GEWU
中

Search articles

Type keywords to search published articles.

    Disciplines / Research questions

    Computer science

    10 research topics · 32 questions

    Whether the emergent abilities of large language models originate from metric scale artifacts3 questions

    1. Extending difficulty-stratified forecasting of emergent abilities beyond multiple-choice tasks: continuous metrics for string-match and generation tasks

    2. Generality, Comparability, and Mechanism of the Pre-Training-Loss Account of Emergent Abilities in Language Models

    3. When do large language models show genuine, metric-invariant emergent abilities? Regimes, token-level error models, and statistical protocols for capability claims

    Geometric explanation of adversarial example transferability4 questions

    1. A predictive theory of adversarial example transferability from weight-space geometry: flatness measures, subspace directions, and deviation scaling of surrogate ensembles

    2. Class-Aware Adversarial Transferability: Do Non-Robust Features Fully Explain Same-Mistake and Different-Mistake Transfer, and Which Non-Robust Features Are Commonly Used Across Diverse DNNs?

    3. When do representation-space attacks transfer between neural networks? Alignment conditions beyond linear misalignment

    4. Benign Overfitting, Regularization Mechanisms, and the Transferability of Adversarial Examples: Extending the Substitute-Generalization Analysis of Wang and Farnia

    Long-tail sources of tail latency in distributed systems3 questions

    1. Provably optimal probabilistic scheduling for tail latency in erasure-coded storage beyond Pareto file sizes and exponential service times

    2. Extending the queueing-Amdahl tail-latency framework beyond M/M/k: realistic distributions, request dependencies, hardware-constrained core scaling, and validation beyond memcached

    3. Unexplained transmit-side TCP flow-control stalls and the remaining sources of 99th-percentile latency in memcached

    Near-distribution leakage problem in out-of-distribution detection benchmarks3 questions

    1. Eliminating Type-1 and Type-2 benchmark contamination at scale for out-of-distribution object detection: per-box curation policies

    2. ImageNet-scale audit of fit-set leakage in near-OOD benchmarks using the decodable-but-not-detectable fingerprint

    3. Non-circular automated construction of contamination-free out-of-distribution detection test sets

    Incomplete fixes in open-source vulnerability-fixing commits3 questions

    1. Open directions from the Mozilla security-bug-fixing study: pattern-based automated repair, security regression testing, clone-based patch-quality assurance, and typed blocking-bug prediction

    2. Cross-project generalization of process indicators and commit-level prediction of vulnerability-reintroducing fixes in open-source software

    3. Predicting incomplete fixes and recurring multi-location vulnerabilities during multi-patch vulnerability remediation in open-source software

    Stability of the lottery ticket hypothesis across neural architectures and dataset sizes3 questions

    1. Elastic transfer of winning lottery tickets beyond depth: width, mixed topology, and cross-family architectures

    2. Existence conditions and mechanism of rigorous ("Jackpot") winning tickets under the sanity-checked lottery ticket hypothesis

    3. Why pruning-aware data reduction improves the generalization of extremely sparse lottery tickets: a mechanistic account

    Decoupling of Neural Network Calibration Error and Accuracy Improvement2 questions

    1. Is there a method-independent relationship between classification accuracy and calibration error in deep neural network classifiers?

    2. Why decoupled feature-extraction and classifier training improves neural network calibration: a theoretical characterization of TST/V-TST, its variational design choices, and its out-of-distribution limit

    Cascading Effects of Cardinality Estimation Errors in Database Query Optimizers4 questions

    1. Finite-error regime of plan regret: the cardinality-dependent correction to the average-case sub-optimality limit law, and the location of the condition-number/average-case crossover

    2. Generality of cost-guided (plan-aware) training for learned cardinality estimation

    3. End-to-End Planner Integration and Error-Propagation Modeling for EXPLAIN-only Residual Cardinality-Estimation Corrections in TiDB (the TiCard Open Questions)

    4. When do cardinality estimates determine query plan quality, and how can plan construction be made robust to estimation error?

    Finite Sample Mechanism of the Double Descent Phenomenon4 questions

    1. Extending the effective-parameter account of double descent to deep learning: test-time complexity axes and label-free model selection

    2. Non-asymptotic and dynamical completion of the Hessian-based theory of double descent in finite-width neural networks

    3. A rigorous quantitative theory of the three-factor mechanism for test-error divergence at the interpolation threshold

    4. Non-asymptotic rate of the surrogate random design approximation for double descent, and closed-form i.i.d.-design expectations (pseudo-Wishart and MACG moments)

    Distinguishing oversmoothing and oversquashing in graph neural networks3 questions

    1. Extending the Jacobian spectral theory of vanishing gradients in graph neural networks to nonlinear, learned updates and training dynamics

    2. Data-side statistics of label-information localisation and factorisation: the open core of "Position: Don't be Afraid of Over-Smoothing And Over-Squashing" (Kormann, Doerr and Lutzeyer, 2026)

    3. Principled dataset-level predictors of the optimal depth of graph neural networks

    Back to open problems