Structured pruning reduces large language model size and latency.
problem Reducing the size and latency of large language models.
method Parameterizing weights with low-rank factorization and adaptively removing rank-1 components.
result Structured pruning outperforms other methods at various compression levels and speeds up training and inference.
Language models exhibit low-rank structure, which can be used for generation.
problem Understanding the low-dimensional structure of large language models.
method Empirical demonstration and theoretical analysis of the approximate rank of language models' logits.
result Language models can generate responses using linear combinations of unrelated prompts.
Text documents are structured on multiple levels of detail: individual words are related by syntax, but larger units of text are related by discourse structure. Existing language models generally fail to account for discourse structure, but it is crucial if we are to have language models that reward coherence and gener…
SLM models code syntax as trees to generate any programming language code.
problem Generating any piece of code in a given language without restrictions.
method Structural language modeling (SLM) decomposes code into ASTs and estimates probabilities over nodes.
result SLM model generates arbitrary code in any language, outperforming previous methods.
Researchers analyze the structure of token spaces in large language models.
problem Understanding the geometric and topological structure of token spaces in large language models.
method Developed estimators for the dimension and Ricci scalar curvature of the token subspace and applied them to three large language models.
result The token subspace is not a manifold but a stratified manifold with significantly negative Ricci curvature.
The task of translating between programming languages differs from the challenge of translating natural languages in that programming languages are designed with a far more rigid set of structural and grammatical rules. Previous work has used a tree-to-tree encoder/decoder model to take advantage of the inherent tree s…
Transformer models perform slower than convolutional networks in learning hierarchical language structures.
problem Understanding how neural networks learn hierarchical language structures.
method Theoretical scaling laws and empirical validation of neural network performance.
result Convolutional networks outperform transformers in learning hierarchical language structures.
Structured dropout reduces transformer depth for efficient training and inference.
problem Overparameterized transformer networks overfit and require large computation.
method LayerDrop structured dropout for efficient pruning and regularization.
result Sub-networks of any depth can be selected from a large network without performance loss.
Optimizes molecular generation for chemist preferences.
problem Models lack inherent preferences for chemist-desired structures.
method Fine-tuning with Direct Preference Optimization.
result Approach is simple, efficient, and highly effective.
SentenceMIM learns rich latent representations for variable-length language data.
problem Challenges in learning VAEs for variable-length language data, especially posterior collapse.
method Probabilistic auto-encoder trained with Mutual Information Machine (MIM) learning.
result SentenceMIM learns informative latent representations with high mutual information.
Proposes a new RNN for language generation capturing long-range dependencies.
problem Capturing long-range word dependencies and sentence order in text corpora.
method Recurrent Hierarchical Topic-Guided RNN with dynamic deep topic model.
result Outperforms larger-context RNN-based language models and learns interpretable topics.
Language Rectified Flow improves diffusion language generation by simplifying complex steps.
problem Complexity in diffusion language models limits their implementation in NLP applications.
method Reformulates probabilistic flow models to learn neural ODE models for efficient domain transfer.
result Consistently outperforms baselines on fine-grained control tasks and text editing.
APo-VAE generates text in hyperbolic space for better hierarchical representation.
problem Lack of hierarchical structure in Euclidean embeddings for natural language.
method Adversarial Poincare Variational Autoencoder (APo-VAE) in hyperbolic latent space.
result APo-VAE outperforms Euclidean VAEs in capturing latent language hierarchies.
This work offers a broad perspective on probabilistic modeling and inference in light of recent advances in probabilistic programming, in which models are formally expressed in Turing-complete programming languages. We consider a typical workflow and how probabilistic programming languages can help to automate this wor…
Language helps RL agents learn complex relational and causal structures.
problem Learning relational and causal structure in complex environments.
method Training RL agents to predict language descriptions and explanations.
result Language aids agents in learning challenging relational and causal tasks.
Predicting RNA base distances using a large language model.
problem Accurately predicting RNA structural information, especially distance maps.
method Using a large pretrained RNA language model coupled with a transformer.
result The model can accurately infer RNA base distances from sequence data.
