Daily Feed - 2026-02-19

Date: 2026-02-19

3 paper picks + 2 video picks (same bundle for Telegram/email).

Author-talk check: I searched YouTube with exact paper titles for today’s paper picks and did not find clear author/conference talks yet, so I included two high-signal topic-adjacent lectures.

Error Propagation and Model Collapse in Diffusion Models: A Theoretical Study

Domain: ML / Generative Modeling Theory | Time cost: ~15min abstract+setup, ~55min full read

Intuition: The paper studies recursive self-training for diffusion models when each round mixes synthetic samples with a fraction of fresh real data. The key question is when this loop is contractive (stays close to the target distribution) versus when errors compound into model collapse.

Concrete punch: The training pipeline can be written as a mixture update of the next-round data distribution,

where is generated data, is the target distribution, and is the fresh-data fraction. The theory then gives upper/lower divergence bounds whose drift rate depends explicitly on score-estimation error and , identifying regimes where divergence contracts vs accumulates.

Significance: This turns “synthetic-data collapse” from a vague warning into a tunable design rule: fresh-data ratio and score quality become explicit control knobs for long-horizon stability.

Why it matches: Strong mechanism-first analysis, concrete dynamics instead of benchmark anecdotes, and direct relevance to your current generative-model theory focus.

Almost Sure Convergence of Differential Temporal Difference Learning for Average Reward Markov Decision Processes

Domain: RL Theory | Time cost: ~20min theorem skim, ~70min full read

Intuition: Average-reward reinforcement learning is the right lens for continuing tasks, but many convergence results rely on a state-dependent “local clock” learning rate that is rarely used in practice. This paper closes that practice-theory gap for -step differential temporal-difference methods.

Concrete punch: For policy , the differential Bellman relation is

The paper proves almost sure convergence of on-policy -step differential temporal-difference learning (any ) under standard diminishing step sizes (no local clock), and gives sufficient conditions for off-policy convergence in the same practical regime.

Significance: This is exactly the kind of theorem that can upgrade implementation confidence: the learning-rate schedule used in real code now has principled convergence backing.

Why it matches: High mathematical payoff, explicit assumption-level reasoning, and strong fit with your RL + first-principles preference profile.

A Wiener Chaos Approach to Martingale Modelling and Implied Volatility Calibration

Domain: Quant Finance / Stochastic Analysis | Time cost: ~15min abstract+setup, ~65min full read

Intuition: Rather than imposing a narrow parametric volatility process, the paper builds a flexible risk-neutral martingale model by expanding terminal payoffs in Wiener chaos, then recovering intermediate dynamics by conditional expectation.

Concrete punch: The discounted terminal asset can be approximated via truncated chaos expansion,

where are multiple Wiener integrals (Hermite-chaos basis). Dynamics are then reconstructed as

which yields implementable calibration formulas for fitting an implied-volatility surface.

Significance: This is a clean bridge from stochastic-analysis structure to practical calibration speed/flexibility, with direct implications for option-surface modeling choices.

Why it matches: Strong math-structure alignment (martingales, chaos expansions, conditional expectation) and direct relevance to your microstructure/derivatives modeling interests.

Stanford CS236: Deep Generative Models I 2023 I Lecture 13 — Score Based Models

Domain: ML / Deep Generative Modeling (Video) | Time cost: 1h 22m

Intuition: A crisp first-principles treatment of score estimation and denoising objectives that sits directly underneath modern diffusion pipelines.

Concrete punch: Core objective (denoising-score perspective): learn , with weighted objective

This connects estimation error directly to sampling behavior.

Significance: Useful as the clean derivation layer behind today’s model-collapse paper; it clarifies where score error enters the recursion.

Why it matches: High pedagogical quality, equation-level depth, and direct support for your current diffusion-theory thread.

Stanford CS236: Deep Generative Models I 2023 I Lecture 17 — Discrete Latent Variable Models

Domain: ML / Deep Generative Modeling (Video) | Time cost: 1h 14m

Intuition: A strong unifying lecture for discrete latent-variable modeling, linking probabilistic objectives to practical training tricks.

Concrete punch: Variational training is organized around the Evidence Lower Bound (ELBO),

with discrete-latent estimators/relaxations to keep gradients tractable.

Significance: Gives a reusable conceptual bridge between latent-variable modeling and the broader generative unification thread (VAE ↔ diffusion/discrete modeling viewpoints).

Why it matches: High-signal lecture-series quality, strong math-to-method mapping, and concrete transfer value for your active generative research focus.

Source-discovery note

ArXiv: searched recent (6–12 month eligible, prioritizing newest) theory-heavy candidates across diffusion theory, RL convergence, and quantitative finance.
YouTube: searched exact paper-title author talks first; none were clearly available yet for these very recent papers, so selected topic-adjacent high-quality lectures.
Hacker News / Lobsters: scanned recent (<1 week) items; none met today’s mechanism-first + concrete-punch bar.