SwarmWorld: Stigmergic technological evolution in societies of language-model agents

arXiv:2608.260812026-08-26

群体的优势不在一件最强发明，而在更坚韧的技术生态

MIT研究团队把50至200个初始能力相同的LLM智能体放入同一个模拟世界，让它们在没有预设角色和配方的情况下自行探索资源、加工材料、建造设施并编写可执行控制程序。移除智能体后，研究者用8组未见过的扰动测试遗留设施，结果显示共享世界形成的技术组合比独立搜索更广、韧性更强。独立搜索仍可能保留性能最强的单件设施，因此群体优势主要体现在覆盖面和持续性，而不是单项最高分。

METAL LAB 解读图

智能体活动如何积累成持久技术生态

证据状态已报告实测结果

1. 局部探索每个智能体只观察附近资源和已有设施。
2. 建造与代码智能体加工材料、建造结构并安装可执行控制器。
3. 环境记忆后来者观察、继承并修改持久设施及代码谱系。
4. 无智能体测试移除智能体后，用未见扰动测量技术组合的功能和韧性。

这是 METAL LAB 制作的解读图，并非论文作者提供的原图。

他们做了什么

智能体只依据局部观察行动，自行采集和加工资源、建造可持续存在的设施，并编写控制程序，没有预先分配职业或技术路线。
研究比较了四个计算机会匹配的条件：完整文化、取消直接交流、取消显式文化，以及由孤立智能体组成的逐指标best-of-N基线。
功能由确定性模拟器而非智能体自述判定。建造结束后先移除智能体，再用8组留出扰动测试遗留设施。
规模实验让50、100和200个智能体运行800个时间步，每个条件使用4个世界种子；另一个100智能体实验持续到3,200个时间步。

Figure 1: Conceptual lineages leading to SwarmWorld; connectors indicate conceptual convergence rather than direct descent between every adjacent item. (A) Biological collectives and swarm-computing methods contribute decentralized coordination, feedback, and stigmergy. (B) Local-rule models and artificial societies contribute persistent worlds in which local actions alter shared state and generate population-level organization. (C) Distributed-agent and LLM systems contribute explicit agent state, memory-guided planning, embodiment, communication, and reusable executable skills. (D) SwarmWorld combines these elements in a materially constrained multi-agent world where agents build persistent artifacts and inherit or modify their programs. The resulting technological ecology can be evaluated after the agents are removed. Image credits. Ant photograph by Kiran Vati K, CC BY 4.0, cropped and color-adjusted. Boids still from a video by Andrew Magill, CC BY 2.0, with a frame extracted, cropped, and color-adjusted. Game of Life glider-gun diagram by Bryan.burgers, public domain, placed on a grid background. Micropolis screenshot based on the original software by Will Wright and Maxis Software/Electronic Arts, the Unix port by Don Hopkins and DUX Software, screenshot by Tomhannen, and edit by bayo; GPLv3 or later with Section 7 additional terms, cropped and color-adjusted. The Sugarscape view, method schematics, and SwarmWorld rendering were created by the authors.

Figure 2: Simulated world and agent-environment algorithm. (A) Authoritative state at tick 600 from a representative N=200 society. Numbered callouts identify language-model agents, persistent agent-built artifacts, fixed processing stations, resource biomes, an active disturbance field, and tidal resources. The map is the state on which actions are resolved, not a decorative visualization. Agents can move through the terrain, gather and transform matter, inspect or operate artifacts, and encounter local environmental changes produced by both disturbances and other agents. (B) One staggered macroturn begins with a local observation and retrieved private memory, continues through language-model deliberation and a strict structured plan, and ends with transactional validation of actions, targets, permissions, and material preconditions. The deterministic world tick then advances motion, metabolism, resources, fields, disturbances, and every installed artifact program. Resulting events enter agent memory and the shared substrate before the next scheduled observation. The agents decide what to try, but the world decides what actually happens; durable artifacts let one agent’s successful work become part of another agent’s future environment.

Table S1: Glossary of technical terms, implementation-aligned definitions, plain-language interpretations, and broadly accessible examples.
Term	Technical definition	Plain-language meaning	Broad example
World tick	One discrete deterministic state transition, St+1=F⁡(St,At), where St is the complete world state and At is the set of accepted agent actions at tick t. Given the state, actions, and seeded disturbance schedule, F advances movement, metabolism, resources, fields, artifacts, and installed programs.	The basic clock step of the simulated world. The world keeps changing even when no agent is currently asking the language model what to do.	A video game advances one frame at a time; machines, weather, and moving characters continue between a player’s major decisions.
Macroturn or decision opportunity	A scheduled occasion on which one agent receives a local observation and memory context, calls the language model, and proposes a structured multi-action plan. Macroturn phases are staggered and matched across paired conditions.	One chance for an agent to stop, look around, think, and choose its next actions.	A field scientist periodically checks instruments and updates the day’s plan while the experiment continues running between checks.
Persistent artifact	An agent-constructed world object with recorded creator, contributors, materials, geometry, services, provenance, state, and optionally an executable controller. It remains in the world until modified, retired, dismantled, or otherwise changed by simulator rules.	A technology that stays behind and can later be found, tested, operated, or modified by someone else.	A device left on a shared laboratory bench can influence the next researcher even if its builder is absent.
Artifact program	A validated instruction sequence installed on an artifact and executed by the deterministic simulator every tick. Conceptually, ua,t=πa(oa,t): program πa reads the artifact’s permitted local sensors oa,t and emits permitted control operations ua,t.	Code that lets a built object sense local conditions and act without another language-model call.	A thermostat repeatedly reads temperature and turns heating on or off after the installer leaves.
Full culture	The intervention containing one shared physical world, explicit cultural actions, cross-agent executable-program inheritance, and artifact stigmergy. Explicit actions include messages, public records, teaching, trading, task claims, and publication-dependent composition.	Agents can communicate directly, inherit one another’s code, and also coordinate indirectly through the things they build.	A research group shares a laboratory, talks, writes notebooks, and modifies common software and equipment.
Artifact stigmergy	Indirect coordination produced when an agent changes persistent artifacts or the environment and another agent later observes or acts on that changed state. It remains available in the no-explicit-culture condition even when direct cultural channels and cross-agent program inheritance are disabled.	Agents coordinate through traces in the world rather than through direct conversation.	Ants coordinate through modified trails; people can likewise coordinate by adding parts or annotations to a shared workbench.
Independent-search envelope	For endpoint e, checkpoint t, and N isolated one-agent worlds, the control is Yiso(e,t)=maxi∈{1,…,N}⁡Yi(e,t). The maximizing member may differ across endpoints and checkpoints.	The comparison gives isolated search its best available result for each question, rather than forcing one solo agent to win every contest.	A national team may choose one athlete for the sprint and another for the high jump; the envelope records the best specialist in each event.
Discovery-frontier AUC	If p⁡(t) is measured artifact performance, the running frontier is F⁡(t)=maxτ≤t⁡p⁡(τ). The normalized area is DT=T−1∫0TF⁡(t)𝑑t, implemented by trapezoidal integration over recorded samples.	A strong invention scores more if it appears early enough to remain the best-known option for much of the experiment.	Two runners may finish with the same final speed, but the one who led for most of the race has the larger time-averaged frontier.
Current service coverage	For service dimension k at time t, ck(t)=maxa∈𝒜t⁡sak(t) over active, non-retired artifacts. With c¯(t)=K−1∑kck(t) and b⁡(t)=mink⁡ck(t)/c¯(t), balanced coverage is Q⁡(t)=c¯(t)[1+b⁡(t)]/2 when c¯(t)>0.	The society is judged by the best currently available artifact for each need, with a penalty when one need is badly neglected.	An emergency kit with food, water, shelter, and medicine is more balanced than one containing excellent food but nothing else.
Portfolio resilience	The discovery-state score uses lifetime-peak service fingerprints, ckpeak=maxa⁡sakpeak, and the same balance formula P=c¯peak[1+mink⁡ckpeak/c¯peak]/2. Artifact count alone does not increase the score.	A portfolio is valuable when its collection has demonstrated strong and balanced coverage across several functions, not merely because it contains many objects.	A toolbox with a few complementary, proven tools can be more resilient than a warehouse full of duplicate hammers.
Held-out resilience	For unseen schedule s, Rs=T−1∫0TQs(t)𝑑t, where Qs(t) is current balanced service coverage during the agent-free assay. The reported value is R=S−1∑s=1SRs over S=8 paired schedules.	The frozen technology is tested under new disturbances after all agents are removed; the score measures how well its active functions continue to cover multiple needs.	A bridge design is evaluated under earthquakes and wind patterns that were not used while designing it, with no engineer allowed to repair it during the test.
Validated invention	The count is I=∑a𝟏[ra∧da∧ga∧pa≥p∗∧νa≥ν∗]. Artifact a must use a recipe that passed material testing (ra), contain the required design fields (da), include an agent-authored installed program (ga), exceed the performance threshold p∗, and exceed the behavioral-novelty threshold ν∗.	A named idea is not enough. The artifact must be materially grounded, fully specified, executable, functional, and behaviorally distinct.	A patent sketch would not count by itself; a device must have tested materials, a complete design, working controls, and a demonstrably new behavior.
