Tongxuan Liu

Huanyu Liu, Ge Li, Yihong Dong et al. · pku

Large language models (LLMs) based on the Transformer have demonstrated strong performance across diverse tasks. However, current models still exhibit substantial limitations in out-of-distribution (OOD) generalization compared with humans. We investigate this gap through periodicity, one of the basic OOD scenarios. Periodicity captures invariance amid variation. Periodicity generalization represents a model's ability to extract periodic patterns from training data and generalize to OOD scenarios. We introduce a unified interpretation of periodicity from the perspective of abstract algebra and reasoning, including both single and composite periodicity, to explain why Transformers struggle to generalize periodicity. Then we construct Coper about composite periodicity, a controllable generative benchmark with two OOD settings, Hollow and Extrapolation. Experiments reveal that periodicity generalization in Transformers is limited, where models can memorize periodic data during training, but cannot generalize to unseen composite periodicity. We release the source code to support future research.

Yufei Xu, Fanxu Meng, Fan Jiang et al.

Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical token for each query using a lightweight indexer, and then computing attention only over the selected subset. While the downstream sparse attention scales efficiently, the indexer still scans the entire prefix for every query, introducing an O($L^2$) per-layer bottleneck that becomes prohibitive as context length grows. We propose HISA (Hierarchical Indexed Sparse Attention), a drop-in replacement for the indexer that transforms the search process from a flat token scan into a two-stage hierarchical procedure. First, a block-level coarse filter scores pooled block representatives to prune irrelevant regions. Then, a token-level refinement applies the original indexer only within the remaining candidate blocks. HISA preserves the exact token-level top-k sparsity pattern required by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves a 2$\times$ speedup at 32K context length and 4$\times$ at 128K. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 with HISA, without any fine-tuning. HISA closely matches the original DSA in quality while significantly outperforming block-sparse baselines. Moreover, the token selection sets produced by HISA and the original DSA exhibit a mean IoU greater than 99%, indicating that the efficiency gains come with virtually no impact on selection fidelity.

Tongxuan Liu, Xingyu Wang, Weizhe Huang et al.

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse NLP tasks. Extensive research has explored how to enhance the logical reasoning abilities such as Chain-of-Thought, Chain-of-Thought with Self-Consistency, Tree-Of-Thoughts, and multi-agent debates. In the context of multi-agent debates, significant performance improvements can be achieved with an increasing number of agents and debate rounds. However, the escalation in the number of agents and debate rounds can drastically raise the tokens cost of debates, thereby limiting the scalability of the multi-agent debate technique. To better harness the advantages of multi-agent debates in logical reasoning tasks, this paper proposes a method to significantly reduce token cost in multi-agent debates. This approach involves dividing all agents into multiple debate groups, with agents engaging in debates within their respective groups and sharing interim debate results between groups. Comparative experiments across multiple datasets have demonstrated that this method can reduce the total tokens by up to 51.7% during debates and while potentially enhancing accuracy by as much as 25%. Our method significantly enhances the performance and efficiency of interactions in the multi-agent debate.

Tongxuan Liu, Wenjiang Xu, Weizhe Huang et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks but their performance in complex logical reasoning tasks remains unsatisfactory. Although some prompting methods, such as Chain-of-Thought, can improve the reasoning ability of LLMs to some extent, they suffer from an unfaithful issue where derived conclusions may not align with the generated reasoning chain. To address this issue, some studies employ the approach of propositional logic to further enhance logical reasoning abilities of LLMs. However, the potential omissions in the extraction of logical expressions in these methods can cause information loss in the logical reasoning process, thereby generating incorrect results. To this end, we propose Logic-of-Thought (LoT) prompting which employs propositional logic to generate expanded logical information descriptions and utilizes them as an additional augmentation to original contexts, thereby ensuring information completeness and enhancing logical reasoning ability. LoT is orthogonal to existing prompting methods and can be seamlessly integrated with them. Extensive experiments demonstrate that LoT boosts the performance of various prompting methods with a striking margin across five logical reasoning tasks. In particular, LoT enhances Chain-of-Thought's performance on the ReClor dataset by +4.35%, improves Chain-of-Thought with Self-Consistency's performance on the RuleTaker dataset by +3.52%, and boosts performance of Tree-of-Thoughts on the ProofWriter dataset by +8%.

