Searcharxiv⌕ Search

arXiv subjects

Bingning Wang

Publications and source records attributed to Bingning Wang.

At least 19 recordsLinked to original sources
arXiv2026

FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving

Long-context modeling is a pivotal capability for Large Language Models, yet the quadratic complexity of attention remains a critical bottleneck, particularly during the compute-intensive prefilling phase. Our previous work, FlashPrefill, mitigates this cost through instantaneous pattern discovery and max-based dynamic thresholding; however, it remains an algorithmic prototype that is still distant from production deployment. In this paper, we present FlashPrefill V2, which evolves FlashPrefill from a prototype toward practical long-context serving along three dimensions. First, we introduce a mean correction term that effectively suppresses the approximation error, keeping performance degradation manageable even at extreme sparsity levels. Second, we redesign the sparse attention operator with PackGQA memory access, warp specialization, and pingpong pipelining, fully aligning with the latest FlashAttention-3/4 implementations and supporting FP8 inference to meet practical quantization requirements. Third, FlashPrefill V2 natively supports paged KV cache and continuous batching, allowing integration as an attention backend in modern inference frameworks such as SGLang. Extensive evaluations on NVIDIA H20 GPUs---among the most widely deployed inference accelerators---demonstrate that FlashPrefill V2 delivers up to 47.26x and 27.19x speedups over FlashAttention-2 at 128K context length under FP8 and BF16 precision, respectively, and, in FP8, still achieves a 30.49x speedup against an FA3/4-aligned dense baseline.

cs.CL↗
arXiv2026

Post-Trained MoE Can Skip Half Experts via Self-Distillation

Mixture-of-Experts (MoE) scales language models efficiently through sparse expert activation, and its dynamic variant further reduces computation by adjusting the activated experts in an input-dependent manner. Existing dynamic MoE methods usually rely on pre-training from scratch or task-specific adaptation, leaving the practical conversion of fully trained MoE underexplored. Enabling such adaptation would directly alleviate the inference costs by allowing easy tokens to bypass unnecessary expert during serving. This paper introduces Zero-Expert Self-Distillation Adaptation (ZEDA), a low-cost framework that transforms post-trained static MoE models into efficient dynamic ones. To stabilize this architectural conversion, ZEDA injects parameter-free zero-output experts into each MoE layer and adapts the augmented model through two-stage self-distillation, utilizing the original MoE as a frozen teacher and applying a group-level balancing loss. On Qwen3-30B-A3B and GLM-4.7-Flash across 11 benchmarks spanning math, code, and instruction following, ZEDA eliminates over 50% of expert FLOPs at marginal accuracy loss. It outperforms the strongest dynamic MoE baseline by 6.1 and 4.0 points on the two models, and delivers ~1.20$\times$ end-to-end inference speedup.

cs.LG↗
arXiv2026

UniPrefill: Universal Long-Context Prefill Acceleration via Block-wise Dynamic Sparsification

As large language models (LLMs) continue to advance rapidly, they are becoming increasingly capable while simultaneously demanding ever-longer context lengths. To improve the inference efficiency of long-context processing, several novel low-complexity hybrid architectures have recently been proposed, effectively alleviating the computational burden of long-context inference. However, existing research on long-context prefill acceleration remains predominantly focused on sparse attention mechanisms, which achieve their maximum speedup only on full-attention models. When transferred to emerging architectures--such as linear/full attention hybrids or sliding window/full attention hybrids--these prefill acceleration approaches suffer significant performance degradation. Furthermore, such methods are generally incompatible with continuous batching, making them difficult to integrate into modern inference engines such as vLLM. To this end, we propose UniPrefill, a prefill acceleration framework applicable to virtually any model architecture, which directly accelerates the model's computation at the token level. We further implement UniPrefill as a continuous batching operator and extend vLLM's scheduling strategy to natively support prefill-decode co-processing and tensor parallel for UniPrefill, enabling its seamless integration into vLLM. UniPrefill achieves up to 2.1x speedup in Time-To-First-Token (TTFT), with the acceleration becoming increasingly pronounced as the number of concurrent requests grows.

cs.CL↗
arXiv2026

FlashPrefill: Instantaneous Pattern Discovery and Thresholding for Ultra-Fast Long-Context Prefilling

