Research map / Multi-agent systems and games / LLM agents
Automated Stateful Specialization for Adaptive Agent Systems
Myan Vu, Harrish Ayyanar, PANG JIANG, Anwiketh Reddy, Mayank Goel
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes automated design of stateful teams of specialist agents that accumulate knowledge across tasks and can be reconfigured for new tasks without human intervention.
Problem
Automated agent design frameworks either produce static workflows that lack adaptability or per-query optimizers that cannot accumulate agent-level task expertise.
Approach
Creates stateful teams of specialist agents that accumulate knowledge over time and are reconfigured for novel tasks automatically, reconciling static workflow design with per-query optimization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add specialist operators to the operator pool (ablation) | ASpec · GPQA | accuracy: 57.4 -> 62.8 | Improved |
| add specialist operators to the operator pool (ablation) | ASpec · GPQA | inference cost (USD): 2.26 -> 0.88 | Improved |
| keep base operators alongside specialists (ablation) | ASpec · GPQA | accuracy: about 1.5% higher | Improved |
| add the meta-controller retain-then-escalate policy (ablation) | ASpec · GPQA | accuracy: 62.7 vs 62.8 | No clear effect |
| add the meta-controller retain-then-escalate policy (ablation) | ASpec · GPQA | total cost: about 2.3x cost without it | Improved |
| use learned meta-controller instead of coin-flip/cosine alternative control policies (ablation) | ASpec · GPQA | accuracy: 58.3 -> 62.8 | Improved |
| use LLM-as-gate instead of lightweight meta-controller (ablation) | ASpec · GPQA | accuracy / cost: 62.5 vs 62.8, about 4.25x cost | No clear effect |
| add specialist memory (cultivation) (ablation) | ASpec · GPQA | accuracy: about 60.9 -> 62.8 | Improved |
| use k=1 specialist (ablation) | ASpec · GPQA | accuracy: 62.8 -> 58.8 | Hurt |
| use k=10 specialists (ablation) | ASpec · GPQA | accuracy: 62.8 -> 60.9 | Hurt |
| restrict operator pool to specialists trained on a different benchmark (OnlySpec) (ablation) | ASpec · cross-benchmark transfer | accuracy | No clear effect |
| use ASpec instead of vanilla model | Gemini 2.0 Flash · GPQA | accuracy: 56.3 -> 62.8 | Improved |
Closest papers by meaning
- Automated Design of Agentic Systems (ICLR 2025 · similarity 0.88)
- AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.88)
- AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
- REGENT: A Retrieval-Augmented Generalist Agent That Can Act In-Context in New Environments (ICLR 2025 · similarity 0.87)
- SEAgent: Self-Evolving Computer Use Agent with Autonomous Learning from Experience (ICML 2026 · similarity 0.86)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
