Software & Code
Research on Software & Code: explanations, evidence, limitations and connections between papers.
27 results
Future-frame clues help neural codecs compress video
GFSC adds limited future-frame context to video codecs, reducing bitrate at comparable picture quality in tested models and datasets.
Inherited software work exposes a gap in coding agents
SWE-Residual finds lower coding-agent success on inherited work than matched clean starts across repository and machine tasks.
Verifier acceptance can miss weak generated contracts
In the study, only 42.3% of accepted contracts cleared checks for vacuity, test cases and reference agreement.
PACE-Bench checks whether coding agents deliver full requests
PACE-Bench evaluates full PRD delivery in existing repositories; the top tested configuration completed 6 of 43 cases.
Inspect behavior and provenance before reusing a coding capability
A coding team might call a specialist agent that helped before or copy a skill file from another repository. Reuse can spare repeated work, yet a prior passing solution can also carry a way to game a test. The studies suggest a practical first step: check the capability’s behavior in the current task and trace its source. Evidence from an agent library, coder-to-coder imitation and GitHub skill copies supports that judgment in different tested settings.
VulnSAGE checks the code evidence behind security warnings
On 187 labeled reports, VulnSAGE rejected 89 of 97 false alarms and retained 85 of 90 real defects.
Check which failure an automated pass could expose
A passing generated test may look reassuring even when its expected result comes from faulty code. A test checked against one repaired program may also misjudge other valid implementations, while a thread failure may stay hidden until a particular execution order occurs. Across three studies, the useful question is whether the check was placed in a position to distinguish the failure at issue. Their evidence comes from injected code faults, curated implementation panels, and controlled thread executions; each verdict stays tied to those evaluated conditions.
Place review where an uncertain signal would change the dataset
A brain scan can resemble another scan without proving that both belong to one person. A math problem can resemble a benchmark question without being copied from it. In either case, acting on resemblance changes the dataset. Across identity auditing, math-data cleaning, and multilingual example construction, the useful question is where an uncertain signal meets a decision. The studies place review at different points in those workflows and measure different outcomes, so their results support a placement principle rather than one shared estimate of review’s benefit.
Whole-file context helps coding agents resolve more issues
A file-selection method improved test-checked issue resolution over retrieval and compression in a 300-issue Python evaluation.
Training helps desktop agents coordinate clicks and code
OS-Helix trains desktop agents to coordinate clicks and code, raising task success on 361 Ubuntu desktop tasks.
Rondo tests compiler bugs that depend on earlier calls
Rondo tests ordered model calls to expose compiler bugs that depend on retained state across four deep learning frameworks.
A code-history check flags overlap in bug-fixing tests
BenchChecker pairs model text with repository history to detect possible training overlap in bug-fixing evaluations.
Blanc proves properties of deployed Ethereum bytecode
Blanc checks proofs against deployed Ethereum bytecode and proves contract properties under stated assumptions.
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 automation claims to the behavior each check observes
A reconstructed package may pass tests, a controlled run may expose a concurrency bug, and a generated patch may resolve an issue. Each sounds like automation success, but each workflow reaches a different behavior and applies a different check. Read the result through that check. In these studies, native tests assess package behavior, executions reveal triggered failures, and benchmark tests grade applied patches. The claim each result supports stays within its task and study conditions.
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.