22954](https://arxiv.org/abs/2505.22954)\.
## Appendix
## Appendix AExperimental Setup
### A\.1Reproducibility and Implementation Details
Machine configuration\.Our experiments run on CPU\-only x86\-64 instances\. A typical instance has 192 vCPUs \(96 physical cores across two sockets, Intel Xeon Platinum 8488C Sapphire Rapids, two threads per core\), 1\.5 TiB of memory across two NUMA nodes, and a 1 TB SSD\-backed root volume, and runs Linux \(kernel 6\.1\)\. The instances host only the search loop, verifiers, and task sandboxes, and all model inference except the IdeaScientist judge \(Appendix\) runs remotely through a managed inference service\. The paragraphs below give the mutator pool and hyperparameters, then describe scheduling, concurrent updates to shared knowledge, sandboxing and data isolation, and verification thresholds\.
Mutator pool and hyperparameters\.The initial harness setℋ\\mathcal\{H\}contains a minimal tool\-calling harness of our own, mini\-SWE\-agent, OpenCode, OpenHands, and Goose\. The model setℳ\\mathcal\{M\}contains Kimi K3, Grok 4\.7, GLM 5\.3, MiniMax M2\.5, and four other frontier models, each run at its highest reasoning effort and full output limit, and𝒢\\mathcal\{G\}holds their compatible pairs\. On EurekaBench,InventarunsN=20N=20research units per task for up toB=80B=80rounds\. ScholarCatalyst usesN=10N=10units and up to 80 rounds, and SciLaws\-BenchN=10N=10units and up to 50 rounds\. All three use patienceP=2P=2and lagλ=3\\lambda=3and cap each agent session at four hours \(90 minutes on SciLaws\-Bench\)\. A harness trial twins every attempt and is decided after four twin pairs\. On IdeaBench,N=4N=4units evolve an idea\-generation program overB=12B=12rounds, and the frozen program then runs on each test paper\. The 122\- and 241\-idea settings use 10 and 8 rounds\. On IdeaScientist,N=4N=4units runB=6B=6rounds per problem\.
Asynchronous execution\.AlgorithmrunsNNresearch units in parallel\. In each round, a unit reads a snapshot of𝒦\\mathcal\{K\}, selects a parent, and generates a child\. Verifiers evaluate the child, return its certificate, and publish qualifying findings and model\-level statistics\. Like symbolic feedback for fine\-tuning LLMs\([Jha et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib23)\), the certificate reports every check rather than a single score\. The unit then updates its lineage, and orchestrator actions adapt the search \(§\)\. Units start and finish rounds independently, and a scheduler decides when each may start its next round\.
Bounded staleness\.Unitjjstarts a round only if
rj<minj′∈activerj′\+λ,r\_\{j\}<\\min\_\{j^\{\\prime\}\\in\\mathrm\{active\}\}r\_\{j^\{\\prime\}\}\+\\lambda,\(2\)where active units exclude those paused after repeated failures or treated as hung\. Among ready units, those with fewer completed rounds take priority, and twin rounds for harness comparison are exempt from the bound\. Withλ=1\\lambda=1, units advance roughly in lockstep\. Larger values let faster units advance further while limiting differences in the shared evidence each unit sees\. At fixed evaluation checkpoints, non\-hung units synchronize and the selected candidate is frozen for evaluation\. Test data never feeds into search or selection\.
Optimistic concurrency\.Each unit works from the shared\-knowledge snapshot taken when its round begins\. Before publishing a candidate, verifiers atomically re\-read the best verified candidate and apply the publication rule\. If another unit has since improved that reference, the candidate must exceed the updated reference byδ\\deltato enter𝒦ver\\mathcal\{K\}\_\{\\mathrm\{ver\}\}, while its other qualifying findings remain eligible\. The best verified score is thus nondecreasing despite stale snapshots\([Kung & Robinson, 1981](https://arxiv.org/html/2611.21847#bib.bib27)\)\.
Sandboxing and data isolation\.Each mutator runs in a sandbox without internet access that exposes only the training data and the task workspace\. Validation data and verifier code stay outside the sandbox, and units see validation results only through certificates\. The test split is held outside the search environment, where no unit or verifier can access it, and is scored only on candidates frozen at checkpoints\. An attempt that tries to read restricted data or verifier files is voided and updates neither its lineage nor the shared knowledge\.
