12 KiB
| title | excerpt | highlight_image | draft |
|---|---|---|---|
| Aider is SOTA for both the main SWE Bench and SWE Bench Lite | Aider sets SOTA for the main SWE Bench, after recently setting SOTA for the Lite version. | /assets/swe_bench.jpg | true |
Aider is SOTA for both the main SWE Bench and SWE Bench Lite
Aider scored 18.8% on the main SWE Bench benchmark, achieving a state-of-the-art result. The current top leaderboard entry is 13.8% from Amazon Q Developer Agent. The best result reported elsewhere seems to be 13.9% from Devin.
This is in addition to aider's SOTA result on the easier SWE Bench Lite that was reported last week.
Aider was benchmarked on 570 of the 2294 SWE Bench problems. These are the same randomly selected 570 problems that Devin used in their evalulation. Please see the references for more details on the data presented in this chart.
Interactive, not agentic
Aider achieved this result mainly through its existing features that focus on static code analysis, reliable LLM code editing, and pragmatic UX for AI pair programming. Aider intentionally has quite limited and narrow "agentic behavior" to avoid long delays, high token costs and the need for users to repeatedly code review incorrect solutions. It's also worth noting that aider currently does not use RAG, vector search, tools or give the LLM access to search the web or unilaterally execute code.
Aider is first and foremost an interactive tool for engineers to get real work done in real code bases using a chat interface. Aider provides a pair programming UX where users can ask for a change and see the edits performed in real-time. Aider can also offer additional help like fixing lint or test errors, but the user is always in full interactive control. This lets them quickly steer misunderstandings back on course and avoid wasting time and token costs.
Benchmark methodology
Benchmarking was conducted as follows:
- Aider with GPT-4o was launched in each problem's git repository with the problem statement submitted as the opening chat message from "the user".
- After that aider ran as normal, except all of aider's suggestions were always accepted without user approval.
- A simple harness was used to retry the SWE Bench problem if aider produced code that wasn't plausibly correct. Plausibly correct means that aider reported that it had successfully edited the repo without causing syntax errors or breaking any pre-existing tests.
- If the solution from aider with GPT-4o isn't plausible, the harness launches aider to try again from scratch, this time using Claude 3 Opus.
- If no plausible solution is found after those two tries, the harness picks the "most plausible" solution with the fewest edit/lint/test problems.
It's important to be clear that aider and the benchmark harness only had access to the pre-existing tests in each problem's repo. The held out "acceptance tests" were only used after benchmarking to compute statistics on which problems aider correctly resolved.
This is the same methodology that was used for aider's recent SOTA result on SWE Bench Lite. The only difference is that for this result at most two tries were attempted instead of six, due to the increased token costs involved in this benchmark. The SWE Bench problems are more difficult and involve edits to more than one source file, which increased the cost of solving each problem. Further, aider was benchmarked on 570 SWE Bench problems, versus only 300 Lite problems, adding another factor of ~two to the costs.
For a detailed discussion of the methodology, please see the article about aider's SWE Bench Lite results. The aider SWE Bench repository on GitHub also contains the harness and reporting code used for the benchmarks.
The benchmarking process was similar to how a developer might use aider to resolve a GitHub issue:
- They could launch aider in their repo with the command below, which
tells aider they want to accept every suggestion
and to use pytest to run tests.
aider --yes --test-cmd pytest
- They could start the chat by pasting in the URL or text of a GitHub issue. Aider will pull in the URL's content and then try and solve the issue.
- If aider doesn't produce code that lints and tests clean, the user might decide to revert the changes and try again, maybe using aider with a different LLM this time. Aider is tightly integrated with git, so it's always easy to revert AI changes that don't pan out.
Aider with GPT-4o alone was SOTA
Running the benchmark harness only using aider with GPT-4o to find plausible solutions with a single attempt achieved a score of 17.0%. This was itself a state-of-the-art result, before being surpassed by the main result being reported here that used aider with both GPT-4o & Opus.
Aider with GPT-4o & Opus
The benchmark harness started by running aider with GPT-4o once to try and solve the problem. If no plausible solution was found, it then used aider with Opus once to try and solve the problem.
The table below breaks down the proposed solutions that were found for the 570 problems. A proposed solution is either:
- A plausible solution where aider reported no outstanding errors from editing, linting and testing.
- Or, the "most plausible" solution generated by either attempt, with the fewest outstanding editing, linting or testing errors.
