Hackathon

The Fourth Turn

The ESSIR 2026 hackathon, The Fourth Turn, challenged participants to build retrieval-augmented question-answering systems over long scientific documents. The focus was not on producing fluent answers alone, but on building systems that could retrieve relevant evidence, answer follow-up questions, reason across distant parts of a document, and cite the source text accurately.

Five teams worked with five different documents and a shared set of 45 questions. By the end of the week, each team had produced a working system that could place a document in front of a machine and return grounded, cited answers.

Challenge Tasks

The task was structured around three levels of increasing difficulty:

• Level 1: Direct document questions. Systems had to answer questions whose evidence could usually be found in a local part of the document.
• Level 2: Conversational follow-ups. Systems had to keep track of context across turns and resolve questions such as “why does that happen?” before retrieval.
• Level 3: Whole-document reasoning. Systems had to combine evidence from distant sections, tables, and methodological details spread across the document.

Teams were expected to design and implement the core retrieval pipeline, including document parsing, chunking, search, answer generation, citation selection, and evaluation. The strongest submissions were those that measured their own behavior and improved the engineering around retrieval, rewriting, citation checking, and reproducibility.

Results

Teams