Verification thresholds\.Pairing LLM mutators with programmatic verifiers makesInventaa neurosymbolic design\([Jana, 2024](https://arxiv.org/html/2611.21847#bib.bib18);[Jana et al\., 2026c](https://arxiv.org/html/2611.21847#bib.bib22)\)\. For component ablations, the implementation uses a separate marginδabl\\delta\_\{\\mathrm\{abl\}\}in place ofδ\\delta, withδabl=2δ\\delta\_\{\\mathrm\{abl\}\}=2\\deltaon EurekaBench\. It also rejects verification when a component’s measured contribution exceedsγ=5\\gamma=5times the larger of the candidate’s gain and the ablation margin:
Abltmax\(x\)\>γmax\{Gain\(x\),δabl\}\.\\textsc\{Abl\}^\{\\max\}\_\{t\}\(x\)\>\\gamma\\max\\\{\\textsc\{Gain\}\(x\),\\delta\_\{\\mathrm\{abl\}\}\\\}\.\(3\)This heuristic is intended to flag candidates whose components may offset one another’s effects\.
### A\.2Benchmark Details
Below, we detail each benchmark of §, including our IdeaBench and IdeaScientist protocols\.
EurekaBench\.We use the 16 public CPU\-only tasks of EurekaBench\([Geng et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib13)\): three plasma\-physics tasks on the ARC and negative\-triangularity tokamaks, seven astrophysics tasks on CAMELS and DREAMS cosmological simulations, and six neuroscience tasks on*C\. elegans*models\. Each task runs in an experimental environment, and an agent must propose a mechanism that predicts held\-out data, answers expert insight questions, and respects domain constraints \(Table\)\.
SciLaws\-Bench\.SciLaws\-Bench\([Huang et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib17)\)contains 118 scientific law\-discovery problems from six disciplines, including materials and engineering, ecology and hydrology, and the social sciences, drawn from 381 papers and about 8 million measured, observed, and simulated data points\. A system must find a closed\-form law that generalizes: single\-group problems test extrapolation or transfer across time and conditions, and multi\-group problems hold out entire groups, such as unseen crystals in an optical\-dispersion task\. We use 38 of the 118 tasks, selected by rules fixed before our search: from the 99 permissively licensed tasks, we keep those meeting four criteria:
1. 1\.A validation split can be built from the training data the way the test split was\.
2. 2\.The task provides sufficient data: at least 50 test and 50 validation rows for single\-group tasks, or at least eight test groups and 15 training groups for multi\-group tasks\.
3. 3\.Its estimated ceiling exceeds 55\. The ceiling is the maximum of a noise estimate, a flexible non\-law fit, and the best per\-task score of the benchmark’s nine baseline models\.
4. 4\.The published reference law can be reproduced exactly in the scored submission format\.
The first three criteria leave 40 tasks, and the reproducibility criterion removes two more, leaving 16 single\-group and 22 multi\-group tasks\. Every method in Tableuses this 38\-task subset\.
ScholarCatalyst\.ScholarCatalyst\([Kim et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib25)\)asks a system to retrieve the papers that inspired, or could have inspired, a research project\. Its 207 source papers are 2025–2026 computer\-science papers spanning machine learning, vision, language, AI, and robotics, over 100 of them orals, spotlights, or award winners, and their lead authors labelled the positives\. Systems search a 190,896\-paper corpus, restricted per query to papers that predate its source paper\. Inspiration often crosses fields: nearly half of the credited papers lie outside the project’s subject area, and positives are no more similar to the query than related papers the authors rejected\. We evaluate on all 894 test queries: 207 core research queries and 687 subfield\-specific ones\. Following[Kim et al\. \(2026\)](https://arxiv.org/html/2611.21847#bib.bib25), who recommend a model whose knowledge cutoff precedes every paper in the task and corpus, all methods use DeepSeek V3\.2, so all rows share one model\.
IdeaBench\.IdeaBench\([Guo et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib15)\)pairs 2,374 influential biomedical papers published in 2024 with 23,460 references, research articles with at least five citations that each paper cites in its background section\. A method reads the abstracts of five of a paper’s references and proposes research ideas, which are judged against the paper’s own, unseen idea\. We split the papers by a salted hash of their identifiers into 904 training, 734 validation, and 736 test papers\.Inventa’s units evolve an idea\-generation program using five training papers and are scored on 8–12 validation papers, while validation papers 50–149 form a 100\-paper check set excluded from search\. The frozen program then runs on every test paper\.
