Online Control via Counterfactual Tracking
Introduces an online-control method that competes with broad classes of causal policies under adversarial disturbances, expanding beyond standard linear-controller benchmarks.
Open arXiv ↗Reading notes · papers · GitHub projects
Learn With Me
A living page for papers, repositories, and ideas I am following in mathematics, machine learning, and AI agents. I use this as a public learning trail: what looks interesting, why it matters, and what I want to understand next.
Last curated: 2026-07-15
01
Read actively
For each paper or repo, I focus on the core problem, assumptions, method, and what I can reuse.
02
Connect ideas
I group resources by themes so math, ML, and agent systems can cross-pollinate each other.
03
Build small
When a resource looks useful, the next step is a small experiment, reproduction, or note.
Probability, dynamics, optimization, and mathematical foundations
Mathematics
Papers and repositories that connect rigorous mathematical ideas with modeling, simulation, and computational tools.
Low-Precision Rank Compensation for Matrices and Tensor Trains
Studies when saved memory from lower precision can be reinvested into higher rank, giving practical guidance for numerical linear algebra with matrices and tensor trains.
Open arXiv ↗Optimal Parameter-Free First-Order Methods for Convex Optimization with Unknown Growth and Smoothness
Develops anytime, parameter-free bundle-level first-order methods for convex optimization when growth and smoothness constants are not known in advance.
Open arXiv ↗Stabilize-then-optimize: Feedback transformations as preconditioners in optimal control
Uses feedback transformations as preconditioners for optimal-control systems, linking numerical linear algebra with more stable control optimization workflows.
Open arXiv ↗stdlib-js/stdlib
A fundamental numerical library for JavaScript and TypeScript, useful for studying how scientific-computing primitives are engineered for the web ecosystem.
Open GitHub ↗Deep Gaussian Processes on Directed Acyclic Graphs
Extends Gaussian-process modeling to functions composed along DAGs, with theory on prior collapse, graph topology, and information preservation under partial observations.
Open arXiv ↗Deep learning, evaluation, optimization, and applied ML systems
Machine Learning
A running list of ML papers and codebases I am reading to improve model-building, evaluation, and deployment intuition.
The Seriality Gap in Video Diffusion Models
Uses controlled multi-ball dynamics to show where bidirectional video diffusion struggles with long causal chains, sharpening evaluation for physical reasoning in generative models.
Open arXiv ↗LMCache/LMCache
A KV-cache layer for accelerating LLM serving, worth following for practical inference performance, reuse, and memory-management ideas.
Open GitHub ↗Requential Coding: Pushing the Limits of Model Compression with Self-Generated Training Data
Studies model compression through requential coding and self-generated training data, tying shorter codes to the regularities neural networks learn.
Open arXiv ↗Metacognition in LLMs: Foundations, Progress, and Opportunities
Surveys how large language models monitor, assess, and improve their own reasoning, highlighting open problems in reliable self-evaluation.
Open arXiv ↗sgl-project/sglang
A high-performance serving framework for large language and multimodal models, worth tracking for inference systems and structured generation workflows.
Open GitHub ↗Scalable Visual Pretraining for Language Intelligence
Challenges text-only pretraining by showing that visually rich documents, equations, and layouts can improve foundation-model intelligence at scale.
Open arXiv ↗Tool use, memory, autonomy, benchmarks, and agent infrastructure
AI Agents
Resources for understanding how AI agents plan, use tools, evaluate themselves, and operate as long-running systems.
Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution
Argues that LLM agents often over-spend context and effort on simple tasks, motivating complexity-aware execution policies for more efficient automation.
Open arXiv ↗PalmClaw: A Native On-Device Agent Framework for Mobile Phones
Explores agent execution directly on mobile devices, an important direction for privacy-preserving and latency-sensitive personal assistants.
Open arXiv ↗MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations
Benchmarks the full lifecycle of agent memory operations rather than only downstream question answering, making long-horizon memory evaluation more diagnostic.
Open arXiv ↗When Local Monitors Miss Compositional Harm: Diagnosing Distributed Backdoors in Multi-Agent Systems
Shows how monitoring each tool call or message in isolation can miss coordinated failures across multi-agent systems, a practical warning for agent safety design.
Open arXiv ↗Agent Hacks Agent: Autoresearch for Production-Agent Red-Teaming
Focuses on red-teaming production coding agents such as Claude Code and Codex, where untrusted files, commands, and workspace state make failures actionable.
Open arXiv ↗smithersai/smithers
An agent-workflow system with live observability, rewind, fork, and replay, useful for understanding durable execution and debugging of autonomous runs.
Open GitHub ↗