New method uses low logit rank to simplify complex language models.
problem Understanding and learning from modern language models.
method Exploiting the low logit rank structure of language models for efficient learning.
result An efficient algorithm for learning low logit rank models from queries.
Recurrent neural network grammars (RNNG) are generative models of language which jointly model syntax and surface structure by incrementally generating a syntax tree and sentence in a top-down, left-to-right order. Supervised RNNGs achieve strong language modeling and parsing performance, but require an annotated corpu…
PixL2R maps natural language to pixel-based rewards for RL, improving sample efficiency.
problem Sparse reward settings in RL limit applicability to complex problems.
method Directly maps natural language descriptions to pixel-based rewards for guiding RL.
result Language-based rewards significantly improve sample efficiency in policy learning.
ETC improves Transformer models for long and structured inputs.
problem Scaling input length and encoding structured inputs in Transformers.
method Introduces global-local attention, relative position encodings, and CPC pre-training.
result Achieves state-of-the-art results on four natural language datasets.
Study reveals neural scaling laws in random graphs and natural language models.
problem Understanding the origin of neural scaling laws in complex systems.
method Examined scaling laws in transformers trained on random walks and simplified natural language models.
result Neural scaling laws emerge in the absence of power law structure in data correlations.
Optimus pre-trains sentences in a latent space for various NLP tasks.
problem Training large-scale language models for diverse NLP tasks.
method Pre-trained Variational Autoencoder (VAE) on large text corpus, fine-tuned for various tasks.
result Optimus achieves state-of-the-art on VAE language modeling benchmarks.
Language recognition system is typically trained directly to optimize classification error on the target language labels, without using the external, or meta-information in the estimation of the model parameters. However labels are not independent of each other, there is a dependency enforced by, for example, the langu…
Natural language processing often involves computations with semantic or syntactic graphs to facilitate sophisticated reasoning based on structural relationships. While convolution kernels provide a powerful tool for comparing graph structure based on node (word) level relationships, they are difficult to customize and…
VALC provides concept-level interpretations of FLMs, overcoming word-level limitations.
problem Lack of higher-level structure interpretation in FLMs' attention weights.
method Formal definition of conceptual interpretation, variational Bayesian framework (VALC).
result VALC finds optimal language concepts for FLM predictions, providing concept-level interpretations.
Proving that next-token prediction makes language models generate coherent long documents.
problem Understanding why language models generate coherent documents despite focusing on next-token prediction.
method Proving the power of next-token prediction in learning longer-range structure using Recurrent Neural Networks (RNN).
result Optimizing next-token prediction in RNNs yields a model that closely approximates the training distribution, even for long-range coherence.
Deep models learn to parse complex language structures from local data patterns.
problem Understanding how deep models parse and represent language structures.
method Introduced tunable probabilistic context-free grammars and a learning algorithm inspired by deep networks.
result Data correlations across scales enable hierarchical language representations.
Neural language models are a critical component of state-of-the-art systems for machine translation, summarization, audio transcription, and other tasks. These language models are almost universally autoregressive in nature, generating sentences one token at a time from left to right. This paper studies the influence o…
SD-SCMs generate counterfactual data for causal inference benchmarks.
problem Benchmarking causal inference methods with realistic data.
method Sequence-driven structural causal models (SD-SCMs) for causal inference.
result State-of-the-art methods struggle with individual treatment effect estimation.
Enhances Transformer for hierarchical language understanding.
problem Lack of explicit hierarchical structure in Transformer models.
method Inspired by U-Net, integrates hierarchical processing into Transformers.
result Improved performance in chit-chat dialogue tasks.
New model shows natural language exhibits phase transition similar to physics.
problem Understanding critical properties in natural language models.
method Created a context-sensitive random language model.
result Demonstrated a Berezinskii--Kosterlitz--Thouless phase transition.
ProGen models protein sequences for synthetic biology.
problem Generating proteins without structural annotations.
method Trained a 1.2B-parameter language model on 280M protein sequences.
result ProGen generates proteins with fine-grained control and accuracy.