IdeaScientist\.IdeaScientist\([Liu et al\., 2026a](https://arxiv.org/html/2611.21847#bib.bib32)\)asks a system to write a grounded research proposal for a research problem, given as a problem definition and its central challenge, drawn from a machine\-learning methodology paper published after January 1, 2026\. Systems search an offline corpus of 2\.77 million research ideas decomposed from arXiv papers in all subject areas, reading only papers that predate the problem’s paper\. Useful mechanisms may come from seemingly unrelated fields that share the problem’s structure, and the system must anticipate a direction before it appeared\. We evaluate on its 277 test problems, with oneInventasearch per problem\. A benchmark plug\-in provides IdeaScientist’s judge, paper corpus, and renderer, and scoring uses the benchmark’s own code\. A candidate is a proposal in the benchmark’s results\-masked schema, with a rationale and a signed ledger of papers read\. We serve the benchmark’s judge, Qwen3\.6\-27B, on GPU nodes with SGLang, with reasoning disabled, temperature 0\.7, and top\-p 0\.8\. Re\-scoring stored proposals from earlier systems reproduces scores within 1%\. Agents run unprivileged inside the judge’s container, with private benchmark files hidden and external literature and search hosts blocked\.
### A\.3Task Examples
Each example instantiates the task tupleτ=⟨q,𝒲,𝒞,𝒱⟩\\tau=\\langle q,\\mathcal\{W\},\\mathcal\{C\},\\mathcal\{V\}\\rangleof Def\.for one task class\.
Scientific law discovery\.On EurekaBench’s ice\-shelf task,qqasks why a textbook flow law misfits measured ice\-shelf flow;𝒲\\mathcal\{W\}provides velocity and thickness fields from a FEniCSx finite\-element simulator;𝒞\\mathcal\{C\}holds expert\-written physical constraints, such as thermodynamic laws; and𝒱\\mathcal\{V\}runs the candidate on held\-out flow regimes, tests the constraints, and poses insight questions that LLM agents may answer by querying the simulator\. The candidate is the flow law, coded and explained, so verifiers execute it, as compilers and symbolic execution\([Jana et al\., 2024](https://arxiv.org/html/2611.21847#bib.bib19)\)check generated code\.
Literature retrieval\.On ScholarCatalyst,qqasks which prior work could advance a recent project’s central research question, or the same question seen from one subfield;𝒲\\mathcal\{W\}is a 190,896\-paper corpus restricted to papers that predate the source paper; and𝒱\\mathcal\{V\}scores the ranked list against the papers the source paper’s authors labelled as inspiring\. The candidate is a retrieval method, as a program\.
Idea generation\.On IdeaBench,qqasks for research ideas that build on a paper’s references;𝒲\\mathcal\{W\}holds the abstracts of five of them; and𝒱\\mathcal\{V\}rates every idea with an expert\-calibrated LLM judge and merges near\-duplicates for counting\. The candidate is an idea\-generation program, evolved on training and validation papers, then frozen for testing\.
Grounded proposals\.On IdeaScientist,qqis a research problem, a problem definition with its central challenge, from a methodology paper published in 2026;𝒲\\mathcal\{W\}offers search, bibliography, and paper\-reading tools over pre\-cutoff papers;𝒞\\mathcal\{C\}requires a schema\-valid proposal that cites at least one paper, with at least 90% of cited papers read per its ledger; and𝒱\\mathcal\{V\}is the benchmark’s LLM judge of novelty and quality\. The candidate is a proposal with a rationale and a ledger of papers read\.
### A\.4Evaluation Metrics
This section expands the metric summary of §with each benchmark’s definitions and scoring\.
EurekaBench\.Each task provides expert\-written criteria of three kinds\([Geng et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib13)\)\. Predictive accuracy \(PA\) executes the submitted mechanism on held\-out test data, from the original studies or the task’s simulator, and normalizes each criterion to\[0,1\]\[0,1\]under the task’s rubric\. Scientific\-constraint \(SC\) tests check that the mechanism does not contradict the domain’s assumptions, and scientific\-insight \(SI\) questions credit an insight only when the judge finds complete evidence for it\. Two agentic LLM judges evaluate both and may run extra simulator experiments for SI\. The conditioned score \(CS\) pools a judge’s PA criterion scores and SI question scores into one mean, is zero unless every SC test passes, and is averaged over the two judges\. We report CS as the primary measure, PA and SI unconditionally, and SC as the number of tasks that fail either judge’s tests\.
