AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

Agent Data Protocol

Yueqi Song, Ketan Ramaneti, Zaid Sheikh, Ziru Chen, Boyu Gou, Tianbao Xie, Yiheng Xu, Danyang Zhang, Apurva Gandhi, Fan Yang, Joseph Liu, Tianyue Ou and 9 more

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration

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.

Introduces the Agent Data Protocol (ADP), a light-weight standardized representation that unifies heterogeneous agent training data formats to enable large-scale supervised finetuning of AI agents.

Problem

Large-scale supervised finetuning of AI agents is rarely studied publicly because agent training data is fragmented across heterogeneous formats, tools, and interfaces.

Approach

A light-weight common protocol for agent trajectories that converts diverse existing agent data sources into one unified format, so they can be combined for SFT rather than treating data scarcity as the bottleneck.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Fine-tune on unified ADP dataQwen-2.5-7B-Coder-Instruct with SWE-Agent · SWE-Bench Verifiedresolve accuracy: 0.4% -> 20.2%Improved
Fine-tune on unified ADP dataQwen-2.5-14B-Coder-Instruct with OpenHands · SWE-Bench Verifiedresolve accuracy: 5.8% -> 30.6%Improved
Fine-tune on unified ADP dataQwen-2.5-32B-Coder-Instruct with AgentLab · WebArenasuccess rate: 10.9% -> 22.9%Improved
Fine-tune on unified ADP dataQwen-2.5-7B-Instruct with OpenHands · GAIAaccuracy: 7.3% -> 9.1%Improved
Fine-tune on unified ADP dataQwen-2.5-32B-Coder-Instruct with OpenHands · AgentBench OSaccuracy: 27.8% -> 34.7%Improved
Use ADP mixed corpus instead of SWE-smith only (ablation)Qwen-2.5-7B-Instruct OpenHands · SWE-Bench Verified (cross-task)resolve accuracy: 1.0% -> 10.4%Improved
Use ADP mixed corpus instead of Go-Browse only (ablation)Qwen-2.5-7B-Instruct AgentLab · WebArena (cross-task)success rate: 16.0% -> 20.1%Improved
Use ADP mixed corpus instead of AgentInstruct only (ablation)Qwen-2.5-7B-Instruct OpenHands · GAIA (cross-task)accuracy: 0.6% -> 9.1%Improved
Use ADP mixed corpus instead of AgentInstruct only (ablation)Qwen-3-8B OpenHands · AgentBench OS (cross-task)accuracy: 21.5% -> 25.7%Improved
Use ADP over size-matched SWE-smith data (about 30K) (ablation)Qwen-3-8B OpenHands · SWE-Bench Verified (equal data ~30K)resolve accuracy: 11.0% -> 16.6%Improved
Train on ADP vs SWE-smith, 32B SWE-AgentQwen-2.5-32B-Coder-Instruct SWE-Agent · SWE-Bench Verifiedresolve accuracy: 40.2% -> 40.3%No clear effect
Train on single-domain data (CodeActInstruct + Code-Feedback) instead of ADP (ablation)Qwen-3-8B OpenHands · SWE-Bench Verified (cross-task)resolve accuracy: 16.6% -> 0.2%Hurt

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