Auturnal170 Chinese and English research reviews connecting primary studies, mathematical models and inspectable evidence.
AI and scientific computing · v6
Compare what FNO, DeepONet, and PINN learn, and avoid treating grid errors as guarantees for continuous solutions.
Read the review →
Quantum chemistry · v5
From FermiNet to fixed-node Monte Carlo, deriving the distinctions between low energy, correct wavefunctions, and reliable energy differences.
Read the review →
Efficient inference systems · v6
Separate logical sequences from physical KV blocks, quantify fragmentation, and track sharing, reclamation, and serving latency.
Read the review →
Nanophotonics · v5
From surrogate models to real devices, derive how adjoint gradients, fabrication perturbations, and resonance shifts determine whether a design is trustworthy.
Read the review →
Sequence models · v6
Prove the relative-position inner-product relation, then distinguish length extrapolation from genuine retrieval and composition capabilities.
Read the review →
Alignment and behavioral learning · v10
Deriving connections in preference optimization from a KL-constrained objective, while distinguishing training preferences, actual utility, and safety validation.
Read the review →
Earth science · v5
Understanding how to evaluate ensemble weather models correctly: beyond average error to probabilistic calibration, spatiotemporal dependence, and decision value.
Read the review →
Synthetic biology · v5
Connecting RFdiffusion, sequence design, and experimental screening into an auditable closed loop to analyze success denominators, selection bias, and design diversity.
Read the review →
Efficient inference systems · v6
From rounding error to output perturbations, account separately for weights, the KV cache, activations, and kernel costs.
Read the review →
Quantum information · v5
Using AlphaQubit as a starting point, we explain why decoding accuracy, code-distance scaling, and real-time closed-loop operation must be validated separately.
Read the review →
Knowledge systems and evidence · v10
Study the retriever, evidence set, and answer generator as distinct points of failure, and establish verification methods that do not depend on a single accuracy measure.
Read the review →
Reasoning and verification · v10
Reframing the evaluation of reasoning ability, from success probabilities under independent sampling to correlated errors, missed detections by verifiers, and selection bias.
Read the review →No reviews match these filters.