arXiv subjects

Beren Millidge

Publications and source records attributed to Beren Millidge.

At least 19 recordsLinked to original sources

arXiv2026

ZUNA1.1: A more flexible EEG foundation model for Denoising and Super-resolution

We introduce ZUNA1.1, a 380M-parameter diffusion autoencoder for flexible EEG signal reconstruction. ZUNA1.1 is capable of reconstructing variable length sequences of up to 30s, with an arbitrary number of EEG channels at arbitrary scalp locations, and can reconstruct arbitrary temporal intervals within channels in addition to reconstructing entire channels. We demonstrate that ZUNA1.1 performs at least on par with our earlier ZUNA1 model, while being far more flexible and capable of handling a wide range of reconstruction tasks. ZUNA1.1 continues to substantially outperform standard EEG denoising and reconstruction methods such as spherical spline interpolation, which is ubiquitously deployed in the MNE package. The ZUNA1.1 model is released open source under the permissive Apache 2.0 license.

cs.LG↗

arXiv2026

PUFFER: Incremental Fuzzy Deduplication for Continuously Evolving Corpora

Large language model training corpora grow through successive, often redundant releases, so each release must be deduplicated against both itself and the accumulated history. At trillion-token scale, this requires incremental ingestion, bounded resident memory, deterministic retry, and dataset-scoped lifecycle control without repeated corpus-wide rebuilding. We introduce PUFFER (Provenance-aware Updatable Fuzzy Filtering for Evolving Repositories), a MinHash-LSH fuzzy-deduplication pipeline built around two design choices. First, PUFFER stores each LSH band as immutable, dataset-tagged, memory-mapped sorted segments, enabling exact historical band-key membership checks without RAM proportional to corpus size. Second, T-fanout tiered compaction periodically merges segments to control screening fanout, trading lower query cost against additional index-maintenance writes while preserving membership decisions. Across N ingested keys and K equal-sized releases, PUFFER's cumulative maintenance cost is O(N log N log_T K), compared with Theta(KN) for repeated snapshot rebuilding. Dataset-tagged segments also support dataset-scoped withdrawal: removal is constant-time for uncompacted or protected datasets, while post-compaction withdrawal reconstructs only the affected merged segment, even if the original dataset is unavailable. In our implementation, PUFFER completed cumulative index-stage ingestion for one billion documents in about 1.75 hours in a single process, using 128 bytes per document for a 16-band index. A classical resident MinHash-LSH table required about 6.5 KB per document and exceeded a 900 GiB RAM cap. In a ten-hour comparison capped at one billion documents, PUFFER was 11x faster than LSHBloom and 35x faster than Milvus-LSH. PUFFER is deployed on more than 30 billion documents, and we release it as open-source software at https://github.com/Zyphra/puffer.

cs.DB↗

arXiv2026

ZONOS2 Technical Report

We present ZONOS2 8B, our latest TTS model, which achieves state-of-the-art naturalness, prosody, and voice cloning fidelity. We improve upon Zonos-v0.1 across scale, data, and training recipe. We scale the model from 1.6B to 8B total parameters (900M active) with a novel mixture-of-experts (MoE) backbone, improving inference latency and throughput. We expand our training corpus from 200K to over 6M hours using a new data processing pipeline, and we simplify our post-training and conditioning recipes to improve naturalness and voice cloning fidelity. We evaluate ZONOS2 8B on quality, speaker similarity, WER, and ZTTS1-Eval, our novel TTS benchmark, where it performs competitively with state-of-the-art systems while maintaining good streaming latency. We release our model weights and example inference code under an Apache 2.0 license on GitHub and Hugging Face.

cs.SD↗

arXiv2026

Can Scale Save Us From Plasticity Loss in Large Language Models?

