Data structures and storage · v5
Learned Indexes Under Distribution Change: Prediction, Search, and Correctness
A prediction can accelerate a search without becoming the authority that decides whether a key exists.
Read the review →AUTURNAL / RESEARCH LIBRARY / 2026.10AI meets discovery.
AI meets discovery.
Follow the evidence.
170 Chinese and English research reviews connecting primary studies, mathematical models and inspectable evidence.
170guides526primary sources510illustrations and charts
THE RESEARCH COLLECTION
Explore the research
Data structures and storage · v5
Caching Beyond Recency: Admission, Workload Shift, and the Cost of a Miss
Choosing what may enter a cache can matter as much as choosing what leaves it.
Read the review →
Networks and infrastructure · v5
The Geometry of Tail Latency: Fan-Out, Queueing, and Redundant Work
Tail latency is an emergent property of dependencies, load, and cancellation. Optimizing one percentile on one server can make a distributed request worse.
Read the review →
Data structures and storage · v5
Erasure Coding in Real Storage Systems: Durability, Locality, and Repair Traffic
Storage efficiency is only one coordinate of a code. Repair bandwidth, failure correlation, locality and reconstruction obligations determine whether the design survives contact with an operating system.
Read the review →
Networks and infrastructure · v5
Congestion Control as Delayed Inference: Throughput, Queues, and Fairness
A congestion controller acts on delayed, incomplete evidence. Throughput, queue occupancy and fairness can move in different directions even on the same link.
Read the review →
Networks and infrastructure · v5
Serverless Is a Placement Decision: Cold Starts, Isolation, and State
Moving execution behind an invocation API does not remove state or resource choices. It changes who makes them, what can be observed, and which costs become intermittent.
Read the review →
AI systems and measurement · v5
Serving Language Models Within a Memory Budget: Attention, Paging, and Scheduling
The capacity of an inference service is a joint constraint on weights, attention state, memory movement and scheduling. Tokens per second alone does not describe that constraint.
Read the review →
AI systems and measurement · v5
When a Benchmark Supports a Claim: Experimental Units, Workload Mixtures, and Causal Validity
A benchmark can be repeatable and still answer the wrong question. Experimental units, workload composition and intervention design determine what a measured difference means.
Read the review →
Training & Optimization · v5
What Gradient Noise Actually Controls
Batch size and learning rate jointly set a stochastic process, but neither its stationary variance nor its speed alone determines generalization.
Read the review →
Training & Optimization · v5
Zero Training Error Does Not Determine the Learned Classifier
The direction selected by optimization matters even when every training label is already correct.
Read the review →
Training & Optimization · v5
Normalization Changes the Geometry of Learning
Choosing which coordinates to normalize determines what information, invariance, and cross-example dependence a layer introduces.
Read the review →
Training & Optimization · v5
When Depth Becomes a Dynamical System
Residual connections invite a dynamical interpretation, but stability, expressivity, and numerical accuracy remain separate requirements.
Read the review →No reviews match these filters.