Paper Radar

paper-radar — big tech AI paper radar

Stateless: one run = one window = one report. No paper library is maintained. arXiv is the only source of truth, and re-running the same window necessarily produces the same result — so there is no reason to keep a second copy. Keeping one only adds state that drifts silently and lives outside version control. The cost is re-scraping the whole window every run (~1 min for 3 days, ~3 min for 14 days); what you buy is reproducibility.

Read this first, or you will reach wrong conclusions

arXiv is the only source organized by posting date. Every vendor's own publications page is organized by publication year / venue, which structurally cannot answer "what shipped in the last two weeks" — a data-model difference, not a technical one. Measured:

• Google Research: detail pages carry a year only, and the listing is not chronological; 85 titles across the first 8 pages yielded 0 inside a two-week window, and the most recent entry's arXiv posting date was three months old
• Amazon Science: grouped by year, no per-paper date
• Microsoft: /research/feed/ is an editorially curated blog — ~8% coverage of the same window
• 13 Chinese company sites: 0 date hits across all of them — those are lab landing pages, not paper listings

The other fact you need: arXiv metadata does not carry author affiliation. <arxiv:affiliation> fill rate is 1%, Semantic Scholar 7%, and OpenAlex returns 0 raw_affiliation_strings for preprints (measured: 3/3 empty, while the same endpoint returns 34/34 for published Nature papers). The industry-standard pipeline is "publisher deposits an affiliation string → ML maps it to ROR", and arXiv never deposits that string. So affiliation has to be extracted from the paper's own author block.

Running it

The scripts live in this skill directory and depend on nothing outside it — pure stdlib, no pip packages.

R=<skill-dir>/scripts/paper_radar.py

python3 $R                 # last 3 days
python3 $R --days 14       # last 14 days
python3 $R --from 2026-07-20 --to 2026-08-04
python3 $R --cats cs.AI,cs.CL,cs.LG,cs.CV   # widen categories
python3 $R --no-scrape     # official sources + team names only (seconds — quick company check)
python3 $R -o path.md      # write the report here
python3 $R -o -            # write the report to stdout

Timing: for a two-week window (~3000 papers), metadata takes ~2 min and author-block fetching ~25 s at 6 workers. A 3-day window finishes inside a minute. To cover a longer history, widen the window and re-run — there is no library to query.

Where the report goes — you decide, the script does not guess

The script does not know, and should not know, where the calling project keeps its research. With no -o, it writes to the current working directory.

Decide the destination before each run:

• The user named a path → use it
• This is research worth keeping → write it into that project's research directory, with the window in the filename
• You are just checking one fact ("has Xiaomi published on self-evolving AI?") → use -o - and leave no file. A throwaway question that leaves a file behind will masquerade as a research record later

Files

The only output is one markdown file (path per above), with a "total hits / of which lead" split. The skill directory itself produces no files.

Six stages

0 listing      arXiv API: cat × submittedDate → dedupe by ID
1 full text    arxiv.org/html/{id} + Range first 90KB, cut ltx_authors → ltx_abstract      [85%]
2 structure    scan right-to-left for the last position starting an ascending 1,2,3… run
               extract (author, marks[]) · (mark, affiliation) · emails · role markers
3 entity       (a) email domain  (b) ROR name variants  (c) lab alias table
4 grading      lead_orgs = the first author's affiliations (may be several); intern = marker present
5 supplements  Apple RSS · MSR embedded JSON · arXiv team names
6 report       dedupe by key (one paper can arrive from both the author block and an official
               source), write markdown, keep evidence + matched_on

Five disciplines, every one of them paid for in bugs

1. The author block must be cut strictly at ltx_abstract. Take one character more and the abstract bleeds in. Measured consequences: Oracle 73 papers (all "oracle access"), MiniMax 17 (the minimax algorithm), DeepSeek 58 (papers merely mentioning the model). Cut cleanly, the median author block is ~201 characters.

2. The affiliation list must be found right-to-left. "name + superscript digit" in the author run and "digit + affiliation" in the affiliation run have identical shape, so a left-to-right scan parses Yuchen Zhao 1 Bingxuan Li as 1 → "Bingxuan Li".

3. Never fall back to "only one company matched, so it leads". That only tells you the other affiliations aren't companies you track. Measured misattribution: Filesystem-Based Memory is UIUC-led but was scored Adobe-led because Adobe was the sole match.

4. A company byline is not a company paper. Only * in the report means lead. Measured over a two-week window: Google hit 10 papers but led only 4; Google DeepMind 7 hits, 2 led; Adobe 5 hits, 0 led. Any claim of "N big tech papers" must say which column it means.

5. State the blind spots before the numbers.

• 12% of submissions are PDF-only and are invisible — the hard ceiling of stage 1
• Apple miss rate ~78% (its authors frequently omit the affiliation) → patched via the official RSS feed. This supplement is required, not optional
• When the first author lists several affiliations they cannot be ranked; both get *
• intern is a marker, not a judgment — whether an intern's paper "counts" is a definitional question, surfaced for the caller to decide

Three layers of error — don't conflate them

ROR is a string-to-ID lookup table; you need the string first. Measured on two Xiaomi papers: one had no HTML at all, the other's author block held 8 person names and not one word of affiliation. ROR cannot save either — only supplementary sources can.

Why each supplementary source stays

When a company shows zero — check in this order

1. Wrong window — arXiv indexes 1–2 days late and does not announce on weekends; widen first
2. Wrong categories — the default is cs.AI/cs.CL/cs.LG; robotics is cs.RO, vision cs.CV, systems cs.SE
3. Team-name search — all:"Xiaomi" / "Hunyuan" / "Kimi Team" work; all:"MiniMax" is unusable (5800+ hits from the minimax algorithm) — use au:"MiniMax"
4. Official channels — Apple RSS, Meta's ai.meta.com/results/?content_types[0]=publication (note /research/publications/ returns 400 to plain curl), NVIDIA research.nvidia.com/publications
5. Only when all of the above are empty may you say "this company published nothing in this window"

Roadmap

• Add TeX source as stage 1 tier 2 — untar arxiv.org/e-print/{id} and read \author{}. Verified: papers with no HTML often do have TeX, with superscripts intact. Cheapest available coverage gain
• Within-run health self-check — put stage 1's author-block hit rate, and the count of "official source has it but the author block didn't", straight into the report header. Guards against a silent break when arXiv changes its markup (the 78% Apple miss was found by accident). Do not reintroduce stored historical hit rates for this — the within-run cross-check suffices

One more thing

Paper count is supply, not heat. Measured: of 231 big tech papers in a two-week window, 0 were discussed on Hacker News, while every HN-discussed paper in the same window came from academia or a startup. Judging heat requires a separate social signal.