The loss of plasticity - the ability of a network to learn new information after having already learned older information - is a fundamental challenge in creating artificial neural networks capable of continual learning. Although this phenomenon has been known for decades, it has mostly been studied in older, relatively small architectures and rarely in natural-language domains. To determine whether loss of plasticity remains a problem in the modern transformer-based LLM paradigm, we study plasticity loss in GPT-style Transformer models trained on a multilingual continual learning problem. Consistent with prior work, we find evidence of plasticity loss across models ranging from 5M to 314M non-embedding parameters, as measured by deterioration on a held-out Vietnamese probing task. We further find that the onset of plasticity loss follows a predictable scaling law, growing sublinearly with model size. These results suggest that larger models may delay the measurable effects of plasticity loss, but that increasing parameter count alone is likely to be insufficient to completely prevent it. We also find evidence of plasticity loss under stationary multilingual training, challenging the view that the phenomenon is exclusive to continual learning with abrupt task changes. Overall, our results suggest that even large Transformer language models trained on natural-language will eventually lose the ability to efficiently adapt to new data after sufficiently long training, in both continual and stationary settings.

cs.AI↗

arXiv2026

Scaling Adaptive Depth with Norm-Agnostic Residual Networks

Residual architectures are ubiquitous in deep learning, but they suffer from a subtle structural limitation: the norm of the residual stream can grow rapidly with depth. As a result, updates from later layers become small relative to the accumulated residual state. This reduces their impact on the representation and limits the benefits of scaling models in depth. To address this, we introduce NAG, a norm-agnostic residual architecture that separates magnitude from directional information in the residual stream, preserving meaningful layer contributions throughout depth and preventing later updates from being systematically suppressed by residual-norm growth. Importantly, NAG introduces only a negligible number of additional parameters and relies on simple operations that are easily kernel-fusible, preserving training efficiency in practice. We show that this architecture outperforms baseline Transformers, with gains that increase substantially as depth grows, enabling effective training of much deeper models. The norm-agnostic formulation also leads to an interpretable Mixture-of-Depths (MoD) mechanism that adaptively skips both attention and MLP layers. Beyond serving as a post-training accuracy-compute tradeoff, this mechanism can be used as a pretraining-time scaling strategy: under iso-FLOP training, compute saved by reducing per-token forward-pass cost can be reinvested into training on more tokens while keeping the total parameter count and KV-cache budget fixed. In our experiments, moderate Mixture-of-Depths rates of approximately 20%-25% match full-depth baseline performance under equal training compute while substantially reducing the number of executed layer parameters and forward-pass FLOPs. These results identify sparsity in depth as a new scaling axis for fixed-compute training, enabling very deep yet FLOP-efficient models.

cs.LG↗

arXiv2026

Zamba2-VL Technical Report

We present Zamba2-VL, a suite of vision-language models built on Zamba2, a hybrid language-model architecture combining Mamba2 state-space layers with a small number of shared transformer blocks. Across a broad range of image understanding, reasoning, OCR, grounding, and counting benchmarks, Zamba2-VL is competitive with leading Transformer-based open-weight VLMs of comparable scale, including the Molmo2, Qwen3-VL, and InternVL3.5 families, and substantially outperforms prior SSM-based and hybrid VLMs such as VL-Mamba, Cobra, and mmMamba. Inheriting the near-linear prefill compute and small, near-constant recurrent state of its Zamba2 backbone, Zamba2-VL delivers roughly an order of magnitude lower time-to-first-token (TTFT) than these Transformer baselines at matched parameter scale, with the efficiency gap most pronounced at the smaller 1.2B and 2.7B scales most relevant to on-device and edge deployment. We release three models -- 1.2B, 2.7B, and 7B -- together with inference code at https://huggingface.co/collections/Zyphra/zamba2-vl.

cs.CV↗

arXiv2026

ZAYA1-VL-8B Technical Report

We present ZAYA1-VL-8B, a compact mixture-of-experts vision-language model built upon our in-house language model, ZAYA1-8B. Despite its compact size, ZAYA1-VL achieves performance competitive with leading base models such as Molmo2-4B and InternVL3.5-4B, while surpassing models including Qwen2.5-VL-3B, PLM-3B, and MolmoE-1B across a range of image understanding, reasoning, and counting benchmarks. The architecture incorporates two key innovations: (1) vision-specific LoRA adapters integrated into the LLM to increase modality-specific capacity without increasing the number of experts, and (2) bidirectional attention over image tokens within the LLM to enhance visual understanding. We detail the full training pipeline including data composition at each stage, sequence packing, and the attention masking scheme. The model comprises 9.2B total parameters, with 1.4B active parameters including the vision encoder, and is publicly available at https://huggingface.co/Zyphra/ZAYA1-VL.