Long-context modeling is a pivotal capability for Large Language Models, yet the quadratic complexity of attention remains a critical bottleneck, particularly during the compute-intensive prefilling phase. While various sparse attention mechanisms have been explored, they typically suffer from either significant search latency or insufficient sparsity. In this paper, we propose FlashPrefill, a framework enabling ultra-fast prefilling via instantaneous pattern discovery and thresholding. FlashPrefill leverages a fast block-searching technique to simultaneously locate dynamic vertical, slash, and block-sparse attention patterns. Crucially, it introduces a dynamic thresholding mechanism that bypasses the prohibitive overhead of sorting or accumulating attention scores while effectively eliminating the long-tail distribution to enhance sparsity. Extensive evaluations demonstrate that FlashPrefill achieves a substantial leap in efficiency, delivering an unprecedented 27.78x speedup on 256K sequences. Notably, unlike existing methods that incur efficiency degradation on shorter contexts, FlashPrefill maintains a 1.71x speedup even at a 4K context length, demonstrating its robustness and practical utility across varying sequence scales.

cs.CL↗
arXiv2026

Gradients Must Earn Their Influence: Unifying SFT with Generalized Entropic Objectives

Standard negative log-likelihood (NLL) for Supervised Fine-Tuning (SFT) applies uniform token-level weighting. This rigidity creates a two-fold failure mode: (i) overemphasizing low-probability targets can amplify gradients on noisy supervision and disrupt robust priors, and (ii) uniform weighting provides weak sharpening when the model is already confident. Existing methods fail to resolve the resulting plasticity--stability dilemma, often suppressing necessary learning signals alongside harmful ones. To address this issue, we unify token-level SFT objectives within a generalized deformed-log family and expose a universal gate $\times$ error gradient structure, where the gate controls how much the model trusts its current prediction. By employing the Cayley transform, we map the model's continuously evolving uncertainty onto a continuous focus trajectory, which enables seamless interpolation between scenarios involving uncertain novel concepts and those involving well-established knowledge. We then introduce Dynamic Entropy Fine-Tuning (DEFT), a parameter-free objective that modulates the trust gate using distribution concentration (R\'enyi-2 entropy) as a practical proxy for the model's predictive state. Extensive experiments and analyses demonstrate that DEFT achieves a better balance between exploration and exploitation, leading to improved overall performance.

cs.CL↗
arXiv2025

Towards a Unified View of Large Language Model Post-Training

Two major sources of training data exist for post-training modern language models: online (model-generated rollouts) data, and offline (human or other-model demonstrations) data. These two types of data are typically used by approaches like Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT), respectively. In this paper, we show that these approaches are not in contradiction, but are instances of a single optimization process. We derive a Unified Policy Gradient Estimator, and present the calculations of a wide spectrum of post-training approaches as the gradient of a common objective under different data distribution assumptions and various bias-variance tradeoffs. The gradient estimator is constructed with four interchangeable parts: stabilization mask, reference policy denominator, advantage estimate, and likelihood gradient. Motivated by our theoretical findings, we propose Hybrid Post-Training (HPT), an algorithm that dynamically selects different training signals. HPT is designed to yield both effective exploitation of demonstration and stable exploration without sacrificing learned reasoning patterns. We provide extensive experiments and ablation studies to verify the effectiveness of our unified theoretical framework and HPT. Across six mathematical reasoning benchmarks and two out-of-distribution suites, HPT consistently surpasses strong baselines across models of varying scales and families.

cs.LG↗
arXiv2025

SSRL: Self-Search Reinforcement Learning

We investigate the potential of large language models (LLMs) to serve as efficient simulators for agentic search tasks in reinforcement learning (RL), thereby reducing dependence on costly interactions with external search engines. To this end, we first quantify the intrinsic search capability of LLMs via structured prompting and repeated sampling, which we term Self-Search. Our results reveal that LLMs exhibit strong scaling behavior with respect to the inference budget, achieving high pass@k on question-answering benchmarks, including the challenging BrowseComp task. Building on these observations, we introduce Self-Search RL (SSRL), which enhances LLMs' Self-Search capability through format-based and rule-based rewards. SSRL enables models to iteratively refine their knowledge utilization internally, without requiring access to external tools. Empirical evaluations demonstrate that SSRL-trained policy models provide a cost-effective and stable environment for search-driven RL training, reducing reliance on external search engines and facilitating robust sim-to-real transfer. We draw the following conclusions: 1) LLMs possess world knowledge that can be effectively elicited to achieve high performance; 2) SSRL demonstrates the potential of leveraging internal knowledge to reduce hallucination; 3) SSRL-trained models integrate seamlessly with external search engines without additional effort. Our findings highlight the potential of LLMs to support more scalable RL agent training.

