BASIL: BAYESIAN ADAPTIVE STATE IMAGINATION AND LEARNING WITH LIVE TELEMETRY HALTING AND IN-CONTEXT RULE INDUCTION

Planning in unfamiliar environments requires an agent to infer governing rules, decide whether additional computation is useful, and determine when to stop searching. We study these questions in Multiverse, a procedural benchmark of two-player connection games with disjoint training, validation, and test rule families. A baseline pre-search value-of-computation predictor has zero-valued labels on 73% of sampled states and a held-out correlation of −0.004 with reference planning gain. BASIL combines three mechanisms: arule-conditioned world model with boundary-aware convolutions, a discrete board-state carrier and a symbolic tactical horizon; Live Telemetry Halting (LTH), which predicts a completed-search action-value gap from intermediate search telemetry; and BayesianHypothesis Filtration (BHF), which maintains a posterior over finite rule hypotheses across episodes. In a three-seed validation study, full search scores 0.958 against a greedy opponent and 0.771 against a 100-rollout simulator-search opponent, compared with 0.549 and 0.424 for the fast policy. Against the latter opponent, LTH scores 0.646 using 8.2 rather than 33.9 model expansions per move. On the held-out test split, full-search scores are 0.938 and 0.722, respectively. In six-episode hidden-rule matches, posterior entropy falls from 6.49 to 1.69 bits and posterior-sampled search reaches a score of 0.757 in the final two episodes. These results support adaptive planning within the specified rule grammar, while comparisons with heuristic stopping rules and the cost of symbolic operations delimit the conclusions.Keywords—Model-based reinforcement learning; Adaptive computation; Monte Carlo tree search; Optimal stopping; Bayesian rule induction; World models

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24284779
Primary Topic
Artificial Intelligence in Games
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

BASIL: BAYESIAN ADAPTIVE STATE IMAGINATION AND LEARNING WITH LIVE TELEMETRY HALTING AND IN-CONTEXT RULE INDUCTION

Vishal Joshi, Saurabh Shinde
Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Games
article

BASIL: BAYESIAN ADAPTIVE STATE IMAGINATION AND LEARNING WITH LIVE TELEMETRY HALTING AND IN-CONTEXT RULE INDUCTION

Vishal Joshi, Saurabh Shinde
article en

Abstract

Planning in unfamiliar environments requires an agent to infer governing rules, decide whether additional computation is useful, and determine when to stop searching. We study these questions in Multiverse, a procedural benchmark of two-player connection games with disjoint training, validation, and test rule families. A baseline pre-search value-of-computation predictor has zero-valued labels on 73% of sampled states and a held-out correlation of −0.004 with reference planning gain. BASIL combines three mechanisms: arule-conditioned world model with boundary-aware convolutions, a discrete board-state carrier and a symbolic tactical horizon; Live Telemetry Halting (LTH), which predicts a completed-search action-value gap from intermediate search telemetry; and BayesianHypothesis Filtration (BHF), which maintains a posterior over finite rule hypotheses across episodes. In a three-seed validation study, full search scores 0.958 against a greedy opponent and 0.771 against a 100-rollout simulator-search opponent, compared with 0.549 and 0.424 for the fast policy. Against the latter opponent, LTH scores 0.646 using 8.2 rather than 33.9 model expansions per move. On the held-out test split, full-search scores are 0.938 and 0.722, respectively. In six-episode hidden-rule matches, posterior entropy falls from 6.49 to 1.69 bits and posterior-sampled search reaches a score of 0.757 in the final two episodes. These results support adaptive planning within the specified rule grammar, while comparisons with heuristic stopping rules and the cost of symbolic operations delimit the conclusions.Keywords—Model-based reinforcement learning; Adaptive computation; Monte Carlo tree search; Optimal stopping; Bayesian rule induction; World models

Zenodo (CERN European Organization for Nuclear Research)
Indian Institute of Science Education and Research, Bhopal (IN)
Artificial Intelligence in Games
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