prompt
FreeJoint prompt optimization for multi-agent LLM systems
FreeFree tier
About prompt
MASPO Joint Prompt Optimizer is a prompt optimization system for LLM-based multi-agent systems. It jointly optimizes role-specific prompts across interacting agents using a multi-granularity evaluation that considers local validity, lookahead potential, and global alignment. It employs misalignment-driven generative search with hard-negative mining and evolutionary beam search with adaptive beam refresh. The optimizer does not require ground-truth labels for intermediate agents. It achieves an average improvement of +2.9% over state-of-the-art methods on benchmarks including MATH-500, AQuA, AGIEval-MATH, GPQA-Diamond, MBPP, and HumanEval-ET.
Key Features
Joint optimization across interacting agents in a multi-agent system
Multi-granularity evaluation: Local Validity, Lookahead Potential, Global Alignment
Misalignment-driven generative search with hard-negative mining
Evolutionary beam search with Beam Refresh adaptive dynamics
Trace-guided mutation; no ground-truth labels required for intermediate agents
Demonstrated average +2.9% improvement over SOTA on multiple benchmarks
Pros & Cons
Pros
- Focuses on whole-system success rather than individual agent accuracy
- Uses a principled three-level evaluation to detect coordination failures
- No need for ground-truth labels on intermediate agent outputs
- Empirically outperforms existing prompt optimization methods on math, reasoning, and coding benchmarks
Best For
Optimizing role-specific prompts for every agent in LLM-based multi-agent systemsImproving coordination and final system output in multi-agent LLM pipelinesAutomatically tuning prompts for agent-based reasoning and coding tasks
FAQ
What does MASPO stand for?
MASPO stands for Joint Prompt Optimization for LLM-based Multi-Agent Systems.
What evaluation metrics does MASPO use?
MASPO uses three lenses: Local Validity (does the agent fulfill its role?), Lookahead Potential (does the agent's output set up successor agents to succeed?), and Global Alignment (does the output improve the final system response?).
Does MASPO require labels for intermediate agent outputs?
No, MASPO does not require ground-truth labels for intermediate agents; it works by mining misalignment cases and using evolutionary search.