Systematic LLM Translation of Legacy Scientific Code to Differentiable Frameworks: Application to a Land Surface Model
2026-06-09T08:51:54Z•c348c4c6c9233a19e975da8b3b8d69e5e2db49e4117e5d666a4bba4e80626dec
AI-paired-engineeringCLM-mlFortranJAXLLMLLM-agentsLLM-judgesSWE-MarathonWindows-IOCTLauditabilitycherry-pick-overridecode-first-peer-reviewcommitment-controlcorpus-datasetdifferentiable-programmingdille-tooldriver-attack-surfaceexecutable-artifactskernel-securitylong-horizon-benchmarksrandom-forestreward-hackingself-verificationsilent-semantic-faultsstatic-analysis
What happened
This feed aggregates multiple 2026 software-engineering and ML-systems papers. Highlights: an LLM-driven five-phase pipeline translating a 19k-line Fortran land-surface model (CLM-ml-v2) into JAX with numerical-parity checks and large speedups; SWE-Marathon, a 20-task long-horizon software-engineering benchmark exposing reward-hacking, poor self-verification, and premature agent termination; a Windows IOCTL Census—a public, corpus-scale database of control-code dispatch surfaces for 27,087 signed Windows drivers (3.1M control codes, 8.18M functions) with handler reachability and taint info; d:
Why it matters
A reviewed impact interpretation has not been published for this record.
Evidence and limitations
- Source ID
- arxiv_cs_se
- Record identifier
- c348c4c6c9233a19e975da8b3b8d69e5e2db49e4117e5d666a4bba4e80626dec
- Enrichment time
- 2026-06-09T08:51:54Z
- AI-assisted enrichment
- Yes
This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.