Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
hub
Language models are unsupervised multitask learners
21 Pith papers cite this work. Polarity classification is still indexing.
21
Pith papers citing it
hub tools
representative citing papers
Energy-navigated trajectory shaping during training produces 8-step discrete flow matching students that achieve 32% lower perplexity than 1024-step teachers on 170M language models with unchanged inference cost.
A Markov category framework for language models provides an information-theoretic rationale for speculative decoding and shows that a quadratic surrogate to negative log-likelihood induces generalized CCA alignment in linear-softmax heads after normalization.
PG-DLM applies particle Gibbs sampling over full trajectories in diffusion language models to enable iterative refinement, yielding higher accuracy on reward-guided generation with theoretical convergence guarantees.
K-sparse autoencoders with dead-latent fixes produce clean scaling laws and better feature quality metrics that improve with size, shown by training a 16-million-latent model on GPT-4 activations.
Chronos pretrains transformer models on tokenized time series to deliver strong zero-shot forecasting across diverse domains.
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
Time-LLM reprograms frozen LLMs for time series forecasting via text prototypes and Prompt-as-Prefix, outperforming specialized models in standard, few-shot, and zero-shot settings.
Activation Addition steers language models by adding contrastive activation vectors from prompt pairs to control high-level properties like sentiment and toxicity at inference time without training.
LLaMA-Adapter turns frozen LLaMA 7B into a capable instruction follower using only 1.2M new parameters and zero-init attention, matching Alpaca while extending to image-conditioned reasoning on ScienceQA and COCO.
Gated linear attention enables lower training and test errors in non-stationary in-context learning by adaptively modulating past inputs through a learnable recency bias under an autoregressive model of task evolution.
Loss-based pruning of training data to limit facts and flatten their frequency distribution enables a 110M-parameter GPT-2 model to memorize 1.3 times more entity facts than standard training, matching a 1.3B-parameter model on the full dataset.
Geometry Forcing aligns video diffusion representations with geometric foundation model features via angular cosine and scale regression objectives to improve 3D consistency in generated videos.
Introduces Adaptive Clustering router for MoE models that scales features to identify tight expert clusters, yielding faster convergence, robustness to corruption, and performance gains.
Attention sinks emerge in language models from softmax-induced token dependence on attention scores and do not appear when using sigmoid attention without normalization in models up to 1B parameters.
Router models trained on preference data dynamically select between strong and weak LLMs, cutting inference costs by more than 2x on benchmarks with no quality loss and showing transfer to new model pairs.
DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.
Chain-of-Verification reduces hallucinations in large language models by drafting responses, planning independent verification questions, answering them separately, and generating a final verified output.
Improved ViT-VQGAN enables autoregressive Transformer pretraining on ImageNet tokens to reach IS 175.1 and FID 4.17 for generation plus 73.2% linear-probe accuracy, beating prior iGPT models.
Toeplitz MLP Mixers replace attention with masked Toeplitz multiplications for sub-quadratic complexity while retaining more sequence information and outperforming on copying and in-context tasks.
RAFT aligns generative models by ranking samples with a reward model and fine-tuning only on the top-ranked outputs, reporting gains on reward scores and automated metrics for LLMs and diffusion models.
citing papers explorer
-
Progress measures for grokking via mechanistic interpretability
Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
-
Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation
Energy-navigated trajectory shaping during training produces 8-step discrete flow matching students that achieve 32% lower perplexity than 1024-step teachers on 170M language models with unchanged inference cost.
-
A Markov Categorical Framework for Language Modeling
A Markov category framework for language models provides an information-theoretic rationale for speculative decoding and shows that a quadratic surrogate to negative log-likelihood induces generalized CCA alignment in linear-softmax heads after normalization.
-
Inference-Time Scaling of Diffusion Language Models via Trajectory Refinement
PG-DLM applies particle Gibbs sampling over full trajectories in diffusion language models to enable iterative refinement, yielding higher accuracy on reward-guided generation with theoretical convergence guarantees.
-
Scaling and evaluating sparse autoencoders
K-sparse autoencoders with dead-latent fixes produce clean scaling laws and better feature quality metrics that improve with size, shown by training a 16-million-latent model on GPT-4 activations.
-
Chronos: Learning the Language of Time Series
Chronos pretrains transformer models on tokenized time series to deliver strong zero-shot forecasting across diverse domains.
-
Self-Rewarding Language Models
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
-
Time-LLM: Time Series Forecasting by Reprogramming Large Language Models
Time-LLM reprograms frozen LLMs for time series forecasting via text prototypes and Prompt-as-Prefix, outperforming specialized models in standard, few-shot, and zero-shot settings.
-
Steering Language Models With Activation Engineering
Activation Addition steers language models by adding contrastive activation vectors from prompt pairs to control high-level properties like sentiment and toxicity at inference time without training.
-
LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention
LLaMA-Adapter turns frozen LLaMA 7B into a capable instruction follower using only 1.2M new parameters and zero-init attention, matching Alpaca while extending to image-conditioned reasoning on ScienceQA and COCO.
-
Learning to Adapt: In-Context Learning Beyond Stationarity
Gated linear attention enables lower training and test errors in non-stationary in-context learning by adaptively modulating past inputs through a learnable recency bias under an autoregressive model of task evolution.
-
Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts
Loss-based pruning of training data to limit facts and flatten their frequency distribution enables a 110M-parameter GPT-2 model to memorize 1.3 times more entity facts than standard training, matching a 1.3B-parameter model on the full dataset.
-
Geometry Forcing: Marrying Video Diffusion and 3D Representation for Consistent World Modeling
Geometry Forcing aligns video diffusion representations with geometric foundation model features via angular cosine and scale regression objectives to improve 3D consistency in generated videos.
-
Tight Clusters Make Specialized Experts
Introduces Adaptive Clustering router for MoE models that scales features to identify tight expert clusters, yielding faster convergence, robustness to corruption, and performance gains.
-
When Attention Sink Emerges in Language Models: An Empirical View
Attention sinks emerge in language models from softmax-induced token dependence on attention scores and do not appear when using sigmoid attention without normalization in models up to 1B parameters.
-
RouteLLM: Learning to Route LLMs with Preference Data
Router models trained on preference data dynamically select between strong and weak LLMs, cutting inference costs by more than 2x on benchmarks with no quality loss and showing transfer to new model pairs.
-
Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive
DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.
-
Chain-of-Verification Reduces Hallucination in Large Language Models
Chain-of-Verification reduces hallucinations in large language models by drafting responses, planning independent verification questions, answering them separately, and generating a final verified output.
-
Vector-quantized Image Modeling with Improved VQGAN
Improved ViT-VQGAN enables autoregressive Transformer pretraining on ImageNet tokens to reach IS 175.1 and FID 4.17 for generation plus 73.2% linear-probe accuracy, beating prior iGPT models.
-
Toeplitz MLP Mixers are Low Complexity, Information-Rich Sequence Models
Toeplitz MLP Mixers replace attention with masked Toeplitz multiplications for sub-quadratic complexity while retaining more sequence information and outperforming on copying and in-context tasks.
-
RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment
RAFT aligns generative models by ranking samples with a reward model and fine-tuning only on the top-ranked outputs, reporting gains on reward scores and automated metrics for LLMs and diffusion models.