Artifact-contact AUC	Let q⁡(t) be the fraction of agents within three cells of an active artifact at recorded time t. Contact AUC is C=(tm−t1)−1∫t1tmq⁡(t)𝑑t, evaluated by trapezoidal integration.	The measure records how consistently agents spend time near shared technology, rather than how far they travel in total.	A museum can measure the fraction of visitors near an exhibit throughout the day, not only the number who entered the building.
Behavioral phenotype	A post hoc cluster of complete agent trajectories. For k-means, assignments minimize ∑i∥xi−μzi∥2 in the declared robust-scaled feature space. Condition, identity, and role labels are excluded from fitting.	A recurring style of behavior discovered from what agents actually did, not a role assigned in their prompts.	Travel records might separate commuters who repeatedly visit one workplace from explorers who visit many neighborhoods, without reading their job titles.
Recorded lineage and downstream reach	A directed graph whose edges come from recorded authorship, observation, causal-parent, construction, installation, and fork events. For node v, downstream artifact reach is r(v)=\|{a:v↝a,a is an artifact}\|.	It records which evidence, programs, people, and precursor objects are linked by logged parent and contribution events to later technologies.	A product genealogy can trace a component specification through several revisions into multiple final devices.
Cross-agent program fork and lineage depth	A fork is cross-agent when the child-program author is not in the recorded parent-author set. Executable-lineage depth is the longest directed path in the content-addressed program-fork directed acyclic graph up to a checkpoint.	One agent edits another agent’s controller, and depth counts how many successive generations of such inherited code accumulate.	A programmer forks a colleague’s repository; later colleagues fork the modified versions, forming a software family tree.
Interaction modularity	Weighted Newman-Girvan modularity, Q=(2m)−1∑ij[Aij−kikj/(2m)]δ(gi,gj), computed on the weighted bipartite agent-artifact graph using Louvain communities at resolution 1. Event-type weights are multiplied by log⁡(1+n) for repeated events.	High modularity means interactions are concentrated within recognizable technological neighborhoods rather than spread uniformly across the whole society.	A university is modular when laboratory members collaborate mostly within their labs, with fewer links between labs.
Community persistence	Adjusted mutual information between consecutive agent community assignments, AMI=[MI−𝔼⁡(MI)]/[(HU+HV)/2−𝔼⁡(MI)]. Adjustment removes similarity expected by chance.	The measure asks whether the same agents remain grouped together from one checkpoint to the next, allowing community labels themselves to change.	If project teams retain most of the same members next semester, their community persistence is high even if team numbers are renamed.
Participation coefficient and within-module z-score	For node i, Pi=1−∑c(kic/ki)2 measures how evenly its weighted links span communities. Within its own community, zi=(kiown−μc)/σc measures how unusually strong its internal connectivity is relative to peers of the same node type.	Participation distinguishes local specialists from cross-community connectors; the z-score distinguishes ordinary members from local hubs.	A scientist collaborating only within one lab is a specialist, one collaborating across many labs is a connector, and a highly connected member within one lab is a local hub.
NODF nestedness	Binary nestedness is 100 times the mean \|Ni∩Nj\|/min⁡(ki,kj) over unequal-degree pairs, evaluated for both agent rows and artifact columns. High values mean the smaller neighborhood is largely contained in the larger one.	Nestedness asks whether specialists mostly use subsets of the technologies used by generalists.	A small shop is nested within a department store if nearly everything it stocks also appears in the larger store.
Densification exponent	Cumulative unique agent-artifact ties are fit to Ecum∝Vα, or log⁡Ecum=αlog⁡V+c, where V is agents plus constructed artifacts. The plotted α is an ordinary least-squares descriptive fit across seeds and checkpoints for one condition.	If α>1, relationships accumulate faster than the number of participating agents and artifacts.	A growing professional community densifies when adding people and projects creates disproportionately many new collaborations.
Relationship reuse	For consecutive analysis intervals, reuse is ρt=\|Et−1∩Et\|/\|Et\|, where Et is the set of agent-artifact pairs active in the current interval. It is distinct from the symmetric Jaccard index.	The measure is the fraction of current working relationships that were already active in the preceding period.	A shop with many returning customers has high relationship reuse even if it also attracts new customers.
Cross-agent adoption and time to first reuse	An artifact is adopted when a noncreator first observes, causally references, programs, or repairs it; co-construction does not count. Time to first reuse is Ta=tafirstnoncreatorreuse−tacreated. Unreused artifacts are right-censored at tick 3,200.	Adoption means someone other than the inventor actually engages with the technology, and the delay measures how quickly that happens.	A neighbor borrowing and using a tool counts as adoption; merely helping build it does not count as later reuse.
Structural robustness AUC	If L⁡(f) is largest-component size after removing fraction f of agents, robustness is A=fmax−1∫0fmaxL⁡(f)/L⁡(0)𝑑f for fmax=0.5. Removal is random, degree-targeted, or betweenness-targeted; the graph is not rewired.	The score summarizes how much of the recorded network stays connected as agents disappear. It does not measure physical recovery or adaptation after removal.	An internet topology may tolerate random router failures but fragment quickly if its busiest hubs are deliberately disabled.

