AI
New AI System Uses Runtime Tracing to Fix Complex Software Bugs More Efficiently
Researchers introduce DAIRA, an automated repair agent that embeds dynamic analysis into its decision loop to resolve code defects with 79.4% success on challenging benchmarks.
Identifying and fixing intricate code defects from written descriptions has long been a difficult task in software engineering. Recent advances in large language models have enabled the creation of intelligent agent systems capable of automated bug repair, yet these tools often struggle when facing complex problems involving convoluted control flows or subtle type-related issues.
The Challenge with Current Approaches#
Existing automated repair methods typically depend on static code analysis and basic execution feedback. This approach leaves them unable to observe what happens during program execution—the intermediate states, variable values, and call sequences that often hold the key to understanding failure. As a result, these agents frequently engage in trial-and-error exploration, expending enormous numbers of tokens while never pinpointing the actual root cause.
DAIRA's Dynamic Analysis Strategy#
Researchers have unveiled DAIRA (Dynamic Analysis-enhanced Issue Resolution Agent), a framework that fundamentally changes how automated repair systems approach bug fixing. Rather than relying solely on static inspection, DAIRA weaves dynamic analysis—the monitoring of live execution—directly into the agent's reasoning workflow.
The system uses a Test Tracing-Driven methodology that deploys lightweight instrumentation to gather runtime evidence such as call stacks and variable snapshots. This raw data is then transformed into structured, semantically meaningful reports. By capturing the actual execution paths and dependencies between operations, DAIRA can perform precise fault localization without cluttering the agent's context window with irrelevant code snippets. The result is a shift from speculative guessing to more grounded, deterministic problem-solving.
Benchmark Results and Efficiency Gains#
Testing on the SWE-bench Verified benchmark revealed impressive outcomes. When paired with Gemini 3 Flash Preview, DAIRA achieved a 79.4% resolution rate—a new high for the field. Even on the most challenging defects involving deep logical errors, the system managed to fix 44.4% of cases, demonstrating robustness in tackling problems that typically stymie other approaches.
Beyond success rates, DAIRA also delivers practical efficiency improvements. Across multiple language models, it reduces inference costs by roughly and cuts input token consumption by , making automated repair more economical while still delivering better results.
