The paper demonstrates that the transformer architecture naturally arises as the score of last‑token laws in a statistical model of language where sentences are treated as empirical measures of tokens. By applying the maximum‑entropy principle to a prior that assumes no inherent sentence structure, the authors derive a statistic capturing token interactions via Luce’s choice axiom, leading to the softmax choice rule and attention mechanisms. The resulting framework explains key transformer components—query‑key and value matrices as Lagrange multipliers, causal masking, multi‑head attention, and residual blocks—as consequences of entropic optimization.
arXiv:2608. 06111v1 Announce Type: cross Abstract: Positional embeddings (PE) in Transformers encode token distance and order but are largely agnostic to \textit{syntactic structure}.
The paper presents a mean‑field analysis of attention in language models, defining an average attention kernel that propagates representations layer by layer. When conditioned on a whole corpus, the kernel predicts the average evolution of representation geometry; when conditioned on a single context, it predicts the expected geometry for that context. The difference between actual attention and the mean‑field prediction—called the mean‑field deviation—captures context‑specific computation, revealing how models diverge from average behavior during training and in few‑shot tasks.
The paper investigates the theoretical expressive power of Transformers in modeling contextual relations. By framing a text as a distribution of representations and attention as a probabilistic relation, it connects attention normalization to optimal transport: softmax yields conditional relations, while Sinkhorn yields joint relations with fixed marginals. The authors prove universal approximation results, showing that Transformers with Sinkhorn normalization can represent any joint probability relation, whereas standard softmax Transformers can represent any conditional probability relation.
arXiv:2607. 18804v1 Announce Type: new Abstract: In the \emph{latent posterior model} of transformer behavior, the next-token distribution arises from a posterior over latent predictive models conditioned on the context, mixed to generate continuations.
The paper investigates how post‑softmax attention weights behave across different language models by marginalizing over query positions to create a joint token‑head space. Across more than 60 diverse LLMs, token‑wise marginal attention reveals a text‑intrinsic signal that is conserved across models, while head‑wise marginal attention yields a model‑private signature that is consistent across documents. This insight enables a practical method for estimating per‑head budgets for key‑value cache eviction, achieving competitive performance on standard benchmarks without per‑document recomputation or training.
arXiv:2607. 20524v1 Announce Type: new Abstract: Mean cross-positional attention degradation is widely reported in transformer interpretability, yet whether it causally limits contextual retrieval remains untested.
arXiv:2608. 09558v1 Announce Type: new Abstract: How expressive is prompting a transformer?
The paper investigates energy-based transformers as predictors of reading difficulty, extending the use of transformer language models in psycholinguistics. It demonstrates that the energy measure from these models robustly predicts reading times across multiple corpora, outperforming traditional metrics like surprisal and attention entropy. In a controlled experiment on relative clause processing, energy captures known asymmetries, suggesting it may unify previously complementary predictors.
arXiv:2410.02343v2 Announce Type: replace Abstract: Large language models (LLMs) routinely fail to output the correct option in multiple-choice question answering (MCQA) while encoding the answer int...
arXiv:2506.19549v2 Announce Type: replace-cross Abstract: Identifying which tokens and activations truly influence model predictions is critical for both the efficiency and interpretability of large...
The paper investigates how entropy over chain‑of‑thought tokens influences policy decisions such as gradient application, pruning, and collapse detection. By separating scaffold tokens from substantive content, the authors analyze entropy, Kullback–Leibler divergence, and entropy velocity for each channel, proving differences between raw and content conventions and bounding answer diversity. Empirical results across 23 configurations show that scaffold tokens can account for up to 41% of high‑entropy positions, with entropy share growing through distillation, while content conventions outperform raw surprisal on compression tasks and reveal significant answer leakage in re‑fed chains.
arXiv:2607. 23050v1 Announce Type: new Abstract: Neural scaling laws describe how loss decreases as models, data, and compute grow, but they do not answer a prior question: for a fixed task, what is the minimum model capacity required to solve it?