> ## Documentation Index
> Fetch the complete documentation index at: https://docs.research.prbe.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Papers

> Record the literature a research project was built from — what each paper says, the repo read alongside it, and where the two disagree.

A literature review is research work, and it leaves as little behind as a training run does unless somebody records it. `probe paper` is where it goes.

Record **during** a review and again at its end — not afterwards from memory.

```bash theme={null}
probe paper add --project folding \
  --source https://arxiv.org/abs/2106.09685 \
  --authors "Hu et al." \
  --summary @lora-summary.md \
  --repo https://github.com/microsoft/LoRA \
  --discrepancies @lora-vs-repo.md \
  --tag adapters --tag peft \
  --via <paper-id> --via-provenance provider_citation \
  --via-reason "cited as the baseline this method improves on"
```

| Field                | What goes in it                                            |
| -------------------- | ---------------------------------------------------------- |
| `--source` / `--url` | The paper URL or a local file path. Required               |
| `--authors`          | Free text: `"Vaswani et al."`                              |
| `--summary`          | **The paper's main idea, in your words.** `@file.md` works |
| `--repo`             | The repository you read alongside it                       |
| `--discrepancies`    | What the repo does that the paper does not say             |
| `--tag`              | Concept tags, repeatable                                   |

`--discrepancies` is the field that earns the entity. "The paper says warmup is linear; the repo's scheduler is cosine with a floor" is the kind of thing that costs a week when it is rediscovered, and nothing else in a review captures it.

## Provenance: how you got there

```bash theme={null}
--via <paper-id>            # the paper you followed to this one
--via none                  # you came to this one directly
--via-provenance observed_call | provider_citation | human | inferred
--via-reason "one sentence on why that paper led to this one"
```

That chain is readable back as a graph:

```bash theme={null}
probe paper edges --project folding
```

Omit `--via` only when you genuinely cannot say. A review whose paths are recorded can be re-walked; one without them is a bag of citations.

## Managing them

```bash theme={null}
probe paper list --project folding --tag adapters     # newest first
probe paper update <id> --summary @revised.md
probe paper tag <id> peft --remove adapters
probe paper remove <id>
```

`update` leaves omitted fields untouched. `""` clears a field — except `--title` and `--source`, which are required and reject an empty value.

## In the dashboard

A project whose kind is `research` leads with **Papers**, right after Overview, and each row opens its own page at `/papers/{id}`.

The page carries the captured **Summary** and **Repo vs paper** documents as full Markdown, and a **Paper** tab holding the provider-extracted abstract. Title, authors, publication year, venue and category chips are filled in from extraction where it succeeds; a pending extraction shows a spinner for about a minute and then goes quiet rather than pretending to still be working.

<Note>
  The Papers tab is **read-only**. Recording a paper is an act by whoever did the reading — through the CLI or an agent session — exactly as attaching a subproject is. An empty tab names the capture path rather than offering a form.
</Note>

## Finding papers to record

Agents have a `find_papers` tool over 40M+ abstracts, with `search`, `read` and `similar` modes.

<Tip>
  `find_papers` wants **natural language**, not a keyword bag. A question-shaped description retrieves substantially more relevant papers than a bag of terms. Its `categories` filter is an AND, and its score rises as relevance falls — read the results, do not rank by the number.
</Tip>

See [the MCP reference](/agents/mcp).