Pith. sign in

REVIEW 3 major objections 4 minor 26 references

When Does Trace-Driven Evaluation Mislead MoE Expert Caching? Replay Semantics, Workload Contamination, and Operating Regimes

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Trace-driven evaluation of MoE expert caching is fragile in three specific ways; after correction, the remaining 44–46% offline-optimal gap is mostly future-victim knowledge that a representative causal predictor cannot supply.

desk verdict A genuinely careful measurement-fragility study with a portable checklist; the headline numbers rest on one 4-bit model, but the paper scopes its claims honestly and deserves a real referee. read the letter →

arxiv 2608.07911 v1 pith:5ODLSPPQ submitted 2026-08-08 cs.LG cs.PF

Yu Zhang This is my paper · ORCID
classification cs.LGcs.PF
keywords mixture-of-expertsexpertcachingtrace-drivenevaluationreplaysemanticsworkloadcontaminationoffline-optimalgapnext-usepredictioncacheeviction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the standard way of measuring how much an expert-cache policy can help MoE inference is fragile enough to produce reversed conclusions, and that the honest remainder is mostly not actionable. Three evaluation choices are isolated: replay semantics (whether a fused (step, layer) cache event is replayed as individual accesses), workload construction (whether probe categories share instruction templates), and operating regime (the ratio of per-step expert union to per-layer capacity). Each axis is shown to change conclusions, not just numbers: sequential replay inflates recency-based policies by 27–29% and inverts their ranking; single-template probe sets manufacture apparent workload locality; aligning models on batch size instead of the union-to-capacity ratio hides a 36.8-percentage-point spread. After all corrections, a stable 44.2–45.9% gap to the offline optimum remains at rho = 40% and B = 8, but the forced-admission decomposition attributes 84.3–96.6% of it to knowing which resident expert is used furthest in the future. A linear next-use predictor fitted on causal features recovers -11.4% of the gap, choosing an optimal victim barely more often than random, which supports the paper's narrow central claim: reporting a large offline-optimal gap without decomposing it overstates the gains recoverable by lightweight causal mechanisms.

What carries the argument

The argument is carried by four instruments. (1) Event-atomic replay (Definition 2): the cache treats the batch-wide expert union $E(s,l)$ at each (step, layer) as one committed unit, classifies all members against a start-of-event snapshot, and applies admission and eviction only at the event boundary; this is what makes sequential replay's intra-event eviction visible as an artifact. (2) A matched-pair probe design in which the same source records are rendered with diverse versus fixed templates, turning prompt-surface contamination into a measurable treatment. (3) The regime ratio $r(s,l) = |E(s,l)|/c_l$, the per-step per-layer expert union divided by the per-layer cache quota, with the claim that its distribution must be reported for cross-model comparison even though its mean is not sufficient. (4) The Belady-forced-admit decomposition, which splits the offline-optimal gap into a bypass-admission share and a future-victim-selection share, plus a causal next-use-distance predictor substituted into the same eviction and admission machinery.

What would settle it

Collect a decode-routing trace from a frontier-scale model (896 routed experts, top-16) at a batch size where $r_{\mathrm{bar}} \approx 0.3$, replay it under the paper's event-atomic protocol, and compute the forced-admission decomposition. If the future-victim share is well below 84% there, or if a next-use predictor trained on that trace recovers a positive fraction of the gap, the paper's central claim would fail. Separately, an end-to-end engine that serializes expert loads within a layer and shows no 27–29% inflation for LRU under per-access replay would falsify the replay-semantics axis for that execution contract.

Watch

Extended reading notes

Core claim

Under an event-atomic replay protocol—where the expert set touched by the whole batch at each (step, layer) is one committed unit, hits and misses are classified against the snapshot at event start, and admission/eviction is deferred to the event boundary—the paper finds that the offline-optimal gap to Belady is large and stable across workload compositions, but that most of it is not reachable by the causal mechanisms a practitioner would use. The decomposition via Belady-forced-admit shows bypass admission accounts for 15.7% at B = 8 and 3.4% at B = 2 of the gap, leaving 84.3% and 96.6% to future-victim selection. A next-use-distance predictor with strictly more features than the best causal baseline transfers at $R^2 = 0.24$ and still performs worse than LFRU, recovering -11.4% of the gap; at eviction decision points it picks an optimal victim 3.39% of the time versus 2.42% for random and 20.6–22.1% for LRU and LFRU. The paper states the claim narrowly: a large offline-optimal gap substantially overstates the gains recovered by representative lightweight causal mechanisms in these evaluated settings, so a gap reported without the forced-admission decomposition invites an unwarranted inference.