Figure 3: Mechanism-resolved conditions and the agent-free held-out resilience assay. (A) Rows define the four interventions and columns identify the mechanisms available in each. A filled circle means the mechanism is present. Full culture combines a shared physical world, explicit cultural actions, cross-agent program inheritance, and artifact stigmergy. No communication removes nearby messages, public records, teaching, trading, task claims, and publication-dependent composition while preserving the shared world and executable inheritance. No explicit culture also removes cross-agent program forking and measured skill inheritance, leaving agents able to coordinate only through persistent artifacts and environmental changes. Independent search consists of N isolated one-agent worlds and therefore has none of the three collective mechanisms. (B) At a discovery checkpoint, the complete state is frozen into eight exact clones. Each clone receives a paired unseen schedule of contamination, drought, and storm, with new centers, timings, and orderings. Agents take no actions during evaluation, but the world physics and installed artifact programs continue. Balanced service coverage Qs(t) is integrated over assay time and averaged across schedules to produce held-out resilience. The assay asks whether the technology left behind can protect the habitat under new stresses after its inventors have gone.

Figure 4: Population scaling and mechanism-resolved collective effects over 800 discovery ticks. Points and heavy lines in panels A-D are means across four matched world seeds; error bars are paired seed-bootstrap 95% intervals. (A) Discovery-frontier AUC integrates the running best measured artifact performance, rewarding discoveries that occur early and remain on the frontier. (B) Held-out resilience AUC measures agent-free service coverage under eight unseen disturbance schedules. (C) Portfolio resilience quantifies the breadth and redundancy of functional service across the final artifact collection. (D) Validated inventions count technologies that pass the simulator’s objective validation criteria. (E) Each cell is a shared-world condition minus the paired endpoint-wise independent-search envelope. Printed values are raw paired effects, asterisks mark intervals excluding zero, and color is normalized within each outcome row because the four outcomes have different units. The shared world yields the most consistent gains in held-out and portfolio resilience and in invention count, whereas discovery AUC depends on condition and scale. Societies become better at building useful collections of technologies, but more communication does not automatically make them better at finding the single fastest early winner.