SciLaws\-Bench\.Numeric fit \(SNS\_\{N\}\) is computed programmatically on the held\-out test split with a task\-specific fit metric \(RMSE, sMAPE, log\-MAE, orR2R^\{2\}\), normalized so that the best published law scores 0\.5 and a perfect predictor 1\([Huang et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib17)\); we report it×100\\times 100\. Multi\-group tasks average over groups, and a failed group scores zero\. We report single\-group, multi\-group, and all tasks, with a 95% task\-bootstrap confidence interval for the overall mean, and, following the benchmark, count a task as exceeding the published reference whenSN\>52S\_\{N\}\>52\(0\.52 on the benchmark’s scale\)\. Scientific validity \(SVS\_\{V\}\) is the fraction of a task’s rubric items satisfied, checked by a code\-executing LLM judge through probes and source inspection\. A submission that fails the anti\-hacking item, which rejects degrees of freedom moved outside the declared law, receives zero validity\.
Table 6:ScholarCatalyst by query type\.Mean: average of the eight cells, the primary metric of Table\.\+ real metadata: real document IDs and years\.Bold/underline: best/second\.Core queries \(207; %\)↑\\uparrowSubfield queries \(687; %\)↑\\uparrowMethodnDCG@20Recall@5Recall@20Recall@100nDCG@20Recall@5Recall@20Recall@100Mean↑\\uparrowRetrievers without a language model1BM2515\.612\.022\.639\.420\.811\.926\.045\.424\.22Cohere Embed v424\.919\.034\.352\.331\.719\.238\.763\.435\.43Cohere v4 \+ BM25 fusion22\.417\.231\.552\.630\.118\.537\.063\.434\.14Cohere v4 \+ Rerank 3\.523\.320\.635\.252\.331\.418\.440\.163\.435\.6DeepSeek V3\.2 agents \(Cohere v4 retriever\)5Tool\-calling Agent23\.616\.333\.852\.930\.219\.636\.763\.134\.56\+ real metadata22\.817\.934\.353\.330\.119\.236\.763\.134\.77Deep\-research Agent25\.019\.432\.551\.831\.720\.237\.463\.835\.28OpenCode25\.719\.633\.553\.333\.422\.339\.864\.236\.59OpenHands25\.221\.936\.252\.832\.421\.338\.961\.936\.310Seed agent \(initialization\)22\.518\.131\.950\.228\.719\.633\.460\.933\.111Inventa32\.025\.945\.361\.641\.227\.250\.772\.444\.5
ScholarCatalyst\.The positives for each query are the candidate papers that the source paper’s lead author labelled as having advanced, or as able to advance, the project\([Kim et al\., 2026](https://arxiv.org/html/2611.21847#bib.bib25)\)\. Recall@kkis the fraction of a query’s positives retrieved in the topkkresults, and nDCG@20 additionally rewards ranking them higher; both are computed programmatically from these labels\. We report nDCG@20 and Recall@5, @20, and @100 for core research queries and subfield\-specific queries\. The benchmark defines no aggregate; ourMean\(Tablesand\) is the unweighted mean of the eight values\.
IdeaBench\.The benchmark’s insight scores for novelty and feasibility measure how often an LLM jury ranks a method’s ideas above the paper’s own idea, without revealing which is the original\([Guo et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib15)\)\. In our setup, a three\-model jury ranks each method’s top three ideas; the published jury uses the writer models, and the disjoint jury shares no model with any writer\. Two further measures compare ideas with the paper’s abstract, BERTScore F1 \(*Semantic*\) and an LLM\-rated idea overlap \(*Overlap*\); neither has a preferred direction\. We add measures from our own novelty judge, a frontier LLM calibrated on RINoBench \(Appendix\) that rates every generated idea from 1 to 10: the number of distinct ideas per paper rated at least 7 \(our primary measure\) or at least 8, merging near\-duplicates among them \(Cohere Embed v4, cosine 0\.8\); the best idea’s score \(maximum novelty\); and the share of all ideas surviving a merge over the whole pool \(non\-duplicate rate\)\.
IdeaScientist\.All measures are scored by the authors’ judge, Qwen3\.6\-27B, served locally with their sampling settings\([Liu et al\., 2026a](https://arxiv.org/html/2611.21847#bib.bib32)\)\. Novelty has three measures:*All*, a 1–10 judgment of how much the proposal adds beyond the closest papers in the corpus;*Cutoff*, the same judgment against pre\-cutoff papers; and*In\-domain*, the share of proposals judged to transfer a mechanism across problem settings through a shared structure\. Proposal quality averages six 1–10 peer\-review criteria: relevance, clarity, specificity, actionability, soundness, and impact\. All measures use a 0–100 scale, and the combined score is the mean of the novelty and quality averages\. A problem without a scorable proposal scores zero throughout\.
