Xiao Liu · page 2
Papers
- Glyph: Scaling Context Windows via Visual-Text Compression (2026 · 6 citations)
Long-context LLMs usually scale by handling more text tokens, which raises compute and memory costs. Glyph takes a different route. Instead of enlarging the token window, it packs more source text into each token: documents are rendered as images and fed to a vision-language model. Built on GLM-4.1V-9B-Base, the… - Data Mixing Agent: Learning to Re-weight Domains for Continual Pre-training (2026 · 1 citation)
Continual pre-training of an LLM on a narrow target field, such as math, often causes catastrophic forgetting of general skills. - Attribution-Based Analysis and Optimization of Modular Agentic Workflows (2026)
Agent systems built from separate modules (planning, reasoning, action, reflection) are usually judged only by final task success. - COSMOS: Connectivity-Oriented Submodular Maximization for Optimal Subgraph Retrieval (2026)
The paper tackles evidence retrieval for Knowledge Base Question Answering (KBQA), where an LLM needs a compact and connected set of knowledge-graph facts rather than isolated triplets. The authors identify weaknesses in three existing families of retrievers: - DynamixSFT: Dynamic Mixture Optimization of Instruction Tuning Collections (2026)
Instruction-tuning collections such as TÜLU combine many heterogeneous datasets, yet their mixing ratios are usually fixed by hand or need extra proxy models and validation sets to adjust. - EXCEEDS: Extracting Complex Events via Nugget-based Grid Modeling in Scientific Domain (2026)
Scientific writing has received little attention in event extraction (EE). Few datasets or methods are designed for it, even though research abstracts pack in many triggers and arguments and often have complicated structures: nested events, discontinuous arguments, and arguments that sit far from their trigger. The… - GATE: Graph-based Adaptive Tool Evolution Across Diverse Tasks (2026)
LLM-based tool-making systems tend to accumulate bloated, repetitive tool libