A recent arXiv paper tackles a known weakness in LLM-based directed test input generation: the models can produce inputs that miss their intended targets when their underlying assumptions about how the target program behaves are incorrect. The abstract describes this as a problem of "off-target" test inputs stemming from faulty hypotheses.

To address this, the authors propose refining the generation process using runtime value feedback. Rather than relying solely on static reasoning, the method appears to use actual program execution values to ground the LLM's decisions, helping it adjust its hypotheses and steer inputs toward the desired code paths.

The paper's abstract is brief, so details on implementation and evaluation are not available from the source. Still, the idea of closing the loop between LLM generation and runtime behavior offers a plausible path toward more reliable automated testing.