Organizations: Flowers AI & CogSci Lab, Inria, France · IT University of Copenhagen, Denmark · Independent Researcher, UK · Institute for Advanced Study in Toulouse, France
Human collective intelligence depends on transmission processes: who shares what with whom, how, and when. While these processes emerge from individual cognition, they can also be directed by deliberate top-down protocols. Prior work has studied how transmission shapes collective outcomes primarily through the lens of network structure, varying who shares with whom and when. But networks are state-agnostic: they cannot condition transmission on what agents know or on the state of the collective. Here, we formalize transmission protocols as state-aware programs that route information and resources based on agent and collective states, and we use LLM-guided evolutionary search to design effective protocols in a collective discovery task. Evolved protocols increase collective performance over standard baselines from the literature by up to 37%. Ablations confirm that state-awareness drives this advantage: removing content-dependence while preserving network topology and timing eliminates performance gains. We find that evolved protocols also transfer across domain variations and agent populations. These results demonstrate that effective and generalizable transmission protocols can be discovered in silico, suggesting a path toward AI-assisted design of coordination infrastructure that enhances human collective intelligence.
Figures & tables
Figure 1: Discovering adaptive transmission protocols in silico . (a) State-aware transmission protocols (TPs) route information and resources between agents based on agent and collective states. (b) We search for effective TPs with an evolutionary algorithm that leverages large language models (LLMs) to generate and refine candidate protocols expressed as stateful programs. (c) Best evolved protocols increase collective performance in simulation by 37% over baselines (mean ± sem, N=5, p<0.001).
Figure 2: LLM-guided search uncovers high-performing transmission protocols . (a) Collective performance of best protocols (one per evolutionary run) and baselines across game steps (mean ± sem over 30 simulations). (b) Comparison of performance measures at 150 game steps. Stars indicate significance of one-sided t-tests (all p<0.001 ).
Figure 3: State-awareness ablation . Collective performance (mean ± sem) for top-20 and bottom-20 protocols per evolutionary run (100 in each set), under original (full) and ablated (random content) conditions. Ablation eliminates the performance advantage of top protocols, reducing them to baseline levels, while bottom protocols are unaffected.
Figure 4: Generalization across population and domain configurations . Evolved protocols (blue) outperform baseline protocols (gray) across all population configurations and domain variants. Mean ± sem, N=5, p<0.01 (), p<0.001 (*).
Among successful recipes, 1−(fraction of agents who own r) . Higher values indicate sharing rarer items.
Item novelty
Among successful recipes, fraction where r is not in the learner’s inventory.
Actionable has- k
Fraction of transmissions where the learner owns exactly k ingredients AND the result item is novel to the learner. Combines routing with content novelty.
Success bias
Whether the shared recipe is successful ( r=null ) minus the teacher’s base success rate at that step.
Routing
Has- k fraction
Fraction of transmissions where the learner owns exactly k ingredients of the shared recipe.
Appendix
Table 1: Behavioral features computed from transmission logs. Each feature is averaged across all transmissions (or per-step aggregations) and then across evaluation seeds. Notation: recipe =(i1,i2,r) where i1,i2 are ingredients and r is the result ( r=null for failures). Features indexed by k∈{0,1,2} are computed separately for each value.
Figure 5: Fitness of best evolved protocol as a function of search step .
Figure 6: LLM-guided search uncovers high-performing protocols for populations of 10 empowerment agents . Fitness of best protocols (one per search run) and extended set of baselines across 300 game steps (mean ± sem over 30 simulations).
Figure 7: Program search discovers transmission protocols that diverge from the generative prior . (left) t-SNE projections of discovered protocols, with per-run best-performing protocols marked with stars. (right) Maximum mean discrepancy (MMD) measuring distributional divergence across search runs; evolved protocols are more distant from generated protocols than from each other.
Figure 8: Transmission protocol effectiveness is not correlated with transmission volume . The quantity of information transmitted is measured as the number of memories received by an agent that had not experienced this memory on its own at the time of reception.
Group
Feature
ρ
Sig.
Content
Recency bias
+0.64
Routing
Cond. recency has-2
+0.63
Routing
Cond. recency has-1
+0.58
Routing
Targeted failure rate
−0.54
Content
Actionable has-0
+0.49
Routing
Cond. recency has-0
+0.48
Appendix
Table 2: Ablation analysis results across ∼1,000 independently generated protocols. For each feature, we compute the per-protocol difference between original and content-randomized (ablated) values, and correlate this difference with fitness (Spearman ρ ). n indicates the number of protocols for which the feature was defined. Significance after Benjamini–Hochberg FDR correction: ** p<0.01 , *** p<0.001 . Features sorted by ∣ρ∣ .
