Auturnal170 Chinese and English research reviews connecting primary studies, mathematical models and inspectable evidence.
Agents and software systems · v10
Separating “the model proposes an action” from “the external world completes the action” to examine the relationship between recoverable execution and actual task outcomes.
Read the review →
AI and scientific discovery · v6
Distinguish candidate generation, computational stability, and experimental materials, and place each level of results within the scope of its evidence.
Read the review →
AI and the life sciences · v10
Distinguish local structural reliability, relative inter-domain positions, complex configurations, and actual biological function, without letting an attractive structural image replace validation.
Read the review →
Drug discovery · v5
From ranking tens of millions of candidates to measuring a small subset: examining structural novelty, selectivity, activity denominators, and mechanistic attribution.
Read the review →
Automated science · v5
Comparing A-Lab with language-model experimental orchestration, while measuring experiment completion, target confirmation, knowledge gain, and degree of autonomy separately.
Read the review →
Ecological science · v5
From BirdNET to cross-continental validation, distinguishing call classification, species occupancy, abundance, and ecological change.
Read the review →
Molecular science · v5
Distinguishing structural correctness, physical plausibility, thermodynamic binding and cellular function to establish an independently testable chain of molecular evidence.
Read the review →
Trustworthy AI and methodology · v6
Using an exhaustively enumerable structural model to show why observational differences do not equal intervention effects.
Read the review →
Training methods and budgets · v10
Deriving the allocation between model size and data from constrained optimization, then incorporating data quality, training stages, and deployment costs into the same decision problem.
Read the review →
Trustworthy AI and methodology · v6
Deriving the split-calibration quantile and explaining the relationships among coverage, set size, and distribution shift.
Read the review →
Neuroscience · v5
A connectome provides structural constraints subject to measurement error; understanding computation also requires dynamics, functional observations, and falsifiable models.
Read the review →
Data and evaluation · v6
Analyze duplicates, near-duplicate variants, and test leakage to establish a contamination audit in which false positives and false negatives can be reviewed.
Read the review →No reviews match these filters.