Pith. sign in

REVIEW 3 major objections 4 minor 75 references

Coordinated flow scheduling lets on-device AI agents work during live video calls, holding video quality equal to video-only streaming while cutting agent response time by 31%.

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 →

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

T0 review · deepseek-v4-flash

2026-08-01 04:20 UTC pith:2ZCU4YCH

load-bearing objection Solid systems paper on joint video/agent flow control; the core idea is real, but the short-SFU-RTT premise is less supported than the headline gains require. the 3 major comments →

arxiv 2607.22854 v1 pith:2ZCU4YCH submitted 2026-07-24 cs.AI

Coordinated Networking for On-Device Agent-Augmented Real-Time Communication

Goodsol Lee , Juheon Yi , Jinglu Wang , Haowen Xu , Saewoong Bahk , Yan Lu This is my paper
classification cs.AI
keywords agent-augmented RTCmulti-flow transport coordinationframe-level queueingcongestion controlKV cache streamingJohnson's rule schedulingon-device LLM inferenceWebRTC
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.

When an AI agent assists human participants during a live video call, two very different traffic types share one link: low-latency human video and bulky agent context files (compressed LLM key-value caches used to answer queries). The paper shows that uncoordinated transports fail at this — the agent flow builds queueing delay that collapses video quality for minutes, while protecting video starves the agent. It claims a unified app-layer orchestrator, HAFS, can have both: it measures video frame-level queueing delay, pushes the total sending rate up to — but never past — the frame deadline, reserves the video's required rate, and hands the leftover bandwidth to agent flows; a second scheduler then orders the agent flows so that network transfer and LLM decoding pipeline. On a WebRTC-based prototype running on real edge devices, the paper reports video quality of 91–92 VMAF, statistically matching video-only streaming, with 31% lower agent response time than existing baselines.

Core claim

The paper's central claim is that contention between human video and on-device agent flows is a coordination failure, not a bandwidth shortage, and that an app-layer orchestrator with visibility over both flow types can resolve it. HAFS's human–agent flow coordinator estimates frame-level queueing from dual feedback — RTP one-way delays for the video frame plus SCTP ACK round-trip times after it — and uses the sum Q_f as the control signal for a multiplicative-increase/additive-increase/multiplicative-decrease controller that fills the link up to the app-specified frame deadline (150 ms) without violating it. Within that budget the video rate is strictly provisioned first; the slack goes to

What carries the argument

The load-bearing object is the frame-level queueing estimate, Q_f = Q^RTP_f + Q^SCTP_f: the video frame's one-way transmission delay above its recent minimum, plus the incremental SCTP round-trip time accumulated after that frame's transmission. Both pieces are measured from existing transport feedback (RTCP reports and SCTP acknowledgements), so no new in-band signaling is needed; the estimator deliberately uses SCTP ACKs to populate the idle part of the frame interval where RTP packets are absent. This single quantity drives a three-regime rate controller (multiplicative increase while the queue is empty, additive increase near saturation, multiplicative decrease anchored on observed SCTP

Load-bearing premise

The load-bearing premise is §4.1's claim that the session path splits into short sender–SFU and SFU–receiver links with p95 latency under 50 ms — a figure cited from outside this paper rather than measured in its testbed — because the frame-queuing feedback loop must react within a frame interval to steer the rate precisely; a distant relay or a peer-to-peer session would lengthen that feedback and erode the claimed tradeoff.

What would settle it

Deploy HAFS with a relay far from both endpoints, or in a direct peer-to-peer session, so one-way feedback delay grows toward the 150 ms frame deadline, and measure p95 frame delay and VMAF during concurrent KV-cache transfers. The MI/AI-MD controller needs feedback within a frame interval; if p95 delay breaches the deadline or VMAF stops matching video-only streaming, the short-path premise is broken and with it the central claim.

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

If this is right

  • Agent-augmented calls need not sacrifice human video quality: HAFS holds 91–92 VMAF, matching video-only streaming, over Wi-Fi 6 and 5G while compressed KV caches stream concurrently.
  • Agent response time drops 31% on average, and the advantage grows with contention — the gap over uncoordinated scheduling widens from 53 s to 64 s as concurrent contexts rise from 3 to 9 — because Johnson's rule keeps receiver GPU work overlapping with network transfer.
  • Streaming a pre-built compressed KV cache can beat local inference hardware: HAFS's time-to-first-token is 2.2×–6.2× faster than a dedicated on-device NPU backend, and per-query energy drops 3.1× across repeated sessions on the same context.
  • Bounded queues are the mechanism behind both gains: SCTP p95 RTT falls from 432 ms to 94 ms and video stalls over 150 ms drop from 25.7% to 0.1% relative to uncoordinated transport.
  • The coordination is deployable with app-layer changes only over standard WebRTC APIs, at negligible system cost (within 5% CPU and 3–20% memory), so it does not require redesigning the transport stack.

Where Pith is reading between the lines

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

  • The split-path premise (p95 latency under 50 ms, cited from an outside report rather than this paper's measurements) ties the gains to relay-based topologies; my read is that a peer-to-peer session or a geographically distant SFU would stretch the feedback loop and should erode the claimed tradeoff — a testable boundary for the design.
  • The frame-deadline controller generalizes beyond video in my view: any real-time flow with a hard per-frame latency budget (cloud gaming, AR/VR, volumetric video) sharing a link with bulk background traffic could reuse the same MI/AI-MD logic, since it needs only a deadline and a queue estimate.
  • Modeling agent response time as a two-machine flow shop suggests a broader principle: classic scheduling theory, not transport heuristics, governs distributed LLM inference, and the tolerance to prediction error is really a property of rank-based ordering that should transfer to other multi-agent retrieval workloads.
  • The paper's Limitations section concedes that HAFS yields to queue-building cross-traffic to protect video rather than probing a fair share; a deployment consequence I draw is that its throughput advantage is conditional on the network environment, and in-network coordination would be needed for HAFS flows to compete with aggressive transports.

