MLSys Supplementary

Use this when splitting a Conference on Machine Learning and Systems submission between body and appendix. The 2026-cycle mechanics (verified 2026-07-08): appendices are unlimited in length, uploaded as a separate file from the main paper, due at the same time — and reviewers are explicitly not required to read them. That last clause is the design constraint everything below follows from.

The governing rule

The 10-page body must win the paper's acceptance alone. The appendix exists to survive audits — a skeptical reviewer checking one specific configuration — and to serve future readers and artifact evaluators. Anything that must be believed for acceptance goes in the body; anything that must be checkable can go behind it.

Placement decision table

Structuring the appendix file

Because it is a separate upload, the appendix must be navigable without the body open:

• Start with a half-page map: appendix section → the body section and claim it supports.
• Number appendix sections A, B, C... and cross-reference each at least once from the body ("full grid in App. B"); un-pointered appendix content is effectively invisible.
• Restate the experimental setup header (hardware, workloads) at the top rather than assuming the reviewer has the body's Section 5 in memory.
• Keep figures self-captioning: every axis, unit, and trial count readable standalone.

appendix.pdf
├── A. Claims map (body claim -> appendix evidence)
├── B. Full system + baseline configurations
│     B.1 our system   B.2 baseline-1 (tuning grid)   B.3 baseline-2
├── C. Workload characterization (trace statistics, generation scripts described)
├── D. Extended results (per-workload, per-hardware, per-precision)
├── E. Sensitivity and additional ablations
├── F. Anonymized artifact description (layout, run commands, hardware needs)
└── G. Proofs / analytical model derivations (if any)

Anonymity in supplementary material (research track)

• The appendix is part of the double-blind submission: no author names, affiliations, cluster hostnames, internal project codenames, or company-identifying trace fields.
• Scrub trace data specifically — production traces leak identity through service names, datacenter labels, and timestamp patterns. Rename fields and state that you did.
• Code pointers must go to anonymized mirrors; the real repository appears only at camera-ready and artifact evaluation.
• Industrial-track supplements follow that track's looser rule (company identity may remain; author names may not) — do not mix the two tracks' standards.

Worked split: an inference-compiler submission

A fictional paper contributes a compiler pass that fuses attention variants, claiming 1.5-1.9x kernel speedups and 1.2x end-to-end serving gains across three model families.

• Body keeps: the fusion legality insight and the pass algorithm (compressed to one figure plus one column), the end-to-end serving comparison under a stated latency constraint, the kernel-level speedup summary across families, one ablation isolating the fusion from unrelated codegen wins, and the non-win (no gain on short sequences, stated with the mechanism reason).
• Appendix takes: per-operator speedup tables for all 40 fused patterns, the full autotuning search configuration, correctness-test methodology against reference kernels, results on a second GPU generation, and the anonymized artifact description.
• The one judgment call: the correctness methodology gets a two-sentence summary in the body — a compiler paper that never says how it validated correctness invites the objection no appendix can pre-empt, because the appendix may never be opened.

The test of the split: hand a colleague only the 10-page body and ask them to argue for acceptance; every point where they reach for the appendix is a misplacement.

Common failure modes

• The outsourced paper: body asserts, appendix proves, reviewer reads only the body and scores the assertions as unsupported. Pull the single strongest piece of evidence for each major claim forward.
• The stale appendix: body numbers get regenerated at the deadline, appendix tables do not, and a careful reviewer finds the mismatch. Build both from the same logs in the same script run.
• The dumping ground: forty pages of unreferenced plots signal an unfiltered experiment log, not scholarship. Every appendix element should answer a question a reviewer might actually ask.
• The forgotten upload: the appendix is a separate file with the same deadline — assign its upload explicitly; it is not attached to the paper PDF automatically.

Interaction with later stages

The appendix outlives review. The configuration matrices and workload characterization you file now become the skeleton of the Artifact Appendix at evaluation time (mlsys-artifact-evaluation), and the camera-ready version de-anonymizes rather than rewrites it — so write appendix sections as if a stranger will execute them in six months, because at this venue one probably will. Conversely, anything you promise in the author response ("full sweep in the appendix") must land there by camera-ready; keep a promises ledger.

Cycle-volatility warning

The separate-upload mechanic, unlimited length, and same-deadline rule are 2026-cycle facts; earlier cycles differed and the next may too. Confirm the current Call for Research Papers wording before finalizing the split.

Output format

[Split status] body-sufficient / body-dependent-on-appendix (fix) / not ready
[Claims map] <major claim -> body evidence -> appendix backup>
[Appendix structure] <sections present + cross-reference check>
[Anonymity sweep] <traces/configs/code pointers cleaned?>
[Consistency] <body and appendix generated from same logs? y/n>
[Upload owner] <person responsible for the separate file>