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

Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use

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

Read the paperCheck your idea against 108,967 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.

Formalizes open-world tool-use shifts (OpenAgent), diagnoses SFT and GRPO agents in a synthetic POI sandbox across four tiers, and proposes PAFT, a perturbation-augmented SFT data strategy that recovers much of the lost robustness.

Problem

LLM tool-use agents trained under a static, closed-set assumption degrade when user queries, tool sets, observations, and domains shift at test time; it is unclear whether SFT or RL training generalizes to such open-world shifts.

Approach

Defines the OpenAgent setting with four shift dimensions (query, action, observation, domain) and a four-tier diagnostic sandbox (Perception, Interaction, Reasoning, Internalization); proposes PAFT, which augments clean SFT trajectories with mid-trajectory environmental feedback anomalies, unsolvable (fatal-error) states, and symbolic tool-name perturbations.

Key results

• SFT at 200 steps shows Tier-1 accuracy delta of -67.7 versus -29.8 for base; adding PAFT raises it to +28.6 for that checkpoint.
• At 800 steps, PAFT improves Tier-4 refusal rate from 0.2 to 99.6 while SFT alone stays near 0.
• Ablation: SBP alone restores refusal (RR 99.2) but Tier-1 delta remains -38.9; SRP alone gives best Tier-1 (-14.9) among single components.

Experimental setup

Datasets: Synthetic POI multi-step tool-calling dataset (Beijing POI database, anonymized)

Benchmarks: OpenAgent sandbox tiers (Perception, Interaction, Reasoning, Internalization)

Metrics: accuracy delta tool error rate active exploration score average tool chain length refusal rate

Baselines: SFT GRPO Base Qwen2.5-7B-Instruct

Builds on: GRPO ToolRL LLaMA-Factory Verl-Tool

Closest papers by meaning

• AgentRefine: Enhancing Agent Generalization through Refinement Tuning (ICLR 2025 · similarity 0.89)
• Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.89)
• Don't Just Fine-tune the Agent, Tune the Environment (ICLR 2026 · similarity 0.89)
• AgentNoiseBench: Benchmarking Robustness of Tool-Using LLM Agents Under Noisy Condition (ICML 2026 · similarity 0.88)
• What Do Agents Learn from Trajectory-SFT: Semantics or Interfaces? (ICML 2026 · similarity 0.88)
• Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.88)