Source-linked AI summary
Discovering Adaptive Transmission Programs for Collective Innovation
Cédric Colas, Jérémy Perez, Eleni Nisioti, Akhilesh Mocherla, Pierre-Yves Oudeyer, Clément Moulin-Frier, Maxime Derex
TL;DR
Collective innovation depends on transmission decisions that network-only models cannot condition on agent or collective states. This paper uses LLM-guided evolutionary program search to discover state-aware transmission protocols, finding improved performance, state-dependent advantages, and transfer across settings while identifying important scope limits.
Problem
Network-based transmission models are state-agnostic and cannot condition routing on what agents know or on the collective state, despite the state-dependent nature of real transmission.
Method
The paper formalizes transmission protocols as interpretable state-aware programs and uses LLM-guided evolutionary search to evaluate them across agent populations and domain configurations.
Results
37%: evolved protocols improve collective performance over literature baselines by up to 37%, with ablations linking the advantage to state-aware content routing and protocols transferring across agent types and domain variants.
Takeaways & Limitations
In-silico program search can discover effective and generalizable transmission protocols, suggesting a path toward AI-assisted design of coordination infrastructure for collective intelligence.
Takeaways & Limitations
The study uses a simplified deterministic domain, assumes centralized access to all agent states, and has not tested transfer from in silico protocols to human groups.
Abstract
from arXiv · showhide
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.
Introduction
The paper frames collective innovation as a transmission problem and argues that state-aware protocols can overcome the limits of network-only approaches. It introduces interpretable program search to discover protocols that improve and generalize collective performance.
- Motivation: Collective intelligence depends on how knowledge flows between individuals and generations, including who shares what, how, and when.
- Motivation: Transmission can be directed by top-down protocols such as organizational reporting structures, scientific publication norms, and recommendation-system routing.
- Prior work: Network-based studies vary who shares with whom and generally find that intermediate connectivity preserves diversity better than isolation or full connectivity.
- Research gap: Networks are state-agnostic, whereas real transmission can depend on agent knowledge, collective state, relevance, expertise, and context.
- Approach: The paper formalizes interpretable transmission protocols as state-aware programs and searches for them in silico while testing generalization across agents and task configurations.
- Findings: 37%: discovered protocols outperform literature baselines by up to 37%, with state-awareness and transfer across agent types and domain configurations supporting their effectiveness and generalizability.
Methods
The methods build a collective discovery simulation with stateful transmission programs, bounded-information agents, literature baselines, and LLM-guided evolutionary search. Protocols are evaluated for collective performance and tested under alternative agents and configurations.
- Collective innovation task: The simulation models cumulative innovation as agents combining elements over synchronous steps while sharing memories of past combinations.
- Agent models: Primary empowerment agents choose combinations using a softmax over estimated action empowerment, with learned success probabilities and transition distributions updated through experience.
- Agent models: A social-information bias gives agents probability p_social = 0.5 of restricting candidate combinations to those involving socially received elements.
- Agent models: Alternative stochastic and LLM agents test whether protocol effects generalize beyond the primary empowerment model.
- Transmission constraints: Agents receive at most 20 memories per step, making strategic routing necessary rather than allowing unrestricted broadcasting.
- Transmission protocols (TP): Transmission protocols are executable Python classes whose share method observes agent states and returns memory exchanges, while remaining unable to modify agents or access the recipe graph.
- Transmission protocols (TP): Protocols can condition routing on agent and collective states, including recipient knowledge, failed combinations, and collective progress.
- Protocol search: The search uses an LLM-guided evolutionary archive over 200 iterations, combining baseline-conditioned warmup, parent selection, contrast solutions, candidate refinement, and sandbox validation.
Results
LLM-guided evolutionary search produced high-performing transmission protocols that outperform baselines, rely on state-aware content routing, and generalize across populations and task variants. Qualitative and quantitative analyses indicate that effective protocols adapt what, whom, and when to share.
- LLM-guided search discovers high-performing transmission protocols: +33% average advantage over the best baseline at step 150, with the top protocol reaching +37%.Across five evolutionary runs, all evolved protocols significantly outperformed the baselines; at step 300, the average advantage remained +32% and the top protocol reached +36%.
- LLM-guided search discovers high-performing transmission protocols: High-performing protocols from different evolutionary runs occupied distinct regions of the protocol design space rather than clustering together.This embedding result indicates that the search found multiple distinct high-performing solutions.
