Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis
Multilingual ABSA evaluation across seven languages benchmarks transformer and instruction-tuned models under zero-shot and full-resource settings.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Multilingual ABSA evaluation across seven languages benchmarks transformer and instruction-tuned models under zero-shot and full-resource settings.
Multi-modal transformer for spacecraft celestial orientation via spherical topology, replacing traditional Lost-in-Space algorithms.
Pipeline converting imperative programs to typed graphs using AST parsing and semantic embeddings (SentenceTransformer, CodeBERT) for verification artifact reuse.
Carbon-Taxed Transformers proposes green compression pipeline addressing computational cost, memory overhead, and environmental impact of LLMs in software engineering.
Systematic empirical study of in-context learning scaling behavior on Gaussian-mixture classification, extending prior linear theory.
Analysis shows Transformers with CoT cannot length-generalize beyond TC^0 under standard positional encodings, limiting expressivity gains claimed by theory.
StratFormer: transformer-based meta-agent for opponent modeling and exploitation in imperfect-information games via curriculum learning.
Lecture notes on neural network verification covering feedforward, recurrent, attention mechanisms, transformers, and algorithmic verification techniques.
Learned rotary positional encodings (RoPE) as second dimension of Transformer attention expressivity.
HyLo converts pretrained Transformers to hybrid linear-Transformer models while extending long-context via upcycling.
Transformer-based causal model estimates drug treatment effects on dialysis risk in AKI patients using EHR sequences.
Score-based Variational Flow provides continuous-time theoretical foundation for Transformers via Euler discretization and attention recovery.
Theoretical analysis refines Dong et al. rank collapse findings: layer normalization preserves affine rank, revising understanding of Transformer architectural stability.
In Transformer models, non-GEMM (non-General Matrix Multiplication) operations -- especially Softmax and Layer Normalization (LayerNorm) -- often dominate hardware cost due to their nonlinear nature. To address this, previous approximation studies mainly target rank-oriented…
Transformer-based interpretable models for English reading comprehension with adversarial bias correction and attention visualization for educational use.
BantuMorph transformer recovers 728 noun and 1,525 verb cognate candidates in Bantu languages via encoder embeddings.
AttnLRP explanation method applied to DNABERT-2 genome language model reveals whether Transformer attention captures relevant genomic patterns versus CNNs.
Transformers fail on unseen symbolic reasoning due to unembedding collapse and token copying difficulty, limiting out-of-distribution generalization.
N-gram models match LSTM/Transformer accuracy on event-log prediction with lower resources and better stability than neural baselines.
A-THENA uses time-aware Transformer encoding for early IoT intrusion detection with temporal packet dynamics.
Study shows Transformers, LSTMs, RNNs, and word embeddings converge on period-T periodic features for number representation; proves Fourier sparsity is necessary but insufficient for mod-T separability.
Cognitively-inspired working memory constraints and fixed-width attention improve Transformer performance on low-data grammar tasks.
Empirical study of 40+ transformer compression experiments on GPT-2 and Mistral 7B reveals variance-importance decoupling.
LayerTracer framework analyzes hierarchical representations and robustness bottlenecks across diverse LLM architectures including Transformer, GateDeltaNet, and Mamba.
Deterministic dropout formulation replaces stochastic masking with explicit regularization terms for Transformer training.
Theoretical analysis of adversarial training in Vision Transformers shows benign overfitting under signal-to-noise conditions.
SE(3)-equivariant transformer with contrastive learning for protein-ligand structure encoding in drug discovery.
EXIT multimodal transformer predicts metal-organic framework properties accounting for sample-dependent structural variations.
Sessa: selective state-space attention mechanism improving token influence preservation vs. Transformers in long-context sequence modeling.