What's New

A formula that mentions a dollar amount — a $5 price floor, a $20M volume cut-off — used to break the sentence around it into half-rendered math. It now renders as the formula it is, in the review and in everything you copy from it.

The first-run screen now gives you everything the product can do, at full strength and with no settings to learn. Pick the roast and the review is also updated with the roast's findings that hold up, so you get both versions side by side. Pick the comic and it is written as a script first, then drawn, so panels and lettering come out right. A cover image, with the X thread and LinkedIn post where your plan has them, is one more tick. Behind it, the defaults moved to the newest models: Gemini 4 Argon writes the review, the roast and the translation, Gemini 3.8 Flash handles the fast steps, and Nano Banana 2.1 draws. A new key starts on Flex pricing, about half the cost for a wait of a few minutes per step; the screen says so, and Settings turns it off.

While a paper is being analyzed, its steps are now grouped by what they produce: reading the paper, the review, the roast, visuals, social posts. You see at a glance what is done, what is running right now and for how long, and how far the whole analysis has come, on a wide screen or a phone.

Two versions of a review used to mean reading both and hunting for the difference. Now Compare shows it at a glance: new words in green, removed ones struck through in red, in the finished article with its formulas and links, or in the source. Unchanged paragraphs fold away so a long review comes down to the few places that moved. Compare the review after the roast with the original, any earlier version of a review with the current one, or two versions of a roast, a translation or a post. And the tabs now stay at the top of the screen while you read, so switching to another tab no longer means scrolling back up.

A roast often finds what a polite review missed. Tick “Update review after the roast” and you get a second review that takes in the roast's findings that hold up — checked against the paper again, written in the review's own calm voice, with no jokes and no settling of scores. The original review stays exactly as it was, so the two sit side by side in their own tabs and you choose which one to publish. You can also have the cover image, comic and social posts made from the updated review instead of the original.

The little ladybug in the corner has always had a day coat and a night coat, but it changed them by the clock — at six sharp, like someone who has never met a northern summer or a December afternoon. It now goes by the sun where you are: out in the daylight for as long as there is some, moonlit once it is properly dark, and it changes over at sunset even if the tab has been open since lunch. It does not ask where you live; your time zone is hint enough. Nothing depends on any of this. We just thought it ought to know when to say good evening.

Star, archive and delete say where a paper is — not what you made of it. Now you can label a paper with whatever you like (“Niiice”, “Meh”, “To re-read”), see the labels on the Papers page, and pull up every paper carrying one, archived papers included. So a paper you did not care for can leave your active list and still be on record.

A review can now be copied for the place it is going. The new button beside a review puts the whole post on your clipboard ready to paste into Substack — figures, formulas, and tables too, which Substack's editor cannot hold and used to scatter into loose lines. The review itself keeps its real tables. Every other text tab — the roast, the translations, the social posts — now has the same buttons as the review: copy the markdown, see the source, or copy it for Substack.

A new account now opens on a short start: add your Gemini API key (checked with Google before it is saved), paste an arXiv link, and — if you like — tick "Roast your own paper" to get a brutally honest teardown on top of the review. A new About page says plainly what the preview does, what it doesn't, and your account's limits. And the ladybug button now attaches the state of the paper or workspace you're on (status, stage, error, run id — never the paper's text), so a bug report reaches us with what we need to fix it.

Settings → Your data downloads everything you have in PaperLM — papers, reviews, chats, notes, workspaces, feedback — as one file. Deleting your account removes your API key at once and everything else within 14 days, and you can change your mind until then.

Saving your Gemini API key now checks it with Google first and tells you plainly what is wrong if it doesn't work — a mistyped key, or one that can't use the Gemini API — instead of failing later in the middle of an analysis. Settings shows the last four characters so you know which key is saved, and removing a key stops the analyses using it. A new page, "Your Gemini API key", explains exactly how your key is stored, used and deleted.

