license: other
license_name: per-partition-source-licenses
license_link: https://github.com/aimsresearchlab/seam/blob/main/DATA_LICENSES.md
language:
- en
pretty_name: SEAM
size_categories:
- 1K<n<10K
task_categories:
- text-generation
tags:
- llm-evaluation
- benchmark
- prompt-boundaries
- instruction-data-separation
- editing
configs:
- config_name: permissive
default: true
data_files: benchmark/seam_v4_permissive.jsonl
- config_name: cc-by-sa
data_files: benchmark/restricted/seam_v4_cc-by-sa.jsonl
- config_name: cc-by-nc-sa
data_files: benchmark/restricted/seam_v4_cc-by-nc-sa.jsonl
SEAM: absorption at unmarked prompt seams
A user pastes text into a chat, asks for an edit, and keeps typing a comment below the paste. The model receives one flat message. SEAM measures absorption: the model returns that trailing comment inside the edited text, even though the comment is benign and asks for no insertion.
This is the benchmark from Can LLMs Separate Pasted Artifacts from User Speech? Absorption at Unmarked Prompt Seams (Sugam Panthi, Muhaiminul Yeamin, and Rabab Abdelfattah, arXiv:2610.04210).
- Paper: https://arxiv.org/abs/2610.04210
- Code, scorer, and per-model case-level labels: https://github.com/aimsresearchlab/seam
- Leaderboard: https://aimsresearchlab.com/seam/leaderboard.html
- Companion post: https://spanthi.com/blog/where-does-the-paste-end/
Across 20 models, absorption at a bare newline ranges from 7.7% to 66.7%. Adding a blank line does not significantly reduce absorption in any model, while boundary markers reduce it in 19 of 20 models. Comments that fit the pasted text, such as a code comment typed after code, are absorbed significantly more often in 17 of 20 models.
Structure
300 composition clusters, each under six matched conditions, 1,800 rows in total. A cluster is one task instruction and one pasted artifact; rows that share composition_event_id are matched.
| condition | what follows the pasted artifact |
|---|---|
clean |
nothing; the model's own output for the cluster, used as the comparison |
newline |
the comment after a bare newline |
blank |
the comment after a blank line |
boundary |
the comment after explicit boundary markers |
mitigation |
boundary markers plus one instruction line |
newline_R |
a comment rewritten to fit the genre of the pasted text (called artifact-native in the paper) |
Conditions are scored within a cluster against that cluster's clean output, so a signal counts only when it is absent from the model's unprompted output. The scorer is in the GitHub repository.
Configurations and licenses
The examples reuse text from six public editing corpora, so the data is split by source license. Each row keeps a license field and the identifiers needed to trace it upstream.
| config | rows | sources | license |
|---|---|---|---|
permissive (default) |
1,200 | CanItEdit (MIT), CoEdIT (Apache-2.0), CommitPackFT (MIT dataset; item repository license retained), IteraTeR Human-Doc (Apache-2.0) | per source, permissive |
cc-by-sa |
300 | Stack Exchange | CC BY-SA (version from the item revision date) |
cc-by-nc-sa |
300 | ParaRev | CC BY-NC-SA 4.0, noncommercial |
from datasets import load_dataset
seam = load_dataset("vein05/seam", "permissive", split="train")
Public availability of a source text does not by itself grant redistribution rights, and item-level retrieval dates and URLs are not recorded consistently across all sources. Check the upstream terms before redistributing a partition. DATA_LICENSES.md has the details. The occurrence analysis in the paper uses WildChat and LMSYS-Chat-1M; that conversation text is not included here.
Main fields
| field | meaning |
|---|---|
composition_event_id |
cluster identifier |
condition |
one of the six conditions above |
task_instruction |
typed instruction before the simulated paste |
artifact |
source material designated as pasted |
afterthought |
typed text after the paste; null in clean |
forbidden_proposition |
meaning that must not enter the returned artifact |
witness_tokens |
distinctive afterthought substrings, verified absent from the artifact, reference, and instruction; null in clean |
message |
the flattened user message the model receives |
segments |
typed and pasted origins with exact character offsets |
reference |
upstream reference edit, where the source provides one |
source, license, url |
upstream corpus and licensing metadata |
Scope
The examples are constructed, English, first-turn editing requests with equal source weighting and repeated comment pools. SEAM measures unintended inclusion of benign user speech in a returned artifact. It contains no adversarial instructions or executable payloads, and it does not establish behavior for sensitive content, other languages, or later conversational turns.
Citation
@misc{panthi2026seam,
title = {Can {LLMs} Separate Pasted Artifacts from User Speech? Absorption at Unmarked Prompt Seams},
author = {Panthi, Sugam and Yeamin, Muhaiminul and Abdelfattah, Rabab},
year = {2026},
eprint = {2610.04210},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
doi = {10.48550/arXiv.2610.04210},
url = {https://arxiv.org/abs/2610.04210}
}