Add five papers, from the Qwen2.5-Math technical report and Lemur to Code Llama and MMLU, to ref.bib in the same format as the existing entries, copying the BibTeX from the OpenReview conference page whenever a paper has a conference version.
Please help me add the following article to the ref.bib file in the workspace. Please keep the format consistent with the existing article. If there is a conference version, please copy the bibtex file from the corresponding conference homepage on OpenReview. 1. Scaling relationship on learning mathematical reasoning with large language models 2. Qwen2.5-math technical report 3. Lemur: Harmonizing Natural Language and Code for Language Agents 4. Codellama 5. mmlu
---
name: workspace-filesystem
description: The virtual filesystem layout of this workspace.
---
# Virtual Filesystem
You work in a terminal over a virtual filesystem. Read files with shell
commands; service operations are listed in SKILL.md.
| Path | Description |
|---|---|
| `/workspace` | Agent workspace; current directory and home (`~`). |
| `/tmp` | Temporary files. |
| `/dev` | Device files: `null` and `zero` (mounted automatically). || name | workspace-filesystem |
|---|---|
| description | The virtual filesystem layout of this workspace. |
Virtual Filesystem
You work in a terminal over a virtual filesystem. Read files with shell commands; service operations are listed in SKILL.md.
| Path | Description |
|---|---|
/workspace | Agent workspace; current directory and home (~). |
/tmp | Temporary files. |
/dev | Device files: null and zero (mounted automatically). |
---
name: workspace-commands
description: The shell commands and task tools available in this workspace.
---
# Available commands
Standard Unix commands, among them:
cat head tail ls find grep wc sort uniq cut tr sed awk
diff tee xargs jq unzip curl
with pipes, `>` redirection, `$(...)` substitution and `$'...'`.
The task CLIs below are registered with this shell rather than installed
as files; do not inspect them with `which`, `file` or `/usr/bin`.
## web-search
Use `web-search <command> '<JSON arguments>'`. Run
`web-search <command> --help` to get more information.
| Command | Arguments | Description |
|---|---|---|
| `web-search web-search` | `{query: str, num_results?: int = 10}` | Search the web using Google Serper API with concurrency control and retry mechanisms. Supports various Google search operators. — `query`: Search query with optional Google search operators.; `num_results`: Number of results to return, default 10, max 50 |
## browser
Use `browser <command> '<JSON arguments>'`. Run `browser <command> --help`
to get more information.
| Command | Arguments | Description |
|---|---|---|
| `browser close` | `{}` | Close the page |
| `browser resize` | `{width: number, height: number}` | Resize the browser window — `width`: Width of the browser window; `height`: Height of the browser window |
| `browser console-messages` | `{}` | Returns all console messages |
| `browser handle-dialog` | `{accept: bool, promptText?: str}` | Handle a dialog — `accept`: Whether to accept the dialog.; `promptText`: The text of the prompt in case of a prompt dialog. |
| `browser evaluate` | `{function: str, element?: str, ref?: str}` | Evaluate JavaScript expression on page or element — `function`: () => { /* code */ } or (element) => { /* code */ } when element is provided; `element`: Human-readable element description used to obtain permission to interact with the element; `ref`: Exact target element reference from the page snapshot |
| `browser file-upload` | `{paths: list[str]}` | Upload one or multiple files — `paths`: The absolute paths to the files to upload. Can be a single file or multiple files. |
| `browser install` | `{}` | Install the browser specified in the config. Call this if you get an error about the browser not being installed. |
| `browser press-key` | `{key: str}` | Press a key on the keyboard — `key`: Name of the key to press or a character to generate, such as `ArrowLeft` or `a` |
| `browser type` | `{element: str, ref: str, text: str, submit?: bool, slowly?: bool}` | Type text into editable element — `element`: Human-readable element description used to obtain permission to interact with the element; `ref`: Exact target element reference from the page snapshot; `text`: Text to type into the element; `submit`: Whether to submit entered text (press Enter after); `slowly`: Whether to type one character at a time. Useful for triggering key handlers in the page. By default entire text is filled in at once. |
