Broader Objective

Scientific works are a dense source of information with several signals on the principle contributions of the work. Hypothesis extraction is attractive for science-of-science because it promises machine-readable maps of what papers propose before those claims are tested at scale. In practice, however, the bottleneck is not merely finding plausible hypotheses; it is producing outputs that are structurally valid, evidentially grounded, and disciplined enough to abstain when the abstract does not actually support extraction. Early versions of hypogen showed that simply providing more context did not solve this problem: both abstract-only and focused full-text settings often returned malformed structures or specific-sounding claims untethered from the source text. We therefore reframed the task as reliability-constrained, abstract-grounded claim extraction rather than open-ended summarization.

Author: Akhil Pandey Akella

Affiliation: AllSci Corp

Conference: International Conference on the Science of Science and Innovation (ICSSI 2026), Boulder, Colorado

Program: Poster Session II, poster 108, Tuesday, June 30, 2026, 16:35-18:00, Flatiron 1

Poster: Download the PDF (32 x 45 in.)

Poster

ICSSI 2026 poster: Reliable Hypothesis Extraction with Generate-Judge-Verify, by Akhil Pandey Akella. Diagrams show hypothesis extraction, candidate ranking and verification, and an evidence graph.

Acknowledgement

This work is proprietary and all artifacts are owned by AllSci Corp.