aider/_posts/2024-05-22-swe-bench-lite.md

title	excerpt	highlight_image	draft
Aider scores SOTA 26.3% on SWE Bench Lite	Aider scored 26.3% on SWE Bench Lite, achieving a state of the art result.	/assets/swe_bench_lite.jpg	true

Aider scores SOTA 26.3% on SWE Bench Lite

Aider scored 26.3% on the SWE Bench Lite benchmark, achieving a state of the art result. The current top leaderboard entry is 20.33% from Amazon Q Developer Agent. The best result reported elsewhere online seems to be 22.3% from AutoCodeRover.

Interactive, not agentic

Aider achieved this result mainly through its focus on static code analysis, reliable LLM code editing and pragmatic workflows for interactive pair programming with AI. Aider intentionally has quite limited and narrow "agentic behavior": it doesn't require a highly detailed upfront "spec" from the user, use RAG or vector search, farm out sub-problems to an army of LLMs, allow the LLM to use tools or perform web searches, etc.

Aider is first and foremost a tool for engineers to get real work done in real code bases through a pair programming chat style interface. In normal use, the user is in full interactive control. This lets them quickly steer misunderstandings back on course and avoid wasted time, code reviews and token costs. When a user asks aider for a change, they see the edits performed in real-time. Aider may also then offer additional help like fixing lint or test errors.

Methodology

For the benchmark, aider was launched in each problem's git repository with the problem statement submitted as the opening chat message from "the user". After that aider runs as normal, with the following modifications:

• Aider's suggestions were always accepted. When chatting, aider will suggest which files in the repo may need to be edited based on the conversation. It will offer to lint code that has been edited, and to fix any issues uncovered. Aider has workflows to run the repo's test suite and resolve failing tests. Normally the user is asked to approved such suggestions, but they were always accepted during the benchmark.
• A simple harness was used to retry the SWE Bench problem if aider produced code which wasn't plausibly correct. Plausible means that aider successfully edited the repo without breaking anything. As mentioned, aider has integrated support for linting and testing, so the harness just looks at aider's completion status to see if those operations finished clean. Note that aider only had access to the pre-existing tests in the repo, not the held out "acceptance tests" that are used later to see if the SWE Bench problem was correctly resolved.
• If the solution isn't plausible, the harness launches aider to try again from scratch. The harness alternates between running aider with GPT-4o and Opus up to three times each, until it finds a plausible solution.
• If no plausible solution is found, the harness picks the solution with the least amount of edit/lint/test problems.

This is all roughly equivalent to a user:

• Launching aider in their repo with the something like command below, which tells aider to say yes to every suggestion and use pytest to run tests.
  • aider --yes --test-cmd pytest
• Pasting the text of a GitHub issue into the chat, or adding it via URL with a command in the chat like:
  • /web https://github.com/django/django/issues/XXX
• If aider doesn't produce code that lints and tests clean, the user might decide to revert the changes and try again. Maybe with a different LLM this time. Aider is tightly integrated with git, so it's always easy to undo/revert AI changes that don't pan out.

Of course, outside a benchmark setting it's probably unwise to let any AI agent run unsupervised on your code base. Aider is intended to be used as an interactive pair-programming chat, where the user participates to direct aider's work and approve suggestions. This way the user can offer immediate feedback or corrections if their initial instructions turn out to be ambiguous, or if the AI starts going down a wrong path.

Aider with GPT-4o alone was SOTA

Running the entire SWE Bench Lite benchmark using aider with just GPT-4o achieved a score of 25%. This was itself a state of the art result, before being surpassed by the main result being reported here that uses aider with both GPT-4o & Opus.

GPT-4o vs Opus

The benchmark harness alternated between running aider with GPT-4o and Opus. The harness proceeded in a fixed order, always starting with GPT-4o and then alternating with Opus until a plausible solution was found.