Qingxiao Sun, Tongxuan Liu, Shen Zhang et al.

Recommendation system delivers substantial economic benefits by providing personalized predictions. Generative recommendation (GR) integrates LLMs to enhance the understanding of long user-item sequences. Despite employing attention-based architectures, GR's workload differs markedly from that of LLM serving. GR typically processes long prompt while producing short, fixed-length outputs, yet the computational cost of each decode phase is especially high due to the large beam width. In addition, since the beam search involves a vast item space, the sorting overhead becomes particularly time-consuming. We propose xGR, a GR-oriented serving system that meets strict low-latency requirements under highconcurrency scenarios. First, xGR unifies the processing of prefill and decode phases through staged computation and separated KV cache. Second, xGR enables early sorting termination and mask-based item filtering with data structure reuse. Third, xGR reconstructs the overall pipeline to exploit multilevel overlap and multi-stream parallelism. Our experiments with real-world recommendation service datasets demonstrate that xGR achieves at least 3.49x throughput compared to the state-of-the-art baseline under strict latency constraints.

Zihan Tang, Leqi Shen, Hui Chen et al.

Vision-Language Models (VLMs) have shown strong promise on Optical Character Recognition (OCR), yet the sheer number of visual tokens required to encode dense documents incurs prohibitive inference cost. Existing pruning methods rely on physical eviction, e.g., permanently discarding visual tokens during the prefill stage. While effective for natural images, this strategy fundamentally breaks down on OCR, where virtually every visual token may correspond to a character or structural element, and any irreversible loss leads to catastrophic accuracy degradation. We observe that, although document images appear globally dense and seemingly unprunable, the model's attention to them is in fact temporally sparse: at each decoding step it concentrates on a small region that shifts gradually across steps, much as a human reader fixates on successive words rather than perceiving an entire page at once. Motivated by this Dynamic Visual Fixation phenomenon, we recast the intractable global pruning problem as a tractable local, dynamic one and propose FastOCR, a training-free framework with two complementary modules. Specifically, Focal-Guided Pruning identifies a small set of focal layers and selects the most task-relevant visual tokens from them at each step, while Cross-Step Fixation Reuse exploits the gradual shift of fixation to warm-start each step from the previous one. By dynamically adjusting which tokens are attended rather than evicting any from the cache, FastOCR avoids permanent information loss. Extensive experiments show that FastOCR serves as a plug-and-play acceleration module, generalizing consistently across five VLMs of varying sizes and architectures. On Qwen2.5-VL, FastOCR retains 98% of the unpruned model's accuracy while attending to only 5% of the visual tokens per decoding step, reducing attention latency by 3.0$\times$.

Qingchuan Li, Jiatong Li, Tongxuan Liu et al.

Large Language Models (LLMs) have exhibited remarkable potential across a wide array of reasoning tasks, including logical reasoning. Although massive efforts have been made to empower the logical reasoning ability of LLMs via external logical symbolic solvers, crucial challenges of the poor generalization ability to questions with different features and inevitable question information loss of symbolic solver-driven approaches remain unresolved. To mitigate these issues, we introduce LINA, a LLM-driven neuro-symbolic approach for faithful logical reasoning. By enabling an LLM to autonomously perform the transition from propositional logic extraction to sophisticated logical reasoning, LINA not only bolsters the resilience of the reasoning process but also eliminates the dependency on external solvers. Additionally, through its adoption of a hypothetical-deductive reasoning paradigm, LINA effectively circumvents the expansive search space challenge that plagues traditional forward reasoning methods. Empirical evaluations demonstrate that LINA substantially outperforms both established propositional logic frameworks and conventional prompting techniques across a spectrum of five logical reasoning tasks. Specifically, LINA achieves an improvement of 24.34% over LINC on the FOLIO dataset, while also surpassing prompting strategies like CoT and CoT-SC by up to 24.02%. Our code is available at https://github.com/wufeiwuwoshihua/nshy.

