Researchers have created a benchmark to test whether language models can recover the original research idea behind a published paper based on its pre-publication bibliography. The benchmark, called Reconstruction, withholds the seed paper and all contemporaneous or future literature from the model, requiring it to propose hypotheses that match against the held-out ground-truth idea. Seven frontier models achieved modest Match rates of 3-15% across six scientific domains. However, a reference-only multi-agent pipeline combining cross-model review with Swiss tournament selection raised Match rates to 23-42%, a 2.4x lift over the best single-model baseline. --- Why it matters: This benchmark matters because it sheds light on the limitations of current language models in understanding research ideas and their ability to generalize across different domains. It also highlights the potential benefits of multi-agent pipelines in improving performance. Source: https://arxiv.org/abs/2608.16645