Datasets:
Safe to Omit, Worth Omitting? Keeping the Answer and Cutting Latency When an LLM Agent Skips Another's Message
arXiv | Project Website | Paper | Code | Paper map
Xinyang Li, Chih-Hsuan Yang, Xiaolong Ma, Anjir Ahmed Chowdhury, Vikram Vasudevan, Rajeev Thakur, Feng Yan (University of Houston, Argonne National Laboratory, Oregon State University)
These are the per-question records behind the paper. A receiver agent can answer alone or wait for a helper agent's message, sent either as text (Text) or as a projected KV cache (C2C). The paper certifies, from the receiver's confidence without the message, a threshold below which the message can be skipped with at most 5% of skipped answers changing, and measures the latency this saves. The release contains:
- the receiver's confidence score (ProbeMax) for every question, and the route each deployed policy took;
- the raw outputs of every action (receiver alone, Text, C2C) and their parsed labels;
- per-request timing records of the end-to-end replays;
- every calibration test, including rejected grid points and the settings that fell back;
- the frozen configurations and the exact question IDs of every split.
Model weights and full-vocabulary logits are not included. All models and benchmarks are public
(see configs/model_identities.json and configs/dataset_identities.json for revisions).
Quick start
The code repository downloads this dataset, checks
every file against MANIFEST.json, and re-derives four published results from the records:
git clone https://github.com/Piedpiperrr/safe-to-omit.git && cd safe-to-omit
make setup data verify
To read the records directly:
from huggingface_hub import snapshot_download
import pandas as pd
root = snapshot_download("AgentsSci/ICLR2027-SafeToOmit", repo_type="dataset")
# Routes of the deployed large/ARC/Text policy on the 299 ARC development questions:
# route "R" means the receiver answered alone (the message was skipped).
routes = pd.read_json(f"{root}/records/boundaries_large_small/records/large_arc_T_routes.jsonl", lines=True)
print((routes["route"] == "R").sum(), "of", len(routes), "skipped") # 284 of 299 (large/ARC/Text in the paper's deployed-policy table)
Layout
| path | contents |
|---|---|
records/<experiment>/ |
per-question records: confidence scores (*_probes.jsonl.gz: ProbeMax, p_labels), routes (*_routes.jsonl), raw action outputs and parsed labels, timing (latency_ms, parts_ms) |
results/<experiment>/ |
each experiment's protocol (PREREG*.md with its SHA-256), SUMMARY.md, DEVIATIONS.md, result tables and calibration ledgers |
configs/ |
frozen configurations: probe protocols (label tokens, chat templates, prefix token IDs), deployed thresholds and policies, answer-parser freeze, model and dataset identities |
splits/ |
question IDs of the fit, calibration and development splits, the sealed ARC and QASC tests, and the timing panels |
results/three_agent_qasc/ (and the same name under records/, configs/, splits/) |
three-agent system on QASC: protocol, freeze record, 13 deviations, calibration ledger and label log (40 labels), sealed-test, development and timing results, recheck CSVs, per-question records of every agent and path; DERIVED.md lists what was computed or assembled when the release was built |
results/scale_32b/ (and the same name under records/, configs/, splits/) |
32B helper (Qwen2.5-32B-Instruct) and receiver (Qwen3-32B) on OpenBookQA: protocol, freeze record, deviations, calibration ledger and label log (21 labels), development, held-out and timing results, recheck CSVs, per-question records, rendered prompt examples; DERIVED.md as above |
results/batched_serving/ (and the same name under records/, configs/, splits/) |
batched serving at B = 1, 8, 32 with synchronous, early-return and reference arms on the full OBQA and ARC development sets: protocol, freeze record, deviations, all 36 arm cells and comparisons, per-request and per-batch records, batch composition and replay order; DERIVED.md as above |
results/cpu_reanalysis/ (and the same name under records/, configs/, splits/) |
CPU re-analyses of version-1 records: gain retention on unseen questions, per-discipline MMLU-Pro certification, the medium/OBQA/Text lost-correction test with its ledger and 36-label log, single-model baselines; DERIVED.md as above |
REDACTIONS.json |
the files edited for this release, with the SHA-256 before and after |
MANIFEST.json, CHECKSUMS.txt |
SHA-256 and size of every file (shasum -a 256 -c CHECKSUMS.txt) |
Experiment folders keep the IDs used during the research (E3 … E22, X1 … X4); the four folders added in version 2 have
descriptive names, and each one's SUMMARY.md names its research ID.
docs/PAPER_MAP.md
maps every table and figure in the paper to its folder, and
docs/RECORDS.md lists
the records the paper says are released.