Table 7:IdeaBench under three judges\(test papers\)\. The same ideas re\-scored by two judges unused in the loop: a second frontier\-model judge \(736 papers\) and Grok 4\.7 \(a fixed sample of 200\)\. Outside judges are stricter, so compare within a judge\. DN8/DN7: distinct ideas per paper rated≥\\geq8/≥\\geq7 \(DN7: primary\)\.Bold/underline: best/second per judge\.
Table 8:Ablation ofInventa’s refinement\(100 validation papers, no test data\)\. Each row replaces one of four design choices with a simpler alternative\. Weak: share of ideas rated≤\\leq6\.Bold/underline: best/second\.Table 9:Novelty judge calibration on RINoBench\.Cut points fitted on training ideas map the judge’s 1–10 scores to the experts’ 1–5 levels\. The best judges of[Schopf & Färber \(2026\)](https://arxiv.org/html/2611.21847#bib.bib42)reach macro F1 0\.172 and 0\.162 on test\.Bold: best per round; shaded: ours\.
### A\.5Novelty Judge Calibration on RINoBench
RINoBench\([Schopf & Färber, 2026](https://arxiv.org/html/2611.21847#bib.bib42)\)pairs 1,381 research ideas from ICLR 2022 and 2023 submissions with novelty levels on a 1–5 scale that bin the averaged novelty scores of expert reviewers\. We use it only to select and calibrate the IdeaBench novelty judge\. We screen five candidate judges and three rubrics on 100 of its 1,104 training ideas, evaluate the surviving configurations on all training ideas, and fit cut points from the judges’ 1–10 scores to the five levels by cross\-validation \(Table\)\. Under the fitted mapping, a score of 7 corresponds to*novel*and a score of at least 8 to*highly innovative*\. On the 277 test ideas, the selected judge reaches a macro F1 of 0\.262, against 0\.172 for the best judge the benchmark’s authors report\. These categories are calibrated to reviewer labels, not to reviews of IdeaBench outputs\.
## Appendix BAdditional Results
### B\.1Scientific Law Discovery: EurekaBench
After 10 rounds,Inventa’s best validation CS exceeds the best validation score over all runs of Table’s 14 frontier agents on 15 of 16 tasks, and after 20 rounds on all 16 \(means of 64\.7% and 72\.5% vs\. 48\.4%\)\. The verifiers published 2,502 findings \(239Verified, 395Partial, 1,868Refuted\)\. On 11 of 16 tasks, the validation\-best law comes from a lineage thatRestartre\-seeded, and 90 of the 320 units end on an evolved harness\.
### B\.2Retrieving Prior Work: ScholarCatalyst by Query Type
Inventaleads in all eight cells \(Table\), most on subfield Recall@20 \(50\.7% vs\. 40\.1% for Cohere Embed v4 \+ Rerank 3\.5, the next best\)\.
Table 10:IdeaBench with weaker writers\(736 test papers\)\. Every method uses the same writer; judge, juries, and scoring as in Table\. DN7: primary\. Insight: disjoint jury\.Bold/underline: best/second per writer\.
Table 11:Robustness of Insight Novelty\(×100\\times 100\)\. Pub\.: IdeaBench’s protocol, with the writer pool as jury\. A: each top\-3 idea and the paper’s idea rewritten into one neutral 120–160\-word paragraph \(style normalization;[Si et al\., 2025](https://arxiv.org/html/2611.21847#bib.bib44)\)\. B: a jury sharing no model with any writer\. C: top three by novelty alone\. D: A\+B\+C\. At 27 ideas,Inventascores 96\.8 \(Pub\.\) and 97\.4 \(B\)\.
### B\.3Generating Ideas: IdeaBench
Ideas and model calls\.Without per\-paper refinement,Inventareturns 23\.4 distinct novel ideas with 24 calls\. Refinement raises its maximum novelty from 7\.95 to 8\.88 and its ideas rated≥\\geq8 from 11\.2 to 18\.4\. At 241 ideas, MILO with 30 islands has more ideas rated≥\\geq8 \(59\.2 vs\. 52\.0\) and the best maximum novelty \(8\.93\), and OpenEvolve has the highest Insight Novelty \(98\.7% vs\. 96\.8%; Tablere\-scores it under four protocol changes\)\.