Load-bearing premise

The load-bearing premise is that MoE serving obeys the fused-event traffic contract—each (step, layer) sees the batch-wide expert union as one committed unit with retention deferred to the event boundary—and that traces from a 4-bit 128-expert model plus two smaller models represent the operating regime where expert caching will be deployed; if a real engine serializes expert loads inside a layer, the replay-semantics results do not apply, and the paper concedes the frontier-scale regime is the one its data covers least well.

Editorial extensions

If this is right

  • Any simulator that claims fused-event accounting but replays per-access eviction will inflate LRU and LFRU by 27–29% and invert policy rankings; such results should be re-checked under event-atomic replay before being used to motivate a policy.
  • Cross-model MoE cache comparisons should align on the per-step union-to-capacity ratio $r_{\mathrm{bar}}$ rather than batch size; the three models here agree to within 1.1–4.6 percentage points when aligned on $r_{\mathrm{bar}}$, versus a 36.8-percentage-point spread on batch size.
  • Report the forced-admission decomposition alongside any offline-optimal gap: at the two held-out operating points 84.3–96.6% of the gap is future-victim knowledge, so the raw gap is not available headroom.
  • Workload-conditioned routing studies should use multiple template forms and matched-pair controls; single-template probe sets manufactured an apparent 57.2% cache benefit that fell to 5.9% after correction and reversed the category ordering.
  • The reported gap is regime-specific: the 44.18–45.93% range applies only to rho = 40%, B = 8, and must not be pooled with other operating points.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The checklist's likely effect, if adopted, is to make a raw offline-optimal gap alone an insufficient motivation for a cache controller; future work would need to name the recoverable share, which should push effort toward admission policies, prefetching, and scheduling rather than generic eviction learning.
  • The union-to-capacity ratio suggests a testable scaling rule: the recoverable gap should collapse as $r_{\mathrm{bar}}$ crosses 1, so workloads with larger batches or smaller per-layer caches will have less real headroom; a frontier-scale trace would let someone check whether this holds at 896 experts.
  • The paper's negative next-use result is for victim ranking, not for the different task of predicting which experts will be routed next; a prefetch policy aimed at the latter could in principle capture part of the gap that eviction cannot, and is a natural next experiment.
  • If real engines with serialized expert loads within a layer do not show the replay inflation, then the replay-semantics axis reduces to a simulator-correctness requirement rather than a system property; measuring transferred block counts on such an engine would settle the scope.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper studies trace-driven evaluation of MoE expert caching and argues that three evaluation choices—replay semantics, workload probe construction, and operating-regime alignment—can change conclusions rather than merely shift numbers. Under an event-atomic replay contract it shows that flattened per-access replay inflates recency-based policies by 27–29% and inverts policy rankings. It demonstrates that single-template workload probes generate verbatim shared prefixes that inflate apparent expert locality, and it introduces a matched-pair rendering control plus diversity-controlled probe set. It shows that normalized miss fractions do not transfer across models and that the per-step expert-union-to-capacity ratio is necessary but not sufficient, since temporal reordering of an identical event stream changes the offline-optimal gap. After correcting these axes, a stable offline-optimal gap of 44.2–45.9% remains at rho=40%, B=8; a forced-admission oracle attributes 84.3–96.6% of this gap to future-victim knowledge, and a linear next-use predictor recovers -11.4% of the gap. The paper closes with a reporting checklist, four scoped negative results, one invalidated design, and a detailed self-audit. The central claim is deliberately narrow: in the authors' evaluated settings, a large offline-optimal gap substantially overstates the gains recovered by representative lightweight causal mechanisms.

