Probing the Geometry of Diffusion Models with the String Method

This is a research paper with accompanying reference implementation (0 stars, 4 forks, 41 days old, 0 velocity) introducing the string method adapted to diffusion models. The core contribution—using string method from computational chemistry to probe DM geometry—is a novel application, but the underlying techniques (string method, score-based sampling) are established. The work is primarily theoretical with a reference implementation rather than a production system. Defensibility is low (4/10) because: (1) it's an early-stage research artifact with minimal adoption, (2) the method is algorithm-implementable by anyone with DM knowledge, (3) no ecosystem or community has formed, (4) dependency on pretrained models creates no lock-in. Frontier risk is HIGH because: (1) frontier labs (OpenAI, Anthropic, Google DeepMind) are actively researching diffusion model interpretability and geometry, (2) this directly supports model understanding/improvement which is core to their work, (3) could be trivially integrated into model analysis pipelines as an interpretability tool, (4) solves a problem (DM landscape analysis) that frontier labs care deeply about for their own model development. Anthropic and OpenAI have published on diffusion model mechanistic interpretability; Google has string-method adjacent work in manifold learning. A frontier lab would likely build or integrate this rather than adopt the reference implementation. Novelty is 'novel_combination': applies established string method (from computational chemistry/materials science) to a new domain (diffusion models) with meaningful results, but neither the string method nor score-based evolution are novel individually.