Dong Dong

Duling Xu, Zheng Chen, Zaifeng Pan et al.

Recently, skills have been widely adopted in large language model (LLM)-based agent systems across various domains. In existing frameworks, skills are typically injected into the agent reasoning loop as contextual guidance once matched to a runtime task, enabling specialized task-solving capabilities. We find that this execution paradigm introduces two major sources of redundancy: irrelevant context injection and repeated skill-specific reasoning and planning. To this end, we propose SkillSmith, a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces. By extracting fine-grained operational boundaries from skills, SkillSmith enables agents to dynamically access and execute only the relevant components at runtime, thereby minimizing unnecessary context injection and redundant reasoning overhead. In the evaluation on SkillsBench benchmark, SkillSmith reduces solve-stage token usage by 57.44%, thinking iterations by 42.99%, solve time by 50.57% (2.02x faster), and token-proportional monetary cost by 57.44% compared with using raw-skills. Moreover, compiled artifacts produced by a stronger model can be reused by a smaller or more efficient runtime model, improving task accuracy in cases where raw skill interpretation fails. The source code and data are available at https://github.com/AetherHeart-AI/Aeloon.

Enhao Huang, Zhiyu Zhang, Tianxiang Xu et al.

Complex-valued signals encode both amplitude and phase, yet most deep models treat attention as real-valued correlation, overlooking interference effects. We introduce the Holographic Transformer, a physics-inspired architecture that incorporates wave interference principles into self-attention. Holographic attention modulates interactions by relative phase and coherently superimposes values, ensuring consistency between amplitude and phase. A dual-headed decoder simultaneously reconstructs the input and predicts task outputs, preventing phase collapse when losses prioritize magnitude over phase. We demonstrate that holographic attention implements a discrete interference operator and maintains phase consistency under linear mixing. Experiments on PolSAR image classification and wireless channel prediction show strong performance, achieving high classification accuracy and F1 scores, low regression error, and increased robustness to phase perturbations. These results highlight that enforcing physical consistency in attention leads to generalizable improvements in complex-valued learning and provides a unified, physics-based framework for coherent signal modeling. The code is available at https://github.com/EonHao/Holographic-Transformers.

Dong Dong, Weijie Su

We introduce a new tokenizer for language models that minimizes the average tokens per character, thereby reducing the number of tokens needed to represent text during training and to generate text during inference. Our method, which we refer to as the Length-MAX tokenizer, obtains its vocabulary by casting a length-weighted objective maximization as a graph partitioning problem and developing a greedy approximation algorithm. On FineWeb and diverse domains, it yields 14--18\% fewer tokens than Byte Pair Encoding (BPE) across vocabulary sizes from 10K to 50K, and the reduction is 13.0\% when the size is 64K. Training GPT-2 models at 124M, 355M, and 1.3B parameters from scratch with five runs each shows 18.5\%, 17.2\%, and 18.5\% fewer steps, respectively, to reach a fixed validation loss, and 13.7\%, 12.7\%, and 13.7\% lower inference latency, together with a 16\% throughput gain at 124M, while consistently improving on downstream tasks including reducing LAMBADA perplexity by 11.7\% and enhancing HellaSwag accuracy by 4.3\%. Moreover, the Length-MAX tokenizer achieves 99.62\% vocabulary coverage and the out-of-vocabulary rate remains low at 0.12\% on test sets. These results demonstrate that optimizing for average token length, rather than frequency alone, offers an effective approach to more efficient language modeling without sacrificing -- and often improving -- downstream performance. The tokenizer is compatible with production systems and reduces embedding and KV-cache memory by 18\% at inference.

Yusong Zhang, Dong Dong, Chi-tim Hung et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in language understanding and generation. Advanced utilization of the knowledge embedded in LLMs for automated annotation has consistently been explored. This study proposed to develop an emotion lexicon for Cantonese, a low-resource language, through collaborative efforts between LLM and human annotators. By integrating emotion labels provided by LLM and human annotators, the study leveraged existing linguistic resources including lexicons in other languages and local forums to construct a Cantonese emotion lexicon enriched with colloquial expressions. The consistency of the proposed emotion lexicon in emotion extraction was assessed through modification and utilization of three distinct emotion text datasets. This study not only validates the efficacy of the constructed lexicon but also emphasizes that collaborative annotation between human and artificial intelligence can significantly enhance the quality of emotion labels, highlighting the potential of such partnerships in facilitating natural language processing tasks for low-resource languages.

Yanfei Song, Bangzheng Pu, Peng Wang et al.

Segment Anything Model (SAM) has garnered significant attention in segmentation tasks due to their zero-shot generalization ability. However, a broader application of SAMs to real-world practice has been restricted by their low inference speed and high computational memory demands, which mainly stem from the attention mechanism. Existing work concentrated on optimizing the encoder, yet has not adequately addressed the inefficiency of the attention mechanism itself, even when distilled to a smaller model, which thus leaves space for further improvement. In response, we introduce SAM-Lightening, a variant of SAM, that features a re-engineered attention mechanism, termed Dilated Flash Attention. It not only facilitates higher parallelism, enhancing processing efficiency but also retains compatibility with the existing FlashAttention. Correspondingly, we propose a progressive distillation to enable an efficient knowledge transfer from the vanilla SAM without costly training from scratch. Experiments on COCO and LVIS reveal that SAM-Lightening significantly outperforms the state-of-the-art methods in both run-time efficiency and segmentation accuracy. Specifically, it can achieve an inference speed of 7 milliseconds (ms) per image, for images of size 1024*1024 pixels, which is 30.1 times faster than the vanilla SAM and 2.1 times than the state-of-the-art. Moreover, it takes only 244MB memory, which is 3.5\% of the vanilla SAM. The code and weights are available at https://anonymous.4open.science/r/SAM-LIGHTENING-BC25/.

Wojciech Czaja, Dong Dong, Pierre-Emmanuel Jabin et al.

We present a new feature extraction method for complex and large datasets, based on the concept of transport operators on graphs. The proposed approach generalizes and extends the many existing data representation methodologies built upon diffusion processes, to a new domain where dynamical systems play a key role. The main advantage of this approach comes from the ability to exploit different relationships than those arising in the context of e.g., Graph Laplacians. Fundamental properties of the transport operators are proved. We demonstrate the flexibility of the method by introducing several diverse examples of transformations. We close the paper with a series of computational experiments and applications to the problem of classification of hyperspectral satellite imagery, to illustrate the practical implications of our algorithm and its ability to quantify new aspects of relationships within complicated datasets.