arXiv subjects

Bohao Tang

Publications and source records attributed to Bohao Tang.

9 recordsLinked to original sources

arXiv2026

LoGo: Token-Level Dynamic Local-Global Attention

As context lengths scale, attention increasingly becomes a primary computational bottleneck in large language models. Standard Transformers remain powerful but computationally inefficient, as they allocate the same attention budget to every token regardless of its contextual demand. Existing local-global hybrids provide a more efficient alternative by mixing restricted- and full-context attention, but they typically allocate span statically across layers or heads. To address these limitations, we propose LoGo, a token-level dynamic local-global attention mechanism that uses attention span as a direct proxy for attention budget allocation. Each LoGo layer contains coupled local and global branches: all tokens receive efficient local attention over a restricted context window, while a learned gate activates global attention with full-context access only for tokens requiring long-range information. A threshold-based budget controller maintains a target global ratio without auxiliary losses, and a progressive masking schedule stabilizes training before sparse routing takes effect. We further implement query-sparse Triton kernels that convert reduced global-attention computation into practical speedups. Extensive experiments validate LoGo's effectiveness, showing that it preserves the scaling behavior of full-attention Transformers across model sizes. In controlled comparisons, LoGo improves over the full-attention Transformer and matched-budget static local-global hybrids, with clear gains on long-range retrieval. Analysis further shows that LoGo learns interpretable span allocation patterns. These results suggest that learned token-level span allocation is an effective and scalable way to improve the long-context performance-compute trade-off.

cs.CL↗

arXiv2026

RIBOSPAN: A Long-Context RNA Foundation Model for Versatile RNA Modeling

Full-length RNAs, particularly messenger RNAs, often exceed the context lengths used to pretrain existing RNA foundation models, limiting complete-transcript modeling at single-nucleotide resolution. We present RIBOSPAN, a 1.61-billion-parameter bidirectional RNA foundation model natively pretrained with context lengths up to 10,240 nt. RIBOSPAN combines dense bidirectional self-attention, single-nucleotide tokenization, and attention-isolated sequence packing to enable high-resolution modeling of complete long RNAs. Native 10K pretraining preserves strong reconstruction at 10,240 tokens and, in a controlled long-context benchmark, maintains strong contextual responsiveness and context-specific representation separation while keeping perturbation-induced changes highly localized. Inference-time YaRN scaling recovers much of the contextual organization lost by direct short-context extrapolation, but induces substantially greater distal representation diffusion. Frozen RNA-type evaluations show that RIBOSPAN learns state-of-the-art RNA representations, with a particularly clear advantage on long RNAs. Across downstream biological benchmarks, RIBOSPAN emerges as the strongest encoder-only RNA foundation model, achieving state-of-the-art performance in both full-transcript biological property prediction and zero-shot mutation-fitness modeling. Building on the same backbone, we develop a multidimensionally conditioned discrete-diffusion framework for full-length mRNA generation and redesign, including synonymous-codon diffusion for protein-preserving CDS optimization. Together, RIBOSPAN establishes a powerful long-context foundation for transferable RNA representation learning, biological prediction, and full-transcript mRNA design.

cs.LG↗

arXiv2026

Modular TTT: Rethinking Test-Time Training as Composable Modules

Test-time training (TTT) views sequence modeling as an online learning problem in which fast weights are updated by an internal learning rule. Despite the growing number of TTT variants, existing approaches typically hard-code each variant separately, which makes it difficult to design new TTT methods and to isolate the role of each component. To address this, we propose Modular TTT, a framework that represents the inner learner as a directed acyclic graph and exposes the fast-weight network, loss function, learning rate, weight decay, and normalization as explicit design dimensions. Modular TTT automatically composes primitive-level train-view forward, train-view backward, and causal query-view rules into the full graph-level TTT computation, including the fast-weight state transition. Using Modular TTT, we systematically ablate the components of TTT and find that small learning-rate initialization, weight decay, and a single-layer nonlinearity improve performance, while MSE and inner-product losses perform similarly. Deeper fast-weight networks and normalization tend to hurt performance because they induce excessively large activations, while residual connections and gating provide little measurable benefit. Guided by these findings, we train the best resulting variant as 410M- and 1.45B-parameter models on 100B tokens, and observe training loss and benchmark performance comparable to Gated DeltaNet.