Qingchuan Li, Mingyue Cheng, Zirui Liu et al.

Logical reasoning is a core challenge in natural language understanding and a fundamental capability of artificial intelligence, underpinning scientific discovery, mathematical theorem proving, and complex decision-making. Despite the remarkable progress of large language models (LLMs), most current approaches still rely on forward reasoning paradigms, generating step-by-step rationales from premises to conclusions. However, such methods often suffer from redundant inference paths, hallucinated steps, and semantic drift, resulting in inefficient and unreliable reasoning. In this paper, we propose a novel framework, Hypothesis-driven Backward Logical Reasoning (HBLR). The core idea is to integrate confidence-aware symbolic translation with hypothesis-driven backward reasoning. In the translation phase, only high-confidence spans are converted into logical form, such as First-Order Logic (FOL), while uncertain content remains in natural language. A translation reflection module further ensures semantic fidelity by evaluating symbolic outputs and reverting lossy ones back to text when necessary. In the reasoning phase, HBLR simulates human deductive thinking by assuming the conclusion is true and recursively verifying its premises. A reasoning reflection module further identifies and corrects flawed inference steps, enhancing logical coherence. Extensive experiments on five reasoning benchmarks demonstrate that HBLR consistently outperforms strong baselines in both accuracy and efficiency.

Xiaoheng Wang, Tongxuan Liu, Zi Gong et al.

Large language models (LLMs) have demonstrated impressive capabilities across a wide range of reasoning tasks, including logical and mathematical problem-solving. While prompt-based methods like Chain-of-Thought (CoT) can enhance LLM reasoning abilities to some extent, they often suffer from a lack of faithfulness, where the derived conclusions may not align with the generated reasoning chain. To address this issue, researchers have explored neuro-symbolic approaches to bolster LLM logical reasoning capabilities. However, existing neuro-symbolic methods still face challenges with information loss during the process. To overcome these limitations, we introduce Iterative Feedback-Driven Neuro-Symbolic (IFDNS), a novel prompt-based method that employs a multi-round feedback mechanism to address LLM limitations in handling complex logical relationships. IFDNS utilizes iterative feedback during the logic extraction phase to accurately extract causal relationship statements and translate them into propositional and logical implication expressions, effectively mitigating information loss issues. Furthermore, IFDNS is orthogonal to existing prompt methods, allowing for seamless integration with various prompting approaches. Empirical evaluations across six datasets demonstrate the effectiveness of IFDNS in significantly improving the performance of CoT and Chain-of-Thought with Self-Consistency (CoT-SC). Specifically, IFDNS achieves a +9.40% accuracy boost for CoT on the LogiQA dataset and a +11.70% improvement for CoT-SC on the PrOntoQA dataset.

Tongxuan Liu, Tao Peng, Peijun Yang et al.

We introduce xLLM, an intelligent and efficient Large Language Model (LLM) inference framework designed for high-performance, large-scale enterprise-grade serving, with deep optimizations for diverse AI accelerators. To address these challenges, xLLM builds a novel decoupled service-engine architecture. At the service layer, xLLM-Service features an intelligent scheduling module that efficiently processes multimodal requests and co-locates online and offline tasks through unified elastic scheduling to maximize cluster utilization. This module also relies on a workload-adaptive dynamic Prefill-Decode (PD) disaggregation policy and a novel Encode-Prefill-Decode (EPD) disaggregation policy designed for multimodal inputs. Furthermore, it incorporates a distributed architecture to provide global KV Cache management and robust fault-tolerant capabilities for high availability. At the engine layer, xLLM-Engine co-optimizes system and algorithm designs to fully saturate computing resources. This is achieved through comprehensive multi-layer execution pipeline optimizations, an adaptive graph mode and an xTensor memory management. xLLM-Engine also further integrates algorithmic enhancements such as optimized speculative decoding and dynamic EPLB, collectively serving to substantially boost throughput and inference efficiency. Extensive evaluations demonstrate that xLLM delivers significantly superior performance and resource efficiency. Under identical TPOT constraints, xLLM achieves throughput up to 1.7x that of MindIE and 2.2x that of vLLM-Ascend with Qwen-series models, while maintaining an average throughput of 1.7x that of MindIE with Deepseek-series models. xLLM framework is publicly available at https://github.com/jd-opensource/xllm and https://github.com/jd-opensource/xllm-service.

