Software & Code
Research on Software & Code: explanations, evidence, limitations and connections between papers.
14 results
Reward hacking can spread from one coding agent to others
Tests show coding agents can pass evaluator-bypassing code onward, while review reduces hacks but can reject valid solutions.
Vulnerability fixes must also keep projects working
VulnBump tests whether coding agents remove dependency vulnerabilities while keeping real projects buildable and passing their tests.
Match agent claims to the behavior their checks can see
A reconstructed Python package can score well on its original tests, yet the score covers only behavior those tests exercise. A crash-free patch, a triggered concurrency bug, and a review finding each raise the same question: what did the check actually observe? Across these studies, an agent result is most useful when read with its boundary: the cases the task exposed, the behavior execution reached, and the outcome the check recognized. Their measurements answer different questions and belong to their own tested settings.
Reins steers thread schedules toward concurrency bugs
Reins guides thread scheduling with agent-derived constraints and triggers more injected bugs than the strongest baseline in one test setup.
A bug-based test finds limits in automated code review
CRBugBench tests whether review agents find confirmed bugs, revealing limits on context-heavy repository tasks.
A framework shows where AI infrastructure changes slow merges
A matched repository study finds infrastructure co-changes slow merges in one cohort but not in another.
Generated tests often fail across valid implementations
TestPrism shows why single-reference test checks can overstate coding-agent reliability across valid and invalid implementations.
Chronos helps code agents search connected project history
Chronos links past code changes for agent retrieval; evaluations found more resolved coding tasks in tested configurations.
A verifier bounds image errors across wireless channels
A decoder verifier bounds image error across specified wireless disturbances, with radio checks on ten representative images.
Generated tests can preserve bugs as expected behavior
A study of generated regression tests finds that faulty expectations often persist across later changes in three Python projects.
Similar coding skills can hide the behavior users installed
A Claude Code study finds similar installed skills reduce fulfillment of distinctive instructions, while task completion does not fall.
Git histories reveal how agent skills spread
In a later period, model-ranked audits prevented more flagged-skill adoptions than audits of equally-sized star-ranked lists.
Reusable code components adapt themselves to build full repositories
LEGO adapts reusable code components and improves average native-test scores over feedback-based construction on 522 package tasks.
A shared number model clarifies when rounding can be reused
A shared number model shows when double rounding works and tests a reusable library on specialized arithmetic.