gavinlaw/causalcache-sparse-history-guiodyssey-v5

CausalCache Sparse-History Corpus

A controlled corpus for testing whether a GUI agent actually reads its visual history — or merely benefits from history being present.

Mobile GUI agents normally see a sliding window of the most recent screenshots. That is a poor fit for long-horizon tasks, where the screen you need to remember may be twenty steps back. This corpus lets you train and, more importantly, falsify adapters that claim to read a sparse, non-contiguous selection of historical screenshots.

The policy stays frozen throughout. Every row in a group shares one target action, so the objective is purely about relative teacher-forced log-probability between conditions — cross-entropy weight is zero.

Why the arms are built this way

The trap in this kind of work is measuring a format effect or a history-presence effect and reporting it as a content effect. So each comparison changes exactly one thing:

quantity	comparison	what it isolates
format_effect	R0 − N0	official multi-turn vs single-turn prompt
frozen_selection_effect	S0 − R0	sparse selection, with no adapter at all
adapter_on_recent	RA − R0	adapter's effect on a recent window
adapter_on_sparse	SA − S0	adapter's effect on a sparse window
main_claim	SA − RA	the adapter's contribution, prompt held fixed
deployment_delta	SA − N0	end-to-end vs the deployed baseline

Two properties make this checkable rather than merely asserted:

• S0 is exactly SA's bypass twin — same messages, same images, only adapter_mode differs. So adapter_on_sparse needs no extra forward pass, and a zero-init adapter must reproduce SA − RA == frozen_selection_effect. That identity is a built-in correctness assertion for any scorer you write.
• Token counts confirm the isolation. Measured on the real processor: R0 and RA tokenize to identical lengths, as do S0, SA and all three negatives. SA − RA therefore cannot be a prompt-length artifact.

Negatives

Three corrupted-history variants, all sparse + adapter-active, let you ask whether the model distinguishes correct history from plausible but wrong history:

negative	corruption
SA_neg_step_shuffled	same images, cyclically shifted so every position is mismatched
SA_neg_irrelevant	images from a different episode in the same split, matched position-by-position on the age vector and preferring the same resolution
SA_neg_duplicate	one history image repeated K times

A cyclic shift is used rather than a reversal because reversing an odd-length sequence leaves the middle image correctly paired.

Composition

Usage

import gzip, json, tarfile
from huggingface_hub import snapshot_download

root = snapshot_download(
    "gavinlaw/causalcache-sparse-history-guiodyssey-v5", repo_type="dataset"
)

with gzip.open(f"{root}/samples.jsonl.gz", "rt", encoding="utf-8") as fh:
    rows = [json.loads(line) for line in fh]

# images live in per-shard tars; each row's paths are relative to the shard root
with tarfile.open(f"{root}/shards/images-part00.tar") as tf:
    tf.extractall("part00/")

group = {}
for r in rows:
    if r["pair_group"] == rows[0]["pair_group"]:
        group[r["arm_slot"]] = r

assert group["SA"]["target_text"] == group["RA"]["target_text"]   # one target per group
assert group["S0"]["adapter_mode"] == "bypass"                    # SA's bypass twin
assert group["SA"]["selected_images"] == group["S0"]["selected_images"]

Row schema

field	meaning
pair_group	"{episode}:{decision_step}" — the unit of comparison
arm_slot	N0/R0/S0/RA/SA/SA_neg_* — the primary key within a group
adapter_mode	"bypass" or "active" — the sole authority on the adapter switch
role	deployment_baseline / reference / measurement / positive / negative
budget	K, the number of history images
selected_steps	1-based, strictly increasing, all < decision_step
selected_images, current_image	paths relative to the shard root
target_text	byte-identical across every arm in the group
split	train / heldout — read this field; do not recompute a hash
negative_kind, negative_scale	present only on negatives

Read adapter_mode; never infer it from a name. Consumers that pattern-match on arm_slot prefixes will silently break the moment an arm is renamed — that exact bug is why the field exists.

Provenance

Everything is pinned. This corpus was rebuilt on a second machine from these public sources alone and produced a byte-identical result — the sorted samples.jsonl sha256 matched exactly, on both machines:

4f6f7065f2053ff0501a2767e98a3656046106c4042b0cc9cf97e9e1c494e063

	repo	revision
trajectories	cua-lite/GUIOdyssey	ea08072b30e523fb4492e4f4597505879ffcd63b
frozen policy used for scoring	mPLUG/GUI-Owl-1.5-8B-Instruct	06d5faecff74840bab2be2425e9c42667a5d04fc

Screenshots and action descriptions originate from GUI-Odyssey (OpenGVLab), released under CC BY 4.0; this derivative keeps the same license and attribution. Please cite GUI-Odyssey alongside this corpus.

Known limitations

Stated plainly, because they change how results must be analysed:

• Groups are not independent. Several decision points come from the same trajectory. Confidence intervals must use episode-cluster bootstrap — treating the 3,252 groups as i.i.d. understated our intervals by roughly half (0.053 vs 0.096 on the same data).
• Short histories admit very few sparse selections. With current_step=9, K=4 there are only 5 legal non-adjacent subsets, so those groups' S0/SA prompts repeat heavily and contribute almost no variance to S0 − R0 or SA − RA. This follows from the non-adjacency constraint, not from a defect — but report those strata separately.
• irrelevant donors are age-matched, not semantics-matched. They come from a different episode at a comparable task stage; they are not adversarially chosen to be confusable.
• The corpus does not settle whether an adapter works. It is built so that a negative result is legible. On our own frozen baseline, the three negative margins are negative — the frozen model scores corrupted history above correct history. Any claim of content sensitivity has to clear that bar first.

Citation

@misc{causalcache_sparse_history_v5,
  title  = {CausalCache Sparse-History Corpus (GUI-Odyssey)},
  author = {Luo, Jiaxuan},
  year   = {2026},
  url    = {https://huggingface.co/datasets/gavinlaw/causalcache-sparse-history-guiodyssey-v5}
}

Please also cite the upstream source:

@article{lu2024guiodyssey,
  title   = {GUIOdyssey: A Comprehensive Dataset for Cross-App GUI Navigation on Mobile Devices},
  author  = {Lu, Quanfeng and Shao, Wenqi and Liu, Zitao and Du, Lingxiao and Meng, Fanqing
             and Li, Boxuan and Chen, Botong and Huang, Siyuan and Zhang, Kaipeng and Luo, Ping},
  journal = {arXiv preprint arXiv:2406.08451},
  year    = {2024}
}