A framework to explain decoder-only sequence classification models using intermediate predictions.
problem Explaining predictions of decoder-only sequence classification models.
method Progressive Inference framework with Single Pass-Progressive Inference and Multi Pass-Progressive Inference methods.
result Significantly better attributions compared to prior work on text classification tasks.
This research compares masked diffusion models to autoregressive language models, focusing on architectural differences.
problem Comparing masked diffusion models to autoregressive language models due to architectural differences.
method Equitably compare MDMs within a decoder-only framework, investigating architectural influences.
result Decoder-only MDMs can achieve significant speedups and comparable perplexity with temperature annealing.
Cisco introduces a new time series model for better forecasting.
problem Improving time series forecasting accuracy.
method Developed a new multiresolution decoder-only model trained on large datasets.
result The new model achieves superior performance on observability datasets.
VAIOM models financial returns using continuous input and categorical output.
problem Modeling continuous, noisy, and heterogeneous financial data.
method VAIOM is a decoder-only Transformer that separates input representation from output likelihood.
result VAIOM models outperform fixed single-bar LightGBM baseline in both Test halves.
Predicts next item in sequential bundles using Transformers.
problem Predicting next item in sequentially consumed bundles.
method Used custom Transformers, GPT-3, LSTM, reinforcement learning, Markov models.
result Custom Transformer with decoder-only architecture most accurate.
New study finds best language model architecture and pretraining objective for zero-shot tasks.
problem Evaluating which language model architectures and pretraining objectives best enable zero-shot generalization.
method Compared three model architectures and two pretraining objectives across 170 billion tokens, with and without finetuning.
result Causal decoder-only models trained on autoregressive language modeling exhibit strongest zero-shot generalization.
Tracr compiles programs into transformer models for interpretability.
problem Uncertainty in understanding transformer model outputs due to unknown learned programs.
method Tracr compiles human-readable programs into known structure transformer models.
result Known structure of Tracr-compiled models serves as ground-truth for interpretability.
We introduce a scalable approach for object pose estimation trained on simulated RGB views of multiple 3D models together. We learn an encoding of object views that does not only describe an implicit orientation of all objects seen during training, but can also relate views of untrained objects. Our single-encoder-mult…
Knowledge graphs enable a wide variety of applications, including question answering and information retrieval. Despite the great effort invested in their creation and maintenance, even the largest (e.g., Yago, DBPedia or Wikidata) remain incomplete. We introduce Relational Graph Convolutional Networks (R-GCNs) and app…
Perplexity fails to distinguish correct predictions from incorrect ones in some cases.
problem Perplexity's limitations in model selection.
method Analyzing Transformer continuity and iso-perplexity plots.
result Perplexity does not always select the correct model.
AutoGraph uses transformers to efficiently generate graphs as sequences.
problem Efficiently generating large, sparse graphs without expensive node features.
method Flattening graphs into sequences and using decoder-only transformers.
result AutoGraph achieves state-of-the-art performance on synthetic and molecular benchmarks.
Spectral measurements reveal hidden representation geometry in language model training.
problem Hidden internal representation in language model training is hard to examine.
method Empirical protocol using activation covariance and per-sample gradient SVD spectra.
result Batch size affects representation geometry, and activation spectra predict token efficiency.
Timer-XL predicts multidimensional time series using a unified Transformer approach.
problem Unified time series forecasting across various tasks and contexts.
method Decoder-only Transformers with a universal TimeAttention mechanism and deft position embedding.
result State-of-the-art performance across multiple forecasting benchmarks.
Study shows Direct Feedback Alignment fails to offer more efficient scaling than backpropagation.
problem Understanding and optimizing training methods for neural networks.
method Use of scaling laws to compare Direct Feedback Alignment (DFA) and backpropagation.
result DFA fails to offer more efficient scaling than backpropagation.
WAVE improves time series forecasting by integrating AR and MA components.
problem Time series forecasting challenges.
method WAVE attention mechanism with AR and MA components.
result WAVE attention consistently improves TSF performance.
A new method directly encodes data into latent space using gradient flow.
problem Suboptimal representations in physical sciences due to encoder inversion.
method Decoder-only approach using gradient flow and ODEs, avoiding integrals.
result Superior data efficiency and explicit encoding compared to traditional autoencoders.
This paper fine-tunes LLMs for stock return prediction using financial news.
problem Improving stock return forecasting accuracy using LLMs.
method Fine-tuning LLMs with text and forecasting modules, comparing encoder-only and decoder-only models, and integrating token-level representations.
result LLMs' aggregated token-level embeddings enhance return predictions for long-only and long-short portfolios.
Novel NAS method balances performance and hardware metrics efficiently.
problem Challenging multi-objective optimization in neural architecture search.
method Parameterizes joint architectural distribution via hypernetwork conditioned on hardware features and preferences.
result Zero-shot transferability to new devices with representative and diverse architectures.
CoT enhances transformer accuracy on serial tasks by enabling serial computation.
problem Improving accuracy of large language models on inherently serial problems.
method Integrating a chain of thought (CoT) into decoder-only transformers to enable serial computation.
result Constant-depth transformers with CoT can solve problems in AC^0, surpassing TC^0 without CoT.
New method prevents forgetting in LLMs by dynamically identifying task-specific subspaces.
problem Catastrophic forgetting in continual learning of LLMs.
method Adaptive Singular Value Decomposition (SVD) for constrained full fine-tuning.
result Achieves state-of-the-art results in continual learning benchmarks.
Improved language models learn complex distributions using Fourier series.
problem Capturing continuous structure in discrete token distributions.
method Introducing a Fourier head layer to model continuous structures.
result Significant improvements in performance across various tasks.
SAGA predicts multi-year earnings with adaptive intervals, improving forecast accuracy.
problem Forecasting long-range nonlinear structure in lifetime earnings.
method Decoder-only transformer for irregular tabular sequences, split conformal calibration.
result Significant improvement in forecast accuracy compared to existing methods.
Gradient flow autoencoder improves data efficiency over traditional autoencoders.
problem Sub-optimal latent space representations in autoencoders.
method Gradient flow through ODE with adaptive step size for optimization.
result Gradient flow autoencoder achieves higher data efficiency.
This paper studies activation sparsity in large language models, finding key trends and implications.
problem Activation sparsity in large language models (LLMs) can be improved for efficiency and interpretability.
method Proposes PPL-p% sparsity, analyzes trends with training data, width-depth ratio, and parameter scale. result ReLU is more efficient for sparsity than SiLU, and deeper architectures can improve sparsity.
Language models fail to process hallucinated responses, and this study diagnoses the failure.
problem Language models fail to process hallucinated responses, leading to over-concentration or diffuse attention.
method The study uses forced scoring of benchmark-labeled responses to compute attention shapes and analyze the symmetric component of the degree-normalized attention operator.
result The study proves that every transpose-invariant spectral diagnostic of the attention operator is orientation-blind and bounds the sensitivity of any Lipschitz diagnostic by the asymmetry coefficient \(G\).