Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →brycewang-stanford/colm-related-work
Use when positioning a COLM paper inside the fastest-moving literature in ML — triaging arXiv-heavy citations, handling concurrent work fairly, citing model and dataset artifacts correctly, distinguishing COLM's three-edition archive from adjacent venues' archives, and keeping self-citation double-blind safe.
Passed 24 of 25 test cases with the skill, and 16 of 25 without it.
and 8 more in the full report
9 fixed15 passed either way0 still failing1 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 9 test cases the skill fixed. Same model, same prompt, with and without the skill.
I am finalizing the related work section for my COLM paper on speculatively decoding long-context LLMs. Could you review my literature review text and draft citations to produce an audit report? I need a breakdown that checks whether I've covered key literature lanes, identified COLM archive neighbors, flagged any improper publication status formatting or laundered arXiv entries, checked concurrent work handling, audited for self-citation/anonymity leaks, and checked overall reference resolution integrity.
| Checked | Without skill | With skill |
|---|---|---|
| Includes the exact header text [Lane coverage] | Missed | Passed |
| Includes the exact header text [COLM-archive neighbors] | Missed | Passed |
| Includes the exact header text [arXiv honesty] | Missed | Passed |
| Includes the exact header text [Concurrent work] | Missed | Passed |
The instructions your model reads when the skill loads.
The LM literature moves faster than any archival process: the work you must position against is substantially on arXiv, some of it appeared after your experiments finished, and some "well-known results" are blog posts. A COLM related-work section is therefore an exercise in versioned honesty — saying precisely what existed, where, and when, without pretending the field is tidier than it is.
Cover each lane that touches your claim; reviewers at this venue are typically active in two or three of them and will notice a missing lane immediately:
| Lane | What to position against | Typical miss |
|---|---|---|
| Method lineage | The 2-4 papers your technique descends from | Citing a survey instead of the source |
| Phenomenon reports | Prior observations of the effect you study, incl. analysis papers | Ignoring negative or contradictory reports |
| Evaluation infrastructure | The benchmarks/harnesses you use or compete with | Treating a benchmark as neutral ground rather than citable, criticizable work |
| Model artifacts | The model families your claims are scoped to | Citing a company blog when a technical report exists (or vice versa — cite what actually exists) |
| Adjacent-venue archive | Related results at ICLR/NeurIPS/ACL/EMNLP and in COLM 2024-2025 | Assuming COLM reviewers only read COLM |
Checking the COLM 2024 and 2025 accepted lists for near neighbors is cheap and high-yield: citing the venue's own archive signals you know where you are submitting, and the archive is only two editions deep (checked 2026-07-08).
State venue status accurately at citation time: published (venue, year), arXiv-only (cite the arXiv ID; do not launder it into a "publication"), or since-published (prefer the archival version, keep the arXiv date if chronology matters to your priority story). For results that exist only as blog posts or model cards, cite them as what they are, with access dates — at a venue that studies industry models, gray literature is unavoidable; disguising it is the sin.
With a March 31 deadline and a field publishing daily, concurrency is guaranteed. The credible pattern:
test cases passed, vs 16 / 25 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.