Researchers Introduce MPC-Patch-Bench, First Security-Focused Benchmark for LLM Code Repair in Secure Multi-Party Computation
Researchers have created MPC-Patch-Bench, a new benchmark for evaluating how well large language models can repair code in secure multi-party computation (MPC) systems used for privacy-preserving applications. The benchmark addresses gaps in existing evaluation frameworks by incorporating cryptographic security checks and numerical-fidelity verification that general-purpose code repair benchmarks lack. This matters because MPC is increasingly deployed in sensitive domains like biomedical collaboration and secure analytics, where code defects could compromise both functionality and security.
MPC-Patch-Bench is a repository-level benchmark designed to evaluate LLM performance on code repair tasks specific to secure multi-party computation frameworks. The researchers identified three structural limitations of applying general benchmarks like SWE-bench to MPC code: MPC repositories contain substantial generic Python infrastructure alongside cryptographic logic, high-value MPC fixes often lack standardized tests required by existing extraction pipelines, and standard fail-to-pass evaluation cannot assess cryptographic safety. The benchmark includes two main components: a data curation framework that filters pull requests through cryptographic layers and uses human-AI collaboration to synthesize missing test cases (yielding 205 verified instances), and an MPC Verifier that performs security and numerical-fidelity checks through dynamic differential testing and static analysis. Testing revealed that the strongest evaluated LLM resolved only 22.9% of tasks functionally, with the security verifier further reducing verified resolution to 17.1%, indicating that up to 40% of functionally-passing patches fail cryptographic or numerical-fidelity requirements.
What's missing
The study does not discuss which specific LLM models were evaluated or provide comparative analysis across different model architectures and sizes. Additionally, the paper does not address the computational cost or time requirements for the verification process, nor does it discuss potential limitations of the differential testing approach against plaintext oracles for detecting all classes of cryptographic vulnerabilities.
What different sources said
- arXiv cs.AICenter
MPC-Patch-Bench: Security-Aware LLM Code Patch for Multi-Party Computation
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.