AI Paper Atlas

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

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add specialist operators to the operator pool (ablation)ASpec · GPQAaccuracy: 57.4 -> 62.8Improved
add specialist operators to the operator pool (ablation)ASpec · GPQAinference cost (USD): 2.26 -> 0.88Improved
keep base operators alongside specialists (ablation)ASpec · GPQAaccuracy: about 1.5% higherImproved
add the meta-controller retain-then-escalate policy (ablation)ASpec · GPQAaccuracy: 62.7 vs 62.8No clear effect
add the meta-controller retain-then-escalate policy (ablation)ASpec · GPQAtotal cost: about 2.3x cost without itImproved
use learned meta-controller instead of coin-flip/cosine alternative control policies (ablation)ASpec · GPQAaccuracy: 58.3 -> 62.8Improved
use LLM-as-gate instead of lightweight meta-controller (ablation)ASpec · GPQAaccuracy / cost: 62.5 vs 62.8, about 4.25x costNo clear effect
add specialist memory (cultivation) (ablation)ASpec · GPQAaccuracy: about 60.9 -> 62.8Improved
use k=1 specialist (ablation)ASpec · GPQAaccuracy: 62.8 -> 58.8Hurt
use k=10 specialists (ablation)ASpec · GPQAaccuracy: 62.8 -> 60.9Hurt
restrict operator pool to specialists trained on a different benchmark (OnlySpec) (ablation)ASpec · cross-benchmark transferaccuracyNo clear effect
use ASpec instead of vanilla modelGemini 2.0 Flash · GPQAaccuracy: 56.3 -> 62.8Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy