Weekly AI Paper Roundup: RL Advances, Tree‑Structured QA, and GraphRAG Breakthroughs

Survey of Reinforcement Learning for Large Reasoning Models

The paper reviews the latest progress of reinforcement learning (RL) in enhancing the reasoning abilities of large language models (LLMs), focusing on developments since DeepSeek‑R1. It analyses the underlying architecture, core challenges, training resource requirements, and downstream applications, aiming to pinpoint future research opportunities.

https://go.hyper.ai/UrAIM

ST‑Raptor: LLM‑Powered Semi‑Structured Table Question Answering

ST‑Raptor introduces a tree‑structured framework that decomposes a user query into simpler sub‑questions, generates a corresponding tree‑operation pipeline, and aligns these operations with the table to execute precise answers.

https://go.hyper.ai/Vp6z2

Visual Representation Alignment for Multimodal Large Language Models (VIRAL)

The authors propose a simple yet effective regularisation strategy called Visual Representation Alignment (VIRAL), which aligns the visual embeddings inside multimodal LLMs with those of a pretrained visual foundation model, enabling more efficient integration of visual information.

https://go.hyper.ai/AGpt3

Youtu‑GraphRAG: Graph Retrieval‑Enhanced Generation

GraphRAG reorganises fragmented knowledge into an explicit, structured graph, dramatically improving LLM performance on complex reasoning tasks. The paper presents a vertically unified agent paradigm—Youtu‑GraphRAG—that tightly integrates the entire pipeline, showing strong adaptability and smooth domain transfer with minimal graph‑mode intervention.

https://go.hyper.ai/UtBzR

CryptoScope: Automated Cryptographic Logic Vulnerability Detection with LLMs

CryptoScope proposes a robust, well‑generalising knowledge‑graph construction framework that combines intelligent document parsing, table‑oriented text chunking, pattern‑guided iterative information extraction, and a reflection‑driven feedback loop to automatically discover cryptographic logic flaws.

https://go.hyper.ai/0oT2b

The roundup concludes with a reminder that more cutting‑edge AI papers are available on the HyperAI "Latest Papers" page, and invites research teams to submit high‑quality work.