New approach improves linear-time attention for language models.
problem Challenges of quadratic attention in long-sequence modelling, especially for discrete data.
method Reinterpreting linear attention through latent probabilistic graphical models, introducing asymmetric structure and recurrent parameterisation.
result Our model achieves competitive performance and outperforms existing linear attention variants on language modelling benchmarks.
A new method improves language model extrapolation without changing training.
problem Improving language model performance on sequences not seen during training.
method Perturbation-based training procedure for language models.
result The method consistently improves out-of-support prediction.
A framework uses free probability to analyze Transformer models.
problem Understanding the dynamics and complexity of Transformer-based language models.
method Formal operator-theoretic analysis using free probability theory.
result Entropy-based generalization bounds derived under freeness assumptions.
TLMG4Eth combines language and graph models for Ethereum fraud detection.
problem Current fraud detection methods fail to consider semantic and similarity patterns in Ethereum transactions.
method TLMG4Eth uses a transaction language model and graph-based methods to capture semantic, similarity, and structural features.
result TLMG4Eth detects anomalies in Ethereum transactions more effectively than existing methods.
NNLMs optimize poorly for word probabilities due to embedding space structure.
problem NNLMs assign suboptimal probabilities to some words.
method Analyzed the inductive bias of NNLMs and the structure of word embeddings.
result Words on the convex hull have bounded probability, affecting others.
The ability to generate natural language sequences from source code snippets has a variety of applications such as code summarization, documentation, and retrieval. Sequence-to-sequence (seq2seq) models, adopted from neural machine translation (NMT), have achieved state-of-the-art performance on these tasks by treating…
HUBERT combines BERT's structure with TPRs to improve NLP task transfer.
problem Improving transferability among NLP tasks.
method Combines BERT's bidirectional Transformer structure with TPRs to learn shared data structures.
result HUBERT outperforms BERT on GLUE and HANS datasets, showing better transferability.
Self-taught optimizer improves code generation using language models.
problem Improving code generation using language models.
method Recursive self-improvement of a scaffolding program that generates code.
result Improved scaffolding program generates programs with significantly better performance.
New measure shows how LSTM models compose hierarchical representations.
problem Understanding how LSTM models capture compositional structure in language.
method Novel measure of interdependence between word meanings in LSTM internal gates.
result High interdependence can hurt generalization and reveals hierarchical structure learning.
The paper proves impossibilities and positive results for universal machine translation.
problem Learning shared sentence representations across multiple language pairs.
method Formal proofs and analysis of natural generative processes.
result Lower bound on translation error and positive results under natural structure.
LLMs will inevitably hallucinate due to their mathematical structure.
problem The inherent limitations of Large Language Models (LLMs).
method Analysis of LLMs using computational theory and Godel's Incompleteness Theorem.
result Hallucinations in LLMs are an inevitable feature, not just errors.
Proposes EM-HRNN model for better hierarchical language representation.
problem Learning hierarchical representations in natural language processing.
method Introduces latent indicator layer and EM algorithm for training.
result EM-HRNN model outperforms other RNN-based models in document classification.
This paper introduces Graph Convolutional Recurrent Network (GCRN), a deep learning model able to predict structured sequences of data. Precisely, GCRN is a generalization of classical recurrent neural networks (RNN) to data structured by an arbitrary graph. Such structured sequences can represent series of frames in v…
We sparsify gated RNNs by simplifying their structure.
problem Improving efficiency of RNNs by reducing their complexity.
method Adjust existing sparsification techniques to gated RNNs, sparsifying preactivations of gates.
result Simplified LSTM structure improves model performance and efficiency.
Let Γ be the fundamental group of a manifold modeled on three dimensional Sol geometry. We prove that Γ has a finite index subgroup G which has a rational growth series with respect to a natural generating set. We do this by enumerating G by a regular language. However, in contrast to most earlier proofs of thi…
This work explains how linear representations in large language models arise from training objectives and gradient descent.
problem Understanding the origins of linear representations in large language models.
method A latent variable model to abstract and formalize concept dynamics, combined with analysis of the softmax cross-entropy objective and gradient descent.
result Linear representations emerge when learning from data matching the latent variable model, and this simple structure suffices to yield linear representations.