arXiv subjects

Jingze Shi

Publications and source records attributed to Jingze Shi.

11 recordsLinked to original sources

arXiv2026

GIST: Generator-state-Informed Supervision for Tabular Foundation Models

Tabular foundation models (TFMs) based on prior-data fitted networks typically use sampled hard query labels, leaving avoidable sampling variability in pretraining. A key observation is that intermediate generator states can reveal label uncertainty beyond a sampled outcome. We propose Generator-state-Informed Supervision for TFMs (GIST), which uses these states to average query losses over conditional label distributions. GIST jointly designs task generation and query supervision, using entropy-calibrated encodings to obtain tractable, input-consistent targets on supported label-dependent branches. We show that this averaging preserves the population objective and yields no larger gradient covariance at fixed parameters within each generating distribution. Under matched pretraining schedules, GIST improves mean rank across three architectures on both TabArena and TALENT without changing architecture or inference. Across these benchmarks and architectures, GIST wins 62.1% of paired dataset-level comparisons against the corresponding hard-label controls.

cs.LG↗

arXiv2026

CoWindow Attention: Full Causal Coverage Is a Collective Property

FullAttn repeatedly exposes the complete causal history to every attention head, creating substantial redundant computation and memory traffic even with IO-efficient dense kernels. We introduce CoWA, a structured attention architecture that distributes access to the causal history across KV heads. All heads share near-diagonal and prefix-sink windows, while complementary long-range windows partition the remaining history. Their union provides full causal coverage although each head attends sparsely to distant tokens. This position-defined attention pattern requires no learned router or indexer, is used consistently during training and inference, and aligns with KV-head tensor parallelism. A window-matched ablation at 8K isolates the effect of complementary long-range allocation: CoWA with 100% collective coverage reaches 89.73% accuracy, compared with 89.97% for FullAttn, while duplicated long-range windows perform substantially worse. Across a broader controlled associative-recall comparison with matched token budgets, CoWA closely tracks FullAttn as the context grows, whereas other sparse patterns lose a substantial fraction of the associations. In an attention-operator benchmark at 128K tokens with tensor parallelism, CoWA reduces forward and backward latency during training by 7.4x and 8.6x and decoding latency during inference by 3.0x over FullAttn. Its per-rank peak operator memory matches FullAttn during training and is 7.6x lower during decoding. Across scaling-law training from 0.6B to 14B parameters, CoWA closely tracks FullAttn in perplexity while reducing total training FLOPs. The resulting 14B models and 32B models from separate continued training achieve comparable knowledge, reasoning, and long-context retrieval scores to FullAttn. These results show that full causal coverage can be a collective property of the head ensemble rather than a duplicated property of every head.

cs.AI↗

arXiv2026

MassAlloc Attention: Let Attention Allocate Its Own Compute

FullAttn often assigns negligible normalized mass to much of the causal score space, yet dense kernels execute the complete post-score path after forming each QK tile. We introduce MALA, a fused attention primitive that preserves score access to every legal causal interaction and uses normalized contribution to allocate post-score computation. Forward uses its evolving online-softmax normalizer, while backward reuses the finalized normalizer to derive nested retained support using only standard attention state. A common tolerance governs training and inference, allowing for adaptive retention of the work. MALA reduces low-contribution post-score computation. A matched-work study at 8K isolates the benefit of distribution-adaptive allocation: under exactly matched total post-score work, MALA approaches a per-instance reference-mass oracle, with mean omitted mass of 0.0188% versus 0.0182%. Across context lengths from 1K to 32K tokens, the same tolerance maintains low output and gradient errors relative to the reference. Across a broader controlled associative-recall comparison, MALA closely tracks FullAttn as context grows, reaching 89.67% accuracy at 8K compared with 89.97% for FullAttn. In an attention-operator benchmark at 128K tokens with tensor parallelism, MALA reduces forward and backward latency during training by 2.2x and 3.0x and decoding latency during inference by 1.6x relative to FullAttn. Across scaling-law training from 0.6B to 14B parameters, MALA closely tracks FullAttn in perplexity while reducing total training FLOPs. The resulting 14B models and 32B models from separate continued training achieve comparable knowledge, reasoning, and long-context retrieval scores to FullAttn. These results indicate that allocating post-score computation according to normalized attention contributions can retain the evaluated capabilities of FullAttn while reducing attention computation.

