Papers › ST-MoE: Designing Stable and Transferable Sparse Expert Models

ST-MoE: Designing Stable and Transferable Sparse Expert Models

17 Feb 2022arXiv:2202.08906archive 2025-07-28

Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, William Fedus

Scale has opened new frontiers in natural language processing -- but at a high cost. In response, Mixture-of-Experts (MoE) and Switch Transformers have been proposed as an energy efficient path to even larger and more capable language models. But advancing the state-of-the-art across a broad set of natural language tasks has been hindered by training instabilities and uncertain quality during fine-tuning. Our work focuses on these issues and acts as a design guide. We conclude by scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder Transformer (Stable and Transferable Mixture-of-Experts or ST-MoE-32B). For the first time, a sparse model achieves state-of-the-art performance in transfer learning, across a diverse set of tasks including reasoning (SuperGLUE, ARC Easy, ARC Challenge), summarization (XSum, CNN-DM), closed book question answering (WebQA, Natural Questions), and adversarially constructed tasks (Winogrande, ANLI R3).

PaperPDFCodeCode Syntology ran

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

For agents, Syntology's MCP tool lists every function and class Syntology harvested from this paper and whether it ran (how to connect): get_harvested_code_for_paper(arxiv_id="2202.08906")

Code

Syntology Ran 5 of 5 code samples harvested from 1 repository linked to this paper; 0 have no recorded run. Of those that ran: 3 ran · honoured contract; 2 ran · our draft was wrong.

By repository: community (archive-listed): 4 samples from 1 repository, 4 ran; 1 identical to code first harvested elsewhere. The run record, sample by sample. “Ran” means executed on a synthesized input, not that the code is correct or reproduces the paper.

tensorflow/mesh officialmentioned in papertf report
xuefuzhao/openmoe mentioned on GitHubpytorch report
yikangshen/megablocks mentioned on GitHubpytorch report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

5 samples harvested; 5 ran; 3 honoured the contract we drafted; 0 have no recorded run. Read from Syntology's graph 2026-09-24; that is when this build read the record, not when the samples ran.

3ran · honoured contract
2ran · our draft was wrong

Licence: 5 of the 5 samples are pointer only, meaning Syntology does not serve that copy's text. This page shows no code text for any sample; each one links to its file in the repository.

Harvested from xuefuzhao/openmoe. Some samples are identical code Syntology first harvested from another repository; for those, this paper's copy is not located and its licence is not recorded. “Ran” means the sample executed on a synthesized input. It does not mean the output is correct, and nothing here reproduces the paper's results. “Honoured” and “violated” refer to a contract Syntology drafted from the code itself; “our draft was wrong” and “fixture could not drive it” are failures of Syntology's instrument, not of the code.

Each sample ends with its code_sha256, Syntology's identity for that exact code. An agent fetches the stored sample with Syntology's MCP tool get_code(code_sha256="…") (how to connect); click an identity to copy that call.

Repository labels, per sample. official repository: The archive marks this repository official for the paper. named in the paper: The archive records that the paper mentions this repository; it is not marked official. community (archive-listed): In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper. found in paper text by Syntology: Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted. community: Not in the archive's code links for this paper; a community repository Syntology harvested. Samples from a repository marked official are listed first. Licence labels name the repository's licence as recorded at harvest. “Pointer only” means Syntology does not serve that copy's text, for one of four reasons: no licence file was found; the licence was not identified; the licence is recorded as permissive but that copy's record is not marked cleared; or the licence is outside the permissive list Syntology serves text under (MIT, Apache-2.0, BSD and similar). Some licences outside that list permit redistribution, such as WTFPL, and GPL-3.0 under its conditions; they are simply not on the list. Hover a licence label for the reason. File links open the file on GitHub at the default branch, which may have changed since the harvest.

apply_llama_chat_template xuefuzhao/openmoe/script/inference_on_multi_devices.py community (archive-listed) ran · our draft was wrong no licence file found · pointer only · 90475ec2cfa11bd6 · report
calculate_num_tokens xuefuzhao/openmoe/eval/plot_bigbench.py community (archive-listed) ran · honoured contract fingerprinted no licence file found · pointer only · 8eb924623b6d2a0f · report
convert_size xuefuzhao/openmoe/eval/plot_bigbench.py community (archive-listed) ran · honoured contract no licence file found · pointer only · 3619322ac2eac1fa · report
read_csv_and_convert_to_dataframe xuefuzhao/openmoe/eval/plot_bigbench.py community (archive-listed) ran · our draft was wrong fingerprinted no licence file found · pointer only · 6152639868e9cdb6 · report
promote_scalar identical code first harvested elsewhere ran · honoured contract fingerprinted licence of this copy not recorded · 7630a3eb1f9be798 · report