The table also provides details on the 107 solutions that were ultimately verified as correctly resolving their issue.
| Attempt | Agent | Number of proposed solutions |
Percent of proposed solutions |
Number of correctly resolved solutions |
Percent of correctly resolved solutions |
Score on SWE Bench Lite |
|---|---|---|---|---|---|---|
| 1 | Aider with GPT-4o | 419 | 73.5% | 87 | 81.3% | 15.3% |
| 2 | Aider with Opus | 151 | 26.5% | 20 | 18.7% | 3.5% |
| Total | 570 | 100% | 107 | 100% | 18.8% |
Non-plausible but correct solutions?
It's worth noting that the first row of the table above only scored 15.3% on the benchmark, which differs from the 17.0% result reported above for aider with just GPT-4o. This is because making additional attempts is not guaranteed to monotonically increase the number of resolved issues. Later attempts may propose solutions which seem "more plausible" than prior attempts, but which are actually worse solutions. Luckily the later attempts usually provide a net increase in the overall number of resolved solutions, as is the case here.
This table breaks down the plausibility of each solution proposed by aider with GPT-4o and with Opus, as well as whether it was actually a correct solution.
| Row | GPT-4o solution plausible? |
GPT-4o solution resolved issue? |
Opus solution plausible? |
Opus solution resolved issue? |
Count |
|---|---|---|---|---|---|
| 1 | plausible | resolved | n/a | n/a | 73 |
| 2 | plausible | not resolved | n/a | n/a | 181 |
| 3 | non-plausible | resolved | plausible | resolved | 1 |
| 4 | non-plausible | resolved | plausible | not resolved | 2 |
| 5 | non-plausible | resolved | non-plausible | resolved | 16 |
| 6 | non-plausible | resolved | non-plausible | not resolved | 5 |
| 7 | non-plausible | not resolved | plausible | resolved | 12 |
| 8 | non-plausible | not resolved | plausible | not resolved | 53 |
| 9 | non-plausible | not resolved | non-plausible | resolved | 4 |
| 10 | non-plausible | not resolved | non-plausible | not resolved | 216 |
| 11 | non-plausible | not resolved | n/a | n/a | 7 |
Rows 1-2 show the case where the first solution found by aider with GPT-4o was plausible. Of those, 73 went on to be deemed as resolving the issue, while 181 were not in fact correct solutions. Opus never got a try at solving these problems, because the harness stopped once a plausible solution was found.
The remaining rows consider cases where aider with GPT-4o did not find a plausible solution, so Opus had a turn to try and solve. Rows 3-6 are cases where GPT-4o's non-plausible solutions were actually found to be correct in hindsight, but in rows 4 we can see that aider with Opus overrides 2 of them with a plausible-but-incorrect solution. The original correct solutions from GPT-4o may not have been plausible because of pre-existing or otherwise unresolved editing, linting or testing errors which were unrelated to the SWE Bench issue or which turned out to be non-fatal.
In rows 5-6 & 9-10 we can see that both GPT-4o and Opus produced non-plausible solutions, and which one was selected has to do with the details about which solution the harness considered "most plausible".
Row 11 contains cases where Opus returned errors due to context window exhaustion or other problems. In these cases aider with Opus was unable to produce any solutions.
Computing the benchmark score
Benchmarking produced one candidate solution for each of the 570 SWE Bench problems.
A separate evaluation script was used to test each of these solutions with the full test suite, including the held out acceptance tests. For this final acceptance testing, any edits that aider made to tests were discarded. This ensured that the correct, unmodified test suite was used for acceptance testing. The evaluation script compared each candidate solution's test results with results from testing the "gold" patch that was developed by a human to correctly solve the issue. If they matched, the candidate solution correctly resolved the issue.
These acceptance tests were only ever run outside of aider and the benchmark harness, and only to compute statistics about the correctly resolved instances. They were never run, used, or even visible during aider's attempts to solve the problems.
Aider correctly resolved 107 out of 570 SWE Bench instances that were benchmarked, or 18.8%.
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.
References
Below are the references for the SWE-Bench results displayed in the graph at the beginning of this article.
- 13.9% Devin (benchmarked on 570 instances)
- 13.8% Amazon Q Developer Agent (benchmarked on 2294 instances)
- 12.5% SWE- Agent + GPT-4 (benchmarked on 2294 instances)
- 10.6% AutoCode Rover (benchmarked on 2294 instances)
- 10.5% SWE- Agent + Opus (benchmarked on 2294 instances)
The graph contains average pass@1 results for AutoCodeRover.
The AutoCodeRover GitHub page
features their pass@3 results
without being clearly labeled.
Table 2 of their
paper
reports an ACR-avg result of 10.59% which is an average pass@1 result.
The official SWE Bench Lite leaderboard only accepts pass@1 results.