cs.AI↗

arXiv2026

OmniMoE: An Efficient MoE by Orchestrating Atomic Experts at Scale

Mixture-of-Experts (MoE) architectures are evolving towards finer granularity to improve parameter efficiency. However, existing MoE designs face an inherent trade-off between the granularity of expert specialization and hardware execution efficiency. We propose OmniMoE, a system-algorithm co-designed framework that pushes expert granularity to its logical extreme. OmniMoE introduces vector-level Atomic Experts, enabling scalable routing and execution within a single MoE layer, while retaining a shared dense MLP branch for general-purpose processing. Although this atomic design maximizes capacity, it poses severe challenges for routing complexity and memory access. To address these, OmniMoE adopts a system-algorithm co-design: (i) a Cartesian Product Router that decomposes the massive index space to reduce routing complexity from O(N) to O(sqrt(N)); and (ii) Expert-Centric Scheduling that inverts the execution order to turn scattered, memory-bound lookups into efficient dense matrix operations. Validated on seven benchmarks, OmniMoE (with 1.7B active parameters) achieves 50.9% zero-shot accuracy across seven benchmarks, outperforming coarse-grained (e.g., DeepSeekMoE) and fine-grained (e.g., PEER) baselines. Crucially, OmniMoE reduces inference latency from 73ms to 6.7ms (a 10.9-fold speedup) compared to PEER, demonstrating that massive-scale fine-grained MoE can be fast and accurate. Our code is open-sourced at https://github.com/flash-algo/omni-moe.

cs.CL↗

arXiv2026

Towards Automated Kernel Generation in the Era of LLMs

The performance of modern AI systems is fundamentally constrained by the quality of their underlying GPU kernels, which translate high-level algorithmic semantics into low-level hardware operations. Achieving near-optimal kernels requires expert-level understanding of hardware architectures and programming models, making kernel engineering a critical but notoriously time-consuming and non-scalable process. Recent advances in large language models and LLM-based agents have opened new possibilities for automating kernel generation and optimization. LLMs are well-suited to compress expert-level kernel knowledge that is difficult to formalize, while agentic systems further enable scalable optimization by casting kernel development as an iterative, feedback-driven loop. Rapid progress has been made in this area. However, the field remains fragmented and lacks a systematic perspective for LLM-driven kernel generation. This survey addresses this gap by providing a structured overview of existing approaches, spanning LLM-based approaches and agentic optimization workflows, and systematically organizing the datasets and benchmarks that underpin learning and evaluation in this domain. Moreover, key open challenges and future research directions are further outlined, aiming to establish a comprehensive reference for the next generation of automated kernel optimization. To keep track of this field, we maintain an open-source GitHub repository at https://github.com/flagos-ai/awesome-LLM-driven-kernel-generation.

cs.LG↗

arXiv2025

Trainable Dynamic Mask Sparse Attention

The increasing demand for long-context modeling in large language models (LLMs) is bottlenecked by the quadratic complexity of the standard self-attention mechanism. The community has proposed sparse attention to mitigate this issue. However, position-aware sparse attention methods rely on static sparse structures that lack adaptability to diverse query contexts, while content-aware sparse attention methods depend on heuristic key-value selection, hindering full differentiability. We introduce a trainable dynamic mask sparse attention mechanism, a method that merges the advantages of both position-aware and content-aware approaches. Dynamic Mask Attention (DMA) achieves this through three key innovations: First, it leverages value vector representations to generate content-aware dynamic masks, enabling the model to adaptively identify and attend to critical information. Second, it computes position-aware sparse weights in a hardware-friendly manner, efficiently skipping unnecessary computational regions. Finally, we demonstrate that the introduced dynamic mask and sparse weights do not obstruct gradients, supporting end-to-end training. We have validated the performance of DMA through comprehensive experiments. A large body of experimental evidence shows that DMA consistently holds a Pareto advantage over state-of-the-art sparse attention baselines in tasks including scaling laws, multi-query associative recall, standard benchmarks, and needle in a haystack tests, while also delivering up to a 10x overall speedup. These results highlight its ability to effectively balance model efficiency with long-context modeling capabilities. Our computational kernel code is now open-source at https://github.com/SmallDoges/flash-dmattn to encourage further research and application by the community.

cs.AI↗

arXiv2025

TransXSSM: A Hybrid Transformer State Space Model with Unified Rotary Position Embedding

Transformers exhibit proficiency in capturing long-range dependencies, whereas State Space Models (SSMs) facilitate linear-time sequence modeling. Notwithstanding their synergistic potential, the integration of these architectures presents a significant challenge, primarily attributable to a fundamental incongr inuity their respective positional encoding mechanisms: Transformers rely on explicit Rotary Position Embeddings (RoPE), while SSMs leverage implicit positional representations via convolutions. This divergence often precipitates discontinuities and suboptimal performance.To address this impediment, we propose a unified rotary position embedding (Unified RoPE) methodology, thereby establishing a consistent positional encoding framework for both self-attention and state-space components. Using this Unified RoPE, we introduce TransXSSM, a hybrid architecture that coherently integrates the Transformer and SSM layers under this unified positional encoding scheme. At a 4 sequenceK length, TransXSSM exhibits training and inference speeds that are 42.3% and 29.5% faster, respectively, relative to standard Transformer models. It also delivers higher accuracy: under comparable settings, it surpasses a Transformer baseline by over 4% on language modeling benchmarks.TransXSSM furthermore scales more effectively: TransXSSM-1.3B gains 7.22% in average accuracy over its 320M version (versus about 6% gains for equivalent Transformers or SSMs). Our results show that unified positional encoding resolves positional incompatibility in hybrid models, enabling efficient, high-performance long-context modeling.