The table below breaks down the 79 solutions which were ultimately verified as correctly resolving their task. Some noteworthy observations:

• Aider with GPT-4o immediately found 77% of the valid solutions on the first attempt.
• ~90% of valid solutions were found after one attempt from aider with GPT-4o and Opus.
• A long tail of solutions continued to be found by both models including on the final, 6th attempt.

Attempt	Model	Number resolved	Percent of resolved	Cumulative percent of resolved
1	GPT-4o	61	77.2	77.2
2	Opus	10	12.7	89.9
3	GPT-4o	3	3.8	93.7
4	Opus	2	2.5	96.2
5	GPT-4o	2	2.5	98.7
6	Opus	1	1.3	100.0
Total		79	100%	100%

If we breakdown correct solutions purely by model, we can see that GPT-4o dominates. This isn't a fair comparison, because GPT-4o always took the first attempt at solving. But anecdotal evidence from early runs of the benchmark supports the observation that GPT-4o is significantly stronger than Opus for this endeavor.

Model	Number resolved	Percent of resolved
GPT-4o	66	83.5
Opus	13	16.5
Total	79	100%

Repository map, not RAG

The crucial first step in solving a SWE Bench problem is figuring out which parts of the repo are relevant and which files need to be edited. Most coding agents use some combination of RAG, vector search and arming the LLM with tools to interactively explore the code base.

Aider instead uses a repository map to help the LLM understand the layout, code structure and content of a git repo. The repo map is created from the code's AST and call graph to provide a compact and powerful summary of the entire code base. The map is constantly tailored to show repo context that is relevant to the current state of the chat conversation. This is done by performing a graph optimization on the code's call graph.

When the user asks for a change to their code, the LLM uses the repo map to decide which files to edit. The LLM simply returns a normal text response explaining which files it needs to edit and why. Aider notices when the LLM mentions filenames from the repo, and asks the user if they should be added to the chat. Adding a file to the chat allows the LLM to see the full contents of the file and edit it.

Please add a new /factorial/N endpoint.

To add a new /factorial/N endpoint, the most likely file that needs to be edited is app.py. Please add app.py to the chat so I can proceed with the changes.

app.py
Add these files to the chat? yes

This is a convenient and natural workflow for interactive chat, and it worked well for the SWE Bench tasks. Each task comes with a “gold” patch, which was created by a human developer to solve the issue. Aider successfully identified and added the file from the gold patch in 70.3% of the benchmark tasks.

Of course aider is not able to see or use the gold patch or the files it names in any way. They were only used to compute this statistic after the benchmarking was completed.

Reliable code editing

Once files have been selected for editing, the next step is of course to edit the source code to fix the problem.

Aider has always had a deep focus on ensuring that LLMs can not just write code, but reliably edit code. Aider a collection of prompting strategies and code editing backends which have been honed through extensive benchmarking. These foundational capabilities help ensure that the LLM can not only code up a solution but also properly integrate it into the existing code base and source files.

The repository map helps here too, making sure that the LLM can see relevant classes, functions and variables from the entire repo. This helps ensure that the project's existing APIs and conventions are respected when new code is added.

Linting and fixing

Aider lints code after every LLM edit, and offers to automatically fix any linting errors. Aider includes basic linters built with tree-sitter that support most popular programming languages. These built in linters will detect syntax errors and other fatal problems with the code.

Users can also configure aider to use their preferred linters. This allows aider to check for a larger class of problems, keep the code style aligned with the rest of the repo, etc. But for the benchmark, aider simply used its built-in linters.

