Transformer models rely heavily on the scaled dot-product attention (SDPA) operation, typically implemented as FlashAttention. Characterized by its frequent interleaving of matrix multiplications and softmax operations, FlashAttention fails to fully utilize the compute resources of modern systolic-array-based accelerators designed for consecutive and large matrix multiplications. To fully unleash the performance potential of systolic arrays for FlashAttention, we propose FSA, an enhanced systolic array architecture that runs the entire FlashAttention on the array without external vector units. Combined with SystolicAttention, an optimized kernel for FSA that achieves fine-grained and element-wise overlapping of FlashAttention operations, FSA maximizes array utilization while preserving the original floating-point operation order of FlashAttention. We implement FSA in synthesizable RTL and evaluate its performance against state-of-the-art systolic-array-based accelerators. Our results show that FSA achieves 1.77x and 4.83x higher attention FLOPs/s utilization compared to AWS Neuron-v2 and Google TPUv5e, respectively. We synthesize FSA in a 16 nm technology at 1.5 GHz, and results indicate only a 12% area overhead compared to a standard weight-stationary systolic array.

Recent advance in sparse attention mechanisms has demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs). Native Sparse Attention (NSA), one state-of-the-art approach, introduces natively trainable, hardware-aligned sparse attention that delivers substantial system-level performance boost while maintaining accuracy comparable to full attention. However, the kernel implementation of NSA forces a loop order that is only efficient with a relatively large number of query heads in each Grouped Query Attention (GQA) group, whereas existing LLMs widely adopt much smaller number of query heads in each GQA group -- such an inconsistency significantly limits the applicability of this sparse algorithmic advance. In this work, we propose Flash Sparse Attention (FSA), an alternative kernel implementation that enables efficient NSA computation across a wide range of popular LLMs with varied smaller number of heads in each GQA group on modern GPUs. Compared to vanilla NSA kernel implementation, our empirical evaluation demonstrates that FSA achieves (i) up to 3.5x and on average 1.6x kernel-level latency reduction, (ii) up to 1.25x and 1.09x on average end-to-end training speedup on state-of-the-art LLMs, and (iii) up to 1.36x and 1.11x on average for prefill-phase speedup in LLM generative inference. Github Repo at https://github.com/Relaxed-System-Lab/Flash-Sparse-Attention.

Assisting medical students with clinical reasoning (CR) during clinical scenario training remains a persistent challenge in medical education. This paper presents the design and architecture of the Fuzzy Supervisor Agent (FSA), a novel component for the Multi-Agent Educational Clinical Scenario Simulation (MAECSS) platform. The FSA leverages a Fuzzy Inference System (FIS) to continuously interpret student interactions with specialized clinical agents (e.g., patient, physical exam, diagnostic, intervention) using pre-defined fuzzy rule bases for professionalism, medical relevance, ethical behavior, and contextual distraction. By analyzing student decision-making processes in real-time, the FSA is designed to deliver adaptive, context-aware feedback and provides assistance precisely when students encounter difficulties. This work focuses on the technical framework and rationale of the FSA, highlighting its potential to provide scalable, flexible, and human-like supervision in simulation-based medical education. Future work will include empirical evaluation and integration into broader educational settings. More detailed design and implementation is open sourced here.

Chain-of-Thought (CoT) significantly enhances the performance of large language models (LLMs) across a wide range of tasks, and prior research shows that CoT can theoretically increase expressiveness. However, there is limited mechanistic understanding of the algorithms that Transformer+CoT can learn. In this work, we (1) evaluate the state tracking capabilities of Transformer+CoT and its variants, confirming the effectiveness of CoT. (2) Next, we identify the circuit, a subset of model components, responsible for tracking the world state, finding that late-layer MLP neurons play a key role. We propose two metrics, compression and distinction, and show that the neuron sets for each state achieve nearly 100% accuracy, providing evidence of an implicit finite state automaton (FSA) embedded within the model. (3) Additionally, we explore three realistic settings: skipping intermediate steps, introducing data noise, and testing length generalization. Our results demonstrate that Transformer+CoT learns robust algorithms (FSA), highlighting its resilience in challenging scenarios.

Recently, large language models (LLMs) with hundreds of billions of parameters have demonstrated the emergent ability, surpassing traditional methods in various domains even without fine-tuning over domain-specific data. However, when it comes to financial sentiment analysis (FSA)$\unicode{x2013}$a fundamental task in financial AI$\unicode{x2013}$these models often encounter various challenges, such as complex financial terminology, subjective human emotions, and ambiguous inclination expressions. In this paper, we aim to answer the fundamental question: whether LLMs are good in-context learners for FSA? Unveiling this question can yield informative insights on whether LLMs can learn to address the challenges by generalizing in-context demonstrations of financial document-sentiment pairs to the sentiment analysis of new documents, given that finetuning these models on finance-specific data is difficult, if not impossible at all. To the best of our knowledge, this is the first paper exploring in-context learning for FSA that covers most modern LLMs (recently released DeepSeek V3 included) and multiple in-context sample selection methods. Comprehensive experiments validate the in-context learning capability of LLMs for FSA.

