Inherit-MAS: Test-Time Evolution of Multi-Agent Systems through Workflow and Execution Inheritance
Inherit-MAS evolves multi-agent workflows at test time, beating baselines on WorkBench and HotpotQA while cutting tokens.
Inherit-MAS is a test-time evolution method for LLM multi-agent systems that inherits workflows and cached execution results instead of redesigning or rerunning unchanged steps. A meta-model proposes worker roles, communication, and tool permissions; a judge scores candidates and diagnoses failures, then validated edits remove unhelpful nodes. With GPT-4o-mini workers it reaches 55.4% completion on WorkBench and 49.7% joint F1 on HotpotQA FullWiki, beating EvoAgent, EvoMAS, and TacoMAS; Qwen3-32B workers also exceed those baselines. Execution inheritance cuts worker-token use by 29.1% on WorkBench and 34.6% on HotpotQA, and total tokens by 5.3% and 18.1%.
- Workflow inheritance edits the latest candidate and can drop unhelpful nodes.
- Execution inheritance reuses results only when request and context fully match.
- GPT-4o-mini reaches 55.4% WorkBench completion and 49.7% HotpotQA joint F1.
- Qwen3-32B workers also beat EvoAgent, EvoMAS, and TacoMAS.
- Worker tokens fall 29.1% on WorkBench and 34.6% on HotpotQA.
Full article219 words · extracted from huggingface.co · click to collapse
Multi-agent systems (MAS) built from large language models coordinate specialized agents to tackle complex tasks, but effective workflows are difficult to design in advance. Test-time evolution refines workflows using execution feedback, yet broad revisions can disturb useful components, while re-executing unchanged requests can incur redundant computation. Inspired by the interplay of inheritance and selection in biological evolution, we introduce Inherit-MAS, which makes inheritance explicit at the workflow and execution levels. A meta-model first synthesizes a workflow of worker agents with declared roles, communication inputs, and tool permissions, and a separately prompted judge scores each executed candidate and diagnoses its deficiencies. In ordinary refinement rounds, workflow inheritance starts from the latest completed candidate, may discard removable nodes judged unhelpful, and applies a validated edit to address the diagnosed deficiency. When the new candidate executes, execution inheritance inherits eligible stored results only if the complete resolved request and execution context match, avoiding redundant model and tool calls. With GPT-4o-mini workers, Inherit-MAS achieves 55.4\% completion on WorkBench and 49.7\% joint F1 on HotpotQA FullWiki, outperforming EvoAgent, EvoMAS, and TacoMAS. With Qwen3-32B workers, it also exceeds these evolving-MAS baselines on both benchmarks. Compared with rerunning the same controller with execution inheritance disabled, execution inheritance reduces worker-token usage by 29.1\% on WorkBench and 34.6\% on HotpotQA, and total token usage by 5.3\% and 18.1\%.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2610.02396