lapse / README.md
vein05's picture
LAPSE 1.0: stimuli (dev 4,791, test 1,125), build manifest, dataset card
beb4aed verified
|
Raw History Blame Contribute Delete
3.47 kB
---
license: cc-by-4.0
language:
- en
pretty_name: "LAPSE: Linguistic Aspect Persistence and Stability Evaluation"
size_categories:
- 1K<n<10K
task_categories:
- text-generation
tags:
- llm-agents
- agent-memory
- long-term-memory
- memory-consolidation
- stale-memory
- temporal-reasoning
- tense-and-aspect
- benchmark
configs:
- config_name: default
data_files:
- split: dev
path: data/dev.jsonl
- split: test
path: data/test.jsonl
---
# LAPSE: aspectual flattening in LLM agent memory
LAPSE is a benchmark for one failure of LLM agent memory. A user says "I am
working at Corvida." The memory writer stores "The user works at Corvida." The
progressive told a later reader that the job might be temporary. The stored
note states it as a standing fact, so a later model has less evidence that the
fact may have gone stale. We call this aspectual flattening.
Each LAPSE item is a dated conversation excerpt with one user statement. Matched
items differ only in the statement's temporal form: progressive ("I'm driving"),
simple present ("I drive"), an explicit end date, or a perfect form.
Paper: [Memory Consolidation Flattens the Temporal Shape of User Facts](https://arxiv.org/abs/2609.36457).
Code, scorer, model outputs and labels: [github.com/aimsresearchlab/lapse](https://github.com/aimsresearchlab/lapse).
## What is in this dataset
| Split | Rows |
|---|---:|
| dev | 4,791 |
| test | 1,125 |
`data/manifest.json` is the build manifest (builder version, evaluation date,
prompt templates, checksums).
The `arm` field says what the model is asked to do:
| `arm` | Task |
|---|---|
| `e1` | Write memory notes from the conversation. |
| `l2` | Write memory notes with an instruction to preserve tense and aspect. |
| `behavioral` | Act on the fact directly. |
| `explicit` | Say whether the fact still holds. |
| `anchor` | Choose to send with the remembered value or to check first. |
Other fields: `form` (temporal form), `frame` (content domain: lodging,
workplace, vehicle, class, household, equipment, affiliate_role, project),
`cluster` (rows sharing `frame` and `cluster` differ only in the manipulated
factor), `gap` and `gap_date` (time between statement and evaluation),
`witness` and `wit_token` (the invented entity the scorer searches for), and
linguistic annotations. The GitHub README describes every field and how to
score a new model.
## Use
```python
from datasets import load_dataset
ds = load_dataset("vein05/lapse")
notes = ds["test"].filter(lambda r: r["arm"] == "e1")
```
Score outputs with the rule-based scorer in the GitHub repository
(`tools/score.py`), which labels each note as keeping, flattening or dropping
the temporal form.
## Notes
- All names, companies, institutions, projects and residences are invented.
- The split is fixed by content (`cluster % 5 == 4` is test). Both splits are
public, so a model trained after release may have seen them.
- The paper's natural-text checks used WildChat, LoCoMo and LongMemEval. Those
texts are not included.
## Citation
```bibtex
@misc{panthi2026lapse,
title = {Memory Consolidation Flattens the Temporal Shape of User Facts},
author = {Panthi, Sugam and Yeamin, Muhaiminul and Luo, Siyan and Abdelfattah, Rabab},
year = {2026},
eprint = {2609.36457},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
doi = {10.48550/arXiv.2609.36457},
url = {https://arxiv.org/abs/2609.36457}
}
```