Significance. If the results hold, the paper makes a strong methodological contribution. The machine-checkable reference trace, pre-registered confirmatory thresholds, disjoint discovery/confirmatory splits, matched controls, and candid disclosure of invalidated designs are exemplary and raise the bar for this literature. The forced-admission decomposition of the offline gap into bypass-admission and future-victim components is a genuinely useful diagnostic, and the finding that a next-use predictor with positive transfer R2 is anti-correlated with true victim ranking at decision points is a surprising, falsifiable result. The paper's three scoped negative results and the invalidated partition experiment are reported with unusual honesty. The main limitation is external scope: the primary quantitative claims rest on one 4-bit quantized 128-expert model and one replay contract, and the paper itself concedes that the frontier-scale regime motivating the work is the one its data cover least well. This does not undermine the internal logic, but it does bound how far the headline conclusion can be generalized without additional evidence.

major comments (3)
  1. [§7.4, §11] The central decomposition and predictor-failure claim rest on a single 4-bit quantized 128-expert model (Qwen3-30B-A3B) with greedy decoding, under one replay contract. Section 11 concedes both that quantization may change routing and that no trace was collected at the 896-expert scale that motivates the application. Since the paper's practical message extends beyond this one model—'reporting such a gap without decomposing it invites an unwarranted inference'—the external-validity gap is load-bearing. I ask for either a second full-precision or otherwise non-quantized trace at comparable expert count with the same decomposition, or an explicit scoping of the §7.4 claim to the 128-expert 4-bit fused-event setting in the abstract and conclusion, with the checklist presented as the only generalizable output.
  2. [§2.2, §7.2] The decomposition is defined only for Definition 2's fused-event contract, yet §2.2 explicitly accepts serialized intra-layer loading as a legitimate alternative execution model. The 27–29% replay inflation in Table 2 demonstrates sensitivity to semantics, but the paper never quantifies how m_c, m_B, m_F, or the 84.3–96.6% future-victim share would change under a serialized contract. If serialized execution is legitimate, then the headline statement that most of the gap is future-victim knowledge is contract-conditional. Please add a sensitivity analysis for a serialized intra-event replay, or state in the central claim that all §7 numbers are conditional on the fused-event atomic contract.
  3. [§7.3, Table 16] The causal predictor is a single linear model over eight features, and the negative result is presented as representative of 'representative lightweight causal mechanisms.' Table 16's central exhibit—anti-correlated victim ranking at decision points—is striking, but its force depends on whether the predictor class is representative of what a practitioner would try. The paper acknowledges that a predictor with materially higher transfer accuracy could change the result, but it does not include any stronger baseline on the same features. I request either a second predictor class (for example, gradient-boosted trees or a small MLP with the same features) or a consistent rewording of the abstract and conclusion to say 'a linear next-use predictor fails' rather than 'representative lightweight causal mechanisms fail.'
minor comments (4)
  1. [Throughout] The text contains repeated typographical artifacts such as 'traffic' and 'sufficient' with nonstandard ligatures; these should be normalized to standard 'traffic' and 'sufficient' in the final version.
  2. [§10, Appendix B] The section numbering jumps from §10.4 to §10.6, and Appendix B.2 refers to 'the reporting block of §10.5,' which has no heading in the main text. Add the missing §10.5 heading or renumber the checklist subsections consistently.
  3. [Table 8, Table B1] Table 8 is badly corrupted in the supplied text: the tool agent row and the interval notation are unreadable as printed, and Table B1 contains a garbled row label. These tables must be repaired before publication because they support the workload-contamination and static-vs-dynamic claims.
  4. [§9.2] The text says the static identity predicts a miss ratio of 1−0.723 = 27.99% of union accesses, but 1−0.723 = 27.7%. If the simulator's 27.99% comes from unrounded shares, please state that explicitly; otherwise correct the arithmetic.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: central quantities are measured against an external Belady benchmark or on held-out data, and the gap decomposition is an explicit arithmetic identity, not a fitted prediction.

