Pith. sign in

REVIEW 4 major objections 5 minor 23 references

A single routing model can dispatch queries to the best LLM or agent and, on the tested pool, outperform the strongest single LLM while spending about 31 percent less.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-04 21:57 UTC pith:PDKFA4FC

load-bearing objection Unified model-and-agent routing is a real idea, but the paper's own SFT baseline contradicts its headline cost-efficiency claim and the agent evaluation is missing. the 4 major comments →

arxiv 2509.07571 v2 pith:PDKFA4FC submitted 2025-09-09 cs.MA cs.AI

Towards Generalized Routing: Model and Agent Orchestration for Adaptive and Efficient Inference

classification cs.MA cs.AI
keywords LLM routingagent selectionmixture of models and agentscost-performance tradeoffPareto frontierTOPSISintent recognitionstate machine routing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that the usual split between LLM routers and agent-selection systems is unnecessary: one trained router can decide, for each query, whether a specialized agent should handle it or which LLM should. The authors build MoMA, a router that learns a per-model capability profile from LLM-judged pairwise comparisons and chooses execution units from a performance-cost Pareto frontier, with agents selected by a two-layer state machine. The empirical claim is that this joint routing beats existing routers and matches or exceeds the best single model in its pool at a lower cost, with a roughly 31 percent saving in performance-priority mode and a larger saving in auto mode.

Core claim

MoMA's central claim is that routing decisions should be made jointly over models and agents, not separately. For LLM routing the router encodes the query with an instruction-tuned LLM, passes the hidden state through a mixture-of-experts head, and outputs a predicted performance score for every candidate model; pairwise comparisons judged by an LLM provide training labels. At inference these scores are combined with per-model prices through a Pareto frontier, and TOPSIS picks the model closest to the ideal point of lowest cost and highest score. For agent routing, a first layer maps the query to coarse categories by embedding similarity, and a context-aware finite state machine with token-l

What carries the argument

The load-bearing machinery is the combination of an LLM encoder with a mixture-of-experts (MoE) scoring head trained on LLM-judged pairwise comparisons, plus a TOPSIS-based Pareto selection step. The MoE head maps a query to a vector of predicted scores, one per candidate model; pairwise judge labels (win/lose/tie, with strong/weak variants) define a five-way soft target, and a categorical cross-entropy loss trains the scorer. The Pareto frontier over predicted score and price, followed by TOPSIS relative closeness, converts that score vector into one chosen model. On the agent side, the context-aware finite state machine (a 4-tuple state machine with rule-based and embedding-based transitio

Load-bearing premise

Everything depends on the LLM judge's pairwise preferences being a faithful stand-in for true answer quality: the router is trained on those judgments, but the paper never shows they correlate with the correctness scores used in the benchmarks, so if the judge favors fluent or stylish answers over correct ones, the reported savings would not transfer to real accuracy.

What would settle it

Take a held-out set of queries from AIME2024, LiveCodeBench, and SimpleQA, get the router's predicted scores and the judge's pairwise preferences for each pair of models, and compare them with the benchmark correctness of the two model outputs. If models the judge ranks higher lose on correctness in a substantial fraction of pairs, or if the router's chosen model underperforms the always-best model when the reported TOPSIS weights are used, the central claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the benchmark results hold, a service that always calls its best single model can switch to MoMA in performance-priority mode and get equal or better average scores while spending about 31 percent less per query.
  • Auto-routing mode spends 37.19 percent less than performance-priority mode and still scores above deepseek-v3, giving operators a concrete knob between accuracy and price.
  • Because the router can route math queries to jiutian-math-8b and code queries to jiutian-code-8b, small specialized models become economically viable even when their standalone benchmark average is low.
  • Agent selection scales with the number of agents: the two-layer design and masking keep context bounded, so adding agents does not force the router to compare all agent descriptions at once.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The percentage savings are tied to the specific API prices in Table 2; the durable structural claim is that a pool of heterogeneous models has a performance-cost frontier that a judge-trained router can navigate, so the size of the saving will shift as prices move.
  • A natural testable extension is to measure the agreement between the LLM judge's pairwise preferences and the three benchmarks' correctness scores on a held-out query set; that correlation predicts how much of the reported saving would survive on new domains.
  • The same FSM-plus-masking selection mechanism could be reused inside a single agent to constrain tool calls, not just to pick among agents, since it only requires a state set and a logits mask over valid next tokens.
  • If the router's per-domain capability scores are reliable, they could be used to choose which new small model to fine-tune, closing a loop between routing data and model development.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes MoMA, a routing framework that jointly directs user queries to either an LLM or an agent. It constructs a 2.25M-instance training corpus, profiles LLM capabilities with pairwise LLM-as-a-judge labels, trains an LLM-encoder with a mixture-of-experts head to predict per-model performance scores, and selects models via a Pareto frontier plus TOPSIS. Agent routing is handled by hierarchical category retrieval followed by a context-aware finite state machine with token-logit masking. Experiments on AIME2024, LiveCodeBench, and SimpleQA compare MoMA against single models, an SFT classification router, and a contrastive-learning router. In performance-priority mode, MoMA achieves an average score of 70.1 at cost 10.04, versus 68.6 at cost 14.65 for the best single model (qwen3-235b-a22b), which is the paper's main quantitative evidence for a cost-performance advantage.

Significance. If the reported results hold, the joint model-and-agent routing architecture and the 31% cost saving in performance-priority mode are useful contributions to LLM routing. The paper also contributes a large capability-profiling dataset and a systematic comparison of three router architectures. However, the headline claim of "superior cost-efficiency" is not supported by the paper's own auto-routing comparison, and key elements of the routing mechanism are not reported or validated. Because these issues affect the central claim, the manuscript needs substantial additional evidence before it can be accepted.

major comments (4)
  1. [§5.2.4, Table 3] The auto-routing row directly contradicts the "superior cost-efficiency" claim. SFT auto-routing achieves average score 62.7 at cost 8.667 (score/cost ≈ 7.23), while MoMA auto-routing achieves 43.3 at cost 6.306 (score/cost ≈ 6.87). The SFT baseline has both higher absolute score and better score per cost. The only rebuttal, that SFT benefits from "relatively constrained data categories," is not accompanied by any category-count ablation, dataset statistics, or experiment. This is load-bearing: either add an experiment showing SFT degrades as category count grows, or revise the claim.
  2. [§4.1.4, Eq. (9)] The TOPSIS weights w_c and w_s are never reported, but the auto-routing results in Table 3 depend on them. Without these values, the auto-routing comparison is non-reproducible, and the MoMA auto-routing row could reflect favorable weight selection. Report the exact weights and provide a sensitivity analysis over w_c and w_s.
  3. [§4.1.2, Eqs. (1)–(6); §5.2] The router is trained on LLM-as-a-judge pairwise preference labels, but evaluation reports benchmark accuracy. The paper never measures the correlation between judge preferences and benchmark correctness. If the judge rewards style, length, or tone rather than correctness, the predicted score curves and all routing choices built on them are miscalibrated. Add a held-out validation study comparing predicted score orderings with actual benchmark performance, or clearly report the correlation.
  4. [§4.2, §5.3] The agent-routing component has no quantitative evaluation. Section 5.3 only states that MoMA is deployed with over 20 agents; there are no accuracy, cost, latency, or ablation results for agent selection. Since the paper claims a "generalized routing" framework covering both LLMs and agents, the agent-routing claims are unsupported. Add agent-selection benchmarks or explicitly scope the experimental claims to LLM routing.
minor comments (5)
  1. [§5.2.2] Typo: "can bes seen" should be "can be seen."
  2. [Table 3] Typo "cost-proirity" should be "cost-priority." Also, the table lacks clear units for cost and formatting for missing values (e.g., jiutian-math-8b rows).
  3. [§5.2.3] The text mentions combining "LLMs' FLOPS" to form the Pareto curve, but Table 2 gives API prices per 1K tokens. Clarify whether cost is measured as FLOPs, price, or both, and keep the terminology consistent.
  4. [§4.1.3–§4.2] Hyperparameters κ, m, θ, α, top-k categories, top-k agents, and Elo rating details (initialization, update rule, number of comparisons) are not reported. Include them for reproducibility.
  5. [Figure 6] The axis labels and text in Figure 6 are garbled/unreadable in the current PDF; please replace with a clearly rendered figure.

Circularity Check

0 steps flagged

No significant circularity: the router's training target (LLM-judge preferences) and evaluation metric (external benchmarks) are distinct; the main issues are correctness/reproducibility risks, not circular reductions.

full rationale

The paper's claimed derivation chain is: (1) construct a 2.25M-query corpus; (2) use LLM-as-a-judge pairwise comparisons to create training labels; (3) train a MoE router to predict judged performance scores; (4) combine predicted scores with costs in a Pareto/TOPSIS selection; (5) evaluate the resulting routing decisions on AIME2024, LiveCodeBench, and SimpleQA. Steps (2)-(4) form a coherent supervised-learning pipeline whose target is the judge's preference, while step (5) measures a different, external quantity (benchmark correctness). No equation defines benchmark accuracy in terms of the judge labels, and no fitted parameter is renamed as a benchmark prediction. The unreported TOPSIS weights w_c, w_s (Eq. 9) and the unmeasured judge-to-benchmark correlation are reproducibility/validity concerns, but they do not make the derivation circular. The SFT auto-routing row in Table 3 (62.7 at cost 8.667 vs MoMA 43.3 at 6.306) contradicts the 'superior cost-efficiency' claim unless the paper's unmeasured category-count explanation holds; this is an internal-consistency problem, not a circularity. There are no self-citation chains, no imported uniqueness theorems, and no ansatz smuggled in via citation. The paper is evaluated against external benchmarks, so it is self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

9 free parameters · 6 axioms · 0 invented entities

The router's central contribution rests on an unvalidated judge-preference training signal and at least nine undisclosed free hyperparameters, including the TOPSIS weights that define the reported cost-performance tradeoff. No new physical or formal entities are introduced; the context-aware FSM and masking strategy are engineering components.

free parameters (9)
  • kappa (comparison strength hyperparameter)
    Eq. (2); no value reported, controls the sharpness of strong-win classification.
  • m (margin hyperparameter)
    Eq. (3); no value reported, shifts the strong-win threshold.
  • theta (dynamic threshold)
    Eq. (1); constrained to theta > 1, but how it is set or computed per pair is not explained.
  • w_c (TOPSIS cost weight)
    Eq. (9); user-set weight whose value is unreported, yet it determines the reported auto-routing cost-performance tradeoff in Table 3.
  • w_s (TOPSIS score weight)
    Eq. (9); user-set weight whose value is unreported.
  • alpha (relevance threshold)
    Eq. (12); first-layer agent category filter threshold, value unreported.
  • top-k categories
    Eqs. (19)-(20); number of categories passed to second-layer routing, unreported.
  • top-k agents
    Eq. (25); candidate agent count for the final LLM decision, unreported.
  • Elo rating parameters
    Section 4.1.2; initialization and update factor for the model ranking used to build training labels, unreported.
axioms (6)
  • domain assumption LLM-as-a-judge pairwise preferences are a faithful proxy for true model quality on the target benchmarks.
    Section 4.1.2 builds all training labels from judge comparisons; Section 5.2 evaluates on correctness-based benchmarks. The correlation between the two is never measured.
  • domain assumption Single-pass decoding on AIME2024, LiveCodeBench, and SimpleQA is the correct operational measure of performance.
    Section 5.2.1 defines the evaluation; no decoding parameters, temperature, or sampling details are given.
  • domain assumption Min-max normalization and TOPSIS with Euclidean distances recover the user's true cost-performance optimum.
    Section 4.1.4 assumes this decision rule is optimal for any user; the weights w_c and w_s are free and unreported.
  • domain assumption The 2.25M-instance training distribution covers the benchmark query distribution.
    Section 4.1.1; no domain-shift or contamination analysis is provided.
  • standard math K-Means over SBERT agent-description embeddings yields semantically coherent agent categories.
    Eq. (15) in Section A.2.1; standard clustering, but cluster validity is not checked.
  • domain assumption Rule-based pre-filtering plus embedding similarity captures user intent states.
    Section A.2.2; no labeled evaluation of state detection accuracy is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 38162 in / 21481 out tokens · 208442 ms · 2026-08-04T21:57:22.464145+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Towards Generalized Routing: Model and Agent Orchestration for Adaptive and Efficient Inference." pith.science (2026). https://pith.science/paper/PDKFA4FC

@misc{pith2026250907571,
  author       = {Pith},
  title        = {Pith review of: Towards Generalized Routing: Model and Agent Orchestration for Adaptive and Efficient Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDKFA4FC}},
  note         = {Machine review of arXiv:2509.07571}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The rapid advancement of large language models (LLMs) and domain-specific AI agents has greatly expanded the ecosystem of AI-powered services. User queries, however, are highly diverse and often span multiple domains and task types, resulting in a complex and heterogeneous landscape. This diversity presents a fundamental routing challenge: how to accurately direct each query to an appropriate execution unit while optimizing both performance and efficiency. To address this, we propose MoMA (Mixture of Models and Agents), a generalized routing framework that integrates both LLM and agent-based routing. Built upon a deep understanding of model and agent capabilities, MoMA effectively handles diverse queries through precise intent recognition and adaptive routing strategies, achieving an optimal balance between efficiency and cost. Specifically, we construct a detailed training dataset to profile the capabilities of various LLMs under different routing model structures, identifying the most suitable tasks for each LLM. During inference, queries are dynamically routed to the LLM with the best cost-performance efficiency. We also introduce an efficient agent selection strategy based on a context-aware state machine and dynamic masking. Experimental results demonstrate that the MoMA router offers superior cost-efficiency and scalability compared to existing approaches.

Figures

Figures reproduced from arXiv: 2509.07571 by Chao Deng, Chunfang Ji, Junlan Feng, Qinglan Li, Shan Wang, Wenhao Xi, Xiyu Guo, Xuefeng Zhao, Yaoyao Liu.

Figure 1
Figure 1. Figure 1: Illustration of the proposed adaptive routing model. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The MoMA routing model framework. The overall framework of MoMA routing model is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training data distribution by category. these samples, we then design pairwise model comparison tasks and collect the corresponding combating results. For evaluation, the LLM-as-a-judge framework is adopted to determine the relative performance of model pairs, resulting in the construction of quadruples in the format Di = [qi , ma, mb, wi] for each query qi , and ma and mb denote LLMs a and b, respectively… view at source ↗
Figure 4
Figure 4. Figure 4: LLM routing network structure. where c k i ′ and s k i ′ denote the normalized cost and score. c k i ′ is expected to be as small as possible, while the s k i ′ is expected to be as large as possible. Then, the ideal point can be denoted as P + = (0, 1) corresponding to the lowest cost and the highest performance, and the anti-ideal point is P − = (1, 0). Given the weights wc and ws for cost and performanc… view at source ↗
Figure 5
Figure 5. Figure 5: Exploring the performance of Jiutian serial LLMs in the mathematics domain. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The Pareto frontiers curve for score-cost. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Model usage percentage across code, mathematical, and general knowledge domains. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Training data distribution in the technology domain. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: The Agent Routing framework. These categories possess the domain and semantic characteristics, and encompass major agent ap￾plication scenarios. Bottom-Up Approach : Let the agents description set be A = {a1, a2, . . . , an}. For each agent j’s functional description aj , we perform vectorization by a pre-trained SBERT embedding model: ej = fembed(aj ) ∈ R d , (14) where d is the embedding vector dimensio… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 4 canonical work pages

  1. [1]

    2024 aime problems and solutions.,

    AIME. 2024 aime problems and solutions.,

  2. [5]

    Graphrouter: A graph-based router for llm selections

    Tao Feng, Yanzhen Shen, and Jiaxuan You. Graphrouter: A graph-based router for llm selections. arXiv preprint arXiv:2410.03834,

  3. [6]

    Prompt-to-leaderboard.arXiv preprint arXiv:2502.14855,

    Evan Frick, Connor Chen, Joseph Tennyson, Tianle Li, Wei-Lin Chiang, Anastasios N Angelopou- los, and Ion Stoica. Prompt-to-leaderboard.arXiv preprint arXiv:2502.14855,

  4. [7]

    Moa: Mixture of sparse attention for automatic large language model compression.arXiv preprint arXiv:2406.14909,

    Tianyu Fu, Haofeng Huang, Xuefei Ning, Genghan Zhang, Boju Chen, Tianqi Wu, Hongyi Wang, Zixiao Huang, Shiyao Li, Shengen Yan, et al. Moa: Mixture of sparse attention for automatic large language model compression.arXiv preprint arXiv:2406.14909,

  5. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    15 Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

  6. [9]

    Joyagents-r1: Joint evolution dynamics for versatile multi-llm agents with reinforcement learning

    Ai Han, Junxing Hu, Pu Wei, Zhiqian Zhang, Yuhang Guo, Jiawei Lu, and Zicheng Zhang. Joyagents-r1: Joint evolution dynamics for versatile multi-llm agents with reinforcement learning. arXiv preprint arXiv:2506.19846,

  7. [10]

    Large language models for code: Security hardening and adver- sarial testing

    Jingxuan He and Martin Vechev. Large language models for code: Security hardening and adver- sarial testing. InProceedings of the 2023 ACM SIGSAC Conference on Computer and Communi- cations Security, pp. 1865–1879,

  8. [12]

    Livecodebench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974,

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974,

  9. [13]

    Llm-blender: Ensembling large language models with pairwise ranking and generative fusion.arXiv preprint arXiv:2306.02561,

    Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. Llm-blender: Ensembling large language models with pairwise ranking and generative fusion.arXiv preprint arXiv:2306.02561,

  10. [16]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437,

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437,

  11. [17]

    Routing to the expert: Efficient reward-guided ensemble of large language models.arXiv preprint arXiv:2311.08692,

    Keming Lu, Hongyi Yuan, Runji Lin, Junyang Lin, Zheng Yuan, Chang Zhou, and Jingren Zhou. Routing to the expert: Efficient reward-guided ensemble of large language models.arXiv preprint arXiv:2311.08692,

  12. [18]

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica

    URLhttps://manus.im/zh-cn/blog/ Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus. Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data.arXiv preprint arXiv:2406.18665,

  13. [19]

    Mixture-of-agents enhances large language model capabilities.arXiv preprint arXiv:2406.04692,

    16 Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Zou. Mixture-of-agents enhances large language model capabilities.arXiv preprint arXiv:2406.04692,

  14. [20]

    Measuring short-form factuality in large language models

    Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. arXiv preprint arXiv:2411.04368,

  15. [21]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

  16. [22]

    Be- yond gpt-5: Making llms cheaper and better via performance-efficiency optimized routing.arXiv preprint arXiv:2508.12631, 2025a

    Yiqun Zhang, Hao Li, Jianhao Chen, Hangfan Zhang, Peng Ye, Lei Bai, and Shuyue Hu. Be- yond gpt-5: Making llms cheaper and better via performance-efficiency optimized routing.arXiv preprint arXiv:2508.12631, 2025a. Yiqun Zhang, Hao Li, Chenxu Wang, Linyao Chen, Qiaosheng Zhang, Peng Ye, Shi Feng, Daling Wang, Zhen Wang, Xinrun Wang, et al. The avengers: A...

  17. [23]

    macro- level category signals

    A APPENDIX A.1 DETAILEDTRAININGDATADISTRIBUTION Figure 3 presents the overall distribution of the constructed training data across different domains, with each domain further divided into multi-level subcategories. Such a hierarchical organization not only ensures comprehensive coverage of diverse user tasks but also provides explicit structural signals t...

  18. [2001]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186,

  19. [2019]

    Hybrid llm: Cost-efficient and quality-aware query routing.arXiv preprint arXiv:2404.14618,

    Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Ruhle, Laks VS Lakshmanan, and Ahmed Hassan Awadallah. Hybrid llm: Cost-efficient and quality-aware query routing.arXiv preprint arXiv:2404.14618,

  20. [2020]

    Rethinking mixture-of-agents: Is mixing differ- ent large language models beneficial?arXiv preprint arXiv:2502.00674,

    Wenzhe Li, Yong Lin, Mengzhou Xia, and Chi Jin. Rethinking mixture-of-agents: Is mixing differ- ent large language models beneficial?arXiv preprint arXiv:2502.00674,

  21. [2023]

    Message type identification of binary network protocols using continuous segment similarity

    Stephan Kleber, Rens W van der Heijden, and Frank Kargl. Message type identification of binary network protocols using continuous segment similarity. InIEEE INFOCOM 2020-IEEE confer- ence on computer communications, pp. 2243–2252. IEEE,

  22. [2024]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capa- bilities.arXiv preprint arXiv:2507.06261,

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capa- bilities.arXiv preprint arXiv:2507.06261,

  23. [2025]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp. 4171–4186,