Conventions: a setting is a model pair (small, medium, large, or a cross-family pair), a
benchmark and a reference path (T = Text, C = C2C). R is the receiver alone. The
uncertainty is u = 1 - max_l p_l over the option labels.
In three_agent_qasc and scale_32b, u is instead the log of the non-top label mass, log of the sum
of p_l over l != argmax (see each folder's SUMMARY.md); a message is still skipped iff u <= tau.
Unreadable outputs parse to a shared
INVALID label. Some older files say "omit" for "skip".
What was changed for release
Process files (progress logs, job-scheduler scripts, draft figures and internal file indexes) are
left out. Eighteen files were edited: Chinese table values and quotations were translated,
internal file names and workflow notes were removed, and scheduler settings and references to
folders that are not released were removed from three configuration files. REDACTIONS.json lists
each file with its SHA-256 before and after. No result was changed. Absolute paths in the stored
files appear as <private_path> or $DATA_DIR. Version 2 applies the same rules to the four new folders (also replacing
scheduler job ids, host and node names, and cluster names by ClusterA and ClusterB, and translating Chinese text written by
the authors; text generated by the models is kept as generated) and adds 121 entries to REDACTIONS.json; no version-1 file
was changed except this card and REDACTIONS.json. Each new folder's DERIVED.md lists the files and columns that were computed
or assembled when the release was built, with the check each passed. Hash lists inside experiment folders (*.sha256,
MANIFEST*.sha256) were written during the research, so files changed since then no longer match
them; the root MANIFEST.json is the reference for this release.
Benchmarks and licences
The records contain question text, options and passages from the benchmarks below, so that every prompt can be reconstructed. That text remains under its source's licence; the share-alike licences of ARC, SQuAD and Natural Questions are why this dataset is released under CC BY-SA 4.0.
| benchmark | source (revision) | used for | licence |
|---|---|---|---|
| OpenBookQA | allenai/openbookqa (388097ea) |
main study | Apache-2.0 (authors' repository; the Hub card says "unknown") |
| ARC-Challenge | allenai/ai2_arc (210d026f) |
main study, sealed test | CC BY-SA 4.0 |
| MMLU-Pro | TIGER-Lab/MMLU-Pro (b189ec76) |
breadth extension | MIT |
| SQuAD v1.1 | rajpurkar/squad (7b6d24c4) |
extractive extension | CC BY-SA 4.0 |
| GSM8K | openai/gsm8k (740312ad) |
math extension | MIT |
| Natural Questions (MRQA version) | mrqa-workshop/mrqa (f3178d98) |
candidate in the SQuAD selection pilot | not stated on the Hub card; Natural Questions is CC BY-SA 3.0 |
| TriviaQA | mandarjoshi/trivia_qa (0f7faf33) |
candidate in the SQuAD selection pilot | not stated on the Hub card ("The University of Washington does not own the copyright of the questions and documents included in TriviaQA.") |
| NQ-Open | google-research-datasets/nq_open (5dd9790a) |
candidate in the SQuAD selection pilot | CC BY-SA 3.0 |
| QASC | allenai/qasc (a34ba204) |
three-agent system | CC BY 4.0 (the official page links to the Hub card, which states CC BY 4.0; the authors' repository says "We release the data under CC BY license") |
Models: Qwen2.5-0.5B/1.5B/7B/32B-Instruct and Qwen3-0.6B/1.7B/8B/32B, OLMo-2-1124-7B-Instruct,
Llama-3.2-1B-Instruct and Llama-3.1-8B-Instruct, and the released C2C fusers
(nics-efc/C2C_Fuser). Model outputs in the records are subject to the terms of the model that
produced them.
Changelog
- Version 2 (November 2026): adds the three-agent system on QASC, the 32B helper and receiver pair, batched serving, and
the CPU re-analyses (gain retention on unseen questions, per-discipline MMLU-Pro certification, single-model baselines and
the medium/OBQA/Text lost-correction test). Version 2 holds 3,527 files, 249.6 MB (version 1: 3,285 files, 233.9 MB), not counting
MANIFEST.jsonandCHECKSUMS.txt. - Version 1 (November 2026): first public release.
Citation
@misc{li2026safetoomit,
title = {Safe to Omit, Worth Omitting? Keeping the Answer and Cutting Latency When an {LLM} Agent Skips Another's Message},
author = {Li, Xinyang and Yang, Chih-Hsuan and Ma, Xiaolong and Chowdhury, Anjir Ahmed and Vasudevan, Vikram and Thakur, Rajeev and Yan, Feng},
year = {2026},
eprint = {2611.23570},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
doi = {10.48550/arXiv.2611.23570},
url = {https://arxiv.org/abs/2611.23570}
}
Contact
Chih-Hsuan (Bella) Yang, bellayang@anl.gov
- Downloads last month
- 87