| `browser navigate` | `{url: str}` | Navigate to a URL — `url`: The URL to navigate to |
| `browser navigate-back` | `{}` | Go back to the previous page |
| `browser navigate-forward` | `{}` | Go forward to the next page |
| `browser network-requests` | `{}` | Returns all network requests since loading the page |
| `browser take-screenshot` | `{raw?: bool, filename?: str, element?: str, ref?: str, fullPage?: bool}` | Take a screenshot of the current page. You can't perform actions based on the screenshot, use browser_snapshot for actions. — `raw`: Whether to return without compression (in PNG format). Default is false, which returns a JPEG image.; `filename`: File name to save the screenshot to. Defaults to `page-{timestamp}.{png\|jpeg}` if not specified.; `element`: Human-readable element description used to obtain permission to screenshot the element. If not provided, the screenshot will be taken of viewport. If element is provided, ref must be provided too.; `ref`: Exact target element reference from the page snapshot. If not provided, the screenshot will be taken of viewport. If ref is provided, element must be provided too.; `fullPage`: When true, takes a screenshot of the full scrollable page, instead of the currently visible viewport. Cannot be used with element screenshots. |
| `browser snapshot` | `{}` | Capture accessibility snapshot of the current page. The snapshot is divided into manageable spans (configurable via --span-size, default 2000 characters, or use -1 to show complete snapshot) and shows the first span by default. Use browser_snapshot_navigate_* tools to view other spans or browser_snapshot_search to find content across all spans. This provides structured element data with ref attributes for interaction, making it better than screenshots for automation. |
| `browser click` | `{element: str, ref: str, doubleClick?: bool, button?: "left" \| "right" \| "middle"}` | Perform click on a web page — `element`: Human-readable element description used to obtain permission to interact with the element; `ref`: Exact target element reference from the page snapshot; `doubleClick`: Whether to perform a double click instead of a single click; `button`: Button to click, defaults to left |
| `browser drag` | `{startElement: str, startRef: str, endElement: str, endRef: str}` | Perform drag and drop between two elements — `startElement`: Human-readable source element description used to obtain the permission to interact with the element; `startRef`: Exact source element reference from the page snapshot; `endElement`: Human-readable target element description used to obtain the permission to interact with the element; `endRef`: Exact target element reference from the page snapshot |
| `browser hover` | `{element: str, ref: str}` | Hover over element on page — `element`: Human-readable element description used to obtain permission to interact with the element; `ref`: Exact target element reference from the page snapshot |
| `browser select-option` | `{element: str, ref: str, values: list[str]}` | Select an option in a dropdown — `element`: Human-readable element description used to obtain permission to interact with the element; `ref`: Exact target element reference from the page snapshot; `values`: Array of values to select in the dropdown. This can be a single value or multiple values. |
| `browser snapshot-navigate-to-span` | `{spanIndex: int}` | Navigate to a specific span in the current page snapshot by index. Each span contains a portion of the page content (limited by span size). Use this when you know the exact span number you want to view, typically after using browser_snapshot_search to locate content. — `spanIndex`: The span index to navigate to (0-based) |
| `browser snapshot-navigate-to-first-span` | `{}` | Navigate to the first span of the current page snapshot. This shows the beginning of the page content and is useful for starting from the top of the page when exploring content sequentially. |
| `browser snapshot-navigate-to-last-span` | `{}` | Navigate to the last span of the current page snapshot. This shows the end of the page content and is useful for accessing footer content, final form elements, or bottom navigation. |
| `browser snapshot-navigate-to-next-span` | `{}` | Navigate to the next span in the current page snapshot. If already at the last span, stays at the last span. Use this for sequential exploration of page content from top to bottom. |
| `browser snapshot-navigate-to-prev-span` | `{}` | Navigate to the previous span in the current page snapshot. If already at the first span, stays at the first span. Use this for sequential exploration of page content from bottom to top. |