full rationale

The paper's derivation chain is: recorded routing traces are replayed under the explicitly defined event-atomic contract of Definition 2; the offline-optimal gap is defined against Belady's algorithm computed on the same fixed access sequence, so the benchmark is external to the policies being evaluated (Section 2.4, Section 7.1). The decomposition into admission and future-victim shares is an arithmetic identity: admission share = (m_F − m_B)/(m_c − m_B) and future-victim share = (m_c − m_F)/(m_c − m_B), stated as such in Section 7.2. This is an accounting decomposition, not a prediction that the future-victim component is recoverable; the paper explicitly warns that neither component is guaranteed to be causally recoverable. The next-use predictor is fitted on a calibration split and evaluated on a disjoint held-out split, and its failure — negative gap recovery and anti-correlated victim ranking — is itself the measured result rather than an input used to produce the central claim. No equation in the paper reduces a predicted quantity to a fitted input: the predictor's output is substituted into the same eviction and admission machinery and compared against the oracle, and the paper separates the definitional claim about the offline bound from the scoped, measured claims in Section 7.4. The same-trace static diagnostic is deliberately fitted on the evaluated trace only as a replay-invariance control, and the paper states explicitly that it is not a deployable baseline (Section 4.3), so it is not a fitted parameter masquerading as a prediction. The analytical reference frame for frontier-scale units is labelled as a projection, never a prediction (Section 3.2, Section 6.4). There is no load-bearing self-citation chain and no imported uniqueness theorem; the paper's limitations about single-model coverage, quantization, and the fused-event replay contract (Sections 2.2 and 11) are external-validity bounds, not circular reasoning. The central quantitative claims therefore stand or fall on the traces and the simulator, not on equivalence between inputs and outputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a small set of choices: the fused-event contract, the per-layer quota rule, the operating point (rho=0.40, B=8), and the representativeness of greedy-decoding traces from three open models. The only fitted objects are the next-use predictor and the static pin list, both used as baselines or as evidence of failure rather than as fitted targets. The invented frontier-scale reference frame is clearly labeled as a projection and does not carry the empirical results.

free parameters (3)
  • Primary operating point (rho, B) = rho=0.40, B=8
    The headline gap and decomposition are stated at this point. Section 6.3 shows the gap moves from 7.42% to 33.32% as rho changes, and Table 12 shows 50.42% at B=2, so the chosen point is load-bearing for the numerical claims.
  • Next-use predictor weights (8 features) = not reported; fitted on calibration split
    The §7.3 negative result uses a linear model fit on calibration traces; exact coefficients are not in the paper. The claim is about failure, so this is a fitted model used as evidence rather than a fitted target.
  • Static pinned set selection = 2,458 blocks from calibration trace
    The static baseline comparison in §9.1 depends on the fitted pin list; the paper audits this protocol, but the list itself is an artifact of the fitting trace.
assumptions (4)
  • domain assumption Fused-event traffic contract (Definition 1) is the target execution model for MoE caching.
    All quantitative conclusions use the batch-wide expert union per (step, layer); the paper acknowledges that a serialized-load engine is legitimate (§2.2, §12.1), so transfer to real systems depends on this contract.
  • domain assumption Per-layer cache scope with equal quotas and remainder to lowest-indexed layers.
    Section 2.3 defines primary scope; global scope is not the basis for the headline numbers.
  • domain assumption Traces collected with greedy decoding, chat template, up to 384 decode steps, and FCFS continuous batching represent production MoE serving workloads.
    Section 11 states no hardware validation, one primary model, 4-bit quantization, and no frontier-scale trace; the paper's own scope section lists this as the dominant limitation.
  • standard math Belady optimality for a fixed access sequence.
    Used as the offline bound; classical result (Belady 1966).
invented entities (1)
  • Frontier-scale analytical reference frame (896 routed experts, top-16, 92 layers)
    purpose: Expresses operating regimes and unit conversions at scale; explicitly no trace or measurement is taken on it.
    Defined in §3.2 from published architecture parameters and used only to state regimes (§6.4). It is hypothetical and clearly labeled, so it does not support the paper's empirical claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Does Trace-Driven Evaluation Mislead MoE Expert Caching? Replay Semantics, Workload Contamination, and Operating Regimes." pith.science (2026). https://pith.science/paper/5ODLSPPQ

@misc{pith2026260807911,
  author       = {Pith},
  title        = {Pith review of: When Does Trace-Driven Evaluation Mislead MoE Expert Caching? Replay Semantics, Workload Contamination, and Operating Regimes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ODLSPPQ}},
  note         = {Machine review of arXiv:2608.07911}
}
read the original abstract

Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard. This makes expert cache management an attractive lever: a policy that raised the hit rate would cut expert traffic per token. Evaluating that is a measurement problem, and we find the measurement fragile. With a trace-driven, event-atomic simulator over three MoE models (40, 64, 128 experts), we isolate three evaluation axes that change conclusions, not just numbers. Replay semantics: under a fused-event traffic contract, an inconsistent per-access replay inflates recency-based policies by 27-29% while leaving frequency-based and static ones within 4%, inverting the policy ranking. Workload contamination: probe sets using one instruction template per category produce verbatim-identical generation prefixes; a matched-pair rendering intervention moves the measured early-window effect by 19.4-31.9 points and reverses which workloads look most cache-friendly. Operating regimes: normalized miss fractions do not transfer across models, so the per-step expert union relative to per-layer capacity must be reported -- yet permuting only the temporal order of an identical event stream moves the offline-optimal gap from 44.9% to 30.8%, so it is not sufficient. Corrected, a stable gap to the offline optimum remains (44.2-45.9% over 13 frozen workload compositions). A forced-admission oracle attributes 84.3-96.6% of it to knowing which resident expert is used furthest in the future. A causal next-use predictor, used as an eviction rule, recovers -11.4% of the gap; it picks an optimal victim 3.4% of the time, against 2.4% for a random resident block and 20.6-22.1% for LRU and LFRU. Our position is narrow: in our evaluated settings a large offline-optimal gap substantially overstates the gains recovered by representative lightweight causal mechanisms.

Figures

Figures reproduced from arXiv: 2608.07911 by the authors.

Figure 1
Figure 1. Replay semantics selectively changes policy measurements. [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Prompt rendering contributes materially to the apparent early-window [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Matching mean union/cache is not sufficient. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Most of the offline-optimal gap is future-victim knowledge in the two held [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 7 canonical work pages

  1. [1]

    Allen Institute for AI. 2024. Public Mixtral Routing Analysis Artifact over C4 . Hugging Face dataset artifact. https://huggingface.co/datasets/allenai/analysis_mixtral

    work page 2024

  2. [2]

    Apache Software Foundation. 2026. Apache OFBiz . Software and sample data. https://ofbiz.apache.org/

    work page 2026

  3. [3]

    Bai, Yushi, Xin Lv, Jiajie Zhang, et al. 2023. LongBench : A Bilingual, Multitask Benchmark for Long Context Understanding . https://arxiv.org/abs/2308.14508

    arXiv 2023

  4. [4]

    Belady, Laszlo A. 1966. ``A Study of Replacement Algorithms for a Virtual-Storage Computer.'' IBM Systems Journal 5 (2): 78--101. https://doi.org/10.1147/sj.52.0078

    doi:10.1147/sj.52.0078 1966

  5. [5]

    Chen, Liangkun, Zijian Wen, Tian Wu, Xiaoxi Zhang, and Chuan Wu. 2025. SP-MoE : Speculative Decoding and Prefetching for Accelerating MoE -Based Model Inference . https://arxiv.org/abs/2510.10302v2

    arXiv 2025

  6. [6]

    Denning, Peter J. 1968. ``The Working Set Model for Program Behavior.'' Communications of the ACM 11 (5): 323--33. https://doi.org/10.1145/363095.363141

    arXiv 1968

  7. [7]

    Du, Zhixu, Shiyu Li, Yuhao Wu, et al. 2024. SiDA-MoE : Sparsity-Inspired Data-Aware Serving for Efficient and Scalable Large Mixture-of-Experts Models . https://arxiv.org/abs/2310.18859v2

    arXiv 2024

  8. [8]

    Eliseev, Artyom, and Denis Mazur. 2023. Fast Inference of Mixture-of-Experts Language Models with Offloading. https://arxiv.org/abs/2312.17238v1

    arXiv 2023

Show all 26 references
  1. [9]

    Fang, Zhiyuan, Zicong Hong, Yuegui Huang, et al. 2025. Fate : Fast Edge Inference of Mixture-of-Experts Models via Cross-Layer Gate . https://arxiv.org/abs/2502.12224v2

    2025 arXiv

  2. [10]

    Fedus, William, Barret Zoph, and Noam Shazeer. 2022. ``Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity.'' Journal of Machine Learning Research 23 (120): 1--39. https://arxiv.org/abs/2101.03961v3

