arXiv subjects

Luca Zancato

Publications and source records attributed to Luca Zancato.

At least 19 recordsLinked to original sources

arXiv2026

Priming: Hybrid State Space Models From Pre-trained Transformers

Hybrid State-Space models combine Attention with recurrent State-Space Model (SSM) layers, balancing eidetic memory from Attention with compressed fading memory from SSMs. This yields smaller Key-Value caches and faster decoding than Transformers, along with a richer architectural design space. Exploring that design space at scale has so far required training from scratch, a barrier that has kept most large-model Hybrid research within a narrow range of architectures. We introduce Priming, a method that turns Hybrid architecture design from a pre-training problem into a knowledge transfer one. Priming initializes a Hybrid model from a pre-trained Transformer and, through short alignment and post-training phases, recovers downstream quality using less than 0.5% of the source model's pre-training token budget. Priming is agnostic to the source Transformer family (e.g., Qwen, Llama, Mistral), model class (dense or Mixture-of-Experts), and model scale. Priming enables us to run the first controlled comparison of SSM layer types at scale under identical conditions. We evaluate, Gated KalmaNet (GKA), Gated DeltaNet (GDN), and Mamba-2, and show that their expressiveness hierarchy, GKA>GDN>Mamba-2, directly predicts downstream performance on long-context reasoning tasks. We scale Priming to 8B/32B reasoning models with native 128K contexts. Our Hybrid GKA 32B improves over its source Qwen3-32B by +3.8 average reasoning points, while staying within 1% of a Transformer post-trained on the same data and enabling up to 2.3x higher decode throughput. To foster research on Hybrid architectures, we release a model zoo of primed Hybrid models for long-context reasoning and instruction following, together with the Priming training and inference code (Sequence Parallelism algorithms for long-context training, optimized GKA kernels, and vLLM serving plugin), all under Apache~2.0 License.

cs.LG↗

arXiv2026

Learning When to Attend: Conditional Memory Access for Long-Context LLMs

Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval. Continued pretraining on long-context data can help but is expensive due to the quadratic scaling of Attention. We observe that most tokens do not require (Global) Attention over the entire sequence and can rely on local context. Based on this, we propose L2A (Learning To Attend), a layer that enables conditional (token-wise) long-range memory access by deciding when to invoke global attention. We evaluate L2A on Qwen 2.5 and Qwen 3 models, extending their effective context length from 32K to 128K tokens. L2A matches the performance of standard long-context training to within 3\% while skipping Global Attention for $\sim$80\% of tokens, outperforming prior baselines. We also design custom Triton kernels to efficiently implement this token-wise conditional Attention on GPUs, achieving up to $\sim$2$\times$ improvements in training throughput and time-to-first-token over FlashAttention. Moreover, L2A enables post-training pruning of highly sparse Global Attention layers, reducing KV cache memory by up to 50\% with negligible performance loss. Our code is released under Apache 2.0 at https://github.com/awslabs/hybrid-model-factory/tree/main/examples/research/L2A.

cs.CL↗

arXiv2025

Gated KalmaNet: A Fading Memory Layer Through Test-Time Ridge Regression

Linear State-Space Models (SSMs) offer an efficient alternative to softmax Attention with constant memory and linear compute, but their lossy, fading summary of the past hurts recall-oriented tasks. We propose Gated KalmaNet (GKA, pronounced "gee-ka"), a layer that accounts for the full past while retaining SSM-style efficiency. We ground our approach in the Kalman Filter (KF), and show that several existing SSM layers (DeltaNet, Gated DeltaNet, Kimi Delta Attention) are approximations to the KF recurrence under an identity error covariance assumption, which ignores how past keys and values should optimally influence state updates. In contrast, GKA maintains the full error covariance and computes the exact Kalman gain. Under a steady-state assumption that enables parallelization, this reduces to an online ridge regression with constant memory and linear compute. The standard KF equations are numerically unstable in low-precision settings (e.g., bfloat16) and hard to parallelize on GPUs. We address this with (1) adaptive regularization via input-dependent gating to control the ridge regression's condition number, and (2) Chebyshev Iteration, which we show is more stable than conventional iterative solvers in low precision. We further develop hardware-aware chunk-wise kernels for efficient training. Empirically, GKA outperforms existing SSM layers (e.g., Mamba2, Gated DeltaNet) on short-context tasks and achieves more than 10\% relative improvement on long-context RAG and LongQA up to 128k tokens. We further show GKA outperforms Mamba when extended to ImageNet classification. Our code, including Triton kernels for training and inference (vLLM), along with a model zoo of GKA-based Hybrid models at 8B and 32B scale on HuggingFace, is released under Apache 2.0.