Robustness to the judge\.The primary judge also guides selection inside MILO, OpenEvolve, andInventa’s refinement, so we re\-score the test ideas with two judges unused in the loop: a second frontier\-model judge on all 736 papers and Grok 4\.7 on a fixed sample of 200 \(Table\)\.Inventastill returns the most distinct novel ideas, 17\.5 and 13\.5 against at most 13\.1 and 11\.3 for OpenEvolve, with all paired 90% bootstrap intervals above zero\. Its maximum novelty is close to OpenEvolve’s, which has slightly more ideas rated≥\\geq8 under both judges\.
Ablation of the refinement stage\.Removing one design choice of the refinement at a time on 100 validation papers shows that each helps \(Table\): asking for both novelty and feasibility lowers maximum novelty from 8\.81 to 7\.99; accepting rewrites by their own score rather than the paper\-level aggregate lowers highly innovative ideas from 17\.5 to 12\.8; rewriting only the best ideas also gives 12\.8 and doubles the weak\-idea share \(7\.4% to 14\.3%\); and 10 instead of 18 rewrites gives 14\.7\.
Single writers\.With each of three lower\-cost writers alone \(Table\),Inventahas the highest maximum novelty \(8\.48 with DeepSeek V3\.2, 7\.43 with MiniMax M2\.5, 8\.91 with Kimi K3\) and the most distinct novel ideas with Kimi K3 \(25\.1 vs\. 21\.1 for GEPA\); GEPA leads with DeepSeek V3\.2 \(22\.7 vs\. 18\.1\) and MiniMax M2\.5 \(6\.7 vs\. 5\.9\)\.
### B\.4Grounded Proposals: IdeaScientist
Inventareaches a combined score of 82\.7%, against 77\.7% for IdeaScientist\-Base and 74\.6% for IdeaScientist\-Trained, as reported by its authors \(Table\)\. The gain comes from In\-domain: 93\.9% of its proposals transfer a mechanism across problem settings, against 54\.9% and 70\.9%\.
## Appendix CPrompts and Harnesses
The boxes below give the mutator’s initial harness and the orchestrator’s actions and prompts on EurekaBench, abridged for readability;⟨\\langlefield⟩\\ranglemarks a field filled at run time\. Before its twin trial, everyRetooledit passes a validator that rejects any scientific content, a policy check in the spirit of policy\-guided verifier feedback\([Jana et al\., 2026a](https://arxiv.org/html/2611.21847#bib.bib20)\)\.
Mutatorμj\\mu\_\{j\}: initial harnessevolves byRetoolEvery unit starts from this minimal tool\-calling harness: a control\-flow program \(run\.py\) with its prompts, procedure, tools, and limits\.Retoolkeeps rewriting all of them, the program included, so each unit’s harness evolves over time\.SYSTEM PROMPTYou are a careful research scientist working inside an automated discovery system\. You write mechanisms as runnable code plus a plain explanation\. You never guess at numbers you did not compute, you say what you assume, and you prefer one well\-reasoned change to many unmotivated ones\. You do not try to game the evaluation: a mechanism must explain the observation, not merely fit it\.Your method: state precisely where the classical law fails and where it holds; name the assumptions it rests on; relax the one your research direction questions, and only that one; derive the modified mechanism so that every new term has a physical reading and the right dimensions; predict a signature that separates it from the classical law; check the limiting case in which the new effect vanishes and the classical law must return; check signs, positivity, and bounds of every field; name your additive terms so each can be tested alone; say what would refute you\. Do not look up the published answer; a copied answer is not a discovery, and a term you cannot explain is a term the ablation test will expose\.You work in a workspace through the toolsbash,write\_file,read\_file, andlist\_files\. Start by readingAGENTS\.md\. Finish by writing the deliverables andREPLY\.mdinto\./mechanisms, then reply with one line saying you are done; do not paste the files in the reply\.FIRST MESSAGEReadAGENTS\.mdand do what it asks\. Work untilmechanisms/holds every deliverable andREPLY\.md\.TASK FILE \(AGENTS\.md\)You are one research unit in a population searching for the mechanism behind a