WMIP October 2026 item banks

Frozen synthetic prompts, reference answers and metadata for evaluating cumulative state tracking in language models.

Paper · Project website · GitHub · Evaluate your model · Scoring and reproduction

Load the items

Install datasets, then load the default 840-item bank:

import json
from datasets import load_dataset

rows = load_dataset("dengzhe-hou/WMIP", split="test", revision="v2026.10.1")

Restore prompts, typed reference answers and optional metadata to the original item schema:

def source_item(row):
    item = json.loads(row["metadata_json"])
    for field in ("item_id", "block", "prompt", "answer_type", "max_tokens"):
        item[field] = row[field]
    item["gold"] = json.loads(row["gold_json"])
    return item

items = [source_item(row) for row in rows]

gold_json encodes integer, word or list answers; metadata_json retains the remaining source fields, including any system message and absent-versus-null distinctions. The nullable system column is also available for browsing; source_file is relative to the source release. The GitHub helper provides the same conversion as wmip.load_items(rows).

Evaluate and score

The quickest route uses a GitHub clone and any OpenAI-compatible endpoint. One command sends all three frozen banks to your model, scores them with the published rules and prints your model among the 18 published models:

git clone --branch v2026.10.4 https://github.com/dengzhe-hou/WMIP.git
cd WMIP/releases/2026-10
python wmip.py run --base-url http://localhost:8000/v1 --model my-model --out-dir my-run

It needs only the Python standard library and resumes when rerun. The published results use thinking-disabled inference; the GitHub README shows how to turn thinking off for common servers.

To run the items yourself instead, send only each item's prompt and optional system message to your model, using its max_tokens budget. Save the full output as one JSON object per line in answers.jsonl:

{"item_id":"A-k3-s2026-p0","response":"12","finish_reason":"stop"}

Only item_id and response are required. Include finish_reason when available to report truncation; a failed request may carry an error. Keep unanswered items missing and report the scored and missing counts with accuracy. For comparison with the released evaluation, preserve its prompts, per-item budgets and thinking-disabled inference mode, and document any changes.

From a GitHub clone, enter releases/2026-10/ and score your actual answers against the matching frozen bank:

python wmip.py score --bank data/item_bank/item_bank.jsonl \
    --responses answers.jsonl --label my-model --out-dir my-scores

The helper writes runner-compatible records, item-level scores and summary CSVs using the release scorer. E/F retain both documented parsers, and I uses its last-integer rule. For another configuration, export its decoded items as JSONL and use that file as --bank:

with open("wmip-items.jsonl", "w", encoding="utf-8") as handle:
    for item in items:
        handle.write(json.dumps(item, ensure_ascii=False) + "\n")

For K/L configurations the same helper selects analysis/score_ext.py, preserving its entity-label and ordered-list rules. It writes K/L-specific scored_items.csv rather than the main scorer's summary CSVs, and rejects synthetic records. For A512, the helper reports the one-integer rule; the release's analysis/score_budget.py retains the paired one-integer/last-integer analyses. Score each configuration separately.

The callable example connects your own Python inference function without an API server, and the lm-evaluation-harness and Inspect tasks run the same banks with the release's scorers.

The local model walkthrough loads a pinned Qwen3 model directly on CPU, saves responses and scores the full bank. Its three-item check and resume instructions are available from software release v2026.10.2 onward; these frozen dataset banks remain at v2026.10.1.

Dataset configurations

Pass a configuration name as the second argument to load_dataset, for example load_dataset("dengzhe-hou/WMIP", "intermediate_multitarget_2026_10", split="test", revision="v2026.10.1"). All use the test split, an evaluation label rather than a new sampling split or a guarantee of unseen items.

Configuration	Rows	Contents
main_2026_10 (default)	840	Main bank, blocks A-I
intermediate_multitarget_2026_10	180	Intermediate-state and multi-target items, K/L
large_budget_2026_10	60	Core prompts with a larger token budget, A512
intermediate_multitarget_hosted70b_2026_10	180	K/L with the hosted Llama 3.3 70B system message
large_budget_hosted70b_2026_10	60	A512 with that hosted system message
intermediate_multitarget_superseded_plural_2026_10	180	Earlier plural wording; provenance only

The 1,500 rows include overlapping protocol variants. The superseded bank differs in 40 single-target L prompts and is excluded from the final analysis. Match hosted variants to their documented runs. Earlier WMF-AM experiments in GitHub's root data/ belong to a separate historical version and must not be pooled with this October evaluation.

Version and citation

The fixed dataset tag v2026.10.1 pairs these banks with WMIP software release 2026.10.1. Pass this tag as revision when loading any configuration to reproduce that version. The six configurations preserve their frozen contents.

Use the release's CITATION.cff or GitHub's Cite this repository button to cite the versioned evaluation materials. Cite the paper, arXiv 2603.27343, with:

@misc{hou2026wmip,
  title         = {{WMIP}: A Working-Memory-Inspired Programmable Probe of Cumulative State Tracking in Large Language Models},
  author        = {Hou, Dengzhe and Jiang, Lingyu and Li, Deng and Li, Zirui and Lin, Fangzhou and Yamada, Kazunori D},
  year          = {2026},
  eprint        = {2603.27343},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2603.27343}
}

These item banks and this card use CC BY 4.0; see LICENSE-DATA and the original NOTICE. Model responses, analysis outputs and BFCL files remain in the GitHub release under their original terms.