This paper enhances language models with knowledge awareness.
problem Understanding how much knowledge pretrained language models grasp.
method Inserting explicit knowledge layers into pretraining without changing transformer architecture.
result Significantly more knowledge packed into transformer parameters.
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.
This study examines how the size and alignment of pretraining data affect the performance of large language models on downstream tasks.
problem Understanding how the size and alignment of pretraining data impact the performance of large language models on downstream tasks.
method Investigated the scaling behavior of large language models in a transfer learning setting, focusing on machine translation tasks.
result The size of the finetuning dataset and the distribution alignment between pretraining and downstream data significantly influence the scaling behavior of downstream performance.
Unified framework for ICL in causal and masked models.
problem Understanding ICL in masked language models and comparing it to causal models.
method Developed a statistical learning framework representing context by empirical measure and predicting using context and query.
result Upper bounds for masked and autoregressive objectives under Wasserstein-type regularity conditions.
Synthetic continued pretraining enhances model performance with synthetic data.
problem Data inefficiency in pretrained models when adapting to domain-specific documents.
method Synthetic data augmentation using EntiGraph to create a large synthetic corpus.
result Language models can answer questions and follow instructions without access to domain-specific documents.
In natural language processing, it has been observed recently that generalization could be greatly improved by finetuning a large-scale language model pretrained on a large unlabeled corpus. Despite its recent success and wide adoption, finetuning a large pretrained language model on a downstream task is prone to degen…
New research reveals how the pretraining distribution affects in-context learning in large language models.
problem Understanding how the pretraining distribution influences in-context learning in large language models.
method Developed a theoretical framework to characterize the relationship between pretraining distribution properties and in-context learning performance.
result Characterized a fundamental trade-off between robust task selection and generalization in ICL due to the pretraining distribution's statistical properties.
A new approach to protein language models combines latent space prediction with masked language modeling.
problem Improving protein language models by predicting amino acid identities at masked positions.
method A variant of masked language modeling that predicts latent targets only at masked positions, retaining the MLM cross-entropy.
result The new approach outperforms pure masked language modeling on 11 out of 16 downstream tasks.
The ever-growing volume of data of user-generated content on social media provides a nearly unlimited corpus of unlabeled data even in languages where resources are scarce. In this paper, we demonstrate that state-of-the-art results on two Thai social text categorization tasks can be realized by pretraining a language …
Language models help text classification tasks by predicting next words.
problem Lack of theoretical understanding of why language models perform well on downstream tasks.
method Mathematical study of the connection between next word prediction and text classification, formalizing it and quantifying the benefit.
result Language models that are ε-optimal in cross-entropy learn features that can solve classification tasks with linear approximation.
Theoretical analysis of data quality and synergies in LLMs.
problem Understanding why different training methods require different amounts of data.
method Theoretical analysis of transformers trained on a weight prediction task for linear regression.
result SFT excels on smaller datasets challenging for the pretrained model, while RL benefits from large, not overly difficult data.
DatedGPT prevents lookahead bias in financial forecasting models.
problem Lookahead bias in large language models trained on internet-scale data.
method Time-aware pretraining with annual data cutoffs and instruction fine-tuning.
result Models' knowledge is effectively bounded by their data cutoff year, improving forecasting validity.
New framework selects high-quality pretraining data without training LLMs.
problem Slow progress in understanding pretraining data due to costly experiments.
method Statistical framework based on perplexity-benchmark correlations.
result Approach outperforms existing methods on multiple benchmarks.
Transformer models can solve complex math problems with less data.
problem Solving complex symbolic mathematics problems with limited data.
method Pretrain transformer models on language translation tasks and fine-tune for symbolic math.
result Pretrained transformer models achieve comparable accuracy to state-of-the-art models with less data.
Analysis of pretrained models' effectiveness in downstream tasks.
problem Understanding why pretrained models perform well in NLP tasks.
method Analyzed head and prompt tuning approaches using latent variable models.
result Prompt tuning provides stronger guarantees than head tuning.
The study introduces anytime learning schedules for large language models without fixed horizons.
problem Training large language models without knowing the total training horizon.
method Theoretical analysis and weight averaging to create anytime learning schedules.
result Theoretical and empirical evidence shows that weight averaging with simple step sizes can achieve comparable final loss to well-tuned cosine schedules.
