Daily Feed - 2026-02-26

• Cron: 06:45 ET daily
• Source window: arXiv papers (6–12 months) + YouTube evergreen picks + HN/Lobsters <1-week technical discussions

Neural solver for Wasserstein Geodesics and optimal transport dynamics

Domain: ML / Optimal Transport / Control | Time cost: 18 min read

Intuition: This paper turns the classic OT transport problem into a learnable flow problem: instead of only computing a cost value between distributions, it learns the entire geodesic path and the corresponding velocity field that moves particles from source to target. That makes OT not just a static matching criterion, but a reusable map for sampling, simulation, and controlled transport.

Concrete punch: The core optimization is the dynamical OT problem

They solve a neural minimax relaxation of this constrained formulation and recover both the geodesic distribution path and the induced OT map so you can transport new samples consistently, not just post hoc.

Significance: In practice, this gives you a principled way to get a transport map from data to data (or prior to target) in one learned object, which is exactly the missing bridge when you care about controllable generation and downstream simulation budgets.

Why it matches: The formulation is squarely in the variational/dynamical lens you favor (dynamics + geometric objective), with explicit emphasis on mechanism over benchmark chasing.

Probing the Geometry of Diffusion Models with the String Method

Domain: ML / Diffusion / Geometry | Time cost: 16 min read

Intuition: Standard latent interpolations in diffusion models often cut through low-density regions. This paper applies the string method to stay near high-density “transport routes” of the score model, yielding trajectories that are structurally coherent and interpretable as modal transition paths rather than arbitrary coordinate blends.

Concrete punch: They optimize a path by evolving it under score-based forces toward minimum-energy and finite-temperature regimes, then reparameterizing the string to avoid collapse. In other words, they replace naive interpolation by path-evolution dynamics:

so regime choice ( vs finite ) explicitly trades realism for likelihood geometry.

Significance: The method exposes where a diffusion model’s high-likelihood regions are connected (and where they are disconnected), giving concrete guidance for generation controls, mode coverage audits, and model diagnostics.

Why it matches: It’s a concrete, geometry-first diagnostic with the same “how can I make this mechanism concrete?” flavor as your favorite OT/diffusion papers.

Stochastic Discount Factors with Cross-Asset Spillovers

Domain: Finance / Econometrics / ML | Time cost: 20 min read

Intuition: Instead of treating return prediction and asset interdependencies as separate steps, this paper jointly estimates predictive signals and cross-asset spillovers through an SDF that is learned to maximize risk-adjusted return (Sharpe), then reads the resulting influence structure as a directed information network.

Concrete punch: The SDF is identified by a Sharpe-style objective, not raw forecast RMSE:

with spillover-augmented signal channels shaping the effective weight vector. The paper’s key concrete output is the inferred directional network of predictive influence across firms/industries.

Significance: For a market-microstructure workflow, this is a clean way to test whether cross-asset information is stabilizing your predictive stack or simply fitting noise, while preserving interpretable structure.

Why it matches: This is one of the few recent papers that keeps a mechanism-and-structure lens in finance work, instead of dropping into benchmark-heavy “predictive ML” without invariants.

Gabriel Peyré — Diffusion Flows and Optimal Transport in Machine Learning

Domain: ML / OT (Video) | Time cost: ~43 min watch

Intuition: A high-quality survey-style lecture that grounds OT in concrete ML workflows, especially transport duality, entropy regularization, and practical generative modeling tradeoffs.

Concrete punch: The talk repeatedly makes explicit the primal-dual pair (transport cost minimization and dual potential maximization), helping connect abstract OT distance notions back to algorithmic knobs in model training.

Significance: Best viewed as a calibration talk: it can reset your OT intuition before implementing the Wasserstein geodesic paper above.

Why it matches: Strong pedagogy, principled pacing, and topic alignment with current OT-heavy ML work.

Michael Albergo — Non-equilibrium transport and tilt matching for sampling

Domain: ML / Transport sampling (Video) | Time cost: ~38 min watch

Intuition: This presentation reframes sampling as non-equilibrium transport with explicit control of the “tilt” term, which is a useful conceptual bridge if you build sequence models where sampling quality depends on trajectory shaping.

Concrete punch: The central lesson is that careful control of the tilted dynamics (rather than only the target density) changes where mass concentrates along the sampling path, which is a mechanics-level view of variance and mode exploration.

Significance: Useful companion to the first paper: it pushes the same transport intuition from math to implementation-level choices in practical samplers.

Why it matches: High production and pedagogical quality, strong control/dynamics framing, and direct relevance to generative transport workflows.

Source notes

• arXiv papers selected from cs.LG, stat.ML, and q-fin.CP feeds with recency checks and seed/duplicate filters.
• HN/Lobsters scan (<1 week window) did not produce technically deep items that cleared the current quality bar this cycle.
• No direct author talks linked to today’s selected papers were discovered quickly; used high-credibility standalone talks as substitutes (still required to include 1–2 YouTube picks).