cs.LG↗

arXiv2025

Maximally-Informative Retrieval for State Space Model Generation

Given a query and dataset, the optimal way of answering the query is to make use all the information available. Modern LLMs exhibit impressive ability to memorize training data, but data not deemed important during training is forgotten, and information outside that training set cannot be made use of. Processing an entire dataset at inference time is infeasible due to the bounded nature of model resources (e.g. context size in transformers or states in state space models), meaning we must resort to external memory. This constraint naturally leads to the following problem: How can we decide based on the present query and model, what among a virtually unbounded set of known data matters for inference? To minimize model uncertainty for a particular query at test-time, we introduce Retrieval In-Context Optimization (RICO), a retrieval method that uses gradients from the LLM itself to learn the optimal mixture of documents for answer generation. Unlike traditional retrieval-augmented generation (RAG), which relies on external heuristics for document retrieval, our approach leverages direct feedback from the model. Theoretically, we show that standard top-$k$ retrieval with model gradients can approximate our optimization procedure, and provide connections to the leave-one-out loss. We demonstrate empirically that by minimizing an unsupervised loss objective in the form of question perplexity, we can achieve comparable retriever metric performance to BM25 with \emph{no finetuning}. Furthermore, when evaluated on quality of the final prediction, our method often outperforms fine-tuned dense retrievers such as E5.

cs.CL↗

arXiv2025

PICASO: Permutation-Invariant Context Composition with State Space Models

Providing Large Language Models with relevant contextual knowledge at inference time has been shown to greatly improve the quality of their generations. This is often achieved by prepending informative passages of text, or 'contexts', retrieved from external knowledge bases to their input. However, processing additional contexts online incurs significant computation costs that scale with their length. State Space Models (SSMs) offer a promising solution by allowing a database of contexts to be mapped onto fixed-dimensional states from which to start the generation. A key challenge arises when attempting to leverage information present across multiple contexts, since there is no straightforward way to condition generation on multiple independent states in existing SSMs. To address this, we leverage a simple mathematical relation derived from SSM dynamics to compose multiple states into one that efficiently approximates the effect of concatenating raw context tokens. Since the temporal ordering of contexts can often be uninformative, we enforce permutation-invariance by efficiently averaging states obtained via our composition algorithm across all possible context orderings. We evaluate our resulting method on WikiText and MSMARCO in both zero-shot and fine-tuned settings, and show that we can match the strongest performing baseline while enjoying on average 5.4x speedup.

cs.CL↗

arXiv2025

Descriminative-Generative Custom Tokens for Vision-Language Models

This paper explores the possibility of learning custom tokens for representing new concepts in Vision-Language Models (VLMs). Our aim is to learn tokens that can be effective for both discriminative and generative tasks while composing well with words to form new input queries. The targeted concept is specified in terms of a small set of images and a parent concept described using text. We operate on CLIP text features and propose to use a combination of a textual inversion loss and a classification loss to ensure that text features of the learned token are aligned with image features of the concept in the CLIP embedding space. We restrict the learned token to a low-dimensional subspace spanned by tokens for attributes that are appropriate for the given super-class. These modifications improve the quality of compositions of the learned token with natural language for generating new scenes. Further, we show that learned custom tokens can be used to form queries for text-to-image retrieval task, and also have the important benefit that composite queries can be visualized to ensure that the desired concept is faithfully encoded. Based on this, we introduce the method of Generation Aided Image Retrieval, where the query is modified at inference time to better suit the search intent. On the DeepFashion2 dataset, our method improves Mean Reciprocal Retrieval (MRR) over relevant baselines by 7%.

cs.CV↗

arXiv2024

Expansion Span: Combining Fading Memory and Retrieval in Hybrid State Space Models