Multi-agent communication studies how autonomous agents exchange information to coordinate decisions, share knowledge, and act together under constraints.
Large Language Model (LLM)-based agents promise to automate scientific discovery, yet exploring the vast hypothesis space remains costly. Existing principle-evolution methods accelerate this loop, but operate sequentially, which caps exploration breadth and wastes wall-clock time on challenging problems. To address this, we formulate collaborative scientific discovery as evidence transfer between parallel principle-evolution branches. We present COEVOLVE, which realizes this transfer through a coordination core over parallel branches. By integrating value-of-information-gated routing and context-discounted likelihood injection, COEVOLVE enables branches to collaborate through shared measurements while keeping their principle posteriors separate. Across six scientific-discovery tasks under a matched evaluation budget, COEVOLVE attains a mean solution quality of 66.5% versus 57.0% for single-branch principle evolution, with a 1.80x mean wall-clock speedup on the GPT-5.6-Terra backbone; on five auto-research tasks delegated to an autonomous research harness, it is the only arm whose mean stays above the published SOTA anchor on every task. These results establish when evidence sharing accelerates parallel discovery and when transfer safeguards are necessary to limit negative or inert transfers
Yingming Pu, Hongyu Chen, Tao Lin
Zhejiang University, Hangzhou, Zhejiang, China · Westlake University, Hangzhou, Zhejiang, China
We study how teams of AI coding agents coordinate while solving programming tasks. Current evaluations usually report whether the agents complete the task and how much the run costs, leaving the coordination inside the team largely unmeasured. We introduce an instrument to measure this coordination. Each run is represented as a temporal network in which agents and files are nodes, and messages, file writes, and file reads are timestamped directed edges with an associated cost. We apply this instrument to 1902 runs, each evaluated with a fixed test suite, across configurations that vary the team size, the team structure, and the file policy. The resulting networks show how coordination changes as teams grow and as the work changes. Direct messaging initially increases close to quadratically with the number of agents, with much of this growth coming from an early round of introductions. As the teams grow further, this increase levels off in the largest teams we study, where agents increasingly communicate through broadcast messages. The task also shapes the network that emerges. Work built around a shared specification produces dense, highly connected teams, while pipeline tasks produce sparse networks organised around local interfaces. Shared files can replace repeated 1-to-1 communication, cutting output tokens by about 42% at eight agents on message-heavy work, while adding overhead when files already carry the coordination. Naming one agent as coordinator creates no communication hub and provides no reliable improvement in success. We also observe an unprompted tendency for agents to seek out hidden grading material. We repeat the key experimental conditions in a sealed environment, replacing the hidden material with marked placeholder files. Across 244 additional runs, agents still reach for it in four fifths of runs, while the coordinator and file-channel findings reproduce.
Giuseppe Destefanis, Tomaso Aste
Department of Computer Science, University College London, UK
Collective artificial intelligence, where multiple agents work on shared tasks, holds potential to solve expansive problems in fields from medicine to collective governance. But while prescriptive engineering solutions abound, we lack descriptive scientific understanding of artificial collectives, and therefore principles for how to design resource efficient multi-agent systems. Through systematic experiments with optimizing agents, we characterize how agent interpretive abilities, rationality bounds, and task qualities interact to shape collective performance. Agents range from specialists, with narrow interpretive abilities, to generalists, with broad ones. Collectives of specialists correspond to sparse, centralized networks, while collectives of generalists correspond to dense, decentralized ones. We show that interpretive network properties have small performance effects on average (0.07 standard deviations of performance). However, for specific task qualities, these effects are 4.5 times larger (0.33 sd) and can reach much higher for certain task qualities (1.84 sd). This leads collectives of generalists to perform better on tasks that involve generating, choosing, and coordinating, while collectives of specialists with a few generalist mediators perform better on tasks that involve negotiating. Rationality bounds then moderate these relationships. At loose bounds, specialists outperform generalists through more effective sampling of high-dimensional decision spaces. At tight bounds, generalists outperform specialists through better gradient estimation. A fundamental trade-off between performance and convergence speed emerges at moderate bounds. These findings suggest that multi-agent design could benefit from matching interpretive networks to both task demands and agents' computational limits, with implications for the efficiency and energy costs of multi-agent systems.
John Meluso, Laurent Hébert-Dufresne, Christoph Riedl +1
Cornell University · University of Vermont · Santa Fe Institute +1