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

What Do Agents Learn from Trajectory-SFT: Semantics or Interfaces?

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

Proposes PIPE, a protocol that minimally rewrites agent-environment action interfaces (synonym and symbol aliases) to test whether trajectory-SFT agents rely on interface shortcuts, and introduces Interface Reliance (IR) to quantify preference for training-time action names.

Problem

Standard agent benchmarks cannot distinguish semantic tool-use from memorization of training-time interface patterns, so gains from trajectory-based SFT may reflect interface shortcuts rather than environment-invariant capability.

Approach

Rewrite only the action names and formats in the interface specification (synonym- and symbol-based perturbations) while keeping tasks and environment dynamics fixed, then measure the score gap and a counterbalanced alias-based Interface Reliance metric.

Key results

• On AgentBenchPlus, AgentLM-14b drops from 61.2 on WebShop original to 4.44 and 3.08 under perturbations.
• On AgentGymPlus, Gemma3-4b trained on AgentTraj-L drops from 82.23 to 52.99 on SciWorld Perturb 1 and has IR 7.69 on SciWorld, while untrained agents have IR near 1.
• Interface Reliance for AgentLM-13b on ALFWorld reaches 15.50, indicating near-exclusive use of original action names after training.

Surprising findings

• Agent-FLAN is less affected by perturbations than AgentLM models, with positive Delta on some environments, though this is correlational.
• Untrained agents sometimes improve under perturbation (e.g., KnowledgeGraph, Mind2Web), suggesting PIPE does not uniformly increase difficulty.

Limitations

• Excludes many AgentGym environments requiring online API access, limiting coverage.
• Correlational rather than causal attribution for Agent-FLAN's robustness; no controlled ablations.
• Perturbation difficulty neutrality is assumed via design rather than fully verified for all environments.

Experimental setup

Datasets: AgentInstruct AgentTraj-L AgentBench AgentGym

Benchmarks: AgentBench (ALFWorld, OperatingSystem, DataBase, Mind2Web, WebShop, KnowledgeGraph) AgentGym (ALFWorld, BabyAI, MAZE, Movie, SciWorld, SQLGym, TextCraft, Weather, WebShop, Wordle)

Metrics: success rate reward F1 score gap Delta Interface Reliance (IR) legacy action invocation count

Baselines: AgentLM-7b AgentLM-14b AgentLM-70b AgentFLAN GPT-3.5 GPT-4o-mini GPT-5-mini

Builds on: AgentBench AgentGym AgentInstruct Agent-FLAN

Closest papers by meaning

• HATS: Hardness-Aware Trajectory Synthesis for GUI Agents (CVPR 2026 · similarity 0.89)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.89)
• Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks (ICML 2026 · similarity 0.88)
• Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use (ICML 2026 · similarity 0.88)
• Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.88)
• G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.88)