The "state" of State Space Models (SSMs) represents their memory, which fades exponentially over an unbounded span. By contrast, Attention-based models have "eidetic" (i.e., verbatim, or photographic) memory over a finite span (context size). Hybrid architectures combine State Space layers with Attention, but still cannot recall the distant past and can access only the most recent tokens eidetically. Unlike current methods of combining SSM and Attention layers, we allow the state to be allocated based on relevancy rather than recency. In this way, for every new set of query tokens, our models can "eidetically" access tokens from beyond the Attention span of current Hybrid SSMs without requiring extra hardware resources. We introduce a method to expand the memory span of the hybrid state by "reserving" a fraction of the Attention context for tokens retrieved from arbitrarily distant in the past, thus expanding the eidetic memory span of the overall state. We call this reserved fraction of tokens the "expansion span," and the mechanism to retrieve and aggregate it "Span-Expanded Attention" (SE-Attn). To adapt Hybrid models to using SE-Attn, we propose a novel fine-tuning method that extends LoRA to Hybrid models (HyLoRA) and allows efficient adaptation on long spans of tokens. We show that SE-Attn enables us to efficiently adapt pre-trained Hybrid models on sequences of tokens up to 8 times longer than the ones used for pre-training. We show that HyLoRA with SE-Attn is cheaper and more performant than alternatives like LongLoRA when applied to Hybrid models on natural language benchmarks with long-range dependencies, such as PG-19, RULER, and other common natural language downstream tasks.

cs.CL↗

arXiv2024

Marconi: Prefix Caching for the Era of Hybrid LLMs

Hybrid models that combine the language modeling capabilities of Attention layers with the efficiency of Recurrent layers (e.g., State Space Models) have gained traction in practically supporting long contexts in Large Language Model serving. Yet, the unique properties of these models complicate the usage of complementary efficiency optimizations such as prefix caching that skip redundant computations across requests. Most notably, their use of in-place state updates for recurrent layers precludes rolling back cache entries for partial sequence overlaps, and instead mandates only exact-match cache hits; the effect is a deluge of (large) cache entries per sequence, most of which yield minimal reuse opportunities. We present Marconi, the first system that supports efficient prefix caching with Hybrid LLMs. Key to Marconi are its novel admission and eviction policies that more judiciously assess potential cache entries based not only on recency, but also on (1) forecasts of their reuse likelihood across a taxonomy of different hit scenarios, and (2) the compute savings that hits deliver relative to memory footprints. Across diverse workloads and Hybrid models, Marconi achieves up to 34.4$\times$ higher token hit rates (71.1% or 617 ms lower TTFT) compared to state-of-the-art prefix caching systems.

cs.DC↗

arXiv2024

Long-context Protein Language Modeling Using Bidirectional Mamba with Shared Projection Layers

Self-supervised training of language models (LMs) has seen great success for protein sequences in learning meaningful representations and for generative drug design. Most protein LMs are based on the Transformer architecture trained on individual proteins with short context lengths. Such protein LMs cannot extrapolate to longer proteins and protein complexes well. They also fail to account for the underlying biological mechanisms carried out by biomolecular interactions and dynamics i.e., proteins often interact with other proteins, molecules, and pathways in complex biological systems. In this work, we propose LC-PLM based on an alternative protein LM architecture, BiMamba-S, built upon selective structured state-space models, to learn high-quality universal protein representations at the amino acid token level using masked language modeling. We also introduce its graph-contextual variant, LC-PLM, which contextualizes protein-protein interaction (PPI) graphs for a second stage of training. LC-PLM demonstrates favorable neural scaling laws, better length extrapolation capability, and up to 30% and 16% improvements on protein downstream tasks compared to Transformer-based ESM-2 when trained with 100B and 1T tokens, respectively. LC-PLM-G further trained within the context of PPI graphs shows promising results on protein structure and function prediction tasks. Our study demonstrates the benefit of increasing the context size with computationally efficient LM architecture (e.g., structured state space models) in learning universal protein representations and incorporating molecular interaction contexts contained in biological graphs.

q-bio.BM↗

arXiv2024

Compositional Structures in Neural Embedding and Interaction Decompositions

We describe a basic correspondence between linear algebraic structures within vector embeddings in artificial neural networks and conditional independence constraints on the probability distributions modeled by these networks. Our framework aims to shed light on the emergence of structural patterns in data representations, a phenomenon widely acknowledged but arguably still lacking a solid formal grounding. Specifically, we introduce a characterization of compositional structures in terms of "interaction decompositions," and we establish necessary and sufficient conditions for the presence of such structures within the representations of a model.

