Researchers introduce Nyarlathotep, described as the first choreographic programming framework that combines asynchrony with nondeterminism. They define operational semantics using labelled transition systems, prove endpoint projection sound and complete, and mechanise the results in Lean 4.
Software development, languages, tools, and the craft of building software
MetaSieve is a metapath selection layer for relational deep learning that uses SQL join and aggregation statistics to prune uninformative foreign-key metapaths. Evaluated on RelBench with several GNN backbones, it reduced per-epoch training time while maintaining or often improving accuracy.
A study samples 2% of GitHub accounts to examine end-to-end commit signing and how developers manage cryptographic keys. It analyzes 16,112,439 commits across 874,198 repositories, compares findings with prior repository-based research, and proposes a path toward more usable key management.
Biomedical knowledge graphs combine molecular, clinical and population data, yet assembling them into usable knowledge remains hard. The authors argue that limited adoption of software engineering practices, including package management, typed namespaces and reproducible build pipelines, is a contributing cause, and they catalogue eight open integration challenges.
The paper defines a semantic graph-matching distance for comparing no-code and agentic workflows as attributed graphs instead of strings. On 4,705 public n8n exports, evaluations showed near-zero mean distance under surface-preserving changes and higher values under major type or subgraph changes.
A cost-based framework selects an efficient relational schema from an E/R model and representative workload, optimizing mapping decisions such as inheritance and entity or relationship representations. Real-world-inspired experiments show mapping choices affect performance; fixed or heuristic strategies vary widely, while the approach remains competitive.
The paper proposes VerifyPrompt, a method for checking whether third-party text-to-image platforms actually serve the models they advertise. It uses specially designed adversarial prompts and a reinforcement-learning search; experiments report over 90% accuracy across various T2I platforms.
In reinforcement-learning experiments, injected faults at the environment-trainer boundary often produced no error, and learning-curve checks missed most runs unless every value was corrupted. A step ledger kept on both sides detected and located every boundary fault in the reported tests, adding a median 3% training time on Pendulum.
A static-analysis pipeline extracts control-flow and function-call graphs plus behavioral signatures from 3,365 Android proxy apps across four commercial networks, with SGD reaching a 0.985 macro F1 in DEX-grouped cross-validation. Classifier decisions are mapped to Yara rules, and 51.4% of the expanded APKPure sample still contained embedded proxy SDK code.
The paper introduces a framework for validating concurrent correctness witnesses by using semantic information encoded in ghost variables, pairing a checking phase with an exploitation phase. Both phases are cast as thread-modular abstract interpretation, and the authors evaluate resource-invariant and phase-based techniques in a static analyzer using verifier-generated and hand-crafted witnesses.
A new analysis summarizes hierarchical probabilistic cost expressions bottom-up, using an exact short-horizon prefix plus a fitted parametric tail, and reports sound intervals for mean, second moment and variance. Its Dickens tool bounds mean completion time within 1% for quantum-repeater chains of up to 4096 links, where an exact Markov-chain solution runs out of memory at 16 links.
A paper introduces noise fingerprints and an abstract domain with a probabilistic Galois connection to analyze noisy quantum programs. On a synthetic benchmark, fingerprints retained (99 ± 1)% of a full noise model's strength information; on simulated drift, reusing measurements cut fresh rounds for a 0.10 error tolerance by 2.3× at 95% confidence and 27× under a median-error criterion.
The paper argues high-level source code need not remain software’s main persistent artifact if AI agents generate and validate implementations from semantic intent. It reports an experiment producing multiple LLVM artifacts from one frozen semantic specification, checked by an external oracle, showing regeneration coherence rather than proof of correctness.
A systemization-of-knowledge review traces decentralized prediction markets to 2011, covering hundreds of proposals, and reports that modern systems such as Polymarket differ materially from earlier designs including Truthcoin and Augur v1. It outlines an eight-stage modular workflow and assesses design choices and trade-offs in decentralization, expressiveness and manipulation resistance.
PoCoFL separates federated learning type, policy semantics and cryptographic implementation to check client updates and aggregator processing against chosen policies. Its authors formalise admissibility, aggregation correctness and policy-relative privacy and demonstrate proof-of-concept versions for four federated learning settings, assessing verification overhead.
tamserve is an open-source engine that runs one copy of Meta's SAM 2 or EdgeTAM model across multiple live camera feeds, on one GPU or several. It offers a Python API and named-object tracking, with benchmarks reporting lower GPU memory and higher aggregate throughput than separate model copies per camera.
The paper gives matching-logic specifications for initial algebras and final coalgebras of polynomial functors represented as unary container functors. It proves the extracted algebras and coalgebras are W- and M-types relative to shape and position interpretations, derives induction and bisimulation principles, and revises an EPTCS 389 paper with corrected axioms and new proofs.
The article walks through partial evaluation with JavaScript examples, explaining how evaluating known parts of a program can eliminate recursive calls and yield compiled code from an interpreter plus a program. It introduces the first Futamura projection and frames the author's new library as a way to explore the concept.
A SwiftPM evolution pitch proposes lifting restrictions so build tool plugins can generate C, C++ and header files for mixed-language Swift targets. It adds public header search paths and plugin-contributed build settings, enabled with Swift tools version 6.5.
Microsoft published November 2026 patch releases for its Build of OpenJDK that fix CVE-2026-114938, a memory-corruption flaw able to let unauthenticated attackers run code remotely under certain conditions. Microsoft names OpenJDK 17, 21 and 25 as affected and says patching is the primary fix, while firewall mitigations only block known HTTP attacks until upgrades happen.