| `browser snapshot-navigate-to-line` | `{globalLineNumber: int, contextLines?: int}` | Navigate to a specific global line number in the current page snapshot with configurable surrounding context lines. Shows which spans contain the context lines. The target line is marked with ">>>". Use this to quickly jump to specific content found via search results. — `globalLineNumber`: The global line number to navigate to (1-based); `contextLines`: Number of context lines to show before and after the target line (default: 3) |
| `browser snapshot-search` | `{pattern: str, flags?: str}` | Search for a pattern across all spans of the current page snapshot using regular expressions. Returns matches with both global line numbers (across entire snapshot) and in-span line numbers (within each span). Shows which spans contain matches for navigation. Supports regex flags like "gi" for global case-insensitive search. — `pattern`: The regex pattern to search for; `flags`: Optional regex flags (e.g., "gi" for global case-insensitive) |
| `browser tab-list` | `{}` | List browser tabs |
| `browser tab-new` | `{url?: str}` | Open a new tab — `url`: The URL to navigate to in the new tab. If not provided, the new tab will be blank. |
| `browser tab-select` | `{index: number}` | Select a tab by index — `index`: The index of the tab to select |
| `browser tab-close` | `{index?: number}` | Close a tab — `index`: The index of the tab to close. Closes current tab if not provided. |
| `browser wait-for` | `{time?: number, text?: str, textGone?: str}` | Wait for text to appear or disappear or a specified time to pass — `time`: The time to wait in seconds; `text`: The text to wait for; `textGone`: The text to wait for to disappear |
## fetch
Use `fetch <command> '<JSON arguments>'`. Run `fetch <command> --help` to
get more information.
| Command | Arguments | Description |
|---|---|---|
| `fetch html` | `{url: str, headers?: dict}` | Fetch a website and return the content as HTML — `url`: URL of the website to fetch; `headers`: Optional headers to include in the request |
| `fetch markdown` | `{url: str, headers?: dict}` | Fetch a website and return the content as Markdown — `url`: URL of the website to fetch; `headers`: Optional headers to include in the request |
| `fetch txt` | `{url: str, headers?: dict}` | Fetch a website, return the content as plain text (no HTML) — `url`: URL of the website to fetch; `headers`: Optional headers to include in the request |
| `fetch json` | `{url: str, headers?: dict}` | Fetch a JSON file from a URL — `url`: URL of the JSON to fetch; `headers`: Optional headers to include in the request |
## scholarly
Use `scholarly <command> '<JSON arguments>'`. Run
`scholarly <command> --help` to get more information.
| Command | Arguments | Description |
|---|---|---|
| `scholarly search-arxiv` | `{keyword: str}` | Search arxiv for articles related to the given keyword. |
| `scholarly search-google-scholar` | `{keyword: str}` | Search google scholar for articles related to the given keyword. |
## git
Use `git --help` for Git operations on workspace repositories.
## python3
Execute Python code directly under the agent workspace and return stdout
and stderr. Code can be pasted into a `.py` file and run with `python3`.
Python cannot launch subprocesses or make direct network requests. Run the
documented service commands directly in the shell. Use Python for local
file processing and generating JSON arguments.| name | workspace-commands |
|---|---|
| description | The shell commands and task tools available in this workspace. |
Available commands
Standard Unix commands, among them:
cat head tail ls find grep wc sort uniq cut tr sed awk
diff tee xargs jq unzip curlwith pipes, > redirection, $(...) substitution and $'...'.
The task CLIs below are registered with this shell rather than installed as files; do not inspect them with which, file or /usr/bin.
web-search
Use web-search <command> '<JSON arguments>'. Run web-search <command> --help to get more information.
| Command | Arguments | Description |
|---|---|---|
web-search web-search | {query: str, num_results?: int = 10} | Search the web using Google Serper API with concurrency control and retry mechanisms. Supports various Google search operators. — query: Search query with optional Google search operators.; num_results: Number of results to return, default 10, max 50 |
browser
Use browser <command> '<JSON arguments>'. Run browser <command> --help to get more information.
