Training & Optimization · v5
Compute-Optimal Learning Depends on the Accounting
An optimal parameter count is conditional on the loss model, data supply, and the costs included in the budget.
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
Representation & Architecture · v5
Contrastive Learning Between Alignment and Information Loss
A representation can become excellent at identifying augmented pairs while discarding the variable a future decision requires.
Read the review →
Representation & Architecture · v5
Reconstruction Is Not a Certificate of Semantic Understanding
Predicting missing pixels rewards conditional predictability; semantic usefulness depends on what is hidden, what is visible, and which target is reconstructed.
Read the review →
Representation & Architecture · v5
Where Vision Transformers Put Their Inductive Bias
Patch boundaries, position signals, attention connectivity and supervision all constrain what a visual transformer can learn efficiently.
Read the review →
Representation & Architecture · v5
Equivariance Is a Contract About Transformations
A symmetry constraint is useful only when the input transformation, output transformation and target behavior agree.
Read the review →
Representation & Architecture · v5
Graph Depth, Smoothing, and the Loss of Distinctions
Greater message-passing depth can both mix away local distinctions and compress distant evidence; these are different mechanisms.
Read the review →
Generation & Adaptation · v5
Diffusion Generation Through the Lens of Score Error
Noise prediction, score estimation, and sampler accuracy are different error budgets; a solvable Gaussian example makes their relationship explicit.
Read the review →
Generation & Adaptation · v5
Flow Matching and the Geometry of Transport
Straight conditional paths, marginal velocity fields, and optimal transport are related constructions with different guarantees.
Read the review →
Generation & Adaptation · v5
When a Latent Variable Stops Carrying Information
Posterior collapse is a statement about information flow, not simply a small KL term or an unattractive sample.
Read the review →
Generation & Adaptation · v5
What Low-Rank Adaptation Can and Cannot Represent
An update's rank bounds a matrix approximation, while task quality depends on input geometry, layer placement, and optimization.
Read the review →
Generation & Adaptation · v5
Sparse Experts Under Capacity and Routing Constraints
Conditional computation separates stored capacity from active work, but routing correlations decide whether the promised capacity is actually available.
Read the review →
Reliability & Interpretation · v5
Robustness Claims Need an Explicit Threat Model
Adversarial evaluation becomes meaningful only after input coordinates, attacker access, optimization strength, and certificate scope are fixed.
Read the review →No reviews match these filters.