cs.CL↗
arXiv2025

Surrogate Signals from Format and Length: Reinforcement Learning for Solving Mathematical Problems without Ground Truth Answers

Large Language Models (LLMs) have achieved remarkable success in natural language processing tasks, with Reinforcement Learning (RL) playing a key role in adapting them to specific applications. In mathematical problem solving, however, the reliance on ground truth answers poses significant challenges due to their high collection cost and limited availability. This work explores the use of simple surrogate signals, format and length, to guide RL training. We find that early training is dominated by format learning, where structural feedback alone accounts for most performance gains. Incorporating length-based rewards further refines outputs by discouraging overly long or short responses, enabling a GRPO approach with format-length signals to approximate, and in some cases surpass, ground-truth-based optimization. For example, our method achieves 40.0% accuracy on AIME2024 with a 7B base model, and generalizes across different model sizes and series. Beyond practical efficiency, these findings provide an inspirational perspective on RL: rather than imparting new knowledge, RL primarily activates reasoning capabilities already embedded in pre-trained models. This insight suggests that lightweight, label-efficient strategies can complement pre-training to unlock LLMs' latent potential in reasoning-intensive tasks.

cs.CL↗
arXiv2025

Amplify Adjacent Token Differences: Enhancing Long Chain-of-Thought Reasoning with Shift-FFN

Recently, models such as OpenAI-o1 and DeepSeek-R1 have demonstrated remarkable performance on complex reasoning tasks through Long Chain-of-Thought (Long-CoT) reasoning. Although distilling this capability into student models significantly enhances their performance, this paper finds that fine-tuning LLMs with full parameters or LoRA with a low rank on long CoT data often leads to Cyclical Reasoning, where models repeatedly reiterate previous inference steps until the maximum length limit. Further analysis reveals that smaller differences in representations between adjacent tokens correlates with a higher tendency toward Cyclical Reasoning. To mitigate this issue, this paper proposes Shift Feedforward Networks (Shift-FFN), a novel approach that edits the current token's representation with the previous one before inputting it to FFN. This architecture dynamically amplifies the representation differences between adjacent tokens. Extensive experiments on multiple mathematical reasoning tasks demonstrate that LoRA combined with Shift-FFN achieves higher accuracy and a lower rate of Cyclical Reasoning across various data sizes compared to full fine-tuning and standard LoRA. Our data and code are available at https://anonymous.4open.science/r/Shift-FFN

cs.CL↗
arXiv2025

Efficient Motion-Aware Video MLLM

Most current video MLLMs rely on uniform frame sampling and image-level encoders, resulting in inefficient data processing and limited motion awareness. To address these challenges, we introduce EMA, an Efficient Motion-Aware video MLLM that utilizes compressed video structures as inputs. We propose a motion-aware GOP (Group of Pictures) encoder that fuses spatial and motion information within a GOP unit in the compressed video stream, generating compact, informative visual tokens. By integrating fewer but denser RGB frames with more but sparser motion vectors in this native slow-fast input architecture, our approach reduces redundancy and enhances motion representation. Additionally, we introduce MotionBench, a benchmark for evaluating motion understanding across four motion types: linear, curved, rotational, and contact-based. Experimental results show that EMA achieves state-of-the-art performance on both MotionBench and popular video question answering benchmarks, while reducing inference costs. Moreover, EMA demonstrates strong scalability, as evidenced by its competitive performance on long video understanding benchmarks.

