Continual Search
An iterative framework for root-cause attribution in AI agent failures. It drives an LLM to continuously search for unexamined diagnostic evidence through multi-turn dialogue, reformulating RCA as a directed search process to counteract premature convergence.
Timeline (last 90 days)
The paper proposes an iterative RCA solution called Continual Search.
Continual Search continuously drives the LLM serving as judge to keep searching for unresolved diagnostic evidence through multiple rounds of iteration.
The research team evaluated Continual Search on four existing RCA benchmarks.
On MegaRCA-Mix, Continual Search improved GPT-5.5's F1 score from 0.349 to 0.498, an increase of over 40%.
Within the same model family, a lower-tier model using Continual Search can surpass a higher-tier model.
Research shows that effective search strategies can outperform raw model scale.
Continual Search does not modify model weights, but instead changes the model's reasoning approach through a multi-turn dialogue protocol.
All Facts (7)
The paper proposes an iterative RCA solution called Continual Search.
50%UnverifiedContinual Search continuously drives the LLM serving as judge to keep searching for unresolved diagnostic evidence through multiple rounds of iteration.
50%UnverifiedThe research team evaluated Continual Search on four existing RCA benchmarks.
50%UnverifiedOn MegaRCA-Mix, Continual Search improved GPT-5.5's F1 score from 0.349 to 0.498, an increase of over 40%.
50%UnverifiedWithin the same model family, a lower-tier model using Continual Search can surpass a higher-tier model.
50%UnverifiedResearch shows that effective search strategies can outperform raw model scale.
50%UnverifiedContinual Search does not modify model weights, but instead changes the model's reasoning approach through a multi-turn dialogue protocol.
50%