Qingchuan Li, Jiatong Li, Zirui Liu et al.

Logical reasoning with large language models (LLMs) has received growing attention. One mainstream approach translates natural language into formal logic and then applies symbolic solvers for deduction. While effective in many tasks, these LLM-based translators often fail to generate consistent symbolic representations when the same concept appears in different linguistic forms. Such inconsistencies break logical coherence and lead to solver errors. However, most existing benchmarks lack this type of linguistic variation, which frequently occurs in real-world text, leaving the problem underexplored. To address this gap, we present SoLT, a benchmark that systematically rewrites reasoning datasets into diverse yet logically equivalent forms across multiple levels. Beyond evaluation, SoLT also provides a general method to enrich any dataset with linguistic diversity while preserving both meaning and logic. To further enhance the stability of LLM-based reasoning, we propose MenTaL, which explicitly guides models to build a concept-symbol mapping table during translation. By linking equivalent expressions to shared symbols, MenTaL maintains consistency and mitigates symbol drift. Experiments on SoLT demonstrate that LLMs indeed suffer from inconsistent symbol mapping under linguistic variation, leading to significant drops in reasoning accuracy. Meanwhile, applying MenTaL brings clear and stable performance improvements across diverse inputs. Overall, our findings reveal that overlooking linguistic diversity hides key weaknesses in LLM-based translators, and our work offers a step toward more reliable logical reasoning in varied real-world scenarios. Our code is available at https://github.com/wufeiwuwoshihua/LinguDiver.

Ruijie Zhou, Fanxu Meng, Yufei Xu et al.

DeepSeek Sparse Attention (DSA) sets the state of the art for fine-grained inference-time sparse attention by introducing a learned token-wise indexer that scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses many query heads (for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of Indexer Sparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool of mixture-of-experts. A lightweight router uses cheap block-level statistics to pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligible router term computed on a small set of pooled keys. We further introduce a hierarchical variant of MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully green Needle-in-a-Haystack heatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. Our TileLang kernel delivers roughly a 3.82 times speedup over DSA's original indexer kernel on a single NVIDIA H200 GPU.

Ben Wan, Yan Feng, Zihan Tang et al.

DeepSeek-OCR leverages visual-text compression to reduce long-text processing costs and accelerate inference, yet visual tokens remain prone to redundant textual and structural information. Moreover, current token pruning methods for conventional vision-language models (VLMs) fail to preserve textual fidelity due to improper compression mechanisms. By analyzing the decoding process of DeepSeek-OCR, we find that a distinct two-stage reading trajectory: the model initially prioritizes the majority of high-norm tokens, then subsequently redistributes its attention to the remaining ones. Motivated by this insight, we propose RTPrune, a two-stage token pruning method tailored for DeepSeek-OCR. In the first stage, we prioritize high-norm visual tokens that capture salient textual and structural information. In the second stage, the remaining tokens are paired and merged based on optimal transport theory to achieve efficient feature aggregation. We further introduce a dynamic pruning ratio that adapts to token similarity and textual density for OCR tasks, enabling a better efficiency-accuracy trade-off. Extensive experiments demonstrate state-of-the-art performance, as evidenced by 99.47% accuracy and 1.23$\times$ faster prefill on OmniDocBench, achieved with 84.25% token retention when applied to DeepSeek-OCR-Large.

Lei Jiang, Weizhe Huang, Tongxuan Liu et al.

