技能 人工智能 ML系统论文选择指南

ML系统论文选择指南

v20260724
mlsys-topic-selection
该指南旨在帮助研究人员判断其项目是否属于机器学习与计算机系统协同设计范畴。它详细介绍了“共设计测试”和路由表,指导用户判断项目最适合投递至 MLSys,还是更适合其他通用系统或纯ML会议。
获取技能
311 次下载
概览

MLSys Topic Selection

Use this before writing begins. MLSys exists for work where machine learning and computer
systems constrain each other — the CFP spans systems for ML (training, inference,
serving, compilers, runtimes, specialized hardware, hardware-efficient methods,
benchmarks and tooling) and ML for systems (learned scheduling, LLM-driven hardware
design and system optimization), plus adjacent lanes like federated learning, compound AI
and agent systems, privacy/security for ML, and data preparation (2026 topics list,
verified 2026-07-08). Fit failures here are usually routing failures: strong work aimed
at the wrong reviewer pool.

The co-design test

Ask two questions; MLSys wants both answers to be yes.

  1. Does an ML property (model structure, training dynamics, tolerance to approximation,

request patterns of ML workloads) shape the system design?

  1. Does a systems property (memory hierarchy, interconnect, scheduling, cost) shape the

ML method or its evaluation?

One yes is a re-routing signal: pure systems novelty exercised on an ML workload belongs
at a general systems venue; pure ML novelty measured for speed belongs at an ML venue.

Routing table

Project's center of gravityBetter venueWhy
New OS/storage/scheduling abstraction, ML is one workload among severalOSDI / SOSP / ATC / EuroSysTheir reviewers reward general-purpose systems depth
Network protocol or datacenter fabric for training trafficNSDI / SIGCOMMNetworking-first reviewer pool
Microarchitecture or accelerator designASPLOS / ISCA / MICROHardware evaluation norms (simulators, RTL) differ
New learning algorithm; efficiency is one experimentNeurIPS / ICML / ICLRAlgorithmic novelty is the contribution
HPC-scale training on supercomputersSC / HPDCHPC evaluation culture
Mechanism where ML and system constraints genuinely interlockMLSysThis is the house genre
Production ML system's design + deep benchmarks, limited noveltyMLSys industrial track (2026+)Novelty explicitly not required there
Early idea, workshop-scale evidenceAn MLSys-adjacent workshop firstExpanded workshop papers can return with chair approval

Signals your project is MLSys-shaped

  • The abstract's key number is a systems metric (goodput, p99, memory, $/token) achieved

because of an ML-aware design decision — not incidentally.

  • The natural reviewer nightmare is "is this workload representative?" rather than "is

this theorem tight?" or "is this abstraction general?"

  • The related work splits across systems venues and ML venues — no single community owns

the problem (see mlsys-related-work).

  • You can promise an artifact whose evaluation is meaningful on hardware a committee can

access — badge culture is part of this venue's identity.

  • The contribution survives this compression: *"We observed [ML property]; therefore we

built [mechanism]; which yields [measured system payoff] at [stated cost]."*

Track choice within MLSys (2026 structure)

  • Research track: previously unpublished ideas, full double-blind, novelty judged

against both literatures.

  • Industrial track: built-at-scale systems, design methodology and detailed

benchmarks required, novelty not required, company identity may remain visible.
Choosing it for a paper whose real asset is a clever new mechanism undersells the
mechanism; choosing research track for a deployment report gets it rejected for
novelty. Route by the paper's strongest honest asset.

Sharpening moves before committing

Fill these four lines; blanks are routing information.
  ML property exploited:      ______________________
  System constraint engaged:  ______________________
  Named mechanism:            ______________________
  Headline systems metric:    ______  on workload ______  vs baseline ______
If "named mechanism" is blank            -> industrial track or a workshop.
If "ML property" is blank                -> general systems venue.
If "system constraint" is blank          -> ML venue.
If the metric line is blank              -> not ready for any venue; measure first.
  • Check the current cycle's topics list before finalizing — the venue's emphasis has

drifted fast (the 2026 list added compound AI/agent systems and LLM-driven hardware
design; assume continued drift).

  • Consider the calendar honestly: MLSys is annual (October deadline in the 2026 cycle).

If the project peaks in March, a systems venue with a nearer deadline may serve the
work better than a seven-month hold.

Vignette: three borderline projects, routed

  • A learned cost model inside a query optimizer, evaluated on analytics workloads.

ML-for-systems is in scope, but the reviewer pool that can judge query optimization
depth sits at database venues. Route to MLSys only if the contribution is the
learning-systems machinery (training-data collection, drift handling, inference
latency budget); route to a DB venue if it is optimizer quality.

  • An INT4 quantization method with strong accuracy and a fast custom kernel. If the

paper's spine is the accuracy-preservation insight, ML venues fit; it becomes
MLSys-shaped when the kernel/runtime co-design carries equal weight and the
evaluation leads with latency-quality tradeoffs on real serving stacks — the AWQ
pattern in ../../resources/exemplars/library.md.

  • **A cluster scheduler for mixed training/inference jobs, evaluated in simulation

only.** Topic fits; evidence does not — simulation-only evaluation of a systems
claim is a recognized weak pattern here. Either add a real-cluster deployment at
modest scale or target a venue where simulation is the norm.

Anti-patterns that fail the fit test

  • "We ported X to GPUs and it is faster" — engineering without an insight that travels.
  • "Our new attention variant, with a latency table" — ML paper wearing a systems hat.
  • "A survey of serving systems" — MLSys publishes measured contributions, not surveys.
  • "Our company's stack, described" without benchmarks — even the industrial track

demands measurement depth.

  • "Faster than a two-year-old baseline" — in a field whose systems improve monthly,

a stale comparison fails the fit test before the novelty question is even reached.

  • "We propose a benchmark" with no run rules or quality targets — the venue that

published MLPerf holds benchmark papers to benchmark-committee standards.

Output format

[Fit] strong MLSys / MLSys-possible / route elsewhere
[Co-design test] ML->system: yes/no ; system->ML: yes/no
[Track] research / industrial / workshop-first
[Best alternative venue] <venue + reason>
[Contribution compression] observed -> built -> measured -> cost
[Next action] <measurement, mechanism naming, track decision, or reroute>
信息
分类 人工智能
名称 mlsys-topic-selection
版本 v20260724
大小 7.09KB
更新时间 2026-07-28
语言