AI software papers prioritize measurable confidence and controlled agent runtimes

Two papers give concrete ways to score generated code before it enters a larger workflow. Code Is More Than Text combines token entropy spikes, sampled pseudo-code agreement, and self-generated tests. Across five code large language models (LLMs) and four benchmarks, its ensemble raises average area under the receiver operating characteristic curve (AUROC) to 0.776, up 8.1 points over the strongest natural-language-derived baseline.

FASE, short for Fast Adaptive Semantic Entropy, takes a cheaper route. It embeds 10 code samples per task, clusters them with a minimum-spanning-tree rule, and computes entropy over the clusters. On HumanEval and BigCodeBench-hard, the Qwen3-Embedding-8B version reports a 25% average gain in Spearman correlation with Pass@1 and runs at about 0.3% of the cost of LLM-based semantic entropy.

Structured output control adds a boundary case. Template Token Match Generation nearly removes syntax errors for JSON, SQL, code, and function calls, yet schema and value errors remain. Format control helps pipelines parse outputs; it does not certify that the generated artifact is the right one.

Agent papers focus on the runtime layer around the model: what context is kept, which tools can change state, and how enterprise integrations fail. Less Context, Better Agents gives the sharpest measurement. In a 50-task Microsoft Dynamics 365 Finance and Operations hotel-expense benchmark, keeping the last five tool interactions plus a short summary reaches 91.6% completion while using 553,374 tokens. Full history reaches 71.0% completion with 1,480,996 tokens.

The harness definition paper turns a vague term into four operational requirements: an agent loop, environment-changing tools, task-aware context management, and controls that do not rely on model obedience. It applies the test to Claude Code, Codex CLI, Aider, Cline, OpenHands, and SWE-agent.

Enterprise Model Context Protocol (MCP) evidence points to the same control surface. Interviews with 20 practitioners find that all considered MCP important, while all also named fast fault localization as the main troubleshooting obstacle. Context rot adds a maintenance warning: in 356 repositories, 23.0% had at least one stale AI configuration reference.

Several papers build evidence trails for artifacts that AI systems generate or inspect. TestMap records the lifecycle of foundation-model-generated unit tests in C#/.NET repositories. It stores build results, execution results, coverage, mutation signals, test smells, repair attempts, prompts, and model settings. The paper does not report a benchmark win, so its contribution is observability at candidate level.

MLC attacks bug localization cost. It adds a small bug/no-bug head to a frozen code LLM and predicts all buggy lines with one generated token per file. On full-file Defects4J, MLC Qwen1.7B with parameter-efficient tuning reaches Top-5 accuracy of 39.5%, ahead of the listed Ochiai and DeepFL Top-5 baselines.

ATTAIN applies trace evidence to security maintenance. It runs public exploits across historical Java library versions, compares execution divergence, asks an LLM to inspect relevant diffs, and labels affected versions. Its evaluation covers 224 CVEs, 25,943 versions, and 128 libraries, with a reported F1 score of 93.24%.

SIGA shows a practical pattern for specialized scientific software. It keeps the coding-agent harness fixed and adds simulator-specific retrieval, a short procedural memory, an XML validator, and a stop rule. On a representative GEOS task, it produces a complete input deck in about five minutes with TreeSim above 0.90, matching a human expert who spent about three hours. On a harder held-out GEOS set, grounding raises TreeSim from 0.720 to 0.789.

The network-operations paper reports a larger production claim: a multi-agent system for hyperscale incident response resolves more than 90% of well-understood incident categories autonomously, with mean time to resolution reduced from hours to minutes. The excerpt lacks raw incident counts and per-category tables, so the claim is useful as architecture evidence and weaker as an independently checkable benchmark.