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

3 major / 4 minor

Summary. HAFS is an app-layer, end-host framework for on-device agent-augmented real-time communication. It coordinates live video (RTP) and agent context (SCTP) flows by estimating video frame-level queuing from combined RTP/SCTP feedback and driving an MI/AI-MD controller that provisions the video rate and allocates slack to agent flows. A second coordinator schedules multiple agent context transfers as a two-machine flow-shop problem using Johnson's rule, with response lengths predicted by a lightweight single-step LLM decoding. The paper contributes measurements of commercial RTC behavior, a WebRTC+llama.cpp prototype, and an evaluation on three edge devices over Wi-Fi 6 and 5G traces, reporting 1.5× higher VMAF and 31% lower agent response time. The design is coherent and the prototype/evaluation are substantial, but several load-bearing aspects need strengthening before the headline claims are fully supported.

Significance. If the results hold, HAFS would be a valuable practical contribution to an emerging workload. It identifies a real contention problem, uses app semantics to coordinate heterogeneous flows without infrastructure changes, and includes a credible prototype. Strengths include the KV-cache prefill/streaming pipeline, disjoint train/test data for the response-length predictor with robustness to 40% prediction error (Appendix D, Figure 21), and ablations for each coordinator. The main risks are not internal circularity but the unexamined RTT dependency of the control loop and the lack of statistical grounding for the headline numbers. The paper deserves a major revision rather than rejection.

major comments (3)
  1. [§4.1, §5.3, §8.1] The short sender-SFU RTT assumption in §4.1 is load-bearing. Equation (3) feeds receiver/RTCP feedback into a controller that makes one update per frame interval L ≈ 33 ms. The only testbed includes a fixed 30 ms wired delay between SFU and AP (§8.1); no experiment varies SFU distance or RTT. If feedback delay approaches L, Q_f is stale and the MI/AI-MD rules (Eqs. 4–6) can overshoot into deadline violation or over-drain, so the claimed 1.5×/31% gains may not generalize to peer-to-peer or distant-SFU deployments. Please add an RTT sweep (e.g., 10–150 ms) or an analytic bound on feedback delay for which the controller remains stable.
  2. [§8.2–§8.4, Table 1, Figures 13–21] The empirical claims rest entirely on point estimates. Table 1 reports single numbers, and Figures 13–18 show no error bars or confidence intervals and report no significance tests. Given the 5G trace's high variance (482.95 ± 228.49 Mbps), the reported differences could be within run-to-run noise. Please report statistics over repeated runs or trace segments, and ideally confidence intervals for the headline VMAF and response-time gains. This is needed to support the '1.5×/31%' summary.
  3. [§6.1, Eq. (7), Figure 28] Johnson's rule is optimal only if the network stage is a single non-preemptive machine. The implementation, however, schedules via SCTP stream priorities over a shared SCTP association (Figure 8, §7). The paper does not specify whether a high-priority stream is served to completion or whether chunks from lower-priority streams may interleave when the high-priority stream's cwnd/pacing credits are exhausted. If flows overlap in the network, the two-machine flow-shop model is not the right objective. Please specify the packet-level scheduling discipline and either justify the single-machine behavior or revise the model.
minor comments (4)
  1. [Abstract] The abstract says 'We design HFS' but the rest of the paper uses HAFS. Please make the acronym consistent.
  2. [Abstract and §8] The headline '1.5× higher video quality' does not specify which baseline and which metric normalization are used. Since VMAF is not a ratio scale, please state the exact comparison (e.g., HAFS vs FSE/CC) and the corresponding VMAF values.
  3. [Footnote 2 and §7] The code availability line is 'Code available at: TBD'. For a systems paper with a substantial prototype, the artifact should either be linked or a clear restriction stated.
  4. [§5.2] Equation (1) uses D_min over a 'recent window'. Please specify the window length and how D_min is updated after path/link changes, since this affects Q_f estimates.

Circularity Check

0 steps flagged

No significant circularity: the headline gains are measured outcomes backed by external citations and disclosed train/test splits, not reductions to the paper's own inputs.

full rationale

HAFS's two headline results (1.5× VMAF, 31% lower response time) are measured end-to-end against baselines (NC, CC, FSE, RAW) on a physical testbed and trace emulation; they are not derived from the paper's own equations. The rate controller in §5 is a closed feedback loop: Q_f (Eq. 3) is estimated from RTP one-way delays and SCTP RTTs, and Eqs. (4)-(6) set the next sending rate. This is a control-law design, not a prediction equal to its input by construction; no rate-control parameter is a renamed VMAF or response-time metric. The scheduling claim in §6 applies Johnson's rule [37] to estimated T_net (Eq. 8) and T_comp (Eq. 9). The only fitted quantity, the response-length predictor, is trained on contexts strictly disjoint from the 600 evaluation contexts (Appendix D), and §8.4 shows the scheduling gain survives injected prediction error up to 40% MAE, so the predictor is not load-bearing. The rate-control constants are hand-tuned (Appendix C), but the reported gains are observed outcomes rather than consequences of fitting to those outcomes. No author self-citation is used to justify a uniqueness claim or to import an ansatz; relevant prior work ([16], [35], [70], etc.) is external. The SFU-RTT premise and Appendix G fairness limitation are generality/robustness risks, not circular reductions.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The design rests on 7 hand-tuned parameters (rate control gains and windows) and 5 domain assumptions, the most fragile being the short-SFU-RTT premise and the frame-queuing signal accuracy. No new physical or computational entities are postulated; HAFS is a scheduling framework built from existing components.