cs.CV↗

arXiv2026

ZAYA1-8B Technical Report

We present ZAYA1-8B, a reasoning-focused mixture-of-experts (MoE) model with 700M active and 8B total parameters, built on Zyphra's MoE++ architecture. ZAYA1-8B's core pretraining, midtraining, and supervised fine-tuning (SFT) were performed on a full-stack AMD compute, networking, and software platform. With under 1B active parameters, ZAYA1-8B matches or exceeds DeepSeek-R1-0528 on several challenging mathematics and coding benchmarks, and remains competitive with substantially larger open-weight reasoning models. ZAYA1-8B was trained from scratch for reasoning, with reasoning data included from pretraining onward using an answer-preserving trimming scheme. Post-training uses a four-stage RL cascade: reasoning warmup on math and puzzles; a 400-task RLVE-Gym curriculum; math and code RL with test-time compute traces and synthetic code environments built from competitive-programming references; and behavioral RL for chat and instruction following. We also introduce Markovian RSA, a test-time compute method that recursively aggregates parallel reasoning traces while carrying forward only bounded-length reasoning tails between rounds. In TTC evaluation, Markovian RSA raises ZAYA1-8B to 91.9\% on AIME'25 and 89.6\% on HMMT'25 while carrying forward only a 4K-token tail, narrowing the gap to much larger reasoning models including Gemini-2.5 Pro, DeepSeek-V3.2, and GPT-5-High.

cs.AI↗

arXiv2026

Hybrid Associative Memories

Recurrent neural networks (RNNs) and self-attention are both widely used sequence-mixing layers that maintain an internal memory. However, this memory is constructed using two orthogonal mechanisms: RNNs compress the entire past into a fixed-size state, whereas self-attention's state stores every past time step growing its state (the KV cache) linearly with the sequence length. This results in orthogonal strengths and weaknesses. Self-attention layers excel at retrieving information in the context but have large memory and computational costs, while RNNs are more efficient but degrade over longer contexts and underperform for precise recall tasks. Prior work combining these mechanisms has focused primarily on naively interleaving them to reduce computational cost without regard to their complementary mechanisms. We propose the Hybrid Associative Memory (HAM) layer, which combines self-attention and RNNs while leveraging their individual strengths: the RNN compresses the entire sequence, while attention supplements it *only* with information that is difficult for the RNN to predict, which is hence the most valuable information to explicitly store. HAM layers enable data-dependent growth of the KV cache, which can be precisely controlled by the user with a single, continuous threshold. We find that this fine-grained control of the KV cache growth rate has a smooth trade-off with loss and performance. Empirically, we show that our hybrid architecture offers strong, competitive performance relative to RNNs and Transformers even at substantially lower KV-cache usage.

cs.LG↗

arXiv2026

ZUNA: Flexible EEG Superresolution with Position-Aware Diffusion Autoencoders

We present \texttt{ZUNA}, a 380M-parameter masked diffusion autoencoder trained to perform masked channel infilling and superresolution for arbitrary electrode numbers and positions in EEG signals. The \texttt{ZUNA} architecture tokenizes multichannel EEG into short temporal windows and injects spatiotemporal structure via a 4D rotary positional encoding over (x,y,z,t), enabling inference on arbitrary channel subsets and positions. We train ZUNA on an aggregated and harmonized corpus spanning 208 public datasets containing approximately 2 million channel-hours using a combined reconstruction and heavy channel-dropout objective. We show that \texttt{ZUNA} substantially improves over ubiquitous spherical-spline interpolation methods, with the gap widening at higher dropout rates. Crucially, compared to other deep learning methods in this space, \texttt{ZUNA}'s performance \emph{generalizes} across datasets and channel positions allowing it to be applied directly to novel datasets and problems. Despite its generative capabilities, \texttt{ZUNA} remains computationally practical for deployment. We release Apache-2.0 weights and an MNE-compatible preprocessing/inference stack to encourage reproducible comparisons and downstream use in EEG analysis pipelines.

