Science of Computer Programming is dedicated to the distribution of research results in the areas of software systems development, use and maintenance.
Anıl Sezgin introduces a Retrieval-Augmented LLM framework that pairs semantic similarity search with an LLM to estimate story points.
Tested on 23,313 TAWOS tasks across 17 projects, it hit 37.18% Pred(25), beating a vanilla LLM (27.96%).
https://t.co/Ur6TxyKYio
PRIME (Chen et al.) is a semantic-driven framework that prioritizes microbenchmarks by retrieving change context and using an LLM to infer performance impact.
Across 10 Java projects, it improved APFD-P by 14.5% on average over 9 baselines.
https://t.co/vrFrKNYIhc
MDD4CPS (Navarro et al.) is an agent-oriented, model-driven process with a DSL and a https://t.co/gMQGw38xUo tool that transforms models into code, keeping traceability.
A greenhouse automation proof of concept on Arduino units shows its feasibility.
https://t.co/JUGeiou21r
Bergstra and Tucker extend common meadows (fields where division by zero yields a default value) with a binary logarithm and auxiliary operations.
The study explores methods to define entropy and other information measures.
https://t.co/9HyprRRC7d
TDD-TCNN (Yu et al.) combines a CNN and a Transformer to detect method-level technical debt using both code and comments.
It has been trained on 100K+ verified Java methods and achieves state-of-the-art results, with an 82% F1 Score.
https://t.co/RbBPlJkkbb
Nahar Lota et al. systematically reviewed 87 papers (2011–2025) on NLP techniques for bug report analysis, reproduction, and localization.
It presents a taxonomy of methods and best practices and identifies key challenges like scalability and accuracy.
https://t.co/NtTjEBwaGq
Fares et al. introduce reusable patterns for automatically generating Event-B refinements using predicate transformer techniques.
Generated models are auto-simplified for readability, with effectiveness shown on two classic case studies.
https://t.co/iTLtPWxHUs
Justin (Schmitz et al.) brings hybrid CPU+memory auto-scaling to Apache Flink, extending the Flink Kubernetes Operator to fine-tune parallelism and memory per operator instead of one-size-fits-all scaling.
https://t.co/aq3Pu3mdG3
Zhang and Wang propose a Hidden Markov Random Field-based fault detection and RBF NN-based fault tendency prediction method for SaaS application modules.
It achieves high precision/recall for fault judgment and prediction, with low energy consumption.
https://t.co/W5dqv8Uxck
uvl2dimacs (Heradio et al.) converts Universal Variability Language models into DIMACS format using Tseitin transformation and backbone simplification.
It cuts clauses in 99% of cases (median 40%) on 1533 real-world models.
https://t.co/RDGLiDboD3
Li et al. evaluate 24 Graph-GNN combinations (6 program graphs x 4 GNN model types) for software defect prediction across 7 Java projects.
CFG-GCN performs best for within-project prediction, while CPG-GAT excels for cross-project prediction.
https://t.co/kWO68wGytN
Van Cuyck et al. present 3 approaches for checking mutual acceptance in specification models.
These are an algorithmic specification check, a two-procedure conformance test, and an optimized combined version that trades genericity for efficiency.
https://t.co/nS3P1XD1Gy
RebeCaos (Proença and ter Beek) is a web-based front-end tool for Rebeca that allows simulating different (timed) Rebeca semantics.
It offers initial reachability analyses for Rebeca models, e.g., to detect deadlocks or check desirable states.
https://t.co/oKq7GhABj3
ArithHomFA (Waga et al.) uses fully homomorphic encryption for oblivious online monitoring of discrete-time STL, allowing a remote monitor to check signal properties without decrypting the data.
It has been used to monitor vehicle behavior.
https://t.co/iATPibmMSZ
C2VR (Xia et al.) migrates C code with ACSL specs into verifiable Rust with Verus specs, verifying 139/175 programs (92 automatically).
Failures are classified into 15 symptoms, 10 root causes, and 9 repair strategies.
https://t.co/C5QwRkDPBY
Martínez-Castañón et al. propose a DSL to optimize periodic vehicle fleet maintenance scheduling using Model-Driven Engineering.
Usability tests show it reduces time and effort compared to CMMS tools and increases user satisfaction.
https://t.co/H4g1XFAQPt
Kashef and Ghassemi model social network user behavior by clustering users and learning automata (L*, KV, L*P) with SVM classification.
The deterministic models achieved 65–75% accuracy, improving up to 12% with probabilistic strategies.
https://t.co/YtnKfbX8Xp
Rhazi et al. apply contrastive learning with inter-sample attention for defect prediction, capturing relationships between software classes beyond raw metrics.
The method outperformed 5 baselines on 9 projects across 3 benchmarks.
https://t.co/nKe03Z2IE5
Benjamin and Mathew triangulate Stack Overflow discourse, GitHub adoption data, and GitHub Actions telemetry to predict CI/CD reliability risks.
Deep representation learning outperforms linear baselines at early-warning recall for repository instability.
https://t.co/B7F9rbtL2a
AugmenTest (Khandaker et al.) uses LLMs to generate JUnit assertions from natural language docs/comments alone.
It was tested on a curated benchmark of real-world Java test cases.
https://t.co/JnEoFsh1Xq