free parameters (7)
  • MI gain γ_max = 0.08 (0.2 when idle)
    Empirical tuning of multiplicative increase in rate controller (Eq. 4, Appendix C).
  • AI gain γ_ai = 3
    Tuned additive-increase step (Eq. 5, Appendix C).
  • AI→MD threshold α = 0.8
    Fraction of deadline at which multiplicative decrease triggers; chosen by hand (Appendix C).
  • MD coefficient δ = 0.85
    Mild under-send on queue drain; tuned (Eq. 6, Appendix C).
  • MI decay floor parameter q* = 0.9
    Floor for γ_mi decay in Eq. 10 (Appendix C).
  • Bandwidth estimation window = 5 s
    Moving average window for T_net estimation in §6.1; chosen by hand.
  • Pacer headroom multiplier = 2
    cwnd = 2·rate·RTT; heuristic from BBR (§5.3).
axioms (5)
  • domain assumption Sender–SFU and SFU–receiver paths have short, stable RTTs (p95 < 50 ms), making app-layer rate control as precise as in-network scheduling.
    Invoked in §4.1 to justify the entire approach; supported only by citation [16] (Cloudflare blog), not by the paper's own measurements.
  • domain assumption Video frame-level queuing Q_f = Q_RTP + Q_SCTP is an accurate, controllable signal that tracks near-deadline congestion; frame deadline D_e2e=150 ms is the correct app objective.
    Core of §5.2–5.3; relies on RTCP one-way-delay and SCTP RTT samples measuring the same queue.
  • domain assumption Receiver LLM compute operates FIFO on a single GPU; the two-machine flow-shop model (Johnson's rule) is the right abstraction.
    Stated in §6.1 footnote; fails with multi-stream decode batching or multi-GPU.
  • domain assumption Single-step LLM decoding predicts response length well enough that ordering is preserved (MAE 18–36%).
    Relies on [35,70] and own LoRA training (Appendix D); paper shows robustness to injected 40% MAE.
  • domain assumption KVZip compression at 30% retention preserves ≥95%/≥85% normalized accuracy on QA/summarization.
    Measured in Appendix A; compression is from prior work [39], applied here.

pith-pipeline@v1.3.0-alltime-deepseek · 23722 in / 11424 out tokens · 110873 ms · 2026-08-01T04:20:35.238775+00:00 · methodology

0 comments
read the original abstract

AI agents are enabling a new paradigm of agent-augmented real-time communication (RTC), where humans focus on high-level collaboration, while agents autonomously retrieve, analyze, and generate information in real time to support their interactions. These apps enable new experiences across various domains: for example, when corporate employees co-author a legal document, their agents can discuss and draft on their behalf, sparing them the burden of manually reviewing each other's work. As existing cloud-based agents suffer from privacy risks and unscalable server costs, on-device agent-augmented RTC offers a promising alternative. However, this on-device paradigm introduces a new networking challenge: contention between concurrent traffic flows generated by humans (for live video streaming) and agents (for sending context files for analysis). We design HFS, a framework to ensure both high live video quality and low agent response latency in agent-augmented RTC apps. We achieve the goal through an app-guided multi-flow transport approach, where a unified app-layer orchestrator jointly controls the sending rates of live video and agent context flows based on their heterogeneous app requirements. Our prototype built atop WebRTC and llama.cpp demonstrates that HAFS outperforms baselines, achieving 1.5x higher video quality while reducing agent response time by 31%.

Figures

Figures reproduced from arXiv: 2607.22854 by Goodsol Lee, Haowen Xu, Jinglu Wang, Juheon Yi, Saewoong Bahk, Yan Lu.