Pretraining models improves text classification accuracy, but diminishing returns are observed with large datasets.
problem Improving text classification accuracy with pretrained models.
method Examined the benefits of pretrained models on text classification tasks with varying amounts of training data.
result As the number of training examples grows into the millions, the accuracy gap between pretrained BERT-based models and vanilla LSTM narrows to within 1%.
A pretrained LLM and its finetuned version can detect OOD data effectively.
problem Detecting out-of-distribution data in language models.
method Using the likelihood ratio between a pretrained and finetuned LLM.
result The likelihood ratio is an effective criterion for OOD detection.
A new method uses a frozen language model to improve sample efficiency in reinforcement learning.
problem Improving sample efficiency in reinforcement learning with partially observable environments.
method FROZEN Hopfield network and HELM (History Embedding Language Model) method.
result HELM achieves new state-of-the-art results on Minigrid and Procgen environments.
PMI-Masking improves MLM pretraining by masking correlated spans efficiently.
problem Uniform token masking leads to inefficient and suboptimal performance in MLMs.
method PMI-Masking uses Pointwise Mutual Information to mask n-grams with high collocation.
result PMI-Masking reaches half the training time and improves performance.
Universal Language Model for Fine-tuning [arXiv:1801.06146] (ULMFiT) is one of the first NLP methods for efficient inductive transfer learning. Unsupervised pretraining results in improvements on many NLP tasks for English. In this paper, we describe a new method that uses subword tokenization to adapt ULMFiT to langua…
SFAVEL distills features from language models for fact verification without annotations.
problem Fact verification with semantically meaningful and compact features.
method Self-supervised pretraining using contrastive loss with language models.
result Achieved state-of-the-art results on FB15k-237 and FEVER.
This study examines how large language models learn in-context and provides insights into their performance and architecture.
problem Understanding how large language models learn in-context and their performance metrics.
method Bayesian model averaging, parameterization, and statistical analysis of transformer architecture.
result Transformer architecture enables in-context learning through attention mechanisms and fine-grained statistical analysis.
Pretrained masked language models (MLMs) require finetuning for most NLP tasks. Instead, we evaluate MLMs out of the box via their pseudo-log-likelihood scores (PLLs), which are computed by masking tokens one by one. We show that PLLs outperform scores from autoregressive language models like GPT-2 in a variety of task…
Method constructs finance LLMs without instruction data using pretraining and model merging.
problem Developing domain-specific LLMs for finance is resource-intensive.
method Continual pretraining on financial data + model merging of instruction-tuned and domain-specific pretrained vectors.
result Successfully constructs instruction-tuned LLMs for finance without additional instruction data.
The paper studies how search and distillation improve reasoning in large language models.
problem Improving reasoning capabilities of large language models.
method Viewing chain-of-thought generation as a metastable Markov process, proving benefits of search and distillation.
result Search protocol rewards sparse edges, reducing the expected number of steps to reach different clusters.
New model uses pretrained biochemical language models to generate drug compounds.
problem Developing novel compounds targeting specific proteins.
method Exploits pretrained language models to initialize and fine-tune targeted molecule generation models.
result Warm-started models outperform baseline models, with one-stage strategy showing better generalization.
Data selection boosts fact memorization in language models.
problem Language models struggle to accurately memorize factual knowledge.
method Formalizes fact memorization, proposes data selection schemes based on training loss.
result Data selection boosts fact accuracy to model capacity and improves performance.
Bias correction improves language model training performance.
problem Stochastic update bias in preconditioned optimizers.
method Cross-fitted preconditioning and variance-corrected inversion.
result Reduces held-out pretraining loss by 0.15 nats.
FinPT uses large pretrained models to predict financial risks.
problem Outdated algorithms and lack of open financial benchmarks.
method Profile Tuning on large pretrained foundation models.
result Demonstrated effectiveness on FinBench datasets.
NoLBERT avoids lookback and lookahead biases for better econometric inference.
problem Information leakage in language models affects econometric inference.
method Pretrained on text from 1976-1995, avoiding lookback and lookahead biases.
result NoLBERT outperforms domain-specific baselines and predicts higher profit growth.
CERT improves language understanding by contrastively learning sentence-level semantics.
problem Lack of sentence-level semantics in existing pretraining tasks.
method Contrastive self-supervised learning at the sentence level using back-translation augmentations.
result CERT outperforms BERT on 7 out of 11 GLUE benchmark tasks, achieving the same performance as BERT on 2 tasks.