cs.CL↗

arXiv2025

Concise Reasoning, Big Gains: Pruning Long Reasoning Trace with Difficulty-Aware Prompting

Existing chain-of-thought (CoT) distillation methods can effectively transfer reasoning abilities to base models but suffer from two major limitations: excessive verbosity of reasoning traces and inadequate adaptability to problem difficulty. Long reasoning traces significantly increase inference costs, and uniform-length solutions prevent base models from learning adaptive reasoning strategies. To address these issues, we propose a difficulty-aware prompting (DAP) method to dynamically shorten reasoning traces without performance loss. In our approach, a large teacher model first judges each problem's difficulty and then rewrites its reasoning traces to an appropriate shorter length, yielding concise yet complete reasoning traces. Leveraging the DAP pipeline, we curate a distilled dataset called LiteCoT consisting of 100K concise reasoning examples, with solutions averaging only 720 tokens (an order of magnitude shorter than typical CoTs). Using LiteCoT, we distilled a new family of reasoning models called Liter (1.5B, 7B, and 32B) based on the Qwen2.5 architecture. Experiments show that a student model fine-tuned on just 100K of these difficulty-pruned CoT samples outperforms a model distilled on 800K original Long CoT samples, while significantly reducing training and inference costs. Our method also generalizes well: across 11 diverse benchmarks, the shorter difficulty-aware CoTs achieve equal or better accuracy than Long chains, using far fewer tokens. For example, on the challenging AIME24 exam, our approach reaches $74.2\%$ Pass@1 using only about 5K inference tokens, surpassing other methods that consume many more tokens. Our code and data are available at https://github.com/Evanwu1125/LiteCoT.

cs.AI↗

arXiv2024

Wonderful Matrices: Combining for a More Efficient and Effective Foundation Model Architecture

In order to make the foundation model more efficient and effective, our idea is combining sequence transformation and state transformation. First, we prove the availability of rotary position embedding in the state space duality algorithm, which reduces the perplexity of the hybrid quadratic causal self-attention and state space duality by more than 4%, to ensure that the combining sequence transformation unifies position encoding. Second, we propose dynamic mask attention, which maintains 100% accuracy in the more challenging multi-query associative recall task, improving by more than 150% compared to quadratic causal self-attention and state space duality, to ensure that the combining sequence transformation selectively filters relevant information. Third, we design cross domain mixture of experts, which makes the computational speed of expert retrieval with more than 1024 experts 8 to 10 times faster than the mixture of experts, to ensure that the combining state transformation quickly retrieval mixture. Finally, we summarize these matrix algorithms that can form the foundation model: Wonderful Matrices, which can be a competitor to popular model architectures.

cs.LG↗

arXiv2024

Wonderful Matrices: More Efficient and Effective Architecture for Language Modeling Tasks

We prove the availability of inner product form position encoding in the state space dual algorithm and study the effectiveness of different position embeddings in the hybrid quadratic causal self-attention and state space dual algorithms. We propose inner function attention with dynamic mask, which can improve the expressiveness of the attention algorithm and avoid the sequence noise significantly affecting the accuracy of the attention score. We also design cross domain mixture of experts, which can improve the granularity of the sparse activation feedforward network while maintaining the efficiency of parameter utilization and retrieval. The combination of these methods constitutes our foundation model architecture: Wonderful Matrices. We conduct experiments on the language modeling task and find that Wonderful Matrices are more efficient and effective in handling complex language tasks.

cs.LG↗

arXiv2024

OTCE: Hybrid SSM and Attention with Cross Domain Mixture of Experts to construct Observer-Thinker-Conceiver-Expresser

Recent research has shown that combining Mamba with Transformer architecture, which has selective state space and quadratic self-attention mechanism, outperforms using Mamba or Transformer architecture alone in language modeling tasks. The quadratic self-attention mechanism effectively alleviates the shortcomings of selective state space in handling long-term dependencies of any element in the sequence. We propose a position information injection method that connects the selective state space model with the quadratic attention, and integrates these two architectures with hybrid experts with cross-sharing domains, so that we can enjoy the advantages of both. We design a new architecture with a more biomimetic idea: Observer-Thinker-Conceiver-Expresser (OTCE), which can compete with well-known medium-scale open-source language models on a small scale in language modeling tasks.

cs.CL↗