Multi-LLM Systems Exhibit Robust Semantic Collapse

Summary

The strongest existing negative result against the premise that multi-agent LLM systems expand exploration: across model families and simulations of 200-1,000 rounds, inference-only multi-LLM systems converge in semantic representation despite lexical variation, and thirteen interventions (decoding parameters, prompt design, agent composition, activation engineering, RL) all fail to restore diversity. Crucially the authors rule out alignment and conformity bias as the mechanism and attribute collapse to intrinsic properties of autoregressive generation. This is the paper the exploration/novelty construct must be defended against: if collapse is architecture-invariant, the claim that structure moves novelty needs an explicit scope condition, and the open-ended research-forecasting task (not a closed loop) is the natural boundary to argue.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a high-priority candidate for the Architecture as Epistemology (ISR) paper (angle: MAS topology / structure vs. aggregation rule; correlated errors / homogeneity across LLM agents; epistemic quality of LLM outputs). Verified against: Opened https://arxiv.org/abs/2605.17193 (arXiv abstract page; confirmed title, 4 authors, cs.MA, v1 16 May 2026 / v2 1 Sep 2026, full abstract read)

Critique / open questions

Queued — not yet assessed.

Key quotes

Whether machines can originate novel content has been debated for nearly two centuries, from Lovelace's assertion that no engine can "originate anything" to Turing's question of whether a machine can amplify ideas brought in from outside. (abstract)

Systems of multiple interacting LLMs, increasingly deployed for autonomous generation, reopen this question empirically. (abstract)