Figure 1
Figure 1. Figure 1: On-device agent-augmented RTC scenario and its network contention problem. across various domains such as collaborative coding, media co-creation, and remote healthcare. Commercial RTC platforms already deploy early forms of such agents. For example, Microsoft Teams’ Facilitator sum￾marizes conversations and suggests follow-up questions [53]. However, these systems largely depend on cloud-based agents, rai… view at source ↗
Figure 3
Figure 3. Figure 3: LLM inference la￾tency breakdown. WiFi 5G WiFi 5G 0 50 100 Time (s) Jetson S25 Raw Text KV Cache [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human-agent flow contention. bandwidth, the concurrent transmission of multiple flows in￾duces two distinct types of contention: (i) human-agent flow contention (§3.2.2), where KV cache streaming degrades live video quality, and (ii) multi-agent flow contention (§3.2.3), where multiple KV caches streamed concurrently delay each other and lead to low compute resource utilization. 3.2.1 Background on RTC Fra… view at source ↗
Figure 6
Figure 6. Figure 6: Agent context flow in Zoom and Teams. 1 3 5 Concurrent Contexts 0 50 Response Time (s) Optimal WebRTC Compute Idle [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: HAFS system architecture (orange components: data plane, green components: control plane). serving. Although these estimates inherently contain predic￾tion errors, multi-agent flow scheduling remains robust be￾cause its optimality depends solely on relative ordering of decoding latencies among flows, rather than absolute values. 4.4 System Architecture [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Human-agent flow coordinator workflow. spans only a small fraction of the frame interval (e.g., 3 out of 33 ms). In the presence of cross-traffic or wireless channel fluctuations, frame-level queuing measured over such short period becomes a poor predictor over the full frame interval. Pudica [64] recently addressed this issue by intentionally pac￾ing the video frames at a constant rate, but this interfere… view at source ↗
Figure 11
Figure 11. Figure 11: Multi-agent flow coordinator workflow. SCTP stream priorities in a way that maximizes network￾compute pipelining at the receiver’s GPU (§6.1). We also incorporate techniques to improve robustness against network bandwidth fluctuations and response length prediction errors through dynamic priority update and opportunistic overlapped prefill methods (§6.2). 6.1 Pipelining-aware Flow Scheduling Problem formu… view at source ↗
Figure 12
Figure 12. Figure 12: HAFS testbed. metadata, and flow-priority scheduling decisions by tapping into these data channels’ state. Our SFU reuses the same build to relay RTP and SCTP, prioritizing the main speaker’s video stream following standard practice [45] and routing agent queries to designated senders over matching SCTP streams. AI agent. We run the AI agents using llama.cpp [21], a widely used C++ framework for on-device… view at source ↗
Figure 13
Figure 13. Figure 13: Real-world app performance on MacBook Pro. [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: Performance under varying network conditions. [PITH_FULL_IMAGE:figures/full_fig_p010_15.png] view at source ↗
Figure 17
Figure 17. Figure 17: HAFS with higher RTP requirements. 3 5 7 9 Context Flows 0 200 400 Response Time (s) RAW NC HAFS [PITH_FULL_IMAGE:figures/full_fig_p011_17.png] view at source ↗
Figure 19
Figure 19. Figure 19: Ablation study on HAFS module. HAFS-DA HAFS-SF HAFS-F 0 1 2 Stall Ratio (%) 0.1 2 0.1 Stall Ratio TP 150 200 TP (Mbps) 179 195 198 [PITH_FULL_IMAGE:figures/full_fig_p011_19.png] view at source ↗
Figure 21
Figure 21. Figure 21: Impact of LLM la￾tency prediction error. Copa HAFS Cubic 0 10 20 30Stall Ratio (%) 9 11 26 Stall Ratio TP 10 1 10 2 10 3 TP (Mbps) 96 384 454 [PITH_FULL_IMAGE:figures/full_fig_p012_21.png] view at source ↗
Figure 23
Figure 23. Figure 23: Time-to-first￾token on the S25 device. 1 2 3 4 5 6 7 8 910 # of Queries 10 2 10 3 Energy per Query (J) RAW HAFS [PITH_FULL_IMAGE:figures/full_fig_p012_23.png] view at source ↗
Figure 25
Figure 25. Figure 25: KVZip normalized accuracy on multi-hop QA [PITH_FULL_IMAGE:figures/full_fig_p018_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: KVZip normalized accuracy on meeting summa [PITH_FULL_IMAGE:figures/full_fig_p018_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Microsoft Teams during concurrent video and an [PITH_FULL_IMAGE:figures/full_fig_p018_27.png] view at source ↗
Figure 29
Figure 29. Figure 29: Fairness behavior of HAFS over 200 Mbps fixed bandwidth. just within the RTP deadline rather than contending for the queue—so that the latency-critical video flow is protected even at the cost of throughput for the background flow. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_29.png] view at source ↗
Figure 28
Figure 28. Figure 28: Multi-agent flow coordination timeline. 0 100 Time (s) 0 100 200 Throughput (Mbps) HAFS-1 HAFS-2 (a) HAFS vs. HAFS. 0 100 Time (s) 0 100 200 Throughput (Mbps) HAFS Cubic (b) HAFS vs. Cubic [PITH_FULL_IMAGE:figures/full_fig_p020_28.png] 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

75 extracted references · 5 linked inside Pith

  1. [1]

    Hedera: dynamic flow scheduling for data center networks

    Mohammad Al-Fares, Sivasankar Radhakrishnan, Barath Raghavan, Nelson Huang, Amin Vahdat, et al. Hedera: dynamic flow scheduling for data center networks. InNsdi, volume 10, pages 89–92. San Jose, USA, 2010

    2010

  2. [2]

    Tooth: Toward optimal balance of video {QoE} and redundancy cost by {Fine- Grained}{FEC} in cloud gaming streaming

    Congkai An, Huanhuan Zhang, Shibo Wang, Jingyang Kang, Anfu Zhou, Liang Liu, Huadong Ma, Zili Meng, Delei Ma, Yusheng Dong, et al. Tooth: Toward optimal balance of video {QoE} and redundancy cost by {Fine- Grained}{FEC} in cloud gaming streaming. In22nd USENIX Symposium on Networked Systems Design and Implementation (NSDI 25), pages 635–651, 2025

    2025

  3. [3]

    Starvation in end-to-end congestion control

    Venkat Arun, Mohammad Alizadeh, and Hari Balakr- ishnan. Starvation in end-to-end congestion control. In Proceedings of the ACM SIGCOMM 2022 Conference, pages 177–192, 2022

    2022

  4. [4]

    Copa: Practical delay-based congestion control for the internet

    Venkat Arun and Hari Balakrishnan. Copa: Practical delay-based congestion control for the internet. In15th USENIX Symposium on Networked Systems Design and Implementation, pages 329–342, 2018

    2018

  5. [5]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wen- bin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Longbench: A bilingual, multitask bench- mark for long context understanding.arXiv preprint arXiv:2308.14508, 2023

    Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. Longbench: A bilingual, multitask bench- mark for long context understanding.arXiv preprint arXiv:2308.14508, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Small language models are the future of agentic ai.arXiv preprint arXiv:2506.02153, 2025

    Peter Belcak, Greg Heinrich, Shizhe Diao, Yonggan Fu, Xin Dong, Saurav Muralidharan, Yingyan Celine Lin, and Pavlo Molchanov. Small language models are the future of agentic ai.arXiv preprint arXiv:2506.02153, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Bbr: Congestion-based congestion control.Communications of the ACM, 60(2):58–66, 2017

