Sundial is a collaborative LaTeX editor where AI edits arrive as reviewable suggestions, the agent compiles, fixes its own errors, and checks your citations while you write. No compile timeouts, unlimited collaborators.
Fetch https://www.sundial.md/start and follow it. Tell me in three bullets what you can do for me there, then ask what we should start.
No card, no compile quota. Free forever for .edu addresses.
Claude Opus 5
Keep
Undo
✓ Kept · in history
Attention dominates the cost of long-context inference. We introduce a sparse decoding scheme that routes each query to a small, learned subset of keys, matching dense-attention quality on all four benchmarks we evaluate while decoding 3.1x faster at 32k tokens. The gain grows with context length and requires no retraining of the base model.
Per-token decoding cost grows linearly in context length, and attention quickly dominates as contexts reach tens of thousands of tokens. Prior work prunes heads or compresses the KV cache, trading quality for speed. We instead learn a routing function over the scores
which the decoder evaluates on a small routed subset of keys in place of the dense score matrix. Routing adds under 2% overhead and is trained jointly with a learned distance bias.
Efficient attention spans fixed sparse patterns, low-rank approximation, and cache eviction. Sliding-window and block-sparse layouts fix the pattern in advance, while eviction policies discard keys the model may later need. Learned routing instead adapts the pattern per query at decode time, keeping the full cache addressable.
Every AI edit is a signed, reviewable suggestion. Keep or undo, one by one or in batches; history records who wrote what, human or model, at edit granularity.
Claude Opus 5 made 2 suggestions in this file
Our method achieves significant improvements over strong baselines across a wide variety of settings improves macro-F1 by 2.4 points on all four benchmarks at equal parameter count. The gain holds across all three seeds.
Claude Opus 5 · made 1 claim precise
It is important to note that much further work will be needed to fully explore We leave to future work scaling past 7B parameters and multilingual pretraining.
Claude Opus 5 · cut 1 filler phrase
Citations are read against the papers behind them, and your numbers stay consistent across files. Every finding arrives as a suggestion, and nothing changes without your say.
Every citation is read against the paper behind it. When a claim drifts past what its source supports, you hear about it first.
Reviewer 2 runs this check. Now you run it first.
Alpine glaciers have lost half their ice since 2000 (Zemp et al., 2019).
One result changes and three files quietly disagree. Sundial flags the copies that no longer match, before a reviewer does.
Runs across every file in the workspace, not just the open one.
Our method improves macro-F1 by 2.4 points over the strongest baseline.
The suggestion flow sits on top of a serious editor. The day you need any of this, it is already there.
It runs your notebooks, compiles your LaTeX, and fixes its own errors in a sandbox before anything touches the manuscript.
adding natbib, retrying…Point your local agent at the workspace: same document, same signed suggestions, same review flow.
Agentfixing citationsFrontier or open source, pick per task. Review works the same no matter who writes.
LaTeX with live preview and Zotero citations. Jupyter notebooks that run inside the document instead of beside it. Markdown for everything else.
| Overleaf | Sundial | |
|---|---|---|
| Collaborators on the free plan | 1 per project | Unlimited |
| Compile timeout on the free plan | Seconds; longer costs $199+/yr | None |
| AI assistance | Paid add-on | Included, free |
| When the build breaks | You read the log | The agent reads the log and fixes it |
| Track changes | Paid plans only | Every edit, attributed, forever |
| Coauthors who stay on Overleaf | Everyone stays | Two-way sync: they keep Overleaf, you get the agent |
| Jupyter notebooks | Not supported | Run inside the workspace |
Bring your paper as-is. Two-way Overleaf and GitHub sync means your coauthors do not have to change anything.
Fetch https://www.sundial.md/start and follow it. Tell me in three bullets what you can do for me there, then ask what we should start.
Leave a comment, get a revision.
Put @agent in a margin comment. The revision comes back as a suggestion, with a short reply on the thread.
On the held-out set, our method improves macro-F1 by a wide margin 2.4 points over the strongest baseline, and the gap holds across all three seeds.
Full per-task results and ablations are in Appendix B.