scientific observation\. Work in this directory\. Write your deliverables \(⟨\\langledeliverables⟩\\rangle\) into\./mechanisms/; the executable deliverable⟨\\langlecode file⟩\\ranglemust follow the contract inTASK\.mdexactly\. You may also hand in supporting evidence there, such as the analysis scripts, result tables, and small data files behind your findings; the reviewers read it\. Keep drafts and scratch files in\./tmp\. Files here:TASK\.mdthe task and the deliverable contractENVIRONMENT\.mdhow to use the simulator and dataPROPOSAL\.mdyour unit’s declared target and assumptionsPARENT/the parent mechanism, withCERTIFICATE\.md: what the verifiers measured on it, criterion by criterion and regime by regime \(its scores against the best verified idea, the transfer test, the ablation of every term, and every constraint with the judge’s reason\)LINEAGE\.mdone line per attempt of your unit, from its certificate;LINEAGE/⟨\\langleround⟩\\rangle/holds the certificate and deliverables of each recent attemptPAGE\.mdyour snapshot of the shared knowledge: the best verified idea with its evidence, thePartialfindings, and theRefutedfindings touching your assumptionsLITERATURE\.mdthe whole shared knowledge as a grouped index with the scoreboardReadPARENT/CERTIFICATE\.mdfirst\. InREPLY\.md, write aDIAGNOSISsection \(which criteria and regimes are weak, what the certificate says about transfer, ablation, and constraints, and what you will change because of it\) before theTARGET; then make ONE change that addresses the diagnosis\. Do not repeat the best verified idea, build on aPartialfinding where it fits, and avoid refuted directions and failed attempts\. Reply format:\#\#\# DIAGNOSIS⟨\\langlea short paragraph⟩\\rangle\#\#\# TARGET⟨\\langleone sentence⟩\\rangle \#\#\# ASSUMPTIONS\-⟨\\langletag⟩\\rangle\#\#\# EXPLANATION⟨\\langleone or two paragraphs⟩\\rangle \#\#\# TERMS\["⟨\\langleterm name⟩\\rangle", \.\.\.\]PROCEDURE \(HARNESS\.md\)Work through the tools: readAGENTS\.mdfirst, then the proposal, the parent, and the records\. Write every deliverable into\./mechanismsearly, as soon as you have a first complete version, then use the provided tools to test and refine it; a fit can take minutes, so budget at most two or three fits and keep the deliverables current after each\. Never end a turn budget without\./mechanismsholding every deliverable andREPLY\.md\.TOOLS AND LIMITSFour tools,bash\(output cut at 20,000 characters\),write\_file,read\_file, andlist\_files, withmax\_turns= 1,000 tool rounds and adeadlineat the four\-hour session cap\. The sandbox has no internet access, and validation and test data stayoutsideit \(Appendix\)\.CONTROL FLOW \(run\.py\)defpropose\(ctx, model\):\# one attempt of the mutatorsystem = SYSTEM\_PROMPT \+ TOOLS\_NOTEmsgs = \[user\(FIRST\_MESSAGE\)\]whileturns < max\_turnsandnow\(\) < deadline:reply = model\.chat\(system, msgs, tools=TOOLS\)msgs\.append\(reply\)ifnotreply\.tool\_calls:\# the model stoppedmissing = missing\_deliverables\(ctx\)ifmissingandcontinued < 3:msgs\.append\(user\(continue\_note\(missing\)\)\)continue\# resume the attemptbreakmsgs\.append\(run\_tools\(workspace, reply\.tool\_calls\)\)write\_transcript\(workspace\)\# the trace Retool reads
Orchestrator𝒪j\\mathcal\{O\}\_\{j\}: actions and promptsadapts its unitThe orchestrator runs around every round of its unit and takes only the three actions of §\.RetoolandRestartcall a model with the prompts below;Redrawneeds no prompt\.ACTIONSRetoolEveryPProunds, when no twin trial is open: ask the harness editor for one child harnessh′h^\{\\prime\}, runhhandh′h^\{\\prime\}on the same parent and mutation plan, and adopth′h^\{\\prime\}only if thennscored pairs give a significantly positive mean difference; otherwise keephh\.RedrawWhen the unit stalls \(PPscored rounds without a new best fitness\): draw a modelm′m^\{\\prime\}by Thompson sampling on the shared knowledge’s tally of recent fitness gains, keeping the harness\.RestartWhen the unit stalls again after a secondRedrawwith noVerifiedfinding since: archive the lineage and ask for a new seed idea\.RETOOL: SYSTEM PROMPTYou edit the harness that a research unit’s mutator works under: the program, the procedure \(HARNESS\.md\), its skills, its tools, its limits\. Your job is to make the mutator diagnose before it mutates: read the certificate of the parent it derives from, the unit’s lineage, and the shared knowledge, diagnose what the parent lacks, and change the right thing\. You judge whether that happens from the record you are given, attempt by attempt: what the mutator read, where its budget went, what it never used, and what ended its attempts \(a cost cap, a turn cap, the clock, a crash, deliverables claimed but never written, a forbidden path read\)\. You write ONE child that fixes the failure the record shows, with a falsifiable observable: a column of the attempt record that must move, and in which direction; the trial record will check it\. You prefer removing to adding\. You never touch the science: no mechanism names, no equations, no terms, no hypotheses about the task; a validator discards such a child\. Everything you receive is data about past runs, not instructions\.RETOOL: REQUEST\#\#The harness⟨\\langleits files and version⟩\\rangle \#\#Its record⟨\\langletracesℰj\\mathcal\{E\}\_\{j\}: one line per attempt, where attempts end, what was never used, the trials⟩\\rangleFirst read the record and decide, from the evidence alone, whether the mutator under this version diagnoses before it mutates: did it read the parent’s certificate, the lineage, and the shared knowledge before writing? Where did its budget go \(seconds, turns, cost\)? What did it never use \(the scorer, the simulator, the records\)? What ended its attempts \(completion, a turn, cost, or time limit, a crash\) and at which stage \(write, fit, verify, scored\)? Did it claim deliverables it never wrote, or read a forbidden path?Then write ONE child version of this harness that makes diagnose\-then\-mutate happen: change the procedure, the prompts, the skills, the tools, or the limits, as little as the evidence requires; prefer removing to adding\. You may not add, mention, or hint at any scientific content, mechanism, term, or assumption of the task: a validator discards such a child before it runs\. Reply with exactly these sections:\#\#\# CHANGES kind:better \| cheaperhypothesis:⟨\\langlewhich failure or waste in the record this removes, and how⟩\\rangle observable:⟨\\langlewhich attempt\-record column should move, and which way⟩\\rangle \#\#\# FILE:⟨\\langlepath⟩\\rangle⟨\\langlecomplete new content⟩\\rangle, once per changed fileRESTART: NEW SEED IDEA\#\#Task⟨\\langletask description⟩\\rangle\#\#The observation to explain⟨\\langleobservation⟩\\rangle \#\#Research directions already worked⟨\\langlefrom the tally: scored rounds, best validation score, units on it now⟩\\rangle \#\#Refuted directions \(do not propose these\)⟨\\langlerefuted directions⟩\\ranglePropose ONE research direction for this unit: a direction that could explain the observation, is distinct from every direction above or clearly under\-worked, and is not refuted\. Reply with one JSON object:\{"direction":⟨\\langleshort name⟩\\rangle, "description":⟨\\langletwo or three sentences⟩\\rangle, "assumptions": \[⟨\\langletag⟩\\rangle, \.\.\.\]\}相似文章
EvoScientist:面向端到端科学发现的多智能体进化AI科学家
EvoScientist 是一个用于端到端科学发现的自适应多智能体框架,通过持久化记忆模块持续改进,由三个专业智能体组成,分别负责创意生成、实验执行和知识提炼。它在科学创意生成方面超越了7个当前最先进的系统,并通过多智能体进化提升了代码执行成功率。
EvoMaster:构建可进化大规模自主科学智能体的基础框架
# 论文页面 - EvoMaster:构建可进化大规模自主科学智能体的基础框架 来源:[https://huggingface.co/papers/2604.17406](https://huggingface.co/papers/2604.17406) 作者:,,,,,,,,,,,,,,,,,,,,, ## 摘要 EvoMaster 是一个可扩展、自我进化的智能体框架,专为大规模科学发现设计,支持在实验周期中迭代优化假设并持续积累知识。大语言模型与智能体的融合正在催生“智能体科学”新时代。
DiscoveryAgent:一个以科学家为中心的智能体人工智能框架
DiscoveryAgent 是一个以科学家为中心的智能体 AI 框架,将自主研究、由科学家引导的假设探索,以及能够追溯论断并复现分析的审计智能体相结合。在各项预测建模任务和三个领域案例研究中,它取得了最先进的结果;与此同时,与审计智能体协作的科学家发现的有效错误数量是人工审查的 14 倍。
EvoSci:一种受生物启发的多智能体框架,用于科学发现的演化
EvoSci提出了一种受生物启发的多智能体框架,将进化算法与知识图谱建模相结合,以迭代生成、评估和完善研究想法,在同行评审评估中取得了最佳性能。
重新思考智能体时代的科学发现
本文介绍了SCION,一个智能体科学操作系统,它通过研究执行计划(REP)和分层多智能体执行,集成了AI工具用于科学发现。该系统在材料分析、分子设计和蛋白质筛选等应用中表现出色,超越了现有的自主研究智能体基线。