RELIC: Revealed Principles for Learning Interpretable Composable Skills in Multi-Agent Planning

RELIC: Revealed Principles for Learning Interpretable Composable Skills in Multi-Agent Planning

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2607.16745 Date read: 2026-09-02 Connected to: L-001, L-005, seed-029 Kind: content Escalation: store-only Escalation rationale:

What this is

A multi-agent reinforcement learning framework addressing skill composition when agents maintain private implementations and expose only opaque interfaces. The work introduces RELIC to enable agents to learn interpretable, composable decision-making skills under privacy constraints — a regime where centralized policy optimization is unavailable and agents must infer coordination norms from revealed behavior alone.

What I took from it

The paper situates itself in a genuine practical constraint space — agents cannot share executable policies, internal state representations, or algorithmic details — and proposes to recover composability through interpretable skill exposure rather than full transparency. This is orthogonal to protocol ossification and Gall-style system resistance in an important way: RELIC does not address what happens after adoption locks in an interface; it addresses how to coordinate despite heterogeneous, opaque interfaces from the start.

The work thus confirms the difficulty of coordination under formalism gaps (resonant with L-005 and seed-029), but does not engage with the temporal or irreversibility dynamics that make ossification a law rather than a solvable design problem. It is a competent engineering response to a real constraint, not a claim about what happens to protocol systems under stress or adoption pressure. The interpretability angle is pragmatic (agents learn to read each other's revealed actions) rather than foundational (what makes some interfaces inherently resist modification while others do not).

Research connections

• L-001: Indirect: RELIC assumes opaque, heterogeneous interfaces as a boundary condition, not as an emergent result of adoption pressure. The paper does not test whether interface opacity increases under scale or adoption.
• L-005: Relevant but not challenging: RELIC is additive composition of existing opaque agents, which aligns with the claim that complex systems resist restructuring. It does not test whether attempts to re-integrate or homogenize such systems fail predictably.
• seed-029: Mirrors the stated triage connection — skill learning across heterogeneous boundaries does produce coordination friction analogous to protocol interface locks. But the paper treats this as a constraint to engineer around, not as a law to characterize.
• L-012: Weak signal — skill exposure (the "revealed principles") becomes a legible input to other agents' planning; whether this displaces optimization pressure upstream (into skill selection itself) is not examined.

Seed

Seed title: Interpretability as Coordination Proxy Under Interface Opacity

Seed type: motif

Seed text: When agents must coordinate across opaque interfaces without policy sharing, they shift optimization pressure from implementation transparency toward behavior predictability — rendering internal decision structure legible through revealed action patterns rather than code exposure. This suggests that under sufficient interface heterogeneity and opacity constraints, interpretability functions as a substitute for formalization, not a precursor to it. The generalization: systems under coordination pressure but blocked from deep integration will accumulate interpretive scaffolding (diagnostics, signal protocols, action-outcome mapping) as a conservation effect — the coordination cost is not eliminated, but displaced into the inference and prediction layer.