KITE (KV-Invariant Transformer Expansion) is a scaling paradigm that trains a language model from a smaller size to a larger one, saving training costs by upcycling. It places new parameters in regions that do not affect attention KV, so inference only requires prefilling KV from the smaller part, reducing inference costs. The Step Scale Transformer (SST), a two-tower decoder, demonstrates this by achieving lower training loss than comparable MoE Transformers while cutting estimated inference cost by 6.7% and 31.6%.
arXiv:2605. 09877v4 Announce Type: replace-cross Abstract: Recall presents a difficult choice: transformers have a linearly growing memory that slows each successive token, while linear RNNs typically have fixed costs but limited recall.
arXiv:2608. 12385v2 Announce Type: replace Abstract: As large language models serve ever more requests, cumulative inference cost is growing relative to the one-time cost of training.
arXiv:2602.06471v2 Announce Type: replace Abstract: The architectural shape of dense Transformers has remained remarkably stable: narrow-wide-narrow feed-forward networks (FFNs) consume most non-embe...
Scaling Large Language Models (LLMs) has been driven mainly by enlarging the Transformer backbone, but for an already-strong model this requires another round of costly pretraining. We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed.
arXiv:2608.30963v1 Announce Type: cross Abstract: Modern large language model (LLM) serving systems increasingly operate over repeated or shared context, yet each model typically performs its own pre...
arXiv:2606. 15157v1 Announce Type: cross Abstract: KV cache compression is essential for reducing the memory cost of long-context large language model inference.
arXiv:2608. 08888v1 Announce Type: new Abstract: Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth.
The paper demonstrates that architectural changes—specifically looped transformers and boundary operators—can alter scaling exponents in pre‑training, yielding exponential performance gains for a given computational budget. Looping, or recursive depth, enables model growth that matches larger models (e.g., a 7.4B looped architecture matching GPT‑3 13B) with significantly less compute, while boundary operators provide additional, though smaller, efficiency improvements. In data‑constrained, multi‑epoch scenarios, increasing loops with scale serves as a useful regularizer, suggesting that deeper computational depth drives compute‑efficiency gains that grow with model size.
arXiv:2606. 27449v1 Announce Type: new Abstract: Multi-head attention conventionally partitions the hidden dimension equally across all heads at every layer, enforcing an identical representational subspace dimension (dh = dmodel/h) throughout the models depth.
arXiv:2605. 09165v2 Announce Type: replace Abstract: Looped language models repeat a set of transformer layers through depth, reducing memory costs and providing natural early-exit points at loop boundaries.
arXiv:2609.01343v1 Announce Type: new Abstract: Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating archit...
arXiv:2606. 07878v1 Announce Type: new Abstract: The KV cache is the memory bottleneck of long-horizon language model deployment.