| Command | Arguments | Description | ||
|---|---|---|---|---|
browser close | {} | Close the page | ||
browser resize | {width: number, height: number} | Resize the browser window — width: Width of the browser window; height: Height of the browser window | ||
browser console-messages | {} | Returns all console messages | ||
browser handle-dialog | {accept: bool, promptText?: str} | Handle a dialog — accept: Whether to accept the dialog.; promptText: The text of the prompt in case of a prompt dialog. | ||
browser evaluate | {function: str, element?: str, ref?: str} | Evaluate JavaScript expression on page or element — function: () => { /* code */ } or (element) => { /* code */ } when element is provided; element: Human-readable element description used to obtain permission to interact with the element; ref: Exact target element reference from the page snapshot | ||
browser file-upload | {paths: list[str]} | Upload one or multiple files — paths: The absolute paths to the files to upload. Can be a single file or multiple files. | ||
browser install | {} | Install the browser specified in the config. Call this if you get an error about the browser not being installed. | ||
browser press-key | {key: str} | Press a key on the keyboard — key: Name of the key to press or a character to generate, such as ArrowLeft or a | ||
browser type | {element: str, ref: str, text: str, submit?: bool, slowly?: bool} | Type text into editable element — element: Human-readable element description used to obtain permission to interact with the element; ref: Exact target element reference from the page snapshot; text: Text to type into the element; submit: Whether to submit entered text (press Enter after); slowly: Whether to type one character at a time. Useful for triggering key handlers in the page. By default entire text is filled in at once. | ||
browser navigate | {url: str} | Navigate to a URL — url: The URL to navigate to | ||
browser navigate-back | {} | Go back to the previous page | ||
browser navigate-forward | {} | Go forward to the next page | ||
browser network-requests | {} | Returns all network requests since loading the page | ||
browser take-screenshot | {raw?: bool, filename?: str, element?: str, ref?: str, fullPage?: bool} | Take a screenshot of the current page. You can't perform actions based on the screenshot, use browser_snapshot for actions. — raw: Whether to return without compression (in PNG format). Default is false, which returns a JPEG image.; filename: File name to save the screenshot to. Defaults to `page-{timestamp}.{png\ | jpeg} if not specified.; element: Human-readable element description used to obtain permission to screenshot the element. If not provided, the screenshot will be taken of viewport. If element is provided, ref must be provided too.; ref: Exact target element reference from the page snapshot. If not provided, the screenshot will be taken of viewport. If ref is provided, element must be provided too.; fullPage`: When true, takes a screenshot of the full scrollable page, instead of the currently visible viewport. Cannot be used with element screenshots. | |
browser snapshot | {} | Capture accessibility snapshot of the current page. The snapshot is divided into manageable spans (configurable via --span-size, default 2000 characters, or use -1 to show complete snapshot) and shows the first span by default. Use browser_snapshot_navigate_* tools to view other spans or browser_snapshot_search to find content across all spans. This provides structured element data with ref attributes for interaction, making it better than screenshots for automation. | ||
browser click | `{element: str, ref: str, doubleClick?: bool, button?: "left" \ | "right" \ | "middle"}` | Perform click on a web page — element: Human-readable element description used to obtain permission to interact with the element; ref: Exact target element reference from the page snapshot; doubleClick: Whether to perform a double click instead of a single click; button: Button to click, defaults to left |
browser drag | {startElement: str, startRef: str, endElement: str, endRef: str} | Perform drag and drop between two elements — startElement: Human-readable source element description used to obtain the permission to interact with the element; startRef: Exact source element reference from the page snapshot; endElement: Human-readable target element description used to obtain the permission to interact with the element; endRef: Exact target element reference from the page snapshot | ||
browser hover | {element: str, ref: str} | Hover over element on page — element: Human-readable element description used to obtain permission to interact with the element; ref: Exact target element reference from the page snapshot | ||
browser select-option | {element: str, ref: str, values: list[str]} | Select an option in a dropdown — element: Human-readable element description used to obtain permission to interact with the element; ref: Exact target element reference from the page snapshot; values: Array of values to select in the dropdown. This can be a single value or multiple values. | ||