cs.LG↗

arXiv2026

TransText: Alpha-as-RGB Representation for Transparent Text Animation

We introduce the first method, to the best of our knowledge, for adapting image-to-video models to layer-aware text (glyph) animation, a capability critical for practical dynamic visual design. Existing approaches predominantly handle the transparency-encoding (alpha channel) as an extra latent dimension appended to the RGB space, necessitating the reconstruction of the underlying RGB-centric variational autoencoder (VAE). However, given the scarcity of high-quality transparent glyph data, retraining the VAE is computationally expensive and may erode the robust semantic priors learned from massive RGB corpora, potentially leading to latent pattern mixing. To mitigate these limitations, we propose TransText, a framework based on a novel Alpha-as-RGB paradigm to jointly model appearance and transparency without modifying the pre-trained generative manifold. TransText embeds the alpha channel as an RGB-compatible visual signal through latent spatial concatenation, explicitly ensuring strict cross-modal (RGB-and-Alpha) consistency while preventing feature entanglement. Our experiments demonstrate that TransText significantly outperforms baselines, generating coherent, high-fidelity transparent animations with diverse, fine-grained effects.

cs.CV↗

arXiv2025

LiveTalk: Real-Time Multimodal Interactive Video Diffusion via Improved On-Policy Distillation

Real-time video generation via diffusion is essential for building general-purpose multimodal interactive AI systems. However, the simultaneous denoising of all video frames with bidirectional attention via an iterative process in diffusion models prevents real-time interaction. While existing distillation methods can make the model autoregressive and reduce sampling steps to mitigate this, they focus primarily on text-to-video generation, leaving the human-AI interaction unnatural and less efficient. This paper targets real-time interactive video diffusion conditioned on a multimodal context, including text, image, and audio, to bridge the gap. Given the observation that the leading on-policy distillation approach Self Forcing encounters challenges (visual artifacts like flickering, black frames, and quality degradation) with multimodal conditioning, we investigate an improved distillation recipe with emphasis on the quality of condition inputs as well as the initialization and schedule for the on-policy optimization. On benchmarks for multimodal-conditioned (audio, image, and text) avatar video generation including HDTF, AVSpeech, and CelebV-HQ, our distilled model matches the visual quality of the full-step, bidirectional baselines of similar or larger size with 20x less inference cost and latency. Further, we integrate our model with audio language models and long-form video inference technique Anchor-Heavy Identity Sinks to build LiveTalk, a real-time multimodal interactive avatar system. System-level evaluation on our curated multi-turn interaction benchmark shows LiveTalk outperforms state-of-the-art models (Sora2, Veo3) in multi-turn video coherence and content quality, while reducing response latency from 1 to 2 minutes to real-time generation, enabling seamless human-AI multimodal interaction.

