DIANOIA: Diagnostic Decomposition and Joint Optimization for Multi-Agent Reasoning

Researchers introduce DIANOIA, a diagnostic framework for multi-agent LLM systems that decomposes reasoning performance into three measurable channels: coverage, fidelity, and synthesis. The method enables practitioners to identify performance bottlenecks and allocate computational resources more efficiently, achieving significant improvements on multiple benchmarks.

DIANOIA addresses a critical gap in multi-agent AI research: the inability to systematically diagnose why certain architectures succeed or fail on new tasks. Previous work demonstrated that multi-agent LLM systems outperform single-agent baselines, but lacked principled frameworks to guide design choices. This research introduces measurable primitives—coverage (breadth of solution space), fidelity (solution quality), and synthesis (integration effectiveness)—that translate abstract performance gains into actionable diagnostics.

The work emerges from the broader trend toward agent-based reasoning systems as researchers push beyond simple chain-of-thought prompting. Multi-agent approaches have shown promise in mathematics (GSM8K, AIME) and code generation (MBPP), but scaling these systems efficiently remains challenging. DIANOIA's channel-aware decomposition provides a principled methodology for resource allocation, moving beyond trial-and-error experimentation.

For AI developers and practitioners, this framework delivers practical value by enabling task-specific optimization. Rather than applying standardized multi-agent architectures, teams can diagnose bottlenecks and invest tokens accordingly—achieving 5× efficiency gains on some benchmarks. This reduces wasted computation and accelerates deployment of reasoning systems with constrained token budgets, particularly relevant for cost-sensitive production environments.

The release of diagnostic metrics, code, and adapters democratizes this methodology across research teams. Future development likely involves extending DIANOIA to other reasoning domains (code, planning, retrieval) and exploring how bottleneck patterns generalize across model families. The framework's focus on measurable decomposition could influence how multi-agent systems are designed and evaluated industry-wide.

• →DIANOIA decomposes multi-agent reasoning into three measurable channels—coverage, fidelity, synthesis—enabling systematic bottleneck diagnosis
• →The framework achieves 5× token efficiency improvements on MBPP and +4.6 percentage points at matched computational cost across benchmarks
• →Task-specific channel diagnostics allow practitioners to allocate resources strategically rather than applying one-size-fits-all multi-agent architectures
• →Open-source release of diagnostic metrics and adapters enables broader adoption across AI development teams
• →Channel-aware design reframes multi-agent optimization as a resource allocation problem with measurable empirical outcomes