KDD 2026 Conference paper

Test-Time Deep Thinking to Explore Implicit Rules

A framework that helps agents infer hidden environmental rules by separating action execution from deeper exploration.

Wentong Chen, Xin Cong, Zhong Zhang, Yaxi Lu, Siyuan Zhao, Yesai Wu, Qinyu Luo, Haotian Chen, Yankai Lin, Zhiyuan Liu, Maosong Sun

TTExplore framework showing actor and thinker roles, trajectory history, and deep thoughts. — Fig. 2 — Original TTExplore actor and thinker framework.Original source ↗

01 / The question

How can an agent escape repeated trial and error when an environment’s rules are not directly visible?

02 / The approach

TTExplore

TTExplore pairs an actor with a thinker that analyzes interaction history, proposes implicit rules, and revises the plan. The work trains an Exp-Thinker model using a stable reinforcement-learning pipeline and evaluates its guidance across five text-based embodied tasks.

Read the full paper

03 / Citation

Cite this work

BibTeX

@inproceedings{chen2026testtimedeepthinking,
  title = {Test-Time Deep Thinking to Explore Implicit Rules},
  author = {Chen, Wentong and Cong, Xin and Zhang, Zhong and Lu, Yaxi and Zhao, Siyuan and Wu, Yesai and Luo, Qinyu and Chen, Haotian and Lin, Yankai and Liu, Zhiyuan and Sun, Maosong},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
  year = {2026},
  pages = {449--460},
  publisher = {Association for Computing Machinery},
  doi = {10.1145/3770855.3817790},
  url = {https://doi.org/10.1145/3770855.3817790}
}

Select the citation to copy it manually.

04 / Versions & records

Publication record

Earlier or alternate titles

Test-Time Exploration in Unknown Environments

2 Google Scholar records

READ HERE · 12 PAGES

Full paper

Document

Loading document…

Scroll through every page. Select text to copy, or use search to find a passage.

More research

View all

How can an agent escape repeated trial and error when an environment’s rules are not directly visible?

TTExplore

Cite this work

Publication record

Full paper

More research

Agent-World↗

UI-TARS-2↗

Agent-World

UI-TARS-2