Multi-Agent Collaborative Intelligence: Dual-Dial Control for Reliable LLM Reasoning

Multi-agent debate often wastes compute by using a fixed adversarial stance, aggregating without deliberation, or stopping on heuristics. We introduce MACI, an active controller with two independent dials that decouple information from behavior: an information dial that gates evidence by quality, and a behavior dial that schedules contentiousness from exploration to consolidation. A moderator tracks disagreement, overlap, evidence quality, and argument quality, and halts when gains plateau. We provide theory-lite guarantees for nonincreasing dispersion and provable termination, with a budget-feasible scheduler. Across clinical diagnosis and news-bias tasks, MACI improves accuracy and calibration while reducing tokens, and converts residual uncertainty into precision RAG plans that specify what to retrieve next. We use a crossfamily LLM judge (CRIT) as a conservative soft weight and stop signal, validated for order invariance and judge-swap stability; stability depends on using high-capability judges. MACI turns debate into a budget-aware, measurable, and provably terminating controller.

Introduction. Despite intense interest, multi-agent debate has seen limited progress in theory or practice. Unmoderated or ad hoc collaboration lacks provable termination, calibrated uncertainty, and cost control [Cemri et al., 2025]. Common approaches use a fixed adversarial stance [Liang et al., 2024], aggregate without deliberation [Wang et al., 2023], or stop on heuristics, which wastes compute and locks in early errors. Beyond single-dial modulation. Prior work introduced contentiousness modulation [Chang, 2023a,b], showing that behavior matters, but a single dial is insufficient. Systems need concurrent control over information quality and interaction intensity, and they must stop when additional debate yields diminishing returns. The CRIT framework [Chang, 2023a] evaluates arguments but has not been coupled to systematic stopping. MACI: Dual dials with measured stopping. We present MACI (Multi-Agent Collaborative Intelligence), an active controller with two orthogonal dials.

Discussion / Conclusion. Advancing multi-agent orchestration. We presented MACI, a control framework that extends contentiousness modulation Chang [2023a,b] with (1) dualdial control separating information admission from behavioral stance, (2) information-theoretic stopping via relative progress ratios, and (3) adaptive initialization to reduce wasted exploration. Unlike passive aggregation or fixed debate, MACI steers deliberation using measurable signals—disagreement, overlap, evidence quality, and argument quality—and halts when gains plateau. On clinical diagnosis (1,500 cases), MACI improves accuracy (+3.9 pp over majority vote; +3.7 pp over fixed-contentiousness) and calibration (ECE 0.081 vs. 0.103) while using 19% fewer generation tokens; case studies show convergence (disagreement ↓86–96%, CRIT ↑21–26%). CRIT is order-invariant and stable under cross-family judge swaps (2–3% winner flips; Appx. E.7). Theory-lite guarantees bound dispersion and ensure termination in O(1/ε) rounds (improving to O(log(1/ε))), and a budget-feasible UCB scheduler attains ̃O( √ KT) no-regret with zero expected budget violation (Appx. M, L).