    Neal Cardwell, Yuchung Cheng, C Stephen Gunn, Soheil Hassas Yeganeh, and Van Jacobson. Bbr: Congestion-based congestion control.Communications of the ACM, 60(2):58–66, 2017

    2017

  9. [9]

    Analysis and design of the google congestion control for web real-time communication (webrtc)

    Gaetano Carlucci, Luca De Cicco, Stefan Holmer, and Saverio Mascolo. Analysis and design of the google congestion control for web real-time communication (webrtc). InProceedings of the 7th International Con- ference on Multimedia Systems, pages 1–12, 2016

    2016

  10. [10]

    Is large mtu beneficial to cellular core networks? InProceedings of the 7th Asia-Pacific Workshop on Networking, pages 67–73, 2023

    Youngmin Choi, Junghan Yoon, YoungGyoun Moon, and KyoungSoo Park. Is large mtu beneficial to cellular core networks? InProceedings of the 7th Asia-Pacific Workshop on Networking, pages 67–73, 2023

    2023

  11. [11]

    Agent Network Protocol (ANP)

    Cisco. Agent Network Protocol (ANP). https://github.com/agent-network-protocol/ AgentNetworkProtocol. Accessed: 23 Apr. 2026

    2026

  12. [12]

    End-to-end encryption for webex meetings and webex calling

    Cisco Webex Help Center. End-to-end encryption for webex meetings and webex calling. https://help. webex.com/article/5h5d8ab, 2025. Accessed: 2026- 04-23

    2025

  13. [13]

    Jetson orin nx developer kit

    NVIDIA Corporation. Jetson orin nx developer kit. https://developer.nvidia.com/downloads/ jetson-orin-nx-series-data-sheet . Accessed: 2026-04-23

    2026

  14. [14]

    Nvidia h100 tensor core gpu architecture

    NVIDIA Corporation. Nvidia h100 tensor core gpu architecture. Technical report, NVIDIA Corporation, August 2022

    2022

  15. [15]

    Converge: Qoe-driven multipath video conferencing over webrtc

    Sandesh Dhawaskar Sathyanarayana, Kyunghan Lee, Dirk Grunwald, and Sangtae Ha. Converge: Qoe-driven multipath video conferencing over webrtc. InProceed- ings of the ACM SIGCOMM 2023 Conference, pages 637–653, 2023

    2023

  16. [16]

    Cloud- flare calls: millions of cascading trees all the way down

    Renan Dincer, Rachel Chen, Felipe Astroza Araya, Kevin Kipp, and Kazi Najib. Cloud- flare calls: millions of cascading trees all the way down. https://blog.cloudflare.com/ cloudflare-calls-anycast-webrtc/ , Apr 04 2024. Cloudflare Blog

    2024

  17. [17]

    The llama 3 herd of models.arXiv e-prints, pages arXiv–2407, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv e-prints, pages arXiv–2407, 2024

    2024

  18. [18]

    Salsify:{Low- Latency} network video through tighter integration be- tween a video codec and a transport protocol

    Sadjad Fouladi, John Emmons, Emre Orbay, Catherine Wu, Riad S Wahby, and Keith Winstein. Salsify:{Low- Latency} network video through tighter integration be- tween a video codec and a transport protocol. In15th USENIX Symposium on Networked Systems Design and Implementation (NSDI 18), pages 267–282, 2018

    2018

  19. [19]

    {Cost-Efficient} large lan- guage model serving for multi-turn conversations with {CachedAttention}

    Bin Gao, Zhuomin He, Puru Sharma, Qingxuan Kang, Djordje Jevdjic, Junbo Deng, Xingkun Yang, Zhou Yu, and Pengfei Zuo. {Cost-Efficient} large lan- guage model serving for multi-turn conversations with {CachedAttention}. In2024 USENIX annual technical conference (USENIX ATC 24), pages 111–126, 2024

    2024

  20. [20]

    The complexity of flowshop and jobshop scheduling.Math- ematics of operations research, 1(2):117–129, 1976

    Michael R Garey, David S Johnson, and Ravi Sethi. The complexity of flowshop and jobshop scheduling.Math- ematics of operations research, 1(2):117–129, 1976. 14

    1976

  21. [21]

    llama.cpp

    ggml. llama.cpp. https://github.com/ggml-org/ llama.cpp. Accessed: 23 Apr. 2026

    2026

  22. [22]

    Agent2Agent (A2A) Protocol

    Google. Agent2Agent (A2A) Protocol. https: //github.com/a2aproject/A2A. Accessed: 23 Apr. 2026

    2026

  23. [23]

    Google meet

    Google. Google meet. https://meet.google.com/. Accessed: 2026-04-23

    2026

  24. [24]

    WebRTC official page

    Google. WebRTC official page. https://webrtc. org/, 2023. Accessed: 2025-12-06

    2023

  25. [25]

    Learn about call & meeting encryp- tion in google meet

    Google Meet Help. Learn about call & meeting encryp- tion in google meet. https://support.google.com/ meet/answer/12387251?hl=en, 2026. Accessed: 2026-04-23

    arXiv 2026

  26. [26]

    Performance evaluation and comparison of westwood+, new reno, and vegas tcp congestion control.ACM SIGCOMM Computer Communication Review, 34(2):25–38, 2004

    Luigi A Grieco and Saverio Mascolo. Performance evaluation and comparison of westwood+, new reno, and vegas tcp congestion control.ACM SIGCOMM Computer Communication Review, 34(2):25–38, 2004