- State-awareness drives protocol performance: State-awareness drives performance: randomizing transmitted content while preserving each protocol’s schedule reduced top-protocol performance from 88 to 67, matching the asocial baseline.The same ablation had negligible effect on bottom protocols, with Δ = 0.7, indicating that topology and timing alone provided no benefit.
- State-awareness drives protocol performance: Evolved protocols filter and route information using rarity, recipe usefulness, recipient inventories, sharing history, failures, group structure, and game progress.Examples include sharing items owned by fewer than 20–40% of agents, routing failed combinations to agents holding both ingredients, and narrowing recipients as common items saturate the population.
- State-awareness drives protocol performance: Recency bias was the strongest measured behavioral channel, correlating with fitness at ρ= 0.64, while recipient-conditioned recency correlated at ρ= 0.63.Ingredient-based routing also correlated with fitness at ρ= 0.40–0.43, and effective protocols reduced early-game diversity while increasing late-game diversity.
- Generalization across populations and domains: Evolved protocols outperformed baselines across smaller and larger populations, alternative agent types, modified agent parameters, and all tested domain variants.Population transfers included +17% for 6 agents and +49% for 20 agents; domain transfers included +12% for depth-limited and +25% for 3-item-start variants.
Discussion
The discovered protocols rely on state-aware content routing and transfer across agent types and domain variants. The paper identifies open questions about which state dimensions matter, broader-domain transfer, partial observability, and transfer to human groups.
- State-aware content routing, rather than network topology or timing, drives protocol effectiveness.Ablations preserve topology and timing while removing state-dependent routing.
- The same protocols transfer across agent types and domain variants, suggesting they exploit structural properties of collective search.
- Recency bias, ingredient-based routing, and temporal adaptation are the strongest state-dependent predictors in the ablation-based correlational analysis.
- Transfer beyond the simplified deterministic domain remains untested, as does discovery under partial observability and transfer to human groups.The protocols currently assume centralized access to all agent states.
- Program search could extend to organizational structures and opinion-mediation strategies, positioning AI as coordination infrastructure rather than a substitute for human judgment.
A Additional Methods
The methods model empowerment-based agents and use an LLM-guided evolutionary archive to generate, refine, validate, and evaluate executable transmission protocols.
- Empowerment agents: At each game step, agents select combinations via softmax over empowerment values, excluding combinations with zero empowerment.The temperature is β=0.05.
- Empowerment agents: Combination empowerment is the expected empowerment of the resulting element, based on success probability, transition distribution, and result empowerment.
- Empowerment agents: An element’s empowerment is the number of unique elements it can help produce, or zero when the agent already owns it.
- Empowerment agents: Agents update estimates from failed combinations, successful own attempts, and successful social memories as knowledge accumulates.
- Empowerment agents: A social inventory creates a 0.5 probability of restricting candidate combinations to those involving a socially received element, while attention limits reception to 20 memories per step.
- Protocol search: The archive stores executable protocols, descriptions, and fitness scores; warmup seeds diversity before evolution selects parents and contrast solutions for improvement or diversification.
- Protocol search: Candidates are sandbox-validated and evaluated by mean collective performance across 30 simulations with 10 agents over 150 game steps.
A.3 State-awareness ablation details
The ablation preserves each protocol’s transmission schedule while randomizing message content, enabling matched tests of whether performance depends on state-aware routing rather than schedule structure.
- A.3 State-awareness ablation details: The state-agnostic ablation replays who sends to whom, how many memories are sent, and when, while randomly selecting the transmitted memories.
- A.3 State-awareness ablation details: The study selects the top and bottom 20 protocols from each of five generation runs using held-out evaluation seeds.
- A.3 State-awareness ablation details: Matched comparisons use the remaining 15 evaluation seeds, pairing original and schedule-matched ablated trajectories within each seed.
- A.3 State-awareness ablation details: Shadow-log analysis compares behavioral-feature effect sizes and fitness correlations between original and ablated conditions.
- A.3 State-awareness ablation details: Features with Cohen’s d>2.0 and substantial drops in fitness correlation are interpreted as signatures of state-aware programming.
- A.3 State-awareness ablation details: Transmission features are aggregated across events and evaluation seeds, with temporal variants dividing simulations into early, middle, and late periods.