cs.LG↗

arXiv2024

B'MOJO: Hybrid State Space Realizations of Foundation Models with Eidetic and Fading Memory

We describe a family of architectures to support transductive inference by allowing memory to grow to a finite but a-priori unknown bound while making efficient use of finite resources for inference. Current architectures use such resources to represent data either eidetically over a finite span ("context" in Transformers), or fading over an infinite span (in State Space Models, or SSMs). Recent hybrid architectures have combined eidetic and fading memory, but with limitations that do not allow the designer or the learning process to seamlessly modulate the two, nor to extend the eidetic memory span. We leverage ideas from Stochastic Realization Theory to develop a class of models called B'MOJO to seamlessly combine eidetic and fading memory within an elementary composable module. The overall architecture can be used to implement models that can access short-term eidetic memory "in-context," permanent structural memory "in-weights," fading memory "in-state," and long-term eidetic memory "in-storage" by natively incorporating retrieval from an asynchronously updated memory. We show that Transformers, existing SSMs such as Mamba, and hybrid architectures such as Jamba are special cases of B'MOJO and describe a basic implementation, to be open sourced, that can be stacked and scaled efficiently in hardware. We test B'MOJO on transductive inference tasks, such as associative recall, where it outperforms existing SSMs and Hybrid models; as a baseline, we test ordinary language modeling where B'MOJO achieves perplexity comparable to similarly-sized Transformers and SSMs up to 1.4B parameters, while being up to 10% faster to train. Finally, we show that B'MOJO's ability to modulate eidetic and fading memory results in better inference on longer sequences tested up to 32K tokens, four-fold the length of the longest sequences seen during training.

cs.LG↗

arXiv2024

CPR: Retrieval Augmented Generation for Copyright Protection

Retrieval Augmented Generation (RAG) is emerging as a flexible and robust technique to adapt models to private users data without training, to handle credit attribution, and to allow efficient machine unlearning at scale. However, RAG techniques for image generation may lead to parts of the retrieved samples being copied in the model's output. To reduce risks of leaking private information contained in the retrieved set, we introduce Copy-Protected generation with Retrieval (CPR), a new method for RAG with strong copyright protection guarantees in a mixed-private setting for diffusion models.CPR allows to condition the output of diffusion models on a set of retrieved images, while also guaranteeing that unique identifiable information about those example is not exposed in the generated outputs. In particular, it does so by sampling from a mixture of public (safe) distribution and private (user) distribution by merging their diffusion scores at inference. We prove that CPR satisfies Near Access Freeness (NAF) which bounds the amount of information an attacker may be able to extract from the generated images. We provide two algorithms for copyright protection, CPR-KL and CPR-Choose. Unlike previously proposed rejection-sampling-based NAF methods, our methods enable efficient copyright-protected sampling with a single run of backward diffusion. We show that our method can be applied to any pre-trained conditional diffusion model, such as Stable Diffusion or unCLIP. In particular, we empirically show that applying CPR on top of unCLIP improves quality and text-to-image alignment of the generated results (81.4 to 83.17 on TIFA benchmark), while enabling credit attribution, copy-right protection, and deterministic, constant time, unlearning.

cs.CR↗

arXiv2024

Multi-Modal Hallucination Control by Visual Information Grounding

Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as "hallucination" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by 25% and 28%, respectively, and improve the accuracy on VQA benchmarks such as POPE by 21% and 24%.

cs.CV↗

arXiv2023

Leveraging sparse and shared feature activations for disentangled representation learning

Recovering the latent factors of variation of high dimensional data has so far focused on simple synthetic settings. Mostly building on unsupervised and weakly-supervised objectives, prior work missed out on the positive implications for representation learning on real world data. In this work, we propose to leverage knowledge extracted from a diversified set of supervised tasks to learn a common disentangled representation. Assuming each supervised task only depends on an unknown subset of the factors of variation, we disentangle the feature space of a supervised multi-task model, with features activating sparsely across different tasks and information being shared as appropriate. Importantly, we never directly observe the factors of variations but establish that access to multiple tasks is sufficient for identifiability under sufficiency and minimality assumptions. We validate our approach on six real world distribution shift benchmarks, and different data modalities (images, text), demonstrating how disentangled representations can be transferred to real settings.

