Daily Feed - 2026-02-17

Date: 2026-02-17

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

Author-talk check: No exact-title author/conference talks were found yet for today’s very recent papers, so I included two high-signal topic-adjacent YouTube lectures.

Efficient Sampling with Discrete Diffusion Models: Sharp and Adaptive Guarantees

Domain: ML / Generative Modeling Theory | Time cost: ~20min abstract+proof sketch, ~70min full read

Intuition: The paper studies discrete diffusion sampling through a continuous-time Markov chain (CTMC) lens and asks a practical question: how many sampler steps are fundamentally needed to hit a target KL error. The key payoff is that properly designed -leaping can scale with intrinsic problem dimension, not raw token vocabulary size.

Concrete punch: For uniform discrete diffusion, the sampler achieves

iterations to reach -accurate KL divergence, removing linear dependence on vocabulary size . For masking diffusion, convergence is controlled by an effective total correlation term with upper bound , and can be sublinear (or near-constant) on structured data.

Significance: This gives a principled sampler-budget rule for discrete generators (text, graphs, sequences): invest around intrinsic structure, not worst-case state-space size.

Why it matches: Strong mechanism-first theory, explicit KL-rate statements, and direct information-theoretic structure—exactly your preferred “concrete derivation over benchmark folklore” style.

FLAC: Maximum Entropy RL via Kinetic Energy Regularized Bridge Matching

Domain: RL / Generative Control | Time cost: ~15min abstract+algorithm skim, ~60min full method read

Intuition: Diffusion/flow-style policies are expressive for continuous control, but they break classic max-entropy RL pipelines because is often intractable. FLAC reframes policy optimization as a generalized Schrödinger bridge toward a high-entropy reference process, so “stay stochastic while improving return” becomes a geometric control problem.

Concrete punch: The regularizer is path-energy based (kinetic-energy proxy), conceptually of the form

where is the policy flow field. The core claim is that controlling this energy bounds deviation from the high-entropy reference without explicit action-density estimation; FLAC then tunes the regularization weight via a dual (Lagrangian) mechanism.

Significance: Clean route to max-entropy behavior for expressive non-likelihood policies—useful when policy class power outgrows actor-critic assumptions.

Why it matches: Strong variational/bridge perspective, direct RL↔generative unification, and objective-level novelty (not just implementation tweaks).

A unified theory of order flow, market impact, and volatility

Domain: Quant Finance / Microstructure Theory | Time cost: ~20min abstract+model overview, ~75min full read

Intuition: This paper splits order flow into a persistent “core” component plus reaction flow (both Hawkes-based), then derives a scaling limit where multiple empirical stylized facts emerge from one persistence parameter . Instead of fitting each phenomenon separately, it ties them to a single structural driver.

Concrete punch: With estimated , the model predicts:

Plugging gives , consistent with square-root impact.

Significance: A single-parameter bridge between persistent signed flow, rough volatility, and impact-law exponents is high leverage for model design and calibration discipline.

Why it matches: First-principles microstructure modeling with explicit exponents and no-arbitrage constraints—exactly your mechanism + invariants taste.

Generative Flows on Discrete State-Spaces | Andrew Campbell, Jason Yim

Domain: ML / Discrete Generative Models (Video) | Time cost: 52m

Intuition: Research-style talk on transporting diffusion/flow ideas into discrete domains (tokens, graphs, combinatorial objects), which pairs directly with today’s discrete-diffusion paper.

Concrete punch: A central discrete-time/continuous-time modeling backbone is the master-equation view

with the transition-rate generator. This framing clarifies where sampling speed and approximation error enter in -leaping-style methods.

Significance: Gives a practical mental model for when discrete generators need better transition parameterization versus better integrators.

Why it matches: High information density, mathematically grounded exposition, and direct relevance to your generative-model unification thread.

Ciamac Moallemi: High-Frequency Trading and Market Microstructure

Domain: Quant Finance / Market Microstructure (Video) | Time cost: 25m

Intuition: Compact lecture linking inventory risk, adverse selection, and execution frictions to observed market-impact behavior—good complement to today’s unified order-flow theory paper.

Concrete punch: The lecture context is organized around empirical impact scaling of the rough form

where is meta-order size and is market volume scale, alongside order-flow persistence intuition.

Significance: Useful for translating stylized-law intuition into concrete execution and risk-control decisions.

Why it matches: Mechanistic microstructure focus with model-level consequences, not infrastructure/tooling noise.

Source-discovery note

ArXiv: primary source for paper picks (frontier recency + mechanism-first screening).
YouTube: filtered for high pedagogy and direct topic adjacency to today’s papers.
Hacker News/Lobsters: scanned; no <1-week discussion cleared today’s concrete-punch threshold.