Large Vision-Language Models (LVLMs) represent a significant advancement toward achieving superior multimodal capabilities by enabling powerful Large Language Models (LLMs) to understand visual input. Typically, LVLMs utilize visual encoders, such as CLIP, to transform images into visual tokens, which are then aligned with textual tokens through projection layers before being input into the LLM for inference. Although existing LVLMs have achieved significant success, their inference efficiency is still limited by the substantial number of visual tokens and the potential redundancy among them. To mitigate this issue, we propose Focal Pruning (FoPru), a training-free method that prunes visual tokens based on the attention-based token significance derived from the vision encoder. Specifically, we introduce two alternative pruning strategies: 1) the rank strategy, which leverages all token significance scores to retain more critical tokens in a global view; 2) the row strategy, which focuses on preserving continuous key information in images from a local perspective. Finally, the selected tokens are reordered to maintain their original positional relationships. Extensive experiments across various LVLMs and multimodal datasets demonstrate that our method can prune a large number of redundant tokens while maintaining high accuracy, leading to significant improvements in inference efficiency.

Yuting Zeng, Weizhe Huang, Lei Jiang et al.

Large language models (LLMs) have demonstrated remarkable capabilities across various natural language processing (NLP) scenarios, but they still face challenges when handling complex arithmetic and logical reasoning tasks. While Chain-Of-Thought (CoT) reasoning, self-consistency (SC) and self-correction strategies have attempted to guide models in sequential, multi-step reasoning, Multi-agent Debate (MAD) has emerged as a viable approach for enhancing the reasoning capabilities of LLMs. By increasing both the number of agents and the frequency of debates, the performance of LLMs improves significantly. However, this strategy results in a significant increase in token costs, presenting a barrier to scalability. To address this challenge, we introduce a novel sparsification strategy designed to reduce token costs within MAD. This approach minimizes ineffective exchanges of information and unproductive discussions among agents, thereby enhancing the overall efficiency of the debate process. We conduct comparative experiments on multiple datasets across various models, demonstrating that our approach significantly reduces the token costs in MAD to a considerable extent. Specifically, compared to MAD, our approach achieves an impressive reduction of up to 94.5\% in token costs while maintaining performance degradation below 2.0\%.

Chunzhao Xie, Tongxuan Liu, Lei Jiang et al.

Large Vision-Language Models have demonstrated remarkable performance across various tasks; however, the challenge of hallucinations constrains their practical applications. The hallucination problem arises from multiple factors, including the inherent hallucinations in language models, the limitations of visual encoders in perception, and biases introduced by multimodal data. Extensive research has explored ways to mitigate hallucinations. For instance, OPERA prevents the model from overly focusing on "anchor tokens", thereby reducing hallucinations, whereas VCD mitigates hallucinations by employing a contrastive decoding approach. In this paper, we investigate the correlation between the decay of attention to image tokens and the occurrence of hallucinations. Based on this finding, we propose Temporal Attention Real-time Accumulative Connection (TARAC), a novel training-free method that dynamically accumulates and updates LVLMs' attention on image tokens during generation. By enhancing the model's attention to image tokens, TARAC mitigates hallucinations caused by the decay of attention on image tokens. We validate the effectiveness of TARAC across multiple models and datasets, demonstrating that our approach substantially mitigates hallucinations. In particular, TARAC reduces $C_S$ by 25.2 and $C_I$ by 8.7 compared to VCD on the CHAIR benchmark.

Wenqi Jiang, Zhenhao He, Shuai Zhang et al.

Deep neural networks are widely used in personalized recommendation systems. Unlike regular DNN inference workloads, recommendation inference is memory-bound due to the many random memory accesses needed to lookup the embedding tables. The inference is also heavily constrained in terms of latency because producing a recommendation for a user must be done in about tens of milliseconds. In this paper, we propose MicroRec, a high-performance inference engine for recommendation systems. MicroRec accelerates recommendation inference by (1) redesigning the data structures involved in the embeddings to reduce the number of lookups needed and (2) taking advantage of the availability of High-Bandwidth Memory (HBM) in FPGA accelerators to tackle the latency by enabling parallel lookups. We have implemented the resulting design on an FPGA board including the embedding lookup step as well as the complete inference process. Compared to the optimized CPU baseline (16 vCPU, AVX2-enabled), MicroRec achieves 13.8~14.7x speedup on embedding lookup alone and 2.5$~5.4x speedup for the entire recommendation inference in terms of throughput. As for latency, CPU-based engines needs milliseconds for inferring a recommendation while MicroRec only takes microseconds, a significant advantage in real-time recommendation systems.