Generational loop of LLM agents beats one-shot hypothesis generation on cancer targets

single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC

What this means for you

Nothing to act on yet. This is a single unreviewed preprint, judged on two measures its own authors assembled, with no evaluation date and no released code or service mentioned. If you prioritise drug targets, the claim worth holding onto is that iterative revision beat single-pass generation, not the absolute scores.

A preprint on arXiv describes HypoEvolve, which runs specialised LLM agents — separate model calls given distinct roles, such as proposing or critiquing — through generations of a genetic algorithm, keeping the better hypotheses and generating new ones. Tested on drug repurposing against six baselines, it scored highest on two measures built from DepMap and Open Targets, databases of cancer cell-line dependencies and target–disease evidence, across 34 cancer types. DepMap selectivity was 0.171 versus 0.115 for the strongest baseline. Gains held on held-out cancer types. No evaluation date is given, and the work is not peer-reviewed.

Key facts

  • ·The paper is arXiv 2609.15938, posted on 13 September 2026. source
  • ·HypoEvolve coordinates specialised LLM agents through a generational genetic algorithm that retains and revises a population of hypotheses. source
  • ·It was evaluated on drug repurposing across 34 cancer types against six baselines. source
  • ·It reports DepMap selectivity of 0.171, against 0.115 for the strongest baseline. source
  • ·The two external measures used are adapted from DepMap and Open Targets. source
  • ·The abstract gives no evaluation date and the paper is reported as a preprint. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire