FlowingFlowing
← Back to all development cases
This case was judged by an anonymized AI reviewer and has not been independently reviewed by people; inspect the draft, cards, and both outputs to assess it.
AutocompleteWith cards better131 / 280 · beu-76fc92266aed93ca

gpt-oss-120b & gpt-oss-20b Model Card

Computer science · 2508.10925v1

FLOWING EVIDENCE BENCHMARK

How do we compare writing with and without cards?

At the same writing position, the same model produces one first result with cards and one without. All 280 pairs enter the overall anonymized review, which assesses whether a decision not to continue fits the draft.

Overall: 138 with-card wins, 26 ties, 107 no-card wins, 5 pairs where neither was usable, and 4 inconclusive judgments. Both sides wrote text in 218 pairs; 48 had a deliberate stop and 14 included an invalid output.

SAME DRAFT · WITH OR WITHOUT CARDS

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED VERSIONS

SHARED BY BOTH

Draft, writing task, and generation model

A · With cards

Relevant literature cards supplied

B · Without cards

No literature cards supplied

Generation model

Same model and version

A → first output

B → first output

Anonymized AI judge

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Review again with the answer order swapped

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The AI judge sees the same draft and both first outputs without knowing which received cards.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are positions and cards selected?

We selected 40 writing positions in each of seven fields, for 280 comparisons. The position and relevant literature cards were set before each comparison, so both versions began from the same draft.

The main result retains every frozen position, including places where cards might not add a useful next move. This avoids selecting cases based on their outputs, but 49% describes the whole cohort, not a win rate where evidence clearly had an opportunity to help. This cohort did not record that opportunity label for each position before generation, so those two groups cannot yet be separated reliably.

CARD USE · 140/280

Reported card use in continuations with text

The model reports whether a card affected its continuation. In 140 of 280 tests it produced continuation text and reported card use. The previous count of 150 included 10 cases without continuation text and has been corrected.

COMPARISON · 70 / 7 / 30

Which written version was better?

Of the 107 reported-use cases where both versions produced text, the blinded AI judge preferred the with-card version 70 times, found 7 about the same, and preferred the without-card version 30 times.

How are the two table percentages calculated?

ADOPTION COVERAGE

140reported use with text
÷
280all positions
=
50%

WITH-CARD VERSION PREFERRED

70with cards better
÷
107comparable tests
=
65%

The percentages use different groups: 50% counts continuations with text reporting card use among all 280 tests; 65% covers the 107 such cases with text on both sides. The generating model reports card use, while an anonymized AI judge compares writing quality.

Card use is self-reported by the generator. The anonymized AI judge sees the draft and source snippets, so source-aware preference remains possible; there was no independent human review. The overall result describes these 280 frozen positions, not a win probability for any writing task.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

gpt-oss-120b & gpt-oss-20b Model Card

Introduction / methods

E without tools, see Table 3 for the gpt-oss models on AIME without tools 2.3 Tokenizer Across all training stages, we utilize ouro200k_harmony tokenizer, which we open source in our TikToken library. This is a Byte Pair Encoding (BPE) which extends theo200k tokenizer used for other OpenAI models such as GPT-4o and OpenAI o4-mini with tokens explicitly used for our harmony chat format described in Table 18 and has a total of201,088 tokens. 2.4 Pretraining Data: We train the models on a text-only dataset with trillions of tokens, with a focus on STEM, coding, and general knowledge. To improve the safety of the model, we filtered the data for harmful content in pre-training, especially around hazardous biosecurity knowledge, by reusing the CBRN pre-training filters from GPT-4o [18]. Our model has a knowledge cutoff of June 2024. Training: The gpt-oss models trained on NVIDIA H100 GPUs using the PyTorch framework [19] with expert-optimized Triton [20] kernels2. The training run for gpt-oss-120b required 2.1 million H100-hours to complete, with gpt-oss-20b needing almost 10x fewer. Both models leverage the Flash Attention [21] algorithms to reduce the memory requirements and accelerate

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionOverall AI review: With cards better

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Read overall AI review rationale

A = without cards · B = with cards

The draft ends mid-sentence after 'accelerate' in the Training subsection, clearly needing completion. Output B simply adds 'attention computation,' completing the sentence about Flash Attention algorithms reducing memory and accelerating attention computation. This is accurate, fits well, and requires no repair. Output A adds a much longer continuation about 'training, and both use a Mixture-of-Experts architecture with 128 experts per layer.' This is problematic: (1) It grammatically doesn't connect well—the sentence was about Flash Attention, not a general model architecture description; (2) The claim about '128 experts per layer' is a concrete unsupported specific about this manuscript's architecture. The sources mention MoE architectures generally but none specify '128 experts per layer' for gpt-oss models. This is an invented detail that could be false and would require verification; (3) The transition 'training, and both use' is awkward and doesn't flow from discussing Flash Attention. Output B's simpler completion is more accurate, better scoped, and doesn't introduce unsupported claims. However, Output A's unsupported claim about 128 experts is material enough to flag. Despite this, Output A's text is still partially usable as a continuation (the MoE architecture point could be relevant to the section), but B is cleaner and safer. Actually, re-evaluating: Output A's '128 experts per layer' is a very specific claim with no basis in the provided sources or draft, making it a material unsupported assertion. The text also doesn't grammatically fit well. Output B cleanly completes the Flash Attention sentence. I prefer B for accuracy and fit, but both are technically insertable. Wait—the instruction says 'A preferred output must be usable.' Both are usable, but A has an unsupported claim. Let me reconsider: B is more accurate and fits better. A introduces an unsupported specific that could mislead. I'll prefer B.