cs.CV↗
arXiv2025

Baichuan-M1: Pushing the Medical Capability of Large Language Models

The current generation of large language models (LLMs) is typically designed for broad, general-purpose applications, while domain-specific LLMs, especially in vertical fields like medicine, remain relatively scarce. In particular, the development of highly efficient and practical LLMs for the medical domain is challenging due to the complexity of medical knowledge and the limited availability of high-quality data. To bridge this gap, we introduce Baichuan-M1, a series of large language models specifically optimized for medical applications. Unlike traditional approaches that simply continue pretraining on existing models or apply post-training to a general base model, Baichuan-M1 is trained from scratch with a dedicated focus on enhancing medical capabilities. Our model is trained on 20 trillion tokens and incorporates a range of effective training methods that strike a balance between general capabilities and medical expertise. As a result, Baichuan-M1 not only performs strongly across general domains such as mathematics and coding but also excels in specialized medical fields. We have open-sourced Baichuan-M1-14B, a mini version of our model, which can be accessed through the following links.

cs.CL↗
arXiv2025

LongReD: Mitigating Short-Text Degradation of Long-Context Large Language Models via Restoration Distillation

Large language models (LLMs) have gained extended context windows through scaling positional encodings and lightweight continual pre-training. However, this often leads to degraded performance on short-text tasks, while the reasons for this degradation remain insufficiently explored. In this work, we identify two primary factors contributing to this issue: distribution drift in hidden states and attention scores, and catastrophic forgetting during continual pre-training. To address these challenges, we propose Long Context Pre-training with Restoration Distillation (LongReD), a novel approach designed to mitigate short-text performance degradation through minimizing the distribution discrepancy between the extended and original models. Besides training on long texts, LongReD distills the hidden state of selected layers from the original model on short texts. Additionally, LongReD also introduces a short-to-long distillation, aligning the output distribution on short texts with that on long texts by leveraging skipped positional indices. Experiments on common text benchmarks demonstrate that LongReD effectively preserves the model's short-text performance while maintaining comparable or even better capacity to handle long texts than baselines. Our code is available at https://github.com/RUCAIBox/LongReD.

cs.CL↗
arXiv2025

Virgo: A Preliminary Exploration on Reproducing o1-like MLLM

Recently, slow-thinking reasoning systems, built upon large language models (LLMs), have garnered widespread attention by scaling the thinking time during inference. There is also growing interest in adapting this capability to multimodal large language models (MLLMs). Given that MLLMs handle more complex data semantics across different modalities, it is intuitively more challenging to implement multimodal slow-thinking systems. To address this issue, in this paper, we explore a straightforward approach by fine-tuning a capable MLLM with a small amount of textual long-form thought data, resulting in a multimodal slow-thinking system, Virgo (Visual reasoning with long thought). We find that these long-form reasoning processes, expressed in natural language, can be effectively transferred to MLLMs. Moreover, it seems that such textual reasoning data can be even more effective than visual reasoning data in eliciting the slow-thinking capacities of MLLMs. While this work is preliminary, it demonstrates that slow-thinking capacities are fundamentally associated with the language model component, which can be transferred across modalities or domains. This finding can be leveraged to guide the development of more powerful slow-thinking reasoning systems. We release our resources at https://github.com/RUCAIBox/Virgo.

cs.CV↗
arXiv2024

KV Shifting Attention Enhances Language Modeling

The current large language models are mainly based on decode-only structure transformers, which have great in-context learning (ICL) capabilities. It is generally believed that the important foundation of its ICL capability is the induction heads mechanism, which requires at least two layers attention. In order to more efficiently implement the ability of the model's induction, we revisit the induction heads mechanism and proposed a KV shifting attention. We theoretically prove that the KV shifting attention reducing the model's requirements for the depth and width of the induction heads mechanism. Our experimental results demonstrate that KV shifting attention is beneficial to learning induction heads and language modeling, which lead to better performance or faster convergence from toy models to the pre-training models with more than 10 B parameters.

