Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners.
TL;DR: Explore the results, browse human and model replays, or play the games.
NeurIPS 2026 paper · Research code and workflows · Hugging Face dataset
The replay viewer shows gameplay alongside model rationales and the full conversation.
Run python -m http.server 8000 from this directory and open
localhost:8000.
Browser engine and viewer source live in botcs/browser-game-interpreter. This repository serves the website and its bundles.
data-source.js pins the
verified dataset snapshot
used by the catalogue, replays, RDMs and game descriptions. To use a downloaded
dataset, set datasetRoot to its HTTP root, for example "/dataset". Setting it
to null uses the public CDN (or a data/ mirror on localhost). Keep the
catalogue and payloads on the same dataset version.
website-assets/replays/manifest.json indexes compact replays under
website-assets/replays/human/ and website-assets/replays/lrm/. These retain
displayed trajectories and complete conversations; full research recordings
remain under behavior/human/ and behavior/lrm/. Human catalogue entries use
the elaborate prompt condition. RDM metadata and binaries are under
website-assets/rdms/.
The dataset's files catalogue provides sizes and SHA-256 checksums. Original
human data are available from
OpenNeuro ds004323 v1.0.0.
This project builds on the VGDL framework and codebase released by Cedric Colas et al. for "Language and Experience: A Computational Model of Social Learning in Complex Tasks". The Browser Game Interpreter (BGI) is a JavaScript re-implementation of the VGDL engine for interactive visualization and replay.
MIT License. See LICENSE for details.