cs.LG↗

arXiv2023

Meaning Representations from Trajectories in Autoregressive Models

We propose to extract meaning representations from autoregressive language models by considering the distribution of all possible trajectories extending an input text. This strategy is prompt-free, does not require fine-tuning, and is applicable to any pre-trained autoregressive model. Moreover, unlike vector-based representations, distribution-based representations can also model asymmetric relations (e.g., direction of logical entailment, hypernym/hyponym relations) by using algebraic operations between likelihood functions. These ideas are grounded in distributional perspectives on semantics and are connected to standard constructions in automata theory, but to our knowledge they have not been applied to modern language models. We empirically show that the representations obtained from large models align well with human annotations, outperform other zero-shot and prompt-free methods on semantic similarity tasks, and can be used to solve more complex entailment and containment tasks that standard embeddings cannot handle. Finally, we extend our method to represent data from different modalities (e.g., image and text) using multimodal autoregressive models. Our code is available at: https://github.com/tianyu139/meaning-as-trajectories

cs.CL↗

arXiv2023

SemiGPC: Distribution-Aware Label Refinement for Imbalanced Semi-Supervised Learning Using Gaussian Processes

In this paper we introduce SemiGPC, a distribution-aware label refinement strategy based on Gaussian Processes where the predictions of the model are derived from the labels posterior distribution. Differently from other buffer-based semi-supervised methods such as CoMatch and SimMatch, our SemiGPC includes a normalization term that addresses imbalances in the global data distribution while maintaining local sensitivity. This explicit control allows SemiGPC to be more robust to confirmation bias especially under class imbalance. We show that SemiGPC improves performance when paired with different Semi-Supervised methods such as FixMatch, ReMixMatch, SimMatch and FreeMatch and different pre-training strategies including MSN and Dino. We also show that SemiGPC achieves state of the art results under different degrees of class imbalance on standard CIFAR10-LT/CIFAR100-LT especially in the low data-regime. Using SemiGPC also results in about 2% avg.accuracy increase compared to a new competitive baseline on the more challenging benchmarks SemiAves, SemiCUB, SemiFungi and Semi-iNat.

cs.CV↗

arXiv2023

Your representations are in the network: composable and parallel adaptation for large scale models

We propose InCA, a lightweight method for transfer learning that cross-attends to any activation layer of a pre-trained model. During training, InCA uses a single forward pass to extract multiple activations, which are passed to external cross-attention adapters, trained anew and combined or selected for downstream tasks. We show that, even when selecting a single top-scoring adapter, InCA achieves performance comparable to full fine-tuning, at a cost comparable to fine-tuning just the last layer. For example, with a cross-attention probe 1.3% the size of a pre-trained ViT-L/16 model, we achieve performance within 0.2% of the full fine-tuning paragon at a computational training cost of 51% of the baseline, on average across 11 downstream classification. Unlike other forms of efficient adaptation, InCA does not require backpropagating through the pre-trained model, thus leaving its execution unaltered at both training and inference. The versatility of InCA is best illustrated in fine-grained tasks, which may require accessing information absent in the last layer but accessible in intermediate layer activations. Since the backbone is fixed, InCA allows parallel ensembling as well as parallel execution of multiple tasks. InCA achieves state-of-the-art performance in the ImageNet-to-Sketch multi-task benchmark.

cs.LG↗

arXiv2023

Prompt Algebra for Task Composition

We investigate whether prompts learned independently for different tasks can be later combined through prompt algebra to obtain a model that supports composition of tasks. We consider Visual Language Models (VLM) with prompt tuning as our base classifier and formally define the notion of prompt algebra. We propose constrained prompt tuning to improve performance of the composite classifier. In the proposed scheme, prompts are constrained to appear in the lower dimensional subspace spanned by the basis vectors of the pre-trained vocabulary. Further regularization is added to ensure that the learned prompt is grounded correctly to the existing pre-trained vocabulary. We demonstrate the effectiveness of our method on object classification and object-attribute classification datasets. On average, our composite model obtains classification accuracy within 2.5% of the best base model. On UTZappos it improves classification accuracy over the best base model by 8.45% on average.

cs.CV↗