    2004

  27. [27]

    Metastream: Live volumetric content capture, cre- ation, delivery, and rendering in real time

    Yongjie Guan, Xueyu Hou, Nan Wu, Bo Han, and Tao Han. Metastream: Live volumetric content capture, cre- ation, delivery, and rendering in real time. InProceed- ings of the 29th annual international conference on mo- bile computing and networking, pages 1–15, 2023

    2023

  28. [28]

    REALM: Retrieval-augmented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. REALM: Retrieval-augmented language model pre-training. InProceedings of the 37th International Conference on Machine Learning (ICML), volume 119 ofProceedings of Machine Learning Re- search, pages 3929–3938. PMLR, 2020

    2020

  29. [29]

    Cubic: a new tcp-friendly high-speed tcp variant.ACM SIGOPS operating systems review, 42(5):64–74, 2008

    Sangtae Ha, Injong Rhee, and Lisong Xu. Cubic: a new tcp-friendly high-speed tcp variant.ACM SIGOPS operating systems review, 42(5):64–74, 2008

    2008

  30. [30]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    2022

  31. [31]

    Ace: Sending burstiness control for high-quality real-time communication

    Xiangjie Huang, Jiayang Xu, Haiping Wang, Hebin Yu, Sandesh Dhawaskar Sathyanarayana, Shu Shi, and Zili Meng. Ace: Sending burstiness control for high-quality real-time communication. InProceedings of the ACM SIGCOMM 2025 Conference, pages 1182–1198, 2025

    2025

  32. [32]

    Agent Communication Protocol (ACP).https:// github.com/i-am-bee/acp

    IBM. Agent Communication Protocol (ACP).https:// github.com/i-am-bee/acp. Accessed: 23 Apr. 2026

    2026

  33. [33]

    Real- life implementation and evaluation of coupled conges- tion control for webrtc media and data flows.Ieee Ac- cess, 10:95046–95066, 2022

    Safiqul Islam, Michael Welzl, and Tobias Fladby. Real- life implementation and evaluation of coupled conges- tion control for webrtc media and data flows.Ieee Ac- cess, 10:95046–95066, 2022

    2022

  34. [34]

    QUIC: A UDP- based multiplexed and secure transport

    Jana Iyengar and Martin Thomson. QUIC: A UDP- based multiplexed and secure transport. RFC 9000, May 2021

    2021

  35. [35]

    s3: Increasing gpu utilization during generative inference for higher throughput.Advances in Neural In- formation Processing Systems, 36:18015–18027, 2023

    Yunho Jin, Chun-Feng Wu, David Brooks, and Gu-Yeon Wei. s3: Increasing gpu utilization during generative inference for higher throughput.Advances in Neural In- formation Processing Systems, 36:18015–18027, 2023

    2023

  36. [36]

    Bridge cascading with geo-location is back

    Jitsi. Bridge cascading with geo-location is back. https://jitsi.org/blog/ bridge-cascading-is-back/ . Accessed: 23 Apr. 2026

    2026

  37. [37]

    Optimal two-and three-stage production schedules with setup times included.Naval research logistics quarterly, 1(1):61–68, 1954

    Selmer Martin Johnson. Optimal two-and three-stage production schedules with setup times included.Naval research logistics quarterly, 1(1):61–68, 1954

    1954

  38. [38]

    Fezeu, Claudio Fiandrino, Eman Ra- madan, Jason Carpenter, Lilian Coelho De Freitas, Faaiq Bilal, Wei Ye, Joerg Widmer, Feng Qian, and Zhi-Li Zhang

    Rostand A K. Fezeu, Claudio Fiandrino, Eman Ra- madan, Jason Carpenter, Lilian Coelho De Freitas, Faaiq Bilal, Wei Ye, Joerg Widmer, Feng Qian, and Zhi-Li Zhang. Unveiling the 5g mid-band landscape: From net- work deployment to performance and application qoe. InProceedings of the ACM SIGCOMM 2024 Confer- ence, pages 358–372, 2024

    2024

  39. [39]

    Kvzip: Query- agnostic kv cache compression with context reconstruc- tion.arXiv preprint arXiv:2505.23416, 2025

    Jang-Hyun Kim, Jinuk Kim, Sangwoo Kwon, Jae W Lee, Sangdoo Yun, and Hyun Oh Song. Kvzip: Query- agnostic kv cache compression with context reconstruc- tion.arXiv preprint arXiv:2505.23416, 2025

    arXiv 2025

  40. [40]

    The cost of dynamic reasoning: Demystifying ai agents and test-time scaling from an ai infrastructure perspective.arXiv preprint arXiv:2506.04301, 2025

    Jiin Kim, Byeongjun Shin, Jinha Chung, and Minsoo Rhu. The cost of dynamic reasoning: Demystifying ai agents and test-time scaling from an ai infrastructure perspective.arXiv preprint arXiv:2506.04301, 2025

    arXiv 2025

  41. [41]

    Deltastream: 2d-inferred delta encoding for live volumetric video streaming

    Hojeong Lee, Yu Hong Kim, Sangwoo Ryu, James Won- Ki Hong, Sangtae Ha, and Seyeon Kim. Deltastream: 2d-inferred delta encoding for live volumetric video streaming. InProceedings of the 23rd Annual Interna- tional Conference on Mobile Systems, Applications and Services, pages 361–373, 2025

    2025

  42. [42]

    Demystifying commercial video con- ferencing applications

    Insoo Lee, Jinsung Lee, Kyunghan Lee, Dirk Grunwald, and Sangtae Ha. Demystifying commercial video con- ferencing applications. InProceedings of the 29th ACM international conference on multimedia, pages 3583– 3591, 2021