Evidence

Same model and position · retrieved passages supplied · first output

Preferred by blind review

attention computation.

Passages supplied to the Evidence version

Card 1 · FactorLLM: Factorizing Knowledge via Mixture of Experts for Large Language Models

Recent research has demonstrated that Feed-Forward Networks (FFNs) in Large Language Models (LLMs) play a pivotal role in storing diverse linguistic and factual knowledge. Conventional methods frequently face challenges due to knowledge confusion stemming from their monolithic and redundant architec…

Read full passage excerpt

Recent research has demonstrated that Feed-Forward Networks (FFNs) in Large Language Models (LLMs) play a pivotal role in storing diverse linguistic and factual knowledge. Conventional methods frequently face challenges due to knowledge confusion stemming from their monolithic and redundant architectures, which calls for more efficient solutions with minimal computational overhead, particularly for LLMs. In this paper, we explore the FFN computation paradigm in LLMs and introduce FactorLLM, a novel approach that decomposes well-trained dense FFNs into sparse sub-networks without requiring any further modifications, while maintaining the same level of performance. Furthermore, we embed a router from the Mixture-of-Experts (MoE), combined with our devised Prior-Approximate (PA) loss term that facilitates the dynamic activation of experts and knowledge adaptation, thereby accelerating computational processes and enhancing performance using minimal training data and fine-tuning steps. Fact …

Card 2 · LIBMoE: A Library for comprehensive benchmarking Mixture of Experts in Large Language Models

Mixture of experts (MoE) architectures have become a cornerstone for scaling up and are a key component in most large language models such as GPT-OSS, DeepSeek-V3, Llama-4, and Gemini-2.5. However, systematic research on MoE remains severely constrained by the prohibitive computational costs of trai…

Read full passage excerpt

Mixture of experts (MoE) architectures have become a cornerstone for scaling up and are a key component in most large language models such as GPT-OSS, DeepSeek-V3, Llama-4, and Gemini-2.5. However, systematic research on MoE remains severely constrained by the prohibitive computational costs of training and evaluation, restricting large-scale studies accessible to most researchers. We introduce LibMoE, a unified framework for reproducible, efficient, and extensible MoE research that supports both pretraining and sparse-upcycling regimes. Beyond unified implementations, the framework provides transparent analytical tools for probing routing and expert dynamics. Leveraging this foundation, we conduct a comprehensive analysis along three dimensions: (i) routing dynamics, covering expert selection patterns, routing stability and optimality, and how routing entropy reveals task specialization and expert diversity; (ii) the effect of lightweight initialization on load balancing, demonstratin …

Card 3 · Skywork-MoE: A Deep Dive into Training Techniques for Mixture-of-Experts Language Models

In this technical report, we introduce the training methodologies implemented in the development of Skywork-MoE, a high-performance mixture-of-experts (MoE) large language model (LLM) with 146 billion parameters and 16 experts. It is initialized from the pre-existing dense checkpoints of our Skywork…

Read full passage excerpt

In this technical report, we introduce the training methodologies implemented in the development of Skywork-MoE, a high-performance mixture-of-experts (MoE) large language model (LLM) with 146 billion parameters and 16 experts. It is initialized from the pre-existing dense checkpoints of our Skywork-13B model. We explore the comparative effectiveness of upcycling versus training from scratch initializations. Our findings suggest that the choice between these two approaches should consider both the performance of the existing dense checkpoints and the MoE training budget. We highlight two innovative techniques: gating logit normalization, which improves expert diversification, and adaptive auxiliary loss coefficients, allowing for layer-specific adjustment of auxiliary loss coefficients. Our experimental results validate the effectiveness of these methods. Leveraging these techniques and insights, we trained our upcycled Skywork-MoE on a condensed subset of our SkyPile corpus. The evalu …

Card 4 · OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models

To help the open-source community have a better understanding of Mixture-of-Experts (MoE) based large language models (LLMs), we train and release OpenMoE, a series of fully open-sourced and reproducible decoder-only MoE LLMs, ranging from 650M to 34B parameters and trained on up to over 1T tokens.…

Read full passage excerpt

To help the open-source community have a better understanding of Mixture-of-Experts (MoE) based large language models (LLMs), we train and release OpenMoE, a series of fully open-sourced and reproducible decoder-only MoE LLMs, ranging from 650M to 34B parameters and trained on up to over 1T tokens. Our investigation confirms that MoE-based LLMs can offer a more favorable cost-effectiveness trade-off than dense LLMs, highlighting the potential effectiveness for future LLM development. One more important contribution of this study is an in-depth analysis of the routing mechanisms within our OpenMoE models, leading to three significant findings: Context-Independent Specialization, Early Routing Learning, and Drop-towards-the-End. We discovered that routing decisions in MoE models are predominantly based on token IDs, with minimal context relevance. The token-to-expert assignments are determined early in the pre-training phase and remain largely unchanged. This imperfect routing can result …

Card 5 · OLMoE: Open Mixture-of-Experts Language Models

We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all…

Read full passage excerpt

We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all available models with similar active parameters, even surpassing larger ones like Llama2-13B-Chat and DeepSeekMoE-16B. We present various experiments on MoE training, analyze routing in our model showing high specialization, and open-source all aspects of our work: model weights, training data, code, and logs.

Model-reported card use

The model did not report that any card materially affected this completion.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

training, and both use a Mixture-of-Experts architecture with 128 experts per layer.

Earlier card-use comparison

The earlier diagnostic compared only cases with reported card use and text on both sides; its scope differs from the overall AI review above.

This case does not have both with-card continuation text and a model-reported use record. It remains among all 280 tests but does not enter the reported-use comparison.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.