Provides introductions and tutorials on Bilibili-related technologies.
HOMURA introduces a reinforcement learning framework with KL regularization and dynamic syllable-ratio rewards to achieve precise length control in LLM translation for subtitles and dubbing, outperforming baselines on the Sand-Glass benchmark and deployed at scale on Bilibili.
IndexTTS 2.5 adds zero‑sample voice cloning for Chinese, English, Japanese, Spanish and Arabic, doubles inference speed, introduces flexible speech‑rate control, and improves emotion and timbre fidelity, with detailed benchmarks and a systematic multilingual modeling study supporting its design choices.
GeoRK2 introduces a training‑free, plug‑and‑play framework that combines second‑order Runge‑Kutta integration with low‑rank geometric correction of diffusion Transformers, enabling 4‑5× faster image, text‑to‑image, and video generation while preserving high visual quality and stability.
At ACL 2026 in San Diego, Bilibili’s tech team introduced the FATE series—SABER, which reduces overthinking in LLMs with a token‑budgeted switchable training, and CASTER, a community‑aware evaluation system built on Social‑CoT and the MEDEA framework that outperforms GPT‑5.2 and Claude‑4.5‑Opus on the new CASTER‑Bench, while also promoting the B‑UP talent recruitment program.
LeAP replaces costly permutation‑based feature importance with an end‑to‑end learnable module that shuffles features, applies adaptive gradient‑bias regularization, and achieves state‑of‑the‑art sparsity and inference speedups on large‑scale recommender models.
The paper introduces CASTER, a multimodal AI system that uses Social‑CoT reasoning and the MEDEA framework to simulate diverse audience reactions, benchmarked on the large‑scale CASTER‑Bench dataset, and demonstrates superior performance over GPT‑5.2, Claude‑4.5‑Opus, and traditional VQA methods while already being deployed on Bilibili.
This article details how a unified AI assistant framework built for Bilibili's advertising business evolves from plain text output to generating fully interactive UI by leveraging Google’s A2UI protocol, a custom Vue renderer, double‑validation mechanisms, SSE dual‑channel streaming, and a wrapper component system, providing concrete examples and architectural diagrams.
This article details how the Bili‑FE team evolved prompt engineering into a Harness Engineering workflow, creating a structured .workflow knowledge base and a series of AI‑driven commands that automate the entire frontend development lifecycle from requirement preprocessing to testing and mock generation.
The SPICE method leverages Fisher Information Matrix submodularity and a novel gradient‑conflict penalty to select a small, high‑quality subset of instruction‑tuning data, achieving comparable or superior performance to full‑data fine‑tuning while dramatically reducing training cost.
This article examines Neon Vibe Motion, an open‑source platform that lets users describe motion effects in natural language, uses LLMs to generate executable Canvas/WebGL code with adjustable parameters, and details the architecture, workflow, prompt engineering, and export options that enable real‑time, controllable motion graphics.