eess.SP↗

arXiv2026

Online Vector Quantized Attention

Standard sequence mixing layers used in language models struggle to balance efficiency and performance. Self-attention performs well on long context tasks but has expensive quadratic compute and linear memory costs, while linear attention and SSMs use only linear compute and constant memory but struggle with long context processing. In this paper, we develop a sequence mixing layer that aims to find a better compromise between memory-compute costs and long-context processing, which we call online vector-quantized (OVQ) attention. OVQ-attention requires linear compute costs and constant memory, but, unlike linear attention and SSMs, it uses a sparse memory update that allows it to greatly increase the size of its memory state and, consequently, memory capacity. We develop a theoretical basis for OVQ-attention based on Gaussian mixture regression, and we test it on a variety of synthetic long context tasks and on long context language modeling. OVQ-attention shows significant improvements over linear attention baselines and the original VQ-attention, on which OVQ-attention was inspired. It demonstrates competitive, and sometimes identical, performance to strong self-attention baselines up 64k sequence length, despite using a small fraction of the memory of full self-attention.

cs.LG↗

arXiv2025

Equivalence of Personalized PageRank and Successor Representations

The hippocampus appears to implement two core but highly distinct functions in the brain: long term memory retrieval and planning and spatial navigation. Naively, these functions appear very different algorithmically. In this short note, we demonstrate that two powerful algorithms that have each independently been proposed to underlie the hippocampal operation for each function -- personalized page-rank for memory retrieval, and successor representations for planning and navigation, are in fact isomorphic and utilize the same underlying representation -- the stationary distribution of a random walk on a graph. We hypothesize that the core computational function of the hippocampus is to compute this representation on arbitrary input graphs.

cs.NE↗

arXiv2025

Generalising E-prop to Deep Networks

Recurrent networks are typically trained with backpropagation through time (BPTT). However, BPTT requires storing the history of all states in the network and then replaying them sequentially backwards in time. This computation appears extremely implausible for the brain to implement. Real Time Recurrent Learning (RTRL) proposes an mathematically equivalent alternative where gradient information is propagated forwards in time locally alongside the regular forward pass, however it has significantly greater computational complexity than BPTT which renders it impractical for large networks. E-prop proposes an approximation of RTRL which reduces its complexity to the level of BPTT while maintaining a purely online forward update which can be implemented by an eligibility trace at each synapse. However, works on RTRL and E-prop ubiquitously investigate learning in a single layer with recurrent dynamics. However, learning in the brain spans multiple layers and consists of both hierarchal dynamics in depth as well as time. In this mathematical note, we extend the E-prop framework to handle arbitrarily deep networks, deriving a novel recursion relationship across depth which extends the eligibility traces of E-prop to deeper layers. Our results thus demonstrate an online learning algorithm can perform accurate credit assignment across both time and depth simultaneously, allowing the training of deep recurrent networks without backpropagation through time.

cs.LG↗

arXiv2025

Training Foundation Models on a Full-Stack AMD Platform: Compute, Networking, and System Design

We report on the first large-scale mixture-of-experts (MoE) pretraining study on pure AMD hardware, utilizing both MI300X GPUs and Pollara networking. We distill practical guidance for both systems and model design. On the systems side, we deliver a comprehensive cluster and networking characterization: microbenchmarks for all core collectives (all-reduce, reduce-scatter, all-gather, broadcast) across message sizes and GPU counts over Pollara. To our knowledge, this is the first at this scale. We further provide MI300X microbenchmarks on kernel sizing and memory bandwidth to inform model design. On the modeling side, we introduce and apply MI300X-aware transformer sizing rules for attention and MLP blocks and justify MoE widths that jointly optimize training throughput and inference latency. We describe our training stack in depth, including often-ignored utilities such as fault-tolerance and checkpoint-reshaping, as well as detailed information on our training recipe. We also provide a preview of our model architecture and base model - ZAYA1 (760M active, 8.3B total parameters MoE, available at https://huggingface.co/Zyphra/ZAYA1-base) - which will be further improved upon in forthcoming papers. ZAYA1-base achieves performance comparable to leading base models such as Qwen3-4B and Gemma3-12B at its scale and larger, and outperforms models including Llama-3-8B and OLMoE across reasoning, mathematics, and coding benchmarks. Together, these results demonstrate that the AMD hardware, network, and software stack are mature and optimized enough for competitive large-scale pretraining.

cs.CL↗

arXiv2025

Compressed Convolutional Attention: Efficient Attention in a Compressed Latent Space

Multi-headed Attention's (MHA) quadratic compute and linearly growing KV-cache make long-context transformers expensive to train and serve. Prior works such as Grouped Query Attention (GQA) and Multi-Latent Attention (MLA) shrink the cache, speeding decode, but leave compute, which determines prefill and training speed, largely unchanged. We introduce Compressed Convolutional Attention (CCA), a novel attention method which down-projects queries, keys, and values and performs the entire attention operation inside the shared latent space. This simple design dramatically cuts parameters, KV-cache, and FLOPs all at once by the desired compression factor. Because CCA is orthogonal to head-sharing, we combine the two to form Compressed Convolutional Grouped Query Attention (CCGQA), which further tightens the compute-bandwidth Pareto frontier so that users can tune compression toward either FLOP or memory limits without sacrificing quality. Experiments show that CCGQA consistently outperforms both GQA and MLA at equal KV-cache compression on dense and MoE models. Additionally, we show that CCGQA outperforms all other attention methods on MoE models with half the KV-cache of GQA and MLA, achieving an 8x KV-cache compression with no drop in performance compared to standard MHA. CCA and CCGQA also dramatically reduce the FLOP cost of attention which leads to substantially faster training and prefill than existing methods. On H100 GPUs, our fused CCA/CCGQA kernel reduces prefill latency by about 1.7x at a sequence length of 16k relative to MHA, and accelerates backward by about 1.3x.

cs.CL↗

arXiv2025

Tree Attention: Topology-aware Decoding for Long-Context Attention on GPU clusters

Our formulation reveals that the reduction across the sequence axis can be efficiently computed in parallel through a tree reduction. Our algorithm, called Tree Attention, for parallelizing exact attention computation across multiple GPUs enables cross-device decoding to be performed asymptotically faster (up to 8x faster in our experiments) than state-of-the-art approaches such as Ring Attention, while also requiring significantly less communication volume and incurring 2x less peak memory. We demonstrate that Tree Attention speeds up decoding up to 4x on Llama 3.1-8B and can be applied to a variety of hardware and networking setups such as H100 DGX nodes, AMD MI300x nodes, and PCIe connected NVIDIA RTX 4090s. Our code is publicly available here: https://github.com/Zyphra/tree_attention

cs.LG↗

arXiv2024

Mixture-of-PageRanks: Replacing Long-Context with Real-Time, Sparse GraphRAG

Recent advances have extended the context window of frontier LLMs dramatically, from a few thousand tokens up to millions, enabling entire books and codebases to fit into context. However, the compute costs of inferencing long-context LLMs are massive and often prohibitive in practice. RAG offers an efficient and effective alternative: retrieve and process only the subset of the context most important for the current task. Although promising, recent work applying RAG to long-context tasks has two core limitations: 1) there has been little focus on making the RAG pipeline compute efficient, and 2) such works only test on simple QA tasks, and their performance on more challenging tasks is unclear. To address this, we develop an algorithm based on PageRank, a graph-based retrieval algorithm, which we call mixture-of-PageRanks (MixPR). MixPR uses a mixture of PageRank-based graph-retrieval algorithms implemented using sparse matrices for efficent, cheap retrieval that can deal with a variety of complex tasks. Our MixPR retriever achieves state-of-the-art results across a wide range of long-context benchmark tasks, outperforming both existing RAG methods, specialized retrieval architectures, and long-context LLMs despite being far more compute efficient. Due to using sparse embeddings, our retriever is extremely compute efficient, capable of embedding and retrieving millions of tokens within a few seconds and runs entirely on CPU.

cs.IR↗

arXiv2024

The Zamba2 Suite: Technical Report

In this technical report, we present the Zamba2 series -- a suite of 1.2B, 2.7B, and 7.4B parameter hybrid Mamba2-transformer models that achieve state of the art performance against the leading open-weights models of their class, while achieving substantial gains in inference latency, throughput, and memory efficiency. The Zamba2 series builds upon our initial work with Zamba1-7B, optimizing its architecture, training and annealing datasets, and training for up to three trillion tokens. We provide open-source weights for all models of the Zamba2 series as well as instruction-tuned variants that are strongly competitive against comparable instruct-tuned models of their class. We additionally open-source the pretraining dataset, which we call Zyda-2, used to train the Zamba2 series of models. The models and datasets used in this work are openly available at https://huggingface.co/Zyphra

cs.LG↗