browser snapshot-navigate-to-span | {spanIndex: int} | Navigate to a specific span in the current page snapshot by index. Each span contains a portion of the page content (limited by span size). Use this when you know the exact span number you want to view, typically after using browser_snapshot_search to locate content. — spanIndex: The span index to navigate to (0-based) | ||
browser snapshot-navigate-to-first-span | {} | Navigate to the first span of the current page snapshot. This shows the beginning of the page content and is useful for starting from the top of the page when exploring content sequentially. | ||
browser snapshot-navigate-to-last-span | {} | Navigate to the last span of the current page snapshot. This shows the end of the page content and is useful for accessing footer content, final form elements, or bottom navigation. | ||
browser snapshot-navigate-to-next-span | {} | Navigate to the next span in the current page snapshot. If already at the last span, stays at the last span. Use this for sequential exploration of page content from top to bottom. | ||
browser snapshot-navigate-to-prev-span | {} | Navigate to the previous span in the current page snapshot. If already at the first span, stays at the first span. Use this for sequential exploration of page content from bottom to top. | ||
browser snapshot-navigate-to-line | {globalLineNumber: int, contextLines?: int} | Navigate to a specific global line number in the current page snapshot with configurable surrounding context lines. Shows which spans contain the context lines. The target line is marked with ">>>". Use this to quickly jump to specific content found via search results. — globalLineNumber: The global line number to navigate to (1-based); contextLines: Number of context lines to show before and after the target line (default: 3) | ||
browser snapshot-search | {pattern: str, flags?: str} | Search for a pattern across all spans of the current page snapshot using regular expressions. Returns matches with both global line numbers (across entire snapshot) and in-span line numbers (within each span). Shows which spans contain matches for navigation. Supports regex flags like "gi" for global case-insensitive search. — pattern: The regex pattern to search for; flags: Optional regex flags (e.g., "gi" for global case-insensitive) | ||
browser tab-list | {} | List browser tabs | ||
browser tab-new | {url?: str} | Open a new tab — url: The URL to navigate to in the new tab. If not provided, the new tab will be blank. | ||
browser tab-select | {index: number} | Select a tab by index — index: The index of the tab to select | ||
browser tab-close | {index?: number} | Close a tab — index: The index of the tab to close. Closes current tab if not provided. | ||
browser wait-for | {time?: number, text?: str, textGone?: str} | Wait for text to appear or disappear or a specified time to pass — time: The time to wait in seconds; text: The text to wait for; textGone: The text to wait for to disappear |
fetch
Use fetch <command> '<JSON arguments>'. Run fetch <command> --help to get more information.
| Command | Arguments | Description |
|---|---|---|
fetch html | {url: str, headers?: dict} | Fetch a website and return the content as HTML — url: URL of the website to fetch; headers: Optional headers to include in the request |
fetch markdown | {url: str, headers?: dict} | Fetch a website and return the content as Markdown — url: URL of the website to fetch; headers: Optional headers to include in the request |
fetch txt | {url: str, headers?: dict} | Fetch a website, return the content as plain text (no HTML) — url: URL of the website to fetch; headers: Optional headers to include in the request |
fetch json | {url: str, headers?: dict} | Fetch a JSON file from a URL — url: URL of the JSON to fetch; headers: Optional headers to include in the request |
scholarly
Use scholarly <command> '<JSON arguments>'. Run scholarly <command> --help to get more information.
| Command | Arguments | Description |
|---|---|---|
scholarly search-arxiv | {keyword: str} | Search arxiv for articles related to the given keyword. |
scholarly search-google-scholar | {keyword: str} | Search google scholar for articles related to the given keyword. |
git
Use git --help for Git operations on workspace repositories.
python3
Execute Python code directly under the agent workspace and return stdout and stderr. Code can be pasted into a .py file and run with python3.
Python cannot launch subprocesses or make direct network requests. Run the documented service commands directly in the shell. Use Python for local file processing and generating JSON arguments.
| Model | VFS | MCP | VFS Tokens | MCP Tokens | Token Reduction Rate |
|---|---|---|---|---|---|
| Claude Opus 5.5 | 1/1 | 1/1 | 180k | 474k | 62% |
| GPT-6.1 Sol | 0/1 | 0/1 | 315k | 1.40M | 77% |
| GPT-5.6 Luna | 1/1 | 0/1 | 9.52M | 27.1M | 65% |
| Claude Haiku 5.5 | 1/1 | 1/1 | 1.52M | 4.39M | 65% |
| Gemini 3.8 Flash | 0/1 | 0/1 | 6.51M | 6.41M | −2% |