zip2zip: Inference-Time Adaptive Tokenization via Online Compression

This study investigates the impact of tokenizer design on the performance of time series foundation models for forecasting. It emphasizes the significance of scaling and quantization strategies, revealing that the configuration of tokenizers is crucial for the model's representational capacity and stability. The research demonstrates that pretrained models benefit from well-designed tokenizers, especially with smaller vocabularies, while misaligned tokenization can negate the advantages of pretraining.

The paper examines entropy calibration in language models, focusing on whether their entropy aligns with log loss on human text. Previous studies indicated that as text generation lengthens, entropy increases while text quality declines, highlighting a fundamental issue in autoregressive models. The authors investigate whether miscalibration can improve with scale and if calibration without tradeoffs is theoretically feasible, analyzing the scaling behavior concerning dataset size and power law exponents.

The study investigates the ability of transformer models to predict long steps in the Collatz sequence, a complex arithmetic function that maps odd integers to their successors. The accuracy of the models varies significantly depending on the base used for encoding, achieving up to 99.7% accuracy for bases 24 and 32, while dropping to 37% and 25% for bases 11 and 3. Despite these variations, all models exhibit a common learning pattern, accurately predicting inputs with similar residuals modulo 2^p.

Higher-order Neural Additive Models (HONAMs) have been introduced as an advancement over Neural Additive Models (NAMs), which are known for their predictive performance and interpretability. HONAMs address the limitation of NAMs by effectively capturing feature interactions of arbitrary orders, enhancing predictive accuracy while maintaining interpretability, crucial for high-stakes applications. The source code for HONAM is publicly available on GitHub.

Vision-Language Models (VLMs) are emerging models that integrate visual content with natural language. Current methods typically fuse data either early in the encoding process or late through pooled embeddings. This paper introduces a lightweight fusion module utilizing cross-only, bidirectional attention layers to align hidden states from both modalities, enhancing understanding while keeping encoders non-causal. The proposed method aims to improve the performance of VLMs by leveraging the inherent structure of visual and textual data.

A study published on arXiv examines the learning capabilities of infants and language models (LMs) regarding attested versus impossible languages. The research indicates that both groups find attested languages easier to learn than those with unnatural structures. However, the findings reveal that LMs can learn many impossible languages as effectively as attested ones. The study suggests that the complexity of these languages, rather than their impossibility, contributes to the challenges faced by LMs, which lack the human inductive biases essential for language acquisition.

The paper titled 'Are language models rational? The case of coherence norms and belief revision' explores the application of rationality norms, specifically coherence norms, to language models. It distinguishes between logical coherence norms and those related to the strength of belief. The authors introduce the Minimal Assent Connection (MAC), a new framework for understanding credence in language models based on internal token probabilities. The findings suggest that while some language models adhere to these rational norms, others do not, raising important questions about AI behavior and safety.

The paper titled 'Bias-Restrained Prefix Representation Finetuning for Mathematical Reasoning' introduces a new method called Bias-REstrained Prefix Representation FineTuning (BREP ReFT). This approach aims to enhance the mathematical reasoning capabilities of models by addressing the limitations of existing Representation finetuning (ReFT) methods, which struggle with mathematical tasks. The study demonstrates that BREP ReFT outperforms both standard ReFT and weight-based Parameter-Efficient finetuning (PEFT) methods through extensive experiments.