Table S2: Mechanism-resolved experimental conditions.
Condition	Available interaction substrate
Full culture	One shared world with physical stigmergy, messages, publications, teaching, trade, task claims, shared records, program reuse, and program forking.
No communication	One shared world with physical stigmergy. Messaging, publication, teaching, trade, task claims, and publication-dependent composition are removed; physically observable program reuse and forking remain available.
No explicit culture	One shared world with physical stigmergy. Communication, program forking, cross-agent sequence inheritance, the skill library, authored artifact text, and mutation-parent access are removed from the treatment interface.
Independent search	N isolated one-agent copies of the seeded world. Member i matches shared agent i in initial position and macroturn phase. Each endpoint is the maximum over all N members.

Figure 5: Data-driven behavioral phenotypes discovered in the 800-tick scaling study. Each point represents one complete agent trajectory, and the two-group k-means classification is learned from 15 robust-scaled behavioral features without access to condition, population, seed, identity, or an assigned role. (A) PCA gives an auditable linear projection. The numbered arrows are the six largest loading vectors and point toward increasing social interaction, cultural coordination, technology work, artifact proximity, construction/control, and net displacement. Clustering is performed in the full feature space rather than in this two-dimensional display. (B) Cluster-average standardized signatures identify C1 as artifact-centered work and C2 as mobile exploration; red indicates above-corpus values and blue indicates below-corpus values. (C) Stacked bars show the mean N=200 composition, with four seed fractions overlaid. (D) The artifact-centered fraction is followed across N=50, 100, and 200 for each shared-world condition. Initially identical agents divide into a smaller group that stays near and works on technology and a larger group that continues exploring, and explicit culture shifts more agents toward the technology-centered mode.

$Figure 6: Executable technological culture. Seed means and 95% seed-bootstrap intervals are shown across population size. (A) Fraction of constructed artifacts with recorded contributions from multiple agents. Full culture produces the largest collaborative fraction at every population, while appreciable collaboration without explicit culture demonstrates that agents can still meet around and modify persistent objects. (B) Fraction of eligible program forks whose child author differs from the parent author. Cross-agent descent grows with population under full culture and remains possible without direct communication because executable inheritance is still available in that intervention; it is zero by construction when explicit culture and cross-agent inheritance are disabled. (C) Focused content-addressed program lineage centered on Adaptive Chitin Maintenance. Gray boxes are ancestor programs, the green box is the focal program, and purple boxes are descendants. Each node records first tick, number of authors, and installations; each arrow records fork tick, author, and instruction-level edit count. Agents did not merely discuss ideas: they reused, edited, installed, and propagated one another’s executable environmental controllers, producing attributable technological descent.$

Figure 6: Executable technological culture. Seed means and 95% seed-bootstrap intervals are shown across population size. (A) Fraction of constructed artifacts with recorded contributions from multiple agents. Full culture produces the largest collaborative fraction at every population, while appreciable collaboration without explicit culture demonstrates that agents can still meet around and modify persistent objects. (B) Fraction of eligible program forks whose child author differs from the parent author. Cross-agent descent grows with population under full culture and remains possible without direct communication because executable inheritance is still available in that intervention; it is zero by construction when explicit culture and cross-agent inheritance are disabled. (C) Focused content-addressed program lineage centered on Adaptive Chitin Maintenance. Gray boxes are ancestor programs, the green box is the focal program, and purple boxes are descendants. Each node records first tick, number of authors, and installations; each arrow records fork tick, author, and instruction-level edit count. Agents did not merely discuss ideas: they reused, edited, installed, and propagated one another’s executable environmental controllers, producing attributable technological descent.

研究结果

共享世界在留出扰动韧性、技术组合韧性和经验证发明数量上持续优于独立搜索。
在200智能体条件下，取消显式文化但保留物理痕迹的社会，其发现前沿AUC比独立基线高0.069。
在3,200时间步的完整文化实验中，以设施为中心的行为占比相对消融条件增加21.8个百分点，可执行代码的继承深度也持续增长。
约95%的首次技术采用始于物理观察，而最终性能最强的单件设施仍由独立搜索保留。