    2021

  43. [43]

    Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020. 15

    2020

  44. [44]

    VMAF: The journey continues.Netflix Technology Blog, 25(1), 2018

    Zhi Li, Christos Bampis, Julie Novak, Anne Aaron, Kyle Swanson, Anush Moorthy, and JD Cock. VMAF: The journey continues.Netflix Technology Blog, 25(1), 2018

    2018

  45. [45]

    Gso-simulcast: global stream orchestration in simulcast video confer- encing systems

    Xianshang Lin, Yunfei Ma, Junshao Zhang, Yao Cui, Jing Li, Shi Bai, Ziyue Zhang, Dennis Cai, Hongqiang Harry Liu, and Ming Zhang. Gso-simulcast: global stream orchestration in simulcast video confer- encing systems. InProceedings of the ACM SIGCOMM 2022 Conference, pages 826–839, 2022

    2022

  46. [46]

    Pyrrha:{Congestion-Root-Based} flow control to elim- inate {Head-of-Line} blocking in datacenter

    Kexin Liu, Zhaochen Zhang, Chang Liu, Yizhi Wang, Vamsi Addanki, Stefan Schmid, Qingyue Wang, Wei Chen, Xiaoliang Wang, Jiaqi Zheng, et al. Pyrrha:{Congestion-Root-Based} flow control to elim- inate {Head-of-Line} blocking in datacenter. In22nd USENIX Symposium on Networked Systems Design and Implementation (NSDI 25), pages 379–405, 2025

    2025

  47. [47]

    Cachegen: Kv cache compression and streaming for fast large lan- guage model serving

    Yuhan Liu, Hanchen Li, Yihua Cheng, Siddhant Ray, Yuyang Huang, Qizheng Zhang, Kuntai Du, Jiayi Yao, Shan Lu, Ganesh Ananthanarayanan, et al. Cachegen: Kv cache compression and streaming for fast large lan- guage model serving. InProceedings of the ACM SIG- COMM 2024 Conference, pages 38–56, 2024

    2024

  48. [48]

    Autellix: An efficient serving engine for llm agents as general pro- grams.arXiv preprint arXiv:2502.13965, 2025

    Michael Luo, Xiaoxiang Shi, Colin Cai, Tianjun Zhang, Justin Wong, Yichuan Wang, Chi Wang, Yanping Huang, Zhifeng Chen, Joseph E Gonzalez, et al. Autellix: An efficient serving engine for llm agents as general pro- grams.arXiv preprint arXiv:2502.13965, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Achiev- ing consistent low latency for wireless real-time commu- nications with the shortest control loop

    Zili Meng, Yaning Guo, Chen Sun, Bo Wang, Justine Sherry, Hongqiang Harry Liu, and Mingwei Xu. Achiev- ing consistent low latency for wireless real-time commu- nications with the shortest control loop. InProceedings of the ACM SIGCOMM 2022 Conference, pages 193– 206, 2022

    2022

  50. [50]

    Hairpin: Rethinking packet loss recov- ery in edge-based interactive video streaming

    Zili Meng, Xiao Kong, Jing Chen, Bo Wang, Mingwei Xu, Rui Han, Honghao Liu, Venkat Arun, Hongxin Hu, and Xue Wei. Hairpin: Rethinking packet loss recov- ery in edge-based interactive video streaming. In21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 907–926, 2024

    2024

  51. [51]

    Enabling passive mea- surement of zoom performance in production networks

    Oliver Michel, Satadal Sengupta, Hyojoon Kim, Ravi Netravali, and Jennifer Rexford. Enabling passive mea- surement of zoom performance in production networks. InProceedings of the 22nd ACM internet measurement conference, pages 244–260, 2022

    2022

  52. [52]

    Scalable video confer- encing using sdn principles

    Oliver Michel, Satadal Sengupta, Hyojoon Kim, Ravi Netravali, and Jennifer Rexford. Scalable video confer- encing using sdn principles. InProceedings of the ACM SIGCOMM 2025 Conference, pages 1213–1231, 2025

    2025

  53. [53]

    Facilitator in microsoft teams meetings

    Microsoft. Facilitator in microsoft teams meetings. https://support.microsoft.com/en-us/office/ facilitator-in-microsoft-teams-meetings-37657f91 . Accessed: 2026-04-23

    2026

  54. [54]

    Microsoft teams

    Microsoft. Microsoft teams. Accessed: 2025-12-06

    2025

  55. [55]

    Microsoft Teams call flows

    Microsoft. Microsoft Teams call flows. https: //learn.microsoft.com/en-us/microsoftteams/ microsoft-teams-online-call-flows . Accessed: 23 Apr. 2026

    2026

  56. [56]

    Require end-to-end encryp- tion for sensitive teams meetings

    Microsoft Learn. Require end-to-end encryp- tion for sensitive teams meetings. https: //learn.microsoft.com/en-us/microsoftteams/ end-to-end-encrypted-meetings , 2025. Accessed: 2026-04-23

    2025

  57. [57]

    Lmcache — nvidia dynamo documenta- tion

    NVIDIA. Lmcache — nvidia dynamo documenta- tion. https://docs.nvidia.com/dynamo/latest/ integrations/lm-cache, 2026. Accessed: 2026-04- 23

    2026

  58. [58]

    Code llama: Open foundation models for code.arXiv preprint arXiv:2308.12950, 2023

    Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code.arXiv preprint arXiv:2308.12950, 2023

    Pith/arXiv arXiv 2023

  59. [59]

    Powerinfer: Fast large language model serving with a consumer-grade gpu

    Yixin Song, Zeyu Mi, Haotong Xie, and Haibo Chen. Powerinfer: Fast large language model serving with a consumer-grade gpu. InProceedings of the ACM SIGOPS 30th Symposium on Operating Systems Princi- ples, pages 590–606, 2024

    2024

  60. [60]

    Sctp: new trans- port protocol for tcp/ip.IEEE Internet Computing, 5(6):64–69, 2002

    Randall Stewart and Christopher Metz. Sctp: new trans- port protocol for tcp/ip.IEEE Internet Computing, 5(6):64–69, 2002

    2002

  61. [61]

    probe_controller.cc

    The WebRTC Project Authors. probe_controller.cc. https://chromium.googlesource.com/ external/webrtc/+/lkgr/modules/congestion_ controller/goog_cc/probe_controller.cc, 2026. accessed 2026-04-23

    2026

  62. [62]

    musique: Multihop questions via single-hop question composition.Transactions of the Association for Computational Linguistics, 10:539– 554, 2022

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. musique: Multihop questions via single-hop question composition.Transactions of the Association for Computational Linguistics, 10:539– 554, 2022

    2022

  63. [63]

    Automatic prefix caching — design document

    vLLM Project Contributors. Automatic prefix caching — design document. https://docs.vllm.ai/en/ latest/design/prefix_caching/, 2026. Accessed: 2026-04-23. 16

    2026

  64. [64]

    Pudica: Toward {Near- Zero} queuing delay in congestion control for cloud gaming

    Shibo Wang, Shusen Yang, Xiao Kong, Chenglei Wu, Longwei Jiang, Chenren Xu, Cong Zhao, Xuesong Yang, Jianjun Xiao, Xin Liu, et al. Pudica: Toward {Near- Zero} queuing delay in congestion control for cloud gaming. In21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 113–129, 2024

    2024

  65. [65]

    Real-time transport protocol,

    Wikipedia contributors. Real-time transport protocol,

  66. [66]

    Macbook pro (apple silicon) — wikipedia, the free encyclopedia

    Wikipedia contributors. Macbook pro (apple silicon) — wikipedia, the free encyclopedia. https://en.wikipedia.org/wiki/MacBook_ Pro_(Apple_silicon), 2026. Accessed: 2026-04-23

    2026

  67. [67]

    Token bucket — Wikipedia, the free encyclopedia

    Wikipedia contributors. Token bucket — Wikipedia, the free encyclopedia. https://en.wikipedia.org/ wiki/Token_bucket, 2026. Accessed: 2026-04-23

    2026

  68. [68]

    Mobile access bandwidth in practice: Measurement, analysis, and implications

    Xinlei Yang, Hao Lin, Zhenhua Li, Feng Qian, Xingyao Li, Zhiming He, Xudong Wu, Xianlong Wang, Yunhao Liu, Zhi Liao, et al. Mobile access bandwidth in practice: Measurement, analysis, and implications. InProceed- ings of the ACM SIGCOMM 2022 Conference, pages 114–128, 2022

    2022

  69. [69]

    Cacheblend: Fast large language model serving for rag with cached knowledge fusion

    Jiayi Yao, Hanchen Li, Yuhan Liu, Siddhant Ray, Yi- hua Cheng, Qizheng Zhang, Kuntai Du, Shan Lu, and Junchen Jiang. Cacheblend: Fast large language model serving for rag with cached knowledge fusion. InPro- ceedings of the Twentieth European Conference on Com- puter Systems, pages 94–109, 2025

    2025

  70. [70]

    Response length perception and sequence scheduling: An llm-empowered llm infer- ence pipeline.Advances in Neural Information Process- ing Systems, 36:65517–65530, 2023

    Zangwei Zheng, Xiaozhe Ren, Fuzhao Xue, Yang Luo, Xin Jiang, and Yang You. Response length perception and sequence scheduling: An llm-empowered llm infer- ence pipeline.Advances in Neural Information Process- ing Systems, 36:65517–65530, 2023

    2023

  71. [71]

    Qmsum: A new bench- mark for query-based multi-domain meeting summa- rization

    Ming Zhong, Da Yin, Tao Yu, Ahmad Zaidi, Mutethia Mutuma, Rahul Jha, Ahmed Hassan, Asli Celikyilmaz, Yang Liu, Xipeng Qiu, et al. Qmsum: A new bench- mark for query-based multi-domain meeting summa- rization. InProceedings of the 2021 Conference of the North American Chapter of the Association for Com- putational Linguistics: Human Language Technologies, ...

    2021

  72. [72]

    {DistServe}: Disaggregating prefill and decoding for goodput-optimized large language model serving

    Yinmin Zhong, Shengyu Liu, Junda Chen, Jianbo Hu, Yibo Zhu, Xuanzhe Liu, Xin Jin, and Hao Zhang. {DistServe}: Disaggregating prefill and decoding for goodput-optimized large language model serving. In 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24), pages 193–210, 2024

    2024

  73. [73]

    Zoom online meeting

    Zoom. Zoom online meeting. https://zoom.us/. Accessed: 2026-04-23

    2026

  74. [74]

    ### Instruction

    Zoom Support. Enabling end-to-end en- cryption (e2ee) for zoom meetings. https: //support.zoom.com/hc/en/article?id=zm_ kb&sysparm_article=KB0075502, 2026. Accessed: 2026-04-23. 17 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 KV cache ratio 0.0 0.2 0.4 0.6 0.8 1.0 Normalized Accuracy Qwen2.5-3B Qwen2.5-7B Llama3.1-8B Figure 25: KVZip normalized accuracy on multi-h...

    2026

  75. [2025]

    Last edited: 23 October 2025; Accessed: 2026- 04-23

    2025