B.1 Protocol fitness across evolution
Across five generation and five evolution runs, search progressively improves the best protocol’s fitness from baseline performance to substantially higher discovery counts.
- 85–97 elements discovered after 150 game steps represents +21–38% over baselines achieving 66–70 elements.
- Both generation and evolution searches progressively discover better protocols from archives seeded with baseline protocols.
- Figure 5 plots the fitness of the best evolved protocol against search step across five generation and five evolution runs.
B.2 Extended baseline experiments
The study evaluates evolved and generated transmission protocols against diverse literature baselines. None of the baselines reaches performance comparable to the best protocols identified by search.
- B.2 Extended baseline experiments: The evaluation includes asocial, stochastic, paired, dynamic, random-graph, ring, small-world, caveman, multilevel, and intermittent baselines.These baselines vary information sharing, pairing, graph topology, modularity, hierarchy, and interaction timing.
- B.2 Extended baseline experiments: Table 1 summarizes behavioral features computed from transmission logs and averaged across transmissions and evaluation seeds.Features indexed by k are computed separately for k ∈ {0, 1, 2}.
- B.2 Extended baseline experiments: All social baselines use random content selection, isolating network topology from content-selection strategy.This removes state-awareness from the baseline protocols while retaining their differing network structures and schedules.
- B.2 Extended baseline experiments: None of the baselines reaches performance comparable to the best evolved and generated protocols.Figure 6 compares one best protocol per search run with the extended baseline set over 300 game steps and 30 simulations.
B.3 Evolution explores beyond the generative prior.
Evolution moves protocols beyond the shared LLM-generated prior. Best-performing protocols occupy multiple regions of the discovered design space rather than forming one cluster.
- B.3 Evolution explores beyond the generative prior.: Evolved–generated divergence is substantially higher than generated–generated divergence, showing that evolution pushes protocols away from the prior.Generated protocols and evolution warmup share an LLM prior conditioned on baseline protocols.
- B.3 Evolution explores beyond the generative prior.: Evolved–evolved divergence exceeds generated–generated divergence, indicating that different search seeds explore distinct regions.Distributional divergence is measured with maximum mean discrepancy using a cosine similarity kernel.
- B.3 Evolution explores beyond the generative prior.: Best-performing protocols are spread across these regions rather than clustered, suggesting multiple distinct strategies can achieve high performance.Per-run best protocols are marked with stars in the t-SNE projections.
B.4 Transmission effectiveness is not correlated with volume
Protocol effectiveness is not correlated with transmission volume. The analysis measures volume as novel memories received by agents that had not independently experienced them at reception.
- B.4 Transmission effectiveness is not correlated with volume: Across 997 generated protocols, fitness shows no correlation with novel memories received per agent.Novel memories are counted when an agent receives a memory it had not experienced independently at the time of reception.
- B.4 Transmission effectiveness is not correlated with volume: The result is consistent with prior findings that more sharing does not necessarily improve collective performance.The comparison concerns transmission quantity rather than the specific content or routing strategy of each protocol.
B.5 Behavioral feature analyses.
The behavioral analysis examines how evolved transmission protocols use agent roles, inventories, rarity, recency, and collective progress to select memories for sharing. The analyzed strategies combine specialized within-group exchange with selective cross-group transmission and failure information.
- Behavioral features: Other strategies distribute rare discoveries across learners and share feasible failure signals to prevent repeated mistakes.Rare-item relay uses learner-specific offsets, while failure signals are shared only when the learner possesses both ingredients.
- Behavioral features: Protocols assign agents to roles or logical niches and use those assignments to structure transmission.Some implementations create four role-based silos or lineages associated with the basic elements.
- Behavioral features: Transmission prioritizes memories based on recipient inventory, item rarity, recency, or whether discoveries are pivotal.Candidate memories can receive higher scores when the learner already has both ingredients or when the result is used repeatedly as a recipe parent.
- Behavioral features: Within-group sharing is high-volume, while cross-group exchange is selective and focused on rare or potentially useful discoveries.One protocol explicitly contrasts intra-silo high-volume sharing with selective inter-silo crossing.
- Behavioral features: Some protocols allocate separate quotas to intra-lineage sharing and inter-lineage cross-pollination, with the intra-lineage quota reduced later in the simulation.The quota changes from 16 early to 8 later, while inter-lineage sharing fills the remaining capacity.