TurboMamba is a new language‑model architecture that combines a global Mamba‑3 model with a small Transformer to achieve faster inference, faster training, and better token prediction. It introduces three lightweight enhancements—direct access to the previous block’s last token, a small correction to next‑token scores, and a learned mixture with recent‑token frequencies—to improve local predictions. In experiments on FineWeb‑Edu, TurboMamba outperforms Transformer, Mamba‑3, Jamba, and Zamba, achieving the lowest mean test prediction loss and significantly higher training throughput and batch latency speedups.
arXiv:2511. 20849v2 Announce Type: replace-cross Abstract: We introduce a new tokenizer for language models that minimizes the average tokens per character, thereby reducing the number of tokens needed to represent text during training and to generate text during inference.
The paper investigates whether small models distilled from larger ones behave similarly when using byte versus token tokenization. It introduces two methods—Marginalize‑It (approximate) and End‑Of‑Token (exact)—to convert token logits to byte logits, and conducts a large‑scale study on decoder‑only dense transformers ranging from 1 billion to 1 trillion bytes of data. Results show that while token‑based models excel early, byte‑based models eventually surpass them with more compute, achieving higher performance ceilings, greater data efficiency, and lower logit storage costs.
arXiv:2608.22646v1 Announce Type: new Abstract: Diffusion language models can generate many tokens in parallel, but they still require repeated denoising steps during inference. This makes generation...
arXiv:2610.00888v1 Announce Type: cross Abstract: Multi-token prediction (MTP) improves the throughput of autoregressive generation by enabling the language model to draft multiple next tokens per fo...
arXiv:2607. 14427v1 Announce Type: new Abstract: A depth-recurrent transformer applies a weight-tied core a variable number of times, and prior work has shown that training with a randomized recursion count yields one checkpoint usable across a range of inference depths.
The paper demonstrates that a byte‑level BPE tokenizer can be sliced to create multiple vocabulary sizes from a single trained model, preserving exact logits while reducing deployed weights by 66%. Experiments on 30 models show that while sliced models match the full model numerically, they underperform fixed‑cap specialists by a few percentage points in bits‑per‑byte. Multi‑cap training improves robustness to typographical noise, suggesting benefits from training across multiple granularities rather than from control tokens alone.
HyCo is a hybrid drafting method that assigns depth to a server model’s multi‑token prediction head for execution on a mobile NPU, while alternative branches are handled by a small language model in a single pass. By returning only the hidden state, input embedding, and key‑value entries, HyCo reduces server GPU time per token by 38.6% on Spec‑Bench and 29.2% on 32K‑token summarization, and achieves faster delivery and higher savings than an SLM‑only drafter. The approach also incorporates quantization, a restricted vocabulary, a 1K‑token window, and an ε‑threshold rule to relax exact sampling distribution preservation.
Verification-Aware Training (VAT) is a plug‑in framework that improves speculative decoding for large language models by simulating verification during training and using the resulting accept/reject patterns as supervision. VAT adds a lightweight binary verification head to predict whether each draft token will survive sequential verification, and replaces the fixed per‑position weighting with a verification‑adaptive schedule that keeps full weight up to the first rejection point. When applied to EAGLE‑3 and DFlash on Qwen3‑4B, Qwen3‑8B, and LLaMA‑3.1‑8B, VAT increases average acceptance length by up to 11.4% and wall‑clock speedup by up to 8.7%, yielding consistent gains across math, code, and chat benchmarks.
arXiv:2609. 04010v1 Announce Type: new Abstract: Large Language Models (LLMs) owe much of their success to next-token prediction (NTP), but their autoregressive (AR) structure requires slow, sequential token generation.
Instella‑MoE is a fully open Mixture‑of‑Experts language model with 16 billion total parameters and 2.8 billion active parameters per token, trained from scratch on AMD Instinct GPUs. It incorporates a sparsely activated MoE design with Gated Multi‑head Latent Attention and FarSkip‑Collective connectivity, and follows a multi‑stage pipeline that includes pre‑training, long‑context extension, supervised fine‑tuning, direct preference optimization, and reinforcement learning with Multi‑Teacher On‑Policy Distillation. The model achieves an average score of 76.7 on pre‑training benchmarks and 73.2 on instruction‑following, reasoning, math, coding, and chat benchmarks, outperforming comparable fully open and open‑weight models, and its full training pipeline, weights, and code are released for reproducibility.
arXiv:2606. 10932v1 Announce Type: cross Abstract: We present Density Field State Space Models (DF-SSM), a framework for compressing SSMs to a 1-bit scaffold with int8 low-rank correction.
Manacá-1B is a 1.72‑billion‑parameter, open decoder‑only language model trained from scratch for Brazilian Portuguese, released with a fully containerized, reproducible training pipeline and complete logs. The authors evaluate it against nine open baselines on four Portuguese benchmarks, reporting standard errors and paired significance tests, and find that Manacá-1B outperforms smaller models on LAMBADA‑PT while remaining competitive on commonsense completion. They also uncover a tokenizer‑related evaluation pitfall that can drastically lower accuracy and provide a simple fix, releasing all code, logs, and corrected tokenizer for full reproducibility.