initially-identical-llm-agents-act-shared-persistent-world-where CLM-0578When initially identical LLM agents act in a shared persistent world where their artifacts remain and are scored by an external simulator, the shared society builds broader and more resilient technology portfolios than a matched best-of-N population of isolated agents, but isolated search still yields the single best-performing artifact.
Capability: Getting more from several models than from one
Observed on
Simulated spatial world with 50-200 LLM agents, 800-3,200 ticks, four matched world seeds; outcomes scored by a deterministic simulator under unseen disturbances after agents are r.
Sources
- Validated by: Paired per-seed comparison against an endpoint-wise best-of-N isolated-agent envelope on portfolio resilience (0.2474/0.2365 vs 0.1794), validated inventions (5.75/7.00 vs 2.75), held-out resilience, and best single artifact (0.2380 vs 0.3488), across four seeds and five checkpoints.. Measured on four matched world seeds per cell with paired seed-bootstrap intervals; effect sizes rather than population inference, and discovery-frontier AUC ranking flips with population size, so only the portfolio/resilience/invention direction is consistent. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- Splitting a task across several agents helps on long-horizon tasks whose parts depend on each other sparsely, and hurts on tightly coupled sequential workflows, where generic multi-agent methods fall below a single-agent harness because decomposition fragments the dependency chain and coordination overhead outweighs the context isolation it buys; as single agents improve the collaboration benefit shrinks.Getting more from several models than from one · reviewed by AI
- For teams of up to 50 heterogeneous embodied LLM agents in simulated wildfire missions, building the manager hierarchy per mission so that concurrently executable work sits under a parallel-allocating manager and prerequisite-chained work under a phase-advancing manager beats fixed decentralized, single-planner, and hybrid coordination architectures on mission score and on progress accumulated per step, and the advantage does not depend on which base language model is used.Getting more from several models than from one · reviewed by AI
- For CAE simulation tasks (OpenFOAM, FEniCS, PyChrono, COMSOL) run with a current general-purpose coding-agent harness and a strong backbone, a single agent with no multi-agent role decomposition and no scripted self-critique matches or exceeds the published scores of multi-agent specialized CAE systems, because the harness itself supplies the multi-turn reasoning and execution-feedback repair the decomposition was built to compensate for.Getting more from several models than from one · reviewed by AI
- For LLM web agents on cluttered pages containing many concurrent interface components, success falls sharply as component count grows, and having the agent first emit a localized sub-goal plus a task-relevant sub-observation before choosing an action recovers much of the loss, because self-assessment of the action is more reliable in the reduced context.Using the tools it is given · reviewed by AI
- Across 41 models and seven multi-agent benchmarks, gains in single-agent capability translate only weakly into multi-agent competence (understanding other agents, joint planning, efficient communication, adaptation); on the multi-agent planning task accuracy rose only about 30% across a family's full scale range, so coordination ability does not emerge from scaling alone and needs its own data and training.Getting more from several models than from one · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: A matched best-of-N isolated-agent envelope that equals or exceeds the shared society on portfolio resilience and validated-invention counts across seeds would refute it. Drafted stance toward splitting-task-across-several-agents-helps-long-horizon-tasks-whose: supports -- Collective advantage appears for breadth and redundancy across loosely coupled parallel work, not for optimizing one artifact, consistent with the claim that multi-agent gains depend on dependency structure.