State Representation Shapes Long-Horizon Agent Learning and Transfer: An Empirical Study
Using 1.7B, 8B, and 27B Qwen models, we compare supervised learning across workflow horizons and cross the state representation used for training with the one used at evaluation.
ProofPaper ↗
Key points
- An end-to-end success rate alone cannot distinguish failure to learn an action from failure to apply it as task state changes.
- We study this distinction in deterministic tool workflows that provide the current artifact state and candidate action recipes.
- Horizon-matched supervision yields high completion, but an H4-trained 1.7B model transfers poorly to longer workflows under full-state observations.
- An explicit relevance cue improves full-view execution but does not close the gap for the tested checkpoint.
Sources (1)
- [1]State Representation Shapes Long-Horizon Agent Learning and Transfer: An Empirical StudyarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 07:22 AM
Using 1.7B, 8B, and 27B Qwen models, we compare supervised learning across workflow horizons and cross the state representation used for training with the one used at evaluation.
An end-to-end success rate alone cannot distinguish failure to learn an action from failure to apply it as task state changes.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 11, 2026LERR-Nav: Learning to Coordinate Exploration, Recovery, and Reminiscing for Object Navigation
- Oct 9, 2026Taking Tokenization off the Host for Agentic LLM Serving
- Oct 9, 2026Qwen/Qwen-Image-2.1-Turbo
- Oct 9, 2026Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash
- Oct 8, 2026Visual-Wi-Fi Navigation to Signal-Emitting Devices
- Oct 8, 2026nvidia/Qwen3.8-Flash-Next-FP4-Local