An Actionable Diagnosis of Multilingual, Multi-Agent Planning Failures

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.03735 Date read: 2026-09-02 Connected to: L-003, L-015 Kind: content Escalation: store-only Escalation rationale:

What this is

An empirical study of failure modes in LLM-based multi-agent planning systems when operating in non-English languages. The paper derives a taxonomy of planning-grounding failures (request-to-action conversion losses) and correlates failure frequency with language-resource scarcity.

What I took from it

The work confirms L-003 (Formalization Ratchet): as multilingual systems scale, informal task semantics must be formalized into executable plans, and this formalization process itself becomes a locus of failure. The paper shows what is lost in translation (task-critical information), but does not theorize why formalization under resource constraint produces systematic degradation, nor does it examine whether degradation follows a generalizable pattern across domains.

The connection to L-015 (Interpretive Continuity Decay) is present but underdeveloped: the paper documents that institutional knowledge about task context is not preserved across the user-request → LLM-plan → execution chain, but treats this as a technical problem in grounding rather than as a governance or coordination failure. The multilingual angle is a proxy for resource asymmetry, which is interesting but not the theoretical object under study.

Research connections

L-003: Formalization pressure in low-resource languages forces informal semantics into computable form; failure rate increases with formalization load, consistent with ratchet hypothesis, but mechanism not isolated.
L-015: Task-context institutional knowledge is lost across the planning layer even when formal records survive; suggests interpretive continuity decay operating at the semantic level, not just governance level.
seed-062 (Formalization Opacity Collapse): Planning failures concentrate at the formalization boundary — the automation of translation from natural request to formal plan — suggesting that legibility gains in one layer create opacity in another.

Seed

Seed title: Formalization Cost Asymmetry in Resource-Constrained Protocols

Seed type: observation

Seed text: When a protocol must formalize task-critical semantics that are sparse or non-canonical in a low-resource linguistic/computational environment, the cost of that formalization is paid in failure rate rather than in computational overhead. The information loss is not random: it concentrates on task context, intent, and edge-case semantics — precisely the dimensions that are over-represented in high-resource environments because they have been iterated on through prior failures. This suggests a generalization: formal protocols in resource-asymmetric settings do not simply copy high-resource solutions with higher latency; they systematically lose categories of information that were never formalized because they were never scarce in the original domain. This pattern likely holds across multimodal, multilingual, and cross-institutional protocol adoption.