IterSynth search agent improves results across five benchmarks

Researchers report that their 8B-parameter IterSynth search agent scored an average of 50.7 across five long-running search benchmarks, 4.2 points above the strongest previous agent of comparable size. The result is reported in an arXiv preprint. IterSynth assigns planning and answer synthesis to separate roles. It carries an evolving summary between search steps instead of retaining the full search history, and its training method evaluates each role separately. The reported gain is a benchmark result; the source does not establish performance in deployed search services.