cs.CV↗

arXiv2025

Visual Programmability: A Guide for Code-as-Thought in Chart Understanding

Chart understanding presents a critical test to the reasoning capabilities of Vision-Language Models (VLMs). Prior approaches face critical limitations: some rely on external tools, making them brittle and constrained by a predefined toolkit, while others fine-tune specialist models that often adopt a single reasoning strategy, such as text-based chain-of-thought (CoT). The intermediate steps of text-based reasoning are difficult to verify, which complicates the use of reinforcement-learning signals that reward factual accuracy. To address this, we propose a Code-as-Thought (CaT) approach to represent the visual information of a chart in a verifiable, symbolic format. Our key insight is that this strategy must be adaptive: a fixed, code-only implementation consistently fails on complex charts where symbolic representation is unsuitable. This finding leads us to introduce Visual Programmability: a learnable property that determines if a chart-question pair is better solved with code or direct visual analysis. We implement this concept in an adaptive framework where a VLM learns to choose between the CaT pathway and a direct visual reasoning pathway. The selection policy of the model is trained with reinforcement learning using a novel dual-reward system. This system combines a data-accuracy reward to ground the model in facts and prevent numerical hallucination, with a decision reward that teaches the model when to use each strategy, preventing it from defaulting to a single reasoning mode. Experiments demonstrate strong and robust performance across diverse chart-understanding benchmarks. Our work shows that VLMs can be taught not only to reason but also how to reason, dynamically selecting the optimal reasoning pathway for each task.

cs.CV↗

arXiv2024

Direct Bayesian Regression for Distribution-valued Covariates

In this manuscript, we study the problem of scalar-on-distribution regression; that is, instances where subject-specific distributions or densities, or in practice, repeated measures from those distributions, are the covariates related to a scalar outcome via a regression model. We propose a direct regression for such distribution-valued covariates that circumvents estimating subject-specific densities and directly uses the observed repeated measures as covariates. The model is invariant to any transformation or ordering of the repeated measures. Endowing the regression function with a Gaussian Process prior, we obtain closed form or conjugate Bayesian inference. Our method subsumes the standard Bayesian non-parametric regression using Gaussian Processes as a special case. Theoretically, we show that the method can achieve an optimal estimation error bound. To our knowledge, this is the first theoretical study on Bayesian regression using distribution-valued covariates. Through simulation studies and analysis of activity count dataset, we demonstrate that our method performs better than approaches that require an intermediate density estimation step.

stat.ME↗

arXiv2023

Density-on-Density Regression

In this study, a density-on-density regression model is introduced, where the association between densities is elucidated via a warping function. The proposed model has the advantage of a being straightforward demonstration of how one density transforms into another. Using the Riemannian representation of density functions, which is the square-root function (or half density), the model is defined in the correspondingly constructed Riemannian manifold. To estimate the warping function, it is proposed to minimize the average Hellinger distance, which is equivalent to minimizing the average Fisher-Rao distance between densities. An optimization algorithm is introduced by estimating the smooth monotone transformation of the warping function. Asymptotic properties of the proposed estimator are discussed. Simulation studies demonstrate the superior performance of the proposed approach over competing approaches in predicting outcome density functions. Applying to a proteomic-imaging study from the Alzheimer's Disease Neuroimaging Initiative, the proposed approach illustrates the connection between the distribution of protein abundance in the cerebrospinal fluid and the distribution of brain regional volume. Discrepancies among cognitive normal subjects, patients with mild cognitive impairment, and Alzheimer's disease (AD) are identified and the findings are in line with existing knowledge about AD.

stat.ME↗

arXiv2022

Matrix Syncer -- A Multi-chain Data Aggregator For Supporting Blockchain-based Metaverses

Due to the rising complexity of the metaverse's business logic and the low-latency nature of the metaverse, developers typically encounter the challenge of effectively reading, writing, and retrieving historical on-chain data in order to facilitate their functional implementations at scale. While it is true that accessing blockchain states is simple, more advanced real-world operations such as search, aggregation, and conditional filtering are not available when interacting directly with blockchain networks, particularly when dealing with requirements for on-chain event reflection. We offer Matrix Syncer, the ultimate middleware that bridges the data access gap between blockchains and end-user applications. Matrix Syncer is designed to facilitate the consolidation of on-chain information into a distributed data warehouse while also enabling customized on-chain state transformation for a scalable storage, access, and retrieval. It offers a unified layer for both on- and off-chain state, as well as a fast and flexible atomic query. Matrix Syncer is easily incorporated into any infrastructure to aggregate data from various blockchains concurrently, such as Ethereum and Flow. The system has been deployed to support several metaverse projects with a total value of more than $15 million USD.

cs.MM↗