Figure 7: Sixteen ranked, deduplicated agent-invented technologies from the 800-tick study. The colored rule above each panel identifies its experimental condition and the number at upper right is lifetime-peak simulator performance. Panels A-D show the Three Layer Connected Chitin Exchange Lattice, Mycelial Mineral Spring Veil, Fermented Tidal Chitin Panel, and Breathing Cuticle Veil. Panels E-H show the Myco Repair Exchange Panel, Replicated Tri-Phase Healing Exchange Lattice, Chitin Healing Vein Partition, and Myco Mend Lattice. Panels I-L show the Myco Porous Buffer Sheet, Catalyst Springveil Reactive Lattice, Catalyst Venation Habitat Membrane, and Cellulose Nutrient-Moisture Exchange Trellis. Panels M-P show the Stonecap Mycelial Vent Panel, Basalt Mycelium Lattice, Hyphae Pulse Interface, and Kelp Shell Tideveil Panel. Each portrait combines a complete form, partial cutaway, and dominant component specimens inferred from the recorded architecture, geometry, composition, process sequence, and controller. Text below each image reports the input-process-output flow, agent-authored functional claim, sensor-to-actuator control signature, and leading material fractions. The society explored multiple technological families rather than converging on one prewritten object; the images communicate the proposed mechanisms, while the quantitative score comes only from the simulator. See Section S2.13 for details on how these representative technologies were selected.

Figure 8: Representative N=200 movement trajectories from matched seed-3202 societies. For readability, 24 deterministically selected complete paths are shown in each panel over the identical terrain. White circles mark trajectory starts, colored circles mark endpoints, and gold diamonds mark artifacts active at tick 800. (A) Full culture produces paths that repeatedly intersect the central artifact ecology; all-agent summaries report mean path length 36.0 cells, 3.3 regions visited, and artifact-contact AUC 0.31. (B) Without explicit culture, mean path length remains 36.0 cells, but region coverage falls to 2.7 and artifact-contact AUC to 0.14. (C) Without communication, mean path length is 37.0 cells, region coverage is 2.6, and artifact-contact AUC is 0.11. The inset values use all 200 agents, not only the displayed paths. The agents travel similar total distances, but full culture makes their movement more likely to intersect the shared technological infrastructure.

可应用场景

设计长期运行的智能体系统时，可把可检查产物、可执行程序和版本化来源链作为共享记忆，而不只依赖聊天记录。
构建科学发现智能体时，可让探索、建造、维护和协调角色根据结果自然形成，而不是预先固定分工。
评估机器人群体、材料搜索和运营型智能体系统时，可衡量多种需求下的组合稳健性，而不只看一个峰值产出。

Figure 9: Movement linked to scientific work and interaction in representative N=200 societies. Eighteen focal agents per condition are selected by round-robin ranking on distance, artifact work, social interaction, and successful actions so that the display spans the observed behavioral repertoire. Segment color gives instantaneous path speed in cells per tick. Blue circles mark observations or tests, gold diamonds mark artifact construction, purple triangles mark program installation, coral stars mark explicit interaction, and black squares mark active artifacts; square area increases with the number of distinct participants. (A) Full culture combines broad observation with dense construction, installation, and explicit social events around shared artifacts. (B) No explicit culture retains observation and physical artifact work but lacks the direct cultural channel. (C) No communication shows the corresponding spatially mediated activity when messages are unavailable. These are not simply tracks of where agents wandered: they show when travel became measurement, building, code deployment, or coordination, and where repeated work turned artifacts into local activity hubs.

Figure 10: Within-run evolution of a representative no-explicit-culture society with N=200 and seed 3202. (A-C) Authoritative world states nearest ticks 0, 400, and 800; the stored snapshots occur at ticks 0, 408, and 800. Agents are black circles, artifacts are gold diamonds, and processing laboratories are open squares. The artifact count increases from 0 to 25 to 61 as the disturbance field and local activity pattern change. (D) Stepwise accumulation of persistent artifacts across the episode. (E) Running best artifact performance and portfolio resilience. The best-object curve rises rapidly after the first construction wave, whereas portfolio resilience grows more slowly as the society adds complementary technologies. (F) Normalized spatial entropy of agent occupancy on a fixed 10×10 grid declines from roughly 0.86 to 0.76, indicating moderate concentration rather than complete aggregation. Agents progressively convert an initially empty landscape into a persistent technological habitat; the first strong object appears quickly, but a useful portfolio and stable spatial organization take longer to assemble.

