ChainSWE
A benchmark for continuous software engineering tasks in which an agent handles multiple interrelated tasks in sequence on the same codebase, with each task inheriting the repository state left by the preceding tasks.
Timeline (last 90 days)
On the ChainSWE benchmark, CoReMem raises a 27B-parameter actor model's solve rate from 21.0% to 25.7%.
On the ChainSWE benchmark, compared with an untrained actor model without memory, CoReMem reduces total steps by 17% and cache-aware total tokens by 11%, with memory module invocation overhead already included.
ChainSWE is a benchmark for continuous software engineering tasks, designed around having an agent handle multiple interrelated tasks sequentially on the same codebase, with each task inheriting the repository state of the preceding tasks.
All Facts (3)
On the ChainSWE benchmark, CoReMem raises a 27B-parameter actor model's solve rate from 21.0% to 25.7%.
50%UnverifiedOn the ChainSWE benchmark, compared with an untrained actor model without memory, CoReMem reduces total steps by 17% and cache-aware total tokens by 11%, with memory module invocation overhead already included.
50%UnverifiedChainSWE is a benchmark for continuous software engineering tasks, designed around having an agent handle multiple interrelated tasks sequentially on the same codebase, with each task inheriting the repository state of the preceding tasks.
50%