    2022 arXiv

  3. [11]

    Gavhane, Nishant, Arush Mehrotra, Rohit Chawla, and Peter Proenca. 2025. MoE-Beyond : Learning-Based Expert Activation Prediction on Edge Devices . https://arxiv.org/abs/2508.17137v1

    2025 arXiv

  4. [12]

    Ho, et al

    Guha, Neel, Julian Nyarko, Daniel E. Ho, et al. 2023. LegalBench : A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models . https://arxiv.org/abs/2308.11462

    2023 arXiv

  5. [13]

    Hoang, Duc, Ajay Jaiswal, Mohammad Samragh, and Minsik Cho. 2026. SpecMD : A Comprehensive Study on Speculative Expert Prefetching . https://arxiv.org/abs/2602.03921v1

    2026

  6. [14]

    Hwang, Ranggi, Jianyu Wei, Shijie Cao, et al. 2024. Pre-Gated MoE : An Algorithm-System Co-Design for Fast and Scalable Mixture-of-Expert Inference . https://arxiv.org/abs/2308.12066v3

    2024 arXiv

  7. [15]

    IBM Research. 2024. IBM Granite 3.1 : Powerful Performance, Long Context, and More . IBM announcement and model documentation. https://www.ibm.com/new/announcements/ibm-granite-3-1-powerful-performance-long-context-and-more

    2024

  8. [16]

    Lei, Fangyu, Jixuan Chen, Yuxiao Ye, et al. 2024. Spider 2.0 : Evaluating Language Models on Real-World Enterprise Text-to- SQL Workflows . https://arxiv.org/abs/2411.07763

    2024 arXiv

  9. [17]

    MLX Contributors. 2025. MLX : An Array Framework for Apple Silicon . V. 0.32.0. Released. https://github.com/ml-explore/mlx

    2025

  10. [18]

    Muennighoff, Niklas, Luca Soldaini, Dirk Groeneveld, et al. 2025. OLMoE : Open Mixture-of-Experts Language Models . https://arxiv.org/abs/2409.02060v2

    2025 arXiv

  11. [19]

    Raffel, Colin, Noam Shazeer, Adam Roberts, et al. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. https://arxiv.org/abs/1910.10683

    2020 arXiv

  12. [20]

    Shazeer, Noam, Azalia Mirhoseini, Krzysztof Maziarz, et al. 2017. ``Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer.'' International Conference on Learning Representations. https://arxiv.org/abs/1701.06538v1

    2017 arXiv

  13. [21]

    Tang, Peng, Jiacheng Liu, Xiaofeng Hou, et al. 2024. HOBBIT : A Mixed Precision Expert Offloading System for Fast MoE Inference . https://arxiv.org/abs/2411.01433v2

    2024 arXiv

  14. [22]

    V., Celso J

    Vargas, Ricardo E. V., Celso J. Munaro, Patrick M. Ciarelli, et al. 2019. ``A Realistic and Public Dataset with Rare Undesirable Real Events in Oil Wells.'' Journal of Petroleum Science and Engineering 181: 106223. https://doi.org/10.1016/j.petrol.2019.106223

    2019

  15. [23]

    Xue, Leyang, Yao Fu, Zhan Lu, Luo Mai, and Mahesh Marina. 2025. MoE-Infinity : Efficient MoE Inference on Personal Machines with Sparsity-Aware Expert Cache . https://arxiv.org/abs/2401.14361v3

    2025 arXiv

  16. [24]

    Yan, Fei, Huanzhi Mao, Charlie Cheng-Jie Xu, et al. 2024. Berkeley Function Calling Leaderboard. https://arxiv.org/abs/2402.11717

    2024

  17. [25]

    Yang, An, Anfeng Li, Baosong Yang, et al. 2025. Qwen3 Technical Report . https://arxiv.org/abs/2505.09388v1

    2025 arXiv

  18. [26]

    Zhu, Zeyu, Gang Li, Peisong Wang, et al. 2026. DALI : A Workload-Aware Offloading Framework for Efficient MoE Inference on Local PCs . https://arxiv.org/abs/2602.03495v1. CSLReferences document

    2026

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.