局限与待验证事项

每个条件只有4个配对世界种子，因此分析重点是效应大小和种子间一致性，而不是大样本总体推断。
证据来自确定性模拟世界，真实组织或实体机器人群体是否具有同样优势尚未验证。
蛋白质材料世界中的隐藏属性是模拟器预设值，不是新的生化预测或实验测量。

Figure 11: Recorded knowledge lineage from evidence to technology. (A) Society-scale directed ancestry of eight high-performing active artifacts. Circles denote agents, squares denote evidence, small gold nodes denote insights, purple hexagons denote executable programs, and green diamonds denote artifacts. Typed arrows encode authorship or observation, recorded parent references, construction, installation, and program forking; node area increases with betweenness centrality. Labels are selectively placed to preserve readability. (B) Focused ancestry around AdaptiveChitinExchangeScaffold. Multiple agents, evidence records, precursor artifacts, and inherited programs converge on the target artifact; arrow direction is from recorded precursor to downstream record, program, or artifact. (C) Evidence, insight, and program nodes ranked by the number of distinct downstream artifacts they reach. The passive material system has the largest reach, followed by several maintenance and exchange records. Technologies are not isolated model responses: recorded observations and executable ideas become reusable building blocks that pass through several agents and contribute to multiple later artifacts.

Figure 12: Construction and operational matter pathways for the 16 technologies in Figure 7, retained in the same A-P order. The left matrix reports construction feedstocks. Filled circles identify named recipe inputs and circle area increases with input mass; gray open circles mean that the material is absent. The center matrix reports fabrication sequence. Connected numbered circles provide the exact processing order, distinguishing, for example, wash-grind-ferment routes from mineralize-align-weave routes. The right matrix reports realized operational matter flow per 100 active ticks in the authoritative final snapshot. Filled colored circles and printed values encode the realized rate; colored open circles indicate that a controller pathway exists but had zero realized flux; gray open circles indicate that the pathway is unavailable. Water capture transfers local moisture into artifact storage, contamination removal reduces the environmental field, and embodied reserve supports repair, growth, and nutrient release. Construction feedstocks are consumed once at build time and are not automatically replenished during operation. The figure distinguishes what an artifact is made from, how it was fabricated, and what it actually did after construction; only the final columns are evidence of executed simulator function.

为什么重要

这项研究说明，只看多智能体系统能否给出一个最高分答案，会遗漏协作最重要的收益。对于需要长期满足多种需求的系统，可检查的持久设施、可执行代码和可追踪继承关系，可能比对话记录更适合作为共享记忆。

本文术语

痕迹协作（Stigmergy） · 个体通过留在环境中的痕迹或结构间接影响后来者的协作方式
best-of-N搜索 · 进行N次彼此独立的尝试，并为每个指标选择最好结果的强基线
留出扰动 · 建造过程中未使用、在完成后用于检验稳健性的测试条件

论文原文摘要（英文）

Collective intelligence can emerge when individuals coordinate through a shared environment, allowing local actions to accumulate into durable social organization. Language-model agents offer a new substrate for this process, yet most multi-agent systems rely on direct conversation, predefined roles, or centralized workflows. It remains unclear whether decentralized agents can build functional technologies and outperform independent search. Here, initially homogeneous LLM agents in SwarmWorld self-organize without assigned roles or recipes into evolving technological societies. Agents explore a spatial environment, process resources, test materials, construct persistent artifacts, and write executable controllers evaluated by a deterministic simulator under unseen disturbances after the agents are removed. SwarmWorld splits cognition from consequence: agents propose architectures and controllers within fixed action and material schemas, while the simulated world determines function. Shared societies develop broader, more resilient technological portfolios than a strong best-of-N isolated-search baseline, although isolated search remains competitive for the strongest artifact. Agents differentiate into exploration, construction, maintenance, and coordination behaviors, transitioning as the world matures. Technologies accumulate through collaborative construction, executable inheritance, and persistent agent-artifact networks, with most reuse beginning through physical observation rather than communication. Explicit cultural mechanisms amplify collaboration and organization, but functional benefits depend on outcome and timescale. Physical stigmergy alone supports capable societies, while interaction drives persistent technological ecologies rather than universally superior individual inventions.

作者 · Pal, Subhadeep, Wang, Fiona Y., Buehler, Markus J.

在 arXiv 阅读

METAL LAB 最新报道

图片来源： Pal, Subhadeep et al., arXiv:2608.26081, cc-by-nc-nd-4.0