Inductive transfer learning has greatly impacted computer vision, but existing approaches in NLP still require task-specific modifications and training from scratch. We propose Universal Language Model Fine-tuning (ULMFiT), an effective transfer learning method that can be applied to any task in NLP, and introduce tech…
Transformers can learn new tasks from diverse pretraining data but struggle with out-of-domain tasks.
problem Transformer models' ability to learn new tasks in-context is limited by their pretraining data coverage.
method Investigation of transformer models trained on (x,f(x)) pairs, comparing in-context learning capabilities across different task families. result Transformers can identify and learn within task families in their pretraining data but fail with out-of-domain tasks.
Hi-RES framework extracts medical relations from articles and EHRs.
problem Manual annotation bottleneck in relation extraction.
method Labeling sentences, creating improved negative samples, using pretrained language models, and combining EHR embeddings.
result Significant accuracy increases in relation extraction, up to 0.998 for disorder-location relations.
Clinical notes in electronic health records contain highly heterogeneous writing styles, including non-standard terminology or abbreviations. Using these notes in predictive modeling has traditionally required preprocessing (e.g. taking frequent terms or topic modeling) that removes much of the richness of the source d…
Theoretical framework explains why few epochs are enough for LLM fine-tuning.
problem Understanding why few epochs are sufficient for LLM fine-tuning.
method Combining early stopping theory with attention-based Neural Tangent Kernel (NTK) for LLMs.
result Formalizes convergence rate of attention-based fine-tuning with respect to sample size.
Model forgets examples; this research predicts which ones to replay.
problem Language models forget examples during updates, leading to errors.
method Train forecasting models to predict which examples will be forgotten.
result Forecasting models can reduce forgetting of upstream pretraining examples.
GAPA method provides efficient uncertainty quantification for pretrained networks.
problem Reliable uncertainty estimates for pretrained models are challenging.
method Post-hoc Gaussian Process Activations (GAPA) method that shifts Bayesian modeling from weights to activations.
result GAPA method provides efficient uncertainty quantification without altering the backbone's predictions.
Adaptive batch size schedules improve language model training efficiency and generalization.
problem Dilemma of choosing batch sizes in large-scale model training.
method General-purpose adaptive batch size schedules compatible with data and model parallelism.
result Adaptive batch size schedules outperform constant batch sizes and heuristic warmup schedules.
New bounds show large language models can generalize beyond training data.
problem Generalization of large language models beyond training data.
method Compression bound derivation using prediction smoothing and SubLoRA.
result Large language models can discover generalizable regularities.
This work analyzes CoT prompting methods from a statistical estimation perspective.
problem Improving the effectiveness of LLMs in solving multi-step reasoning problems.
method Introducing a multi-step latent variable model to characterize CoT prompting from a statistical estimation viewpoint.
result The CoT estimator is equivalent to a Bayesian estimator when the pretraining dataset is large.
ScheduleFree+ improves large language model training without schedules or learning rates.
problem Scaling up Schedule-Free Learning to large language models.
method Learning-rate-free and schedule-free method for training large language models.
result ScheduleFree+ outperforms SOTA schedules by 31% at 1000 tokens per parameter.
We introduce a pretraining technique called Selfie, which stands for SELFie supervised Image Embedding. Selfie generalizes the concept of masked language modeling of BERT (Devlin et al., 2019) to continuous data, such as images, by making use of the Contrastive Predictive Coding loss (Oord et al., 2018). Given masked-o…
New framework analyzes LLM personalization trade-offs under congestion.
problem Tension between personalization and resource sharing in LLMs.
method Developed a statistical-economic framework to model user incentives.
result Congestion can flip rankings of SFT and ICL, and offers both methods never hurt profits.
Improving optimization for iterate-averaged language models
problem How to optimize the averaged model returned by Language Model pipelines
method Formulating optimizer design as an optimal-control problem
result Proven convergence rate and strict improvement in squared error
WSD schedule improves model training efficiency by adapting learning rates dynamically.
problem Fixed compute budgets limit training efficiency of language models.
method Introduces a WSD schedule that uses a constant learning rate followed by a rapid decay phase.
result WSD schedule generates a non-traditional loss curve with stable and decay phases.
AudioPaLM combines text and speech models to improve speech processing and translation.
problem Improving speech understanding and generation models.
method Unified multimodal architecture combining text and speech models.
result AudioPaLM significantly outperforms existing speech translation systems.