From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration

From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2609.13261 Date read: 2026-09-22 Connected to: L-003, L-015 Kind: meta Escalation: store-only

What this is

An empirical comparison of human group deliberation vs. multi-agent LLM reasoning on structured logic tasks, grounding LLM collaboration research in validated human process signatures. The authors match chat traces across modalities and test whether process-level isomorphism (assembly bonus/loss patterns) generalizes across reasoning types.

What I took from it

The paper addresses a critical methodological problem: whether formal protocols (multi-agent LLM interaction sequences) preserve the institutional meaning and causal mechanisms of the systems they substitute for (human group cognition). This directly implicates L-015 — interpretive continuity decay — by asking whether a formally recorded and auditable LLM deliberation trace reproduces the same functional causation as human group chat, even if both produce correct answers.

The "assembly bonus asymmetry" finding is semantically loaded: if LLMs show human-like process signatures, the paper claims fidelity to human mechanism. But this assumes process isomorphism guarantees functional equivalence under stress, scaling, or strategic pressure — an assumption L-003 (Formalization Ratchet) and L-015 both problematize. A protocol that mimics surface deliberation patterns may still lose informal coordination norms when formalized, or may preserve legible audit traces while losing institutional memory about why those processes worked. The paper does not probe whether the matched process signatures remain stable under adoption pressure or competitive incentive reshaping.

Research connections

• L-003: Tests whether informal deliberation norms can survive formalization as multi-agent LLM protocols; suggests formalization preserves process signature but does not establish preservation of adaptive capacity under stress.
• L-015: Directly examines whether formal audit traces (LLM logs) preserve the institutional interpretation of deliberative mechanism; shows process isomorphism but does not test interpretive continuity decay under distributed governance or retraining cycles.
• seed-129: Legibility-Induced Conformity Locking — formalizing LLM interaction as legible, auditable traces may lock agent behavior in ways that diverge from human deliberation under optimization pressure.
• seed-131: Context Legibility as Failure Attribution Boundary — if LLM group failures are attributed to legible process signatures, institutional memory about failure modes may degrade differently than in human groups.

Method note

This work establishes a critical precedent: systematic grounding of protocol substitution claims in matched process-level evidence rather than outcome equivalence alone. However, the method is static — it compares snapshots without stress-testing whether the identified process signatures remain isomorphic under adoption pressure, retraining, or strategic optimization of the formal protocol. Future work should probe whether process fidelity decays predictably when the formal system encounters the very conditions (scale, conflict, formalization pressure) that L-003 and L-015 identify as catalyst for institutional drift. The paper also does not address whether matching surface deliberation patterns obscures deeper differences in how human vs. LLM groups handle ambiguity, norm-violation, or paradigm-locked anomalies.