The paper demonstrates that Mixture-of-Experts (MoE) decoding can avoid the need for calibration data by using a top‑k channel selector instead of a calibrated threshold. Top‑k selection consistently matches or outperforms threshold‑based methods across various read budgets and tasks, eliminating issues like repetition‑loop collapse. Additionally, the authors propose a reader that uses a ternary sketch and low‑rank mapping to recover skipped channels, achieving 98.1% of full‑model accuracy while reading only 17.8% of routed‑expert weights and speeding up MoE blocks by up to 2.69× compared to optimized vLLM.
arXiv:2609.36222v1 Announce Type: new
Abstract: Large language models are increasingly expensive to serve. In large-scale serving systems, autoregressive decoding is often bottlenecked by transferrin...
By Ali Abbasi, Justin Shi, Soheil Kolouri
Prefill‑only decision models evaluate every candidate in a menu in a single forward pass, avoiding decoding and reducing cost by one to two orders of magnitude compared to generative language models. The paper demonstrates that when only the candidate menu changes, the model’s post‑intervention accuracy can be predicted solely from the cached first‑pass distribution using a simple estimator that renormalizes and selects the argmax, without any labels or second pass. Across seven model families, ten datasets, and two task types, this menu‑only intervention prediction is within 4.2 points of actual accuracy, and in one family it is exact, whereas a probability‑level variant fails by 21 points, indicating the property resides in ranking rather than calibrated probabilities.
By Ran Li, Lei Chen
arXiv:2610.28369v1 Announce Type: cross
Abstract: A growing body of work proposes dynamic expert-skipping methods for Mixture-of-Experts (MoE) models that adjust the number of experts activated per t...
By Jakob Steimle, Thomas Elsken, Daniel Ehrhardt, Lukas Rinder, Daniel Cremers, Xi Wang
arXiv:2610. 29052v1 Announce Type: cross Abstract: A typed decision maps evidence, a question and a fixed set of options to a probability for every option, computed in one forward pass, by reading a language model's logits or a small head over its hidden states, instead of generating and parsing text.
By Yuchuan Weng, Chi Liu, Xin Chen, Srinivasan Manoharan
RaReCache is a framework that allows a large target language model to decode accurately from a KV cache prefilled by a much smaller source model by selectively recomputing only critical token positions. It identifies these positions using a rank disagreement metric that scores tokens based on the energy of their mapped KV in output directions weakly supported by calibration data. Across two model families and five benchmarks, RaReCache achieves up to a 3.04× prefill speedup and significantly reduces prefill latency while maintaining 95–99% of target accuracy with a 30–40% recompute budget.
By Sreetama Sarkar, Saptarshi Mitra, Sitao Huang, Souvik Kundu, Peter A. Beerel
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.
By Liming Liu, Mingze Wang, Tuo Zhao
arXiv:2607. 13205v1 Announce Type: cross Abstract: Attention-based KV cache eviction (H2O and its descendants) compresses the memory-constrained state of a long-context model by ranking tokens on accumulated attention mass, treated here as signal energy, and keeping the heaviest.
By Soumil Mandal
arXiv:2610.15417v1 Announce Type: new
Abstract: Mixture-of-experts (MoE) models activate only a few experts per token, but when their expert weights are offloaded to host memory, host-to-device (H2D)...
By Hongliang Cao, Xiuyu Li, Siqi Wang, Ziming Wang, Jingmin Zhu, Hailong Yang
The paper presents a post‑training method for dynamically routed mixture‑of‑experts (MoE) models that integrates independently optimized specialists efficiently. It introduces calibrated multi‑teacher on‑policy distillation (C‑MOPD) with teacher‑wise loss normalization and clipping, combined with a post‑capacity bounded‑width dispatch, final‑route replay, and asynchronous rollout–training overlap. On the Turing‑20B‑A2B model, the approach achieves a high integration score (0.9988), retains most specialist gains, outperforms shared SFT and Qwen3‑8B on reasoning‑enabled benchmarks, and reduces consolidation time by 2.86×.
By Yue He, Jiahui Hu, Shaman Tang, Hua Zhou, Qiman Wu, Menglin Li, Yizhao Wang, Yuheng Zhang, Da Zhu, Zhou Wang, Anhua Liu, Hanlin Chen, Mingyuan Wang, Rui Xin, Yuhua Wei, Shuang Su, Siqi Liu, Kai Yang, Weichao Huang, Hang Zhang, Honggou Yang, Xianming Liu
arXiv:2607. 24434v1 Announce Type: cross Abstract: Large Mixture-of-Experts (MoE) language models are attractive for end-device deployment because only a small subset of experts is active per token, but their routed expert weights often exceed accelerator memory.
By Dengke Han
arXiv:2608. 07911v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard.
By Yu Zhang