Regenerating a stage now offers exactly the choices a new analysis does — comic mode, image size and aspect ratio, the model for each stage — and starts from the settings and instructions it was last run with, so you adjust instead of retyping. Review reruns take separate instructions for the critique and refinement passes. And every earlier version shows what it was made with: pick one, reload its settings, tweak, and run again.

Under every completed research report in a workspace: save it as a workspace document with one click — then refine it, keep versions, and feed it into other documents like anything else — or turn it straight into an X thread, a LinkedIn post, a cover image or a comic — with the same controls you get everywhere else first: your instructions, and for images the model, style, size and aspect ratio. The workspace's own Generate menu asks the same way now. Everything lands right under the report, per report, so a workspace with several research runs keeps each one's outputs together. Posting somewhere with a size limit? Split the report into posts — by section or to a character limit — where every citation links straight to its source and each post lists the sources it cites, so every piece stands on its own.

A completed research report in a workspace now has a Copy button. Paste it into Substack, Google Docs or Notion and it arrives formatted — headings, tables, links, the numbered reference list — and citation links stay in-post links to that list rather than pointing back here. Paste into a plain-text editor and you get clean Markdown instead. The reference list itself is now worth posting: every source links directly to its page with the page's real title instead of a tracking redirect labelled with a domain name, and a paper cited three ways becomes one entry, with the citations renumbered. Existing reports get this the next time you open them. Clicking a citation in the app now jumps to its reference, too, and the multi-review's Copy button works the same way.

Each review now comes with a handful of suggested tags for your blog post, right under the text — click one to copy it, or copy them all at once. They stay consistent with how you already tag your blog, so its navigation doesn't sprout near-duplicates, and they add fresh tags for whatever the paper itself names: the method, the model, the lab. Tags are searchable in your library, too.

One wrong digit in a paper id and a link opens a real paper from a completely different field. arXiv links in a new review are now checked against arXiv before you see it: a link whose paper doesn't match the one cited is corrected or removed rather than left pointing at an unrelated paper. Better no link than a wrong one. And a citation the paper itself gives no link for — a conference paper listed by title and venue only — now gets one whenever the work can be found, on arXiv or at its publisher.

Tick “Connect to my previous reviews” and a new review is checked against your own library first: it finds the papers you've already reviewed that genuinely relate — a predecessor, the same group's earlier work, a competing approach, a result it contradicts — links them, and draws the comparisons a single paper can't give you. An “In your library” panel under the review takes you straight to each one, with a line on why it's there. It looks through your archive too, stays fast however large your library grows, and is off unless you ask for it — on the new-analysis form, in templates, and in Run stages.

The Papers page now has a search box: type a title, an author, or an arXiv id and results appear as you type — across your archive too, so an archived paper is never lost. Tick “Search inside reviews” to find a paper by something the review said, with the matching passage shown on the card. Filters for status, source, date, prompt set and artifacts, plus sorting, narrow things further. Press / to jump to search.

When a paper builds on previous research by the same group, the review now says so up front, links the predecessor, and explains what the new work adds on top — and the fact-checking pass verifies those links.

Reviews now show the paper's own figures, tables and algorithm listings inline — each one appears under the paragraph that first cites it, with a link to its page in the PDF. Click to enlarge, or copy a figure straight into your post. It copes with the layouts real papers actually use — multi-panel figures, appendix numbering, captions above or below, templates that style their labels their own way. Extraction runs on every new analysis; for papers you analysed earlier, tick “Paper figures” in Run stages. If a figure or table slips past the first look, the extractor goes back for it.

The separate “Rerun review” button is gone: the Run stages dialog now regenerates the review too, with the same template, model, and instruction controls. Better still, you can rerun the review and rebuild translations, social posts, or the roast in one go — and they'll be generated from the new review, not the old one. Previous versions stay restorable, as always.

The settings page now walks you through protecting your own Gemini API key: create it in a dedicated Google Cloud project, set a budget alert, restrict it to the Generative Language API, and rotate it in one click if anything ever looks off. A leaked key can't read your data — but it can spend your money, so two minutes of setup buys a hard ceiling on the damage.

Reviews now consistently identify the specific models behind a paper — exact family, version, and size, and what role each played (backbone, baseline, judge). And when the authors deliberately don't say — increasingly common in safety and security work — the review calls that out explicitly, because a paper that hides its model is telling you something too.

Every paper gets a polite review. Almost none get an honest one. Paper Roast turns the pipeline on the paper itself and says the part everyone only mutters at the poster session: the baseline picked because the others didn't work, the ablation missing exactly where it would hurt, the claim in the abstract the experiments never actually test. It searches the web while it reads, so "this was done in 2019 and they don't cite it" is a receipt, not a hunch. Three intensities — Medium, Spicy, and Nuclear — and if it goes easy on them, tell it to try harder and rerun. Two rules it never breaks: it goes after the work and never the authors, and every hit has to point at a real line in the paper. When it cites the paper's own references it does so the way the paper does — “[52]” only if the paper numbers that work 52, author and year otherwise — and the citation links straight to the work wherever it can be found. Turn it on under Paper Roast when you start an analysis. Your own papers are allowed too. Bring a helmet.

Every paper in your library now has an Archive button next to Star and Delete. Archived papers move out of your main list into the Archived tab — and back again with one click — so you can keep finished work without deleting it.

Paper cards now show how many notes you've taken and how long the chat is, so you can tell which papers you've actually worked through. The header also got a Papers link back to your library, and one clear button for starting a new analysis instead of two.

Comics generated in Advanced mode now show the script that produced them, right under the image. Tweak a line of dialogue, drop a panel, or rename a character, then hit Redraw from script — the comic is redrawn from exactly your edited script, with no rewriting in between. The last mile of getting a comic just right.

Comic generation has a new Advanced mode: it writes the comic script first — cast, panels, and the exact words in every speech bubble — and only then draws it. That means far fewer duplicated panels, garbled lettering, and bubbles pointing at the wrong character. Pick it under Comic generation when you start an analysis, or when you regenerate a comic you're not happy with. Simple one-pass mode is still the default.

Every "additional instructions" box — for new analyses, review reruns, and single-stage regenerations like the X/Twitter thread — now accepts up to 5,000 characters and grows into a resizable multi-line field. Long, detailed prompts no longer get cut off.

Updating a review from chat now always shows the new version right away — previously, if you had opened the review editor, the page could keep showing the old text even though the update succeeded. If you have unsaved edits when a review changes underneath you, a banner now lets you choose between your draft and the latest version.

Got a paper that isn't on arXiv or OpenReview — or a camera-ready copy on your disk? Drop the PDF right on the Analyze a Paper page and it runs through the exact pipeline you configured, no workspace needed. Available on plans with direct PDF upload.

OpenReview recently put automated downloads behind a bot check, which broke analyses of OpenReview papers (including the ICML 2026 award papers). PaperLM now fetches those PDFs through OpenReview's official signed-in API, so forum links analyze normally again.

If a paper analysis ever stalls — usually when an image model has a temporary outage — it no longer hangs forever. It now retries automatically in the background, and the paper page surfaces a clear status with one-tap Re-run and Cancel so you're always in control.

Not happy with a review? Regenerate it with a different template, model, critique on or off, or your own instructions — straight from the paper page. The new review lands as a fresh version, and your previous one is kept and stays restorable, so you can compare and switch back anytime.

Turn a workspace into a blog post, essay, slide outline, or anything custom — generated from the materials you pick (sources, the multi-review, deep-research reports, even other documents), then refined, versioned, and reusable as an input to the next document. You can also save a whole chat answer, or just a selected passage, straight into a document.

No more regenerating from scratch. Steer the multi-review with instructions, refine a finished research report in place (a quick Edit, or send the agent back out for more), and hand-edit the multi-review, reports, or documents yourself when you know better — every change saved as a restorable version. Deep-research runs also got a clear, labeled run switcher.

Attach a specific deep-research report or saved document to a workspace chat so the model works directly against it — summarize it, compare it to your sources, and more.

After untold engineering heroics, sleepless nights, and at least one heated design review, we have done what every serious product must eventually do: added a light/dark theme toggle. Truly, civilization advances. It lives in the header, follows your system preference by default, and remembers your choice — so you can flip between blinding white and cozy dark to your heart's content. You're welcome.

Notes now render full markdown (math, code, lists, tables) and can expand to full screen. Saving a selection from chat keeps its formatting, and the side panel jumps to the prefilled note. Clearing a chat or deleting a note is also undoable now, and the chat-clear button moved so you can't wipe a conversation by mis-clicking.

Combine several papers in one workspace: add analyzed papers, paste arXiv/OpenReview/PDF links, or upload PDFs, then chat across all of them (with optional full-PDF grounding) and generate a cross-paper review that synthesizes the whole set — plus images, comics, and social posts. You can also run Google's Deep Research agent right from a workspace: it proposes a plan you refine, investigates the web, and returns a cited report you can keep chatting about.

Type @ in chat to pull another paper's review into the conversation and ask how the two compare — references stay in context until you remove them. You can also @-mention papers in the review instructions when starting a new analysis, and the new review will compare against and link to them.

Deleted papers now go to a Trash you can restore from; papers left there are deleted for good after 30 days. Mistakes are finally undoable.

Chat now lets you pick the model, toggle live web search and URL context, and renders rich formatted answers — in a cleaner tabbed, expandable side rail. Ask harder questions, get better-grounded answers.

Choose exactly which Gemini model powers each stage of the pipeline — speed where you want it, maximum reasoning where it matters. Reruns now build on your latest edited review.

Save your favorite pipeline configurations as named, reusable templates. Set it up once, reuse it on every paper.

Spotted something off? A floating report button lets you flag it in seconds — and we triage it fast.

Turn on Flex tier in Settings and your paper analyses run on Google's cheaper inference tier — roughly 50% less to produce the same review. The trade is patience, not quality: Flex waits for spare capacity, so a stage can take anywhere from one to fifteen minutes and a full run finishes later than usual. If Flex capacity runs out mid-run we fall back to the standard tier automatically, so nothing stalls or fails. Chat stays on the standard tier either way, because that's the part where waiting actually hurts.

We rebuilt the backend on Go with a dedicated queue-based worker — analyses are faster, survive cold starts, and a retried task can no longer run the same paper twice at once. Failed stages now show clear, human-readable errors instead of cryptic codes.

Paid plans run pipeline stages in parallel — your full review, visuals, and social posts land dramatically faster.

Pick from curated visual style presets per language, or define your own. Make summaries that match your taste.

Analyze papers from custom PDF URLs, not just arXiv and OpenReview.

Reviews are now powered by versioned, tunable prompt bundles — more consistent, higher-quality output.

Read the paper side-by-side with its review, with inline highlights and annotations.

Talk to any paper, keep your own notes, and edit reviews with live preview — every change versioned so you can always go back.

Turn dense papers into shareable comic-panel summaries — perfect for slides and social.

Star, archive, and switch between list and card views with thumbnails. Build a knowledge base you'll actually revisit.

Full LaTeX math, rich markdown, and HTML rendering — papers look the way they should.

The first paper-analysis pipeline goes live — paste an arXiv link, get a structured, multi-agent review. From the creator of ArXivIQ (200+ AI paper reviews).

ArXivIQ starts publishing a review of a new AI paper every day. Hundreds of reviews later, the workflow behind it becomes PaperLM.

Experiments with machine-written paper reviews go public alongside the hand-written ones — first as an entertaining side genre, then good enough to read every day.

GonzoML opens with a review of a paper on unsupervised representation learning. Years of hand-written reviews follow — they set the voice and the bar that PaperLM's reviews are held to.

Long before there was a product, there was a habit: reading research papers, week after week, and working out what each one really says. Everything above grew out of it.