arXiv:2605.27033v1 Announce Type: cross 
Abstract: Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs, but it is not clear that they exploit their full capacity for all inputs. We introduce the s-Trace method to efficiently estimate the subgraph of size s that best approximates a full model output. With this method, we find the computation in a variety of LLMs to be organized in two distinct phases. A small subgraph mostly composed of early-layer nodes can reconstruct the head of the full model output distribution. Adding further nodes, mostly located in later layers and increasingly consisting of attention heads, leads to incremental refinements in approximating the full output distribution. We find moreover that the amount of necessary computation per input correlates with model uncertainty, and that sparser subgraphs encode shallow statistics, such as unigram frequency. Overall, our results suggest a consistent modular organization in effective LLM computation, with a sparse early-layer core providing a rough prediction that is further refined through denser computations in later layers.

قدمت دراسة حديثة طريقة s-Trace لتحليل كثافة الحساب في نماذج اللغة الكبيرة المستندة إلى المحولات (LLMs)، كاشفة أن هذه النماذج تعمل في مرحلتين متميزتين. تستخدم المرحلة الأولية تحت رسم بياني صغير من العقد في الطبقات المبكرة لتقريب مخرجات النموذج، بينما تتضمن المرحلة التالية عقدًا إضافية من الطبقات اللاحقة للتنقيح. وهذا يشير إلى أن نماذج LLMs قد لا تستفيد بالكامل من قدرتها الحسابية لجميع المدخلات.

Un estudio reciente presentó el método s-Trace para analizar la densidad de cálculo en modelos de lenguaje de gran tamaño basados en transformadores (LLMs), revelando que estos modelos operan en dos fases distintas. La fase inicial utiliza un pequeño subgrafo de nodos de las primeras capas para aproximar la salida del modelo, mientras que la fase posterior incorpora nodos adicionales de capas posteriores para un refinamiento. Esto sugiere que los LLMs pueden no estar utilizando completamente su capacidad computacional para todas las entradas.

Une étude récente a introduit la méthode s-Trace pour analyser la densité de calcul dans les modèles de langage de grande taille basés sur des transformateurs (LLMs), révélant que ces modèles fonctionnent en deux phases distinctes. La phase initiale utilise un petit sous-graphe de nœuds des premières couches pour approximer la sortie du modèle, tandis que la phase suivante incorpore des nœuds supplémentaires des couches ultérieures pour un affinage. Cela suggère que les LLMs ne tirent peut-être pas pleinement parti de leur capacité de calcul pour toutes les entrées.

A recent study introduced the s-Trace method to analyze computation density in transformer-based large language models (LLMs), revealing that these models operate in two distinct phases. The initial phase utilizes a small subgraph of early-layer nodes to approximate the model's output, while the subsequent phase incorporates additional nodes from later layers for refinement. This suggests that LLMs may not fully utilize their computational capacity for all inputs.

Tracing Computation Density in LLMs

arXiv:2605.26099v3 Announce Type: replace 
Abstract: Transformer-based large language models are increasingly used for long-horizon tasks; however, their attention mechanism scales poorly with context length. To handle this, we study a sleep-like consolidation mechanism in which a model periodically converts recent context into persistent fast weights before clearing its key-value cache. During sleep, the model performs $N$ offline recurrent passes over the accumulated context and updates the fast weights in its state-space model (SSM) blocks through a learned local rule. During inference, this shifts extra computation to sleep while preserving the latency of wake-time prediction. We test our method on controlled synthetic tasks, including cellular automata and multi-hop graph retrieval, as well as a realistic math reasoning task, on which a regular transformer as well as SSM-attention hybrid models fail. We then show that increasing sleep duration $N$ for our models improves performance, with the largest gains on examples that require deeper reasoning.

تبحث الأبحاث الحديثة في آلية مشابهة للنوم لنماذج اللغة المستندة إلى المحولات (LLMs)، مما يسمح لها بتوحيد السياق الحديث في أوزان سريعة دائمة خلال الفترات غير المتصلة، مما يعزز أداء الاستدلال عبر الإنترنت. تهدف هذه الطريقة إلى معالجة قيود آليات الانتباه في التعامل مع السياقات الطويلة.

Investigaciones recientes exploran un mecanismo similar al sueño para los modelos de lenguaje basados en transformadores (LLMs), permitiéndoles consolidar el contexto reciente en pesos rápidos persistentes durante períodos fuera de línea, mejorando así el rendimiento de la inferencia en línea. Este método busca abordar las limitaciones de los mecanismos de atención en la gestión de contextos largos.

Des recherches récentes examinent un mécanisme de sommeil pour les modèles de langage basés sur des transformateurs (LLMs), leur permettant de consolider le contexte récent en poids rapides persistants pendant les périodes hors ligne, améliorant ainsi les performances d'inférence en ligne. Cette méthode vise à surmonter les limitations des mécanismes d'attention dans la gestion de longs contextes.

Recent research explores a sleep-like mechanism for transformer-based large language models (LLMs), allowing them to consolidate recent context into persistent fast weights during offline periods, thereby enhancing online inference performance. This method aims to address the limitations of attention mechanisms in handling long contexts.

Tracing Computation Density in LLMs

Was this article worth reading? Share it

Ready to build your own newsroom?