Aider shows linting errors to the LLM in a novel format, using the abstract syntax tree (AST) to display relevant code context for each error. This context increases the ability of the LLM to understand the problem and make the correct changes to resolve it.

app.py:23:36: F821 undefined name 'num'
app.py:41:16: F541 f-string is missing placeholders

app.py:
...⋮...
6│class LongNum:
7│ def init(self, num):
8│ """
9│ Initialize the number.
10│ """
...⋮...
19│ def str(self):
20│ """
21│ Render the number as a string.
22│ """
23█ return str(num)
24│
25│
26│@app.route('/subtract/int:x/int:y')
...⋮...
38│@app.route('/divide/int:x/int:y')
39│def divide(x, y):
40│ if y == 0:
41█ return f"Error: Cannot divide by zero"
42│ else:
43│ result = x / y
44│ return str(result)
45│
...⋮...

Attempt to fix lint errors? yes

Testing and fixing

Aider can be configured with the command needed to run tests for a repo. A user working on a python project might do that by launching aider like this:

aider --test-cmd pytest

The repositories that are used in the SWE Bench problems are large open source projects with extensive existing test suites. A repo's test suite can be run in three ways:

1. Run tests as they existed before trying to solve the problem, without any changes.
2. Run tests after aider has modified the repo. So the pre-existing test cases are still present, but may have been modified by aider. Aider may have also added new tests.
3. Run the final "acceptance tests" to judge if the coding agent has successfully resolved the problem. SWE Bench verifies both pre-existing tests and a set of held out acceptance tests (from the so called test_patch) to check that the issue is properly resolved. During this final acceptance testing, any aider edits to tests are discard to ensure a faithful test of whether the issue was resolved.

For the benchmark, aider is configured with a test command that will run the tests as described in (2) above. So testing will fail if aider has broken any pre-existing tests or if any new tests that it created aren't passing. When aider runs a test command, it checks for a non-zero exit status. In this case, aider will automatically share the test output with the LLM and ask it to try and resolve the test failures.

To be clear, aider can not run or even see the "acceptance tests" from the test_patch as described in (3). Those tests are only run outside of aider and the benchmark harness, to compute the final benchmark score.

Finding a plausible solution

As aider executes, it notes the outcome of the editing, linting and testing steps. When aider completes, it returns their final status as either: succeeded with no errors remaining, or ended without resolving all errors.

The benchmark harness uses these outcomes to determine if it has a plausible solution to the current SWE Bench task. A plausible solution is one where aider returns saying that it edited the repo with no outstanding edit, lint or test errors. In this case, aider's changes are taken as the proposed solution and recorded as the SWE Bench model_patch to be evaluated later with the test_patch "acceptance tests".

If the solution is not plausible, another instance of aider is launched again from scratch on the same problem. The harness alternates asking GPT-4o and Opus to solve the problem, and gives each model three attempts -- for a total of six attempts. As soon as a plausible solution is found, it is accepted and the harness moves on to the next SWE Bench instance.

It's worth noting that repositories may have lint or test errors present before aider even starts to edit them. Whether errors are caused by aider or were pre-existing, there will be instances where, after six tries, no plausible solution is obtained.

If all six attempts fail to produce a plausible solution, then the "best" solution available is selected as a the model_patch. Which of the non-plausible solutions to use is determined by ignoring the testing outcome and prioritizing solutions in the following order:

• Pick a solution where editing and linting were completed successfully.
• Pick a solution where editing was at least partially successful and linting succeeded.
• Pick a solution where editing was successful.
• Pick a solution where editing was at least partially successful.

Computing the benchmark score

The benchmark harness produces one "best" solution for each of the 300 SWE Bench Lite instances, and saves it as a model_patch. A separate evaluation script uses the SWE Bench support code to test each of these results with the acceptance tests.

These test_patch acceptance tests are only ever run outside of aider and the benchmark harness, and only to compute the number of correctly resolved instances. They are never run, used or even visible during the attempts to solve the problems.

Aider correctly resolved 79 out of 300 SWE Bench Lite instances, or 26.3%.

Acknowledgments

Much thanks to the team behind the SWE Bench family of AI coding benchmarks. Also thanks to Albert Örwall who has dockerized the SWE Bench evaluation scripts making it faster, easier and more reliable to run the acceptance tests.