Joint State Transition Prediction: Goal Generalization with Sparse Action Labels

Read the full JSTP theory: English translation (34 pages) · 中文原稿 / Chinese original (30 pages). The two companion manuscripts contain all 18 sections and both appendices of Ke Yue's theory of intelligence. They distinguish the author's proposals about action, grouping, language, and the self from the evidence reported so far. The empirical preprint below tests one narrower claim about using partial consequences to act on unfamiliar goals. Public project repository: https://github.com/ykykykk/jstp-intervention-bench. The fixed experimental snapshot remains linked in Appendix A of the paper. Abstract of the empirical preprint. A controller can learn what an action does without being told when to take it. Does that knowledge help once a new goal is specified? Joint State Transition Prediction (JSTP) treats an agent's action and changes in its environment as components of the same transition. We test a network that can predict its action directly or use a latent prediction of a local consequence. Eight methods were trained at three action-label budgets on a fixed bank of 16,384 interactions. Frozen policies operated a four-track simulator on 192 held-out layouts. Each trial pair required the policy to achieve two opposite, previously unseen goals from the same probe history. With 128 labels, JSTP achieved 71.69% paired success, versus 67.45% for a one-step world model and 49.79% for an action-only policy. Disconnecting the consequence-to-action path while retaining both training losses reduced paired success to 66.75% (difference: 4.94 percentage points; 99% paired layout-bootstrap interval: 3.97 to 5.85). The disconnected network scored slightly higher at both larger budgets and had more readily decoded co-motion at 128 labels. The results isolate a benefit of using predicted consequences under sparse action supervision in this apparatus. They do not establish that better control requires more readily decoded object representations. Version 1.1: Adds visible, clickable links on the paper's first page and the full theoretical manuscript in English and Chinese. Experimental results are unchanged from version 1.0.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24366435
Primary Topic
Reinforcement Learning in Robotics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Joint State Transition Prediction: Goal Generalization with Sparse Action Labels

Ke Yue
Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
preprint

Joint State Transition Prediction: Goal Generalization with Sparse Action Labels

Ke Yue
preprint en

Abstract

Read the full JSTP theory: English translation (34 pages) · 中文原稿 / Chinese original (30 pages). The two companion manuscripts contain all 18 sections and both appendices of Ke Yue's theory of intelligence. They distinguish the author's proposals about action, grouping, language, and the self from the evidence reported so far. The empirical preprint below tests one narrower claim about using partial consequences to act on unfamiliar goals. Public project repository: https://github.com/ykykykk/jstp-intervention-bench. The fixed experimental snapshot remains linked in Appendix A of the paper. Abstract of the empirical preprint. A controller can learn what an action does without being told when to take it. Does that knowledge help once a new goal is specified? Joint State Transition Prediction (JSTP) treats an agent's action and changes in its environment as components of the same transition. We test a network that can predict its action directly or use a latent prediction of a local consequence. Eight methods were trained at three action-label budgets on a fixed bank of 16,384 interactions. Frozen policies operated a four-track simulator on 192 held-out layouts. Each trial pair required the policy to achieve two opposite, previously unseen goals from the same probe history. With 128 labels, JSTP achieved 71.69% paired success, versus 67.45% for a one-step world model and 49.79% for an action-only policy. Disconnecting the consequence-to-action path while retaining both training losses reduced paired success to 66.75% (difference: 4.94 percentage points; 99% paired layout-bootstrap interval: 3.97 to 5.85). The disconnected network scored slightly higher at both larger budgets and had more readily decoded co-motion at 128 labels. The results isolate a benefit of using predicted consequences under sparse action supervision in this apparatus. They do not establish that better control requires more readily decoded object representations. Version 1.1: Adds visible, clickable links on the paper's first page and the full theoretical manuscript in English and Chinese. Experimental results are unchanged from version 1.0.

Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon