技能 人工智能 设备端结果可复现性证据

设备端结果可复现性证据

v20260724
mobisys-reproducibility
本指南指导研究人员如何为设备端性能测量(如能耗、延迟、准确率)建立严谨的溯源证据。它强调了在学术论文中,必须记录所有必要的上下文信息(包括设备型号、操作系统、框架版本、电源和热力学状态),确保实验结果在不同硬件或环境中是可验证和可复现的。
获取技能
427 次下载
概览

MobiSys Reproducibility

Use this before submission and again before the artifact-evaluation deadline. On-device results are fragile across hardware and framework versions, so reproducibility at MobiSys is mostly about provenance: recording enough of the device and measurement context that someone on a different phone can rebuild the result.

Evidence map

  • Map each latency, energy, throughput, and accuracy claim to a verifiable location in the paper, appendix, or artifact package.
  • For each on-device result, record device model, SoC, OS build, framework/runtime version, model checkpoint hash, power source, ambient temperature, and thermal state at run start.
  • Record the power-instrument setup and the energy boundary so an energy figure can be re-derived rather than trusted.
  • For stochastic results, report seeds, repeated runs, and the spread; a single number from a single phone is not reproducible evidence.
  • Explain missing data, firmware, or hardware honestly, and describe how a reader could reproduce the result in principle on a different device.
  • Keep the artifact consistent with the manuscript; a number in the PDF that the artifact cannot regenerate is a review-risk multiplier and an AEC failure.

What survives a different device

Provenance item Turnkey answer Common failure caught
Device + SoC + OS build Named per experiment, in a table "on a mobile device" with no model
Framework / runtime version Pinned with hashes Result drifts silently across a minor version
Energy instrument + boundary Instrument, sampling rate, envelope stated "efficient" with no measured joules
Thermal state Ambient + steady-state trace 30-second benchmark hides throttling
Model / cache version Checkpoint hash, cache size Accuracy irreproducible from a different weight file

Marking a device result "reproducible" without its provenance is a recognizable MobiSys red flag, because the AEC and reviewers cross-check the artifact against the PDF and read a gap as carelessness about the rest of the paper.

Vignette: an on-device inference paper

Consider a runtime that keeps CNN latency stable under load, validated on four phones. Its reproducibility spine: the four device models and OS builds, the runtime version and model checkpoint hashes, the power-monitor model and sampling rate, the ambient temperature and the steady-state thermal traces, the workload trace, the seeds and run counts — plus one honest sentence about the device it could not test and why the result may differ there.

Degrees of reproducibility

  • Turnkey: one script regenerates each figure from logged runs on the target device.
  • Scripted: scripts exist but need the specific hardware or manual device setup.
  • Descriptive: prose detailed enough that a competent reader could rebuild the pipeline on a comparable device.

For MobiSys, the on-device measurement path should be as turnkey as the hardware allows, and a hardware-optional downscaled or emulator path should let an evaluator without your exact device reach at least Functional (mobisys-artifact-evaluation). State the achieved level honestly rather than overpromising turnkey behavior that fails on a different phone.

Legal and privacy provenance

  • Decide early what traces, firmware, and datasets can ship; on-device data often carries user, sensor, or vendor-firmware restrictions.
  • Strip identifying content from screenshots, logs, and demo media before release.
  • If data cannot ship, provide a synthetic or public substitute and name which results it cannot reproduce.

Output format

[Claim inventory] <claim -> evidence location>
[Provenance status] complete / partial / missing (device/OS/runtime/energy/thermal)
[Reproducibility gaps] <device pinning / seeds / energy boundary / model version>
[Paper fixes] <must appear in main PDF>
[Artifact fixes] <appendix or artifact additions, incl. hardware-optional path>
信息
Category 人工智能
Name mobisys-reproducibility
版本 v20260724
大小 4.33KB
更新时间 2026-07-28
语言