Large Language Models (LLMs) are often asked to generate structured outputs that obey precise syntactic rules, such as code snippets or formatted data. Grammar-constrained decoding (GCD) can guarantee that LLM outputs matches such rules by masking out tokens that will provably lead to outputs that do not belong to a specified context-free grammar (CFG). To guarantee soundness, GCD algorithms have to compute how a given LLM subword tokenizer can align with the tokens used by a given context-free grammar and compute token masks based on this information. Doing so efficiently is challenging and existing GCD algorithms require tens of minutes to preprocess common grammars. We present a new GCD algorithm together with an implementation that offers 17.71x faster offline preprocessing than existing approaches while preserving state-of-the-art efficiency in online mask computation.

Selective state-space models (SSMs) are an emerging alternative to the Transformer, offering the unique advantage of parallel training and sequential inference. Although these models have shown promising performance on a variety of tasks, their formal expressiveness and length generalization properties remain underexplored. In this work, we provide insight into the workings of selective SSMs by analyzing their expressiveness and length generalization performance on regular language tasks, i.e., finite-state automaton (FSA) emulation. We address certain limitations of modern SSM-based architectures by introducing the Selective Dense State-Space Model (SD-SSM), the first selective SSM that exhibits perfect length generalization on a set of various regular language tasks using a single layer. It utilizes a dictionary of dense transition matrices, a softmax selection mechanism that creates a convex combination of dictionary matrices at each time step, and a readout consisting of layer normalization followed by a linear map. We then proceed to evaluate variants of diagonal selective SSMs by considering their empirical performance on commutative and non-commutative automata. We explain the experimental results with theoretical considerations. Our code is available at https://github.com/IBM/selective-dense-state-space-model.

Language models (LMs) are often expected to generate strings in some formal language; for example, structured data, API calls, or code snippets. Although LMs can be tuned to improve their adherence to formal syntax, this does not guarantee conformance, especially with smaller LMs suitable for large-scale deployment. In addition, tuning requires significant resources, making it impractical for uncommon or task-specific formats. To prevent downstream parsing errors we would ideally constrain the LM to only produce valid output, but this is severely complicated by tokenization, which is typically both ambiguous and misaligned with the formal grammar. We solve these issues through the application of automata theory, deriving an efficient closed-form solution for the regular languages, a broad class of formal languages with many practical applications, including API calls or schema-guided JSON and YAML. We also discuss pragmatic extensions for coping with the issue of high branching factor. Finally, we extend our techniques to deterministic context-free languages, which similarly admit an efficient closed-form solution. In spite of its flexibility and representative power, our approach only requires access to per-token decoding logits and lowers into simple calculations that are independent of LM size, making it both efficient and easy to apply to almost any LM architecture.

Recent work by Hewitt et al. (2020) provides an interpretation of the empirical success of recurrent neural networks (RNNs) as language models (LMs). It shows that RNNs can efficiently represent bounded hierarchical structures that are prevalent in human language. This suggests that RNNs' success might be linked to their ability to model hierarchy. However, a closer inspection of Hewitt et al.'s (2020) construction shows that it is not inherently limited to hierarchical structures. This poses a natural question: What other classes of LMs can RNNs efficiently represent? To this end, we generalize Hewitt et al.'s (2020) construction and show that RNNs can efficiently represent a larger class of LMs than previously claimed -- specifically, those that can be represented by a pushdown automaton with a bounded stack and a specific stack update function. Altogether, the efficiency of representing this diverse class of LMs with RNN LMs suggests novel interpretations of their inductive bias.

Large language models (LLMs) have drastically changed the possible ways to design intelligent systems, shifting the focuses from massive data acquisition and new modeling training to human alignment and strategical elicitation of the full potential of existing pre-trained models. This paradigm shift, however, is not fully realized in financial sentiment analysis (FSA), due to the discriminative nature of this task and a lack of prescriptive knowledge of how to leverage generative models in such a context. This study investigates the effectiveness of the new paradigm, i.e., using LLMs without fine-tuning for FSA. Rooted in Minsky's theory of mind and emotions, a design framework with heterogeneous LLM agents is proposed. The framework instantiates specialized agents using prior domain knowledge of the types of FSA errors and reasons on the aggregated agent discussions. Comprehensive evaluation on FSA datasets show that the framework yields better accuracies, especially when the discussions are substantial. This study contributes to the design foundations and paves new avenues for LLMs-based FSA. Implications on business and management are also discussed.