cs.CL↗
arXiv2024

LLaSA: Large Language and Structured Data Assistant

Structured data, such as tables, graphs, and databases, play a critical role in plentiful NLP tasks such as question answering and dialogue system. Recently, inspired by Vision-Language Models, Graph Neutral Networks (GNNs) have been introduced as an additional modality into the input of Large Language Models (LLMs) to improve their performance on Structured Knowledge Grounding (SKG) tasks. However, those GNN-enhanced LLMs have the following limitations: (1) They employ diverse GNNs to model varying types of structured data, rendering them unable to uniformly process various forms of structured data. (2) The pretraining of GNNs is coupled with specific LLMs, which prevents GNNs from fully aligning with the textual space and limits their adaptability to other LLMs. To address these issues, we propose \textbf{L}arge \textbf{L}anguage and \textbf{S}tructured Data \textbf{A}ssistant (LLaSA), a general framework for enhancing LLMs' ability to handle structured data. Specifically, we represent various types of structured data in a unified hypergraph format, and use self-supervised learning to pretrain a hypergraph encoder, and a G-Former compressing encoded hypergraph representations with cross-attention. The compressed hypergraph representations are appended to the serialized inputs during training and inference stages of LLMs. Experimental results on multiple SKG tasks show that our pretrained hypergraph encoder can adapt to various LLMs and enhance their ability to process different types of structured data. Besides, LLaSA, with LoRA fine-tuning, outperforms previous SOTA method using full parameters tuning.

cs.CL↗
arXiv2024

Beyond Filtering: Adaptive Image-Text Quality Enhancement for MLLM Pretraining

Multimodal large language models (MLLMs) have made significant strides by integrating visual and textual modalities. A critical factor in training MLLMs is the quality of image-text pairs within multimodal pretraining datasets. However, $\textit {de facto}$ filter-based data quality enhancement paradigms often discard a substantial portion of high-quality image data due to inadequate semantic alignment between images and texts, leading to inefficiencies in data utilization and scalability. In this paper, we propose the Adaptive Image-Text Quality Enhancer (AITQE), a model that dynamically assesses and enhances the quality of image-text pairs. AITQE employs a text rewriting mechanism for low-quality pairs and incorporates a negative sample learning strategy to improve evaluative capabilities by integrating deliberately selected low-quality samples during training. Unlike prior approaches that significantly alter text distributions, our method minimally adjusts text to preserve data volume while enhancing quality. Experimental results demonstrate that AITQE surpasses existing methods on various benchmark, effectively leveraging raw data and scaling efficiently with increasing data volumes. We hope our work will inspire future works. The code and model are available at: https://github.com/hanhuang22/AITQE.

cs.CV↗
arXiv2024

Exploring the Design Space of Visual Context Representation in Video MLLMs

Video Multimodal Large Language Models (MLLMs) have shown remarkable capability of understanding the video semantics on various downstream tasks. Despite the advancements, there is still a lack of systematic research on visual context representation, which refers to the scheme to select frames from a video and further select the tokens from a frame. In this paper, we explore the design space for visual context representation, and aim to improve the performance of video MLLMs by finding more effective representation schemes. Firstly, we formulate the task of visual context representation as a constrained optimization problem, and model the language modeling loss as a function of the number of frames and the number of embeddings (or tokens) per frame, given the maximum visual context window size. Then, we explore the scaling effects in frame selection and token selection respectively, and fit the corresponding function curve by conducting extensive empirical experiments. We examine the effectiveness of typical selection strategies and present empirical findings to determine the two factors. Furthermore, we study the joint effect of frame selection and token selection, and derive the optimal formula for determining the two factors. We demonstrate that the derived optimal settings show alignment with the best-performed results of empirical experiments. Our code and model are available at: https://github.com/RUCAIBox/Opt-Visor.

cs.CV↗
arXiv2024

Extracting and Combining Abilities For Building Multi-lingual Ability-enhanced Large Language Models

Multi-lingual ability transfer has become increasingly important for the broad application of large language models (LLMs). Existing work highly relies on training with the multi-lingual ability-related data, which may not be available for low-resource languages. To solve it, we propose a Multi-lingual Abilities Extraction and Combination approach, named as MAEC. Our key idea is to decompose and extract language-agnostic ability-related weights from LLMs, and combine them across different languages by simple addition and subtraction operations without training. Specifically, our MAEC consists of the extraction and combination stages. In the extraction stage, we firstly locate key neurons that are highly related to specific abilities, and then employ them to extract the transferable ability-related weights. In the combination stage, we further select the ability-related tensors that mitigate the linguistic effects, and design a combining strategy based on them and the language-specific weights, to build the multi-lingual ability-enhanced LLM. To assess the effectiveness of our approach, we conduct extensive experiments on LLaMA-3 8B on mathematical and scientific tasks in both high-resource and low-resource lingual scenarios. Experiment results have shown that MAEC can effectively and efficiently extract and combine the advanced abilities, achieving comparable performance with PaLM. Resources are available at https://github.com/RUCAIBox/MAET.

cs.CL↗