Tasks

ARCCommon Sense ReasoningCoreference ResolutionDecoderMixture-of-ExpertsNatural Language InferenceNatural QuestionsQuestion AnsweringTransfer LearningWinograndeWord Sense Disambiguation

1 archive task tag without a task page not shown.

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Common Sense Reasoning ARC (Challenge) ST-MoE-32B 269B (fine-tuned) Accuracy 86.5 #12 of 54 Archive leaderboard report
Common Sense Reasoning ARC (Challenge) ST-MoE-L 4.1B (fine-tuned) Accuracy 56.9 #28 of 54 Archive leaderboard report
Common Sense Reasoning ARC (Easy) ST-MoE-32B 269B (fine-tuned) Accuracy 95.2 #1 of 47 Archive leaderboard report
Common Sense Reasoning ARC (Easy) ST-MoE-L 4.1B (fine-tuned) Accuracy 75.4 #23 of 47 Archive leaderboard report
Common Sense Reasoning ReCoRD ST-MoE-32B 269B (fine-tuned) EM 95.1 #2 of 45 Archive leaderboard report
Common Sense Reasoning ReCoRD ST-MoE-L 4.1B (fine-tuned) EM 88.9 #12 of 45 Archive leaderboard report
Common Sense Reasoning WinoGrande ST-MoE-32B 269B (fine-tuned) Accuracy 96.1 #1 of 77 Archive leaderboard report
Common Sense Reasoning WinoGrande ST-MoE-L 4.1B (fine-tuned) Accuracy 81.7 #13 of 77 Archive leaderboard report
Coreference Resolution Winograd Schema Challenge ST-MoE-32B 269B (fine-tuned) Accuracy 96.6 #5 of 82 Archive leaderboard report
Coreference Resolution Winograd Schema Challenge ST-MoE-L 4.1B (fine-tuned) Accuracy 93.3 #8 of 82 Archive leaderboard report
Natural Language Inference CommitmentBank ST-MoE-L 4.1B (fine-tuned) Accuracy 98.2 #3 of 20 Archive leaderboard report
Natural Language Inference CommitmentBank ST-MoE-32B 269B (fine-tuned) Accuracy 98 #4 of 20 Archive leaderboard report
Natural Language Inference RTE ST-MoE-32B 269B (fine-tuned) Accuracy 93.5% #4 of 90 Archive leaderboard report
Natural Language Inference RTE ST-MoE-L 4.1B (fine-tuned) Accuracy 92.1% #10 of 90 Archive leaderboard report
Question Answering BoolQ ST-MoE-32B 269B (fine-tuned) Accuracy 92.4 #2 of 65 Archive leaderboard report
Question Answering BoolQ ST-MoE-L 4.1B (fine-tuned) Accuracy 88.6 #11 of 65 Archive leaderboard report
Question Answering COPA ST-MoE-32B 269B (fine-tuned) Accuracy 99.2 #3 of 60 Archive leaderboard report
Question Answering COPA ST-MoE-L 4.1B (fine-tuned) Accuracy 91 #14 of 60 Archive leaderboard report
Question Answering MultiRC ST-MoE-32B 269B (fine-tuned) F1 89.6 #2 of 30 Archive leaderboard report
Question Answering MultiRC ST-MoE-L 4.1B (fine-tuned) F1 86 #8 of 30 Archive leaderboard report
Word Sense Disambiguation Words in Context ST-MoE-32B 269B (fine-tuned) Accuracy 77.7 #3 of 37 Archive leaderboard report
Word Sense Disambiguation Words in Context ST-MoE-L 4.1B (fine-tuned) Accuracy 74 #10 of 37 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdamAttentionBPEDense ConnectionsDropoutLabel SmoothingLayer NormalizationLinear LayerMoEMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSoftmaxTransformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections