cloud ai platform teams

Abstract: We study an extension of the Guided Speculative Inference (GSI) algorithm for test-time-scaling of large language models (LLMs) to a novel setting where an ensemble of smaller LLMs proposes draft reasoning steps which are later verified by a larger base model and a process reward model. Using the established framework of multiple importance sampling (MIS), we introduce weighting schemes to aggregate samples from a diverse collection of proposal distributions. We provide new bounds on the KL divergence from the optimal reward-tilted target distribution and the expected reward degradation, then demonstrate that the well-known balance heuristic is optimal. Compared to GSI with a single well-tuned drafter, combining several diverse LLMs while retaining the same total number of samples maintains competitive pass-at-$N$ downstream performance while being mathematically robust to situations where the support of a single draft model does not cover that of the target, where GSI may fail.

Abstract: Federated prompt learning enables vision-language models like CLIP to collaboratively adapt lightweight soft prompts across distributed clients without sharing raw data. However, existing federated prompt learning methods are designed primarily for supervised settings with labeled client data and assumed relatively stable training distributions. At test time, source-free federated prompt tuning remains largely unexplored. Clients may receive unlabeled target samples with heterogeneous, often non-IID distributions that differ substantially across clients and from the source training distribution, causing locally adapted prompts to drift in different directions and reducing the effectiveness of federated aggregation. In this work, we explore Source-Free Federated Test-Time Prompt Tuning (SF-FTTPT), where distributed clients collaboratively adapt soft prompts on unlabeled test streams without access to source data. Our study reveals that conventional federated aggregation methods struggle under heterogeneous client distributions, as locally adapted prompts become increasingly specialized and difficult to combine effectively. To address this challenge, we propose FedRAPP, a personalized aggregation framework that combines reliability-aware client assessment, server-side exponential moving-average prompt memory, and similarity-guided peer aggregation. FedRAPP selectively incorporates information from reliable and compatible clients while preserving local specialization, enabling more robust adaptation under distribution shift. Extensive experiments across nine CLIP benchmark datasets and ImageNet variants demonstrate that FedRAPP consistently outperforms standard federated aggregation baselines under heterogeneous client distributions while maintaining the efficiency advantages of prompt-only adaptation.

Abstract: Frontier AI systems are increasingly natively multimodal, jointly pretrained on multiple modalities, such as text, vision, and audio. However, the scientific understanding of the multimodal data mixtures used to pretrain these systems --- how much of each modality to mix, and at what potential interference cost to the others --- is largely absent or gatekept. In this work, we derive data mixing scaling laws for three modalities: text, vision, and audio. Multimodal Large Language Models (MLLMs) increasingly adopt Mixture-of-Experts (MoE) architectures, as MoEs scale efficiently and partially limit cross-modal interference. We therefore use MoEs to conduct 268 experiments ranging from 457M to 8.3B parameters, on up to 150 billion tokens. Our work leads to multiple actionable insights into multimodal data mixing for training MLLMs. First, compute optima are modality-specific: audio modality is model-heavy, favoring scaling parameters over tokens, while text and vision require balanced scaling of model size and data. Second, repeating vision or audio data beyond 2x yields negligible benefits. Finally, cross-modal interactions are highly asymmetric: vision strongly interferes with audio, whereas audio's interference with learning text and vision is negligible. Overall, we provide a principled foundation for understanding the multimodal data mixtures needed to train frontier MLLMs.

Abstract: Efficient attention mechanisms, such as linear and sliding window attention, aim to reduce the computational complexity of Transformers from quadratic to linear. However, their degraded in-context recall often necessitates interleaving full attention layers, leaving the quadratic bottleneck unresolved in long-context processing. To address this, we introduce Orthros attention, which shifts the efficiency paradigm from structural allocation to phase-aware decoupling. As a unified module, it dynamically switches its computation pattern across different phases of sequence processing, enabling linear complexity prefilling and full attention decoding. Our design grants Orthros attention a lower asymptotic complexity compared to vanilla Transformers in prevalent long-context scenarios characterized by heavy prefilling and light decoding. Beyond theoretical formulation and complexity analysis, extensive empirical evaluations demonstrate Orthros's superiority across multiple dimensions. First, we validate its fundamental language modeling capabilities, demonstrating scalability across varying model sizes and strong competitiveness in comprehensive architectural comparisons. Second, Orthros exhibits exceptional conversion feasibility via a data-efficient Transformer-to-Orthros conversion, yielding performance that even surpasses the original Transformer. Finally, Orthros achieves a $21.9\times$ prefilling speedup over Transformers with end-to-end inference gains, while preserving long-context capabilities, as demonstrated by near-perfect 128K needle-in-a-haystack results and Transformer-level performance on real-world in-context recall tasks. Overall, these advantages in high performance, flexibility, and efficiency position Orthros as a promising candidate to succeed the Transformer in industrial applications.

Abstract: Training a Mixture-of-Experts (MoE) model at long context or large batch size fails as soon as any one component's peak allocation exceeds device memory, so the target is every peak at once, not the average footprint. Four are left unbounded by the parallelism plans in common use, and each grows differently: expert dispatch with the routing matrix, the vocabulary projection with tokens times vocabulary, gradient checkpoint boundaries with depth times sequence length, and optimizer state with parameter count. Which one runs out first changes with the model, the context length, and the device count, so lowering the largest only exposes the next. We bound all four with schedules whose GPU working set is fixed at launch: PipelinedLLEP extends least-loaded expert parallelism with a cap on the tokens each source contributes to a dispatch chunk, Ring-DTP circulates activations or weight shards around a ring at the vocabulary projection and folds each block of logits into an online log-sum-exp, Selective checkpoint offload (SCO) keeps the one long-lived tensor of each checkpoint boundary in CPU memory, and OffloadStreamAdamW turns the serial CPU Adam update of optimizer offload into a bucket pipeline. All four change only the order and granularity of computation and data movement, so the loss and gradients stay exact. In matched component tests, they cut the MoE dispatch peak by up to $59.3\%$ without losing throughput, the vocabulary projection peak by $86.6\%$, and the offloaded optimizer step by $2.05\times$ faster. Composed on MoE models from 120B to 667B parameters, they train at 1M context length, $8$--$32\times$ the reach of a tuned FSDP2 baseline, and up to $10.4\times$ its throughput.

Abstract: Neural Architecture Search (NAS) automates network design, but evaluating a single candidate requires training it to convergence, making exhaustive search intractable. Zero-cost proxies estimate architecture quality at initialization in seconds, yet a single proxy is noisy, and combining several does not straightforwardly help: proxies are strongly correlated, so naive aggregation compounds their shared errors instead of averaging them out. Existing methods exploit either proxy signals or architectural topology - never both within a single active-learning framework. We introduce ZAPS (Zero-cost Active Proxy Search), a four-stage pipeline that closes this gap. ZAPS (i) selects a compact, non-redundant proxy subset offline via ProxyFit, a greedy anti-redundancy criterion; (ii) seeds the search with a hybrid K-means strategy that balances exploitation and exploration; (iii) re-selects proxies at every iteration by a bootstrapped vote as the labeled set grows; and (iv) ranks candidates with an XGBoost ensemble trained jointly on proxy ranks and one-hot topological encodings, queried through an Upper Confidence Bound (UCB) acquisition function. On NAS-Bench-201 under a budget of B=200 evaluations, ZAPS recovers 52.3% of the true top-100 architectures on CIFAR-10 and 65.8% on CIFAR-100, ahead of every baseline we consider - Random Search, Local Search, REA, BANANAS and TPE - and, on CIFAR-10, with less than half the run-to-run standard deviation of the strongest of them. The advantage is largest where evaluations are scarce: on NAS-Bench-201 it narrows as the budget grows, whereas on the harder NAS-Bench-101, which no method comes close to saturating, it widens instead. All methods are scored by a single criterion: how much of the true top-100 lies among the architectures they actually evaluated.