ByteDance’s Pistis team reports that a revised search workflow helped its AI answer nine more questions correctly on a 500-item test requiring visual and web search, while keeping the same maximum of 15 tool interactions per attempt. The result is a small gain in a visual-search benchmark, reported by the team that built the system.
The change was to the program around the model, not its weights. Pistis-Auto-Harnessing (PAH) adjusts how a fixed AI system tracks candidate answers, organizes evidence, chooses search steps and decides when to stop. An optimization agent used 100 development examples to propose changes within set limits; it kept a revision only when it improved a selected measure. The team then froze the code, prompts and settings before testing once on separate examples.
On VDR-testmini, a multimodal benchmark for answering questions that require visual and web search, the baseline workflow scored 26.8% (134 of 500); PAH scored 28.6% (143 of 500). Both had a 15-interaction ceiling, and their average use was nearly identical: 8.68 interactions for the baseline and 8.60 for PAH. The same frozen workflow also posted smaller gains on three other multimodal-search benchmarks, though those remain related test settings rather than evidence of general agent reliability.
The report attributes the result to the harness as a whole, so it does not isolate which change mattered most. It is a ByteDance-led evaluation; the paper reports no independent replication and links no runnable code or model checkpoint.
