Semantic geometry emerges transiently early in next-token prediction training before collapsing to Neural Collapse symmetry in synthetic settings with latent semantic factors.
needle-in-the-haystack problem
9 Pith papers cite this work. Polarity classification is still indexing.
9
Pith papers citing it
representative citing papers
During pretraining, language models exhibit natural ungrokking where learned rules are forgotten based on their support frequency in the corpus, with asymmetric editability of rule survival.
Supervised fine-tuning increases LLM hallucinations via interference among overlapping semantic representations; self-distillation mitigates this by regularizing output-distribution drift while freezing parameters preserves performance when new facts are unnecessary.
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.
Deep sequence models develop geometric memory in embeddings that encodes novel global relationships, transforming l-fold composition tasks into 1-step navigation via a natural spectral bias connected to Node2Vec.
Balanced parametric and in-context knowledge use in LLMs is an emergent property requiring intra-document repetition, moderate inconsistency, and skewed distributions in training data.
Empirical tracing across model families shows verification precedes and outlasts generation for facts, with updates producing simultaneous verification of old and new answers.
FINCH is a loss-adaptive learning-rate schedule that reduces forgetting by 93% on average during LLM fine-tuning while matching standard task performance across several benchmarks.
Activation verbalization methods for LLMs largely reflect the verbalizer model's parametric knowledge rather than privileged information from the target model's activations.
citing papers explorer
-
Structure Before Collapse: Transient semantic geometry in next-token prediction
Semantic geometry emerges transiently early in next-token prediction training before collapsing to Neural Collapse symmetry in synthetic settings with latent semantic factors.
-
Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining
During pretraining, language models exhibit natural ungrokking where learned rules are forgotten based on their support frequency in the corpus, with asymmetric editability of rule survival.
-
Why Fine-Tuning Encourages Hallucinations and How to Fix It
Supervised fine-tuning increases LLM hallucinations via interference among overlapping semantic representations; self-distillation mitigates this by regularizing output-distribution drift while freezing parameters preserves performance when new facts are unnecessary.
-
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.
-
Deep sequence models tend to memorize geometrically; it is unclear why
Deep sequence models develop geometric memory in embeddings that encodes novel global relationships, transforming l-fold composition tasks into 1-step navigation via a natural spectral bias connected to Node2Vec.
-
How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models
Balanced parametric and in-context knowledge use in LLMs is an emergent property requiring intra-document repetition, moderate inconsistency, and skewed distributions in training data.
-
The Future of Facts: Tracing the Factual Generation-Verification Gap
Empirical tracing across model families shows verification precedes and outlasts generation for facts, with updates producing simultaneous verification of old and new answers.
-
Fine-Tuning Without Forgetting via Loss-Adaptive Learning Rates
FINCH is a loss-adaptive learning-rate schedule that reduces forgetting by 93% on average during LLM fine-tuning while matching standard task performance across several benchmarks.
-
Do Activation Verbalization Methods Convey Privileged Information?
Activation verbalization methods for LLMs largely reflect the verbalizer model's parametric knowledge rather than privileged information from the target model's activations.