Gather — 27 July 2026

Empirical Studies: Maintainability, Security, and Pair Programming #

• Is LLM-Generated Code More Maintainable & Reliable than Human-Written Code? (arXiv 2508.00700) — Direct empirical comparison study asking the maintainability/reliability question head-on rather than assuming AI code is worse or better by default; complements GitClear’s aggregate longitudinal data (already tracked) with a controlled comparison framing.
• Assessing the Quality and Security of AI-Generated Code: A Quantitative Analysis (arXiv 2508.14727) — Quantitative analysis pairing code-quality metrics with security posture in the same study, rather than treating correctness/maintainability and security as separate literatures.
• How to Scale Code Quality for AI-Generated Code (Sonar, 2026) — Follow-up to Sonar’s “Great Toil Shift” report (already tracked) shifting from diagnosis to prescription: organizational-scale static-analysis and quality-gate practices specifically for teams where a large share of new code is AI-generated.
• The Impact of AI-Pair Programmers on Code Quality and Developer Satisfaction: Evidence from TiMi studio (ACM, DOI 10.1145/3665348.3665383) — Industrial case study inside TiMi Studio (Tencent games division) measuring code-quality and satisfaction effects of AI pair programming; identifies six distinct developer-experience failure modes — reliability-questioning, explainability-questioning, trust-lacking, communication-lacking, autonomy-losing, motivation-losing — alongside the quality gains, i.e., quality improvement and developer friction rise together rather than trading off.

Benchmarks and Evaluation Methodology #

• ProjDevBench: Benchmarking AI Coding Agents on End-to-End Project Development (arXiv 2602.01655) — Benchmark targeting whole-project development rather than isolated function/snippet generation, addressing a known gap in HumanEval/MBPP-style benchmarks that this topic’s existing entries (e.g. arXiv 2605.09059) have already flagged as too narrow.
• CodeJudgeBench: Benchmarking LLM-as-a-Judge for Coding Tasks (arXiv 2507.10535) — Benchmarks how reliable LLM-as-a-judge scoring is specifically for code (as opposed to general text), relevant given how many quality studies in this space now lean on LLM judges rather than human review.
• Are They All Good? Evaluating the Quality of CoTs in LLM-based Code Generation (arXiv 2507.06980) — Shifts the evaluation target from final generated code to the quality of the reasoning trace (chain-of-thought) that produced it — a distinct signal from output-only correctness benchmarks.

Testing, Regression, and Verification Practices #

• TDAD: Test-Driven Agentic Development — Reducing Code Regressions in AI Coding Agents via Graph-Based Impact Analysis (arXiv 2603.17973) — Graph-based impact analysis method that reduced regressions by 70% versus a vanilla baseline on SWE-bench Verified; a concrete, benchmarked technique rather than a general call for “more testing.”
• AgentAssay: Token-Efficient Regression Testing for Non-Deterministic AI Agent Workflows (arXiv 2603.02601) — Proposes agent-specific coverage metrics and mutation testing with domain-specific operators for stochastic agent systems, arguing traditional deterministic test-coverage tooling doesn’t transfer cleanly to agent workflows.
• From Prompting to Verification: How Experience Shapes Vibe Coding Practices (arXiv 2605.24521) — Studies how developer experience level shifts practice away from prompt-only reliance toward active verification habits; the verification-practice angle is squarely a code-quality concern even though the paper is framed in vibe-coding terms (see cross-link).

Cross-links #

• AI Code Review The "AI code review" best practices keyword now returns almost exclusively vendor tool-marketing content (Graphite, CodeAnt, BSWEN, DataAnnotation, exceeds.ai, collinwilkins) plus one substantive academic benchmark — Benchmarking and Studying the LLM-based Code Review (arXiv 2509.01494) — which belongs with the newly split-out ai-code-review topic rather than here.
• AI Agent Accountability The AI generated code postmortem OR incident keyword returned no new incident material this cycle — results were dominated by unrelated “AI writes your postmortem” SRE tooling (Rootly, ilert, Blameless), confirming Harper Foley’s catalogue (already captured 2026-07-26) is the sole substantive source in this vein so far; new hits on that thread should route to ai-agent-accountability.
• Vibe Coding Approaches “From Prompting to Verification” (arXiv 2605.24521, included above) is framed by its authors as a vibe-coding practices study; flagging here in case vibe-coding’s own gather surfaces it independently, to avoid duplicate treatment.

Meta-observations #

• Noise pattern: "AI code review" best practices surfaces almost entirely vendor tool-marketing blogs post-split (Graphite, CodeAnt, BSWEN, DataAnnotation, exceeds.ai, collinwilkins) — confirms this keyword’s signal now belongs to ai-code-review, not here.
• Method note: AI generated code postmortem OR incident returns near-100% off-topic results (SRE postmortem-generation products, not AI-code-generation incidents) since the accountability split — Harper Foley’s piece remains the only substantive hit found across two gather cycles.
• Keyword suggestion: Consider retiring "AI code review" best practices from this topic’s config now that ai-code-review exists as a sibling topic — it is producing near-zero new code-quality-proper signal per this cycle’s results.
• Gap: Vendor blogs (e.g. Tembo) continue to cite a “CMU SEI: AI code introduces up to 35% more technical debt” statistic with no traceable primary source — same unverified-vendor-stat pattern flagged last cycle for Diffblue/CodeRabbit; still unresolved.
• Emerging pattern: Two new arXiv papers this cycle (TDAD, AgentAssay) independently argue traditional test-coverage tooling doesn’t fit agent-generated/non-deterministic code, proposing graph-based regression-impact analysis and agent-specific mutation testing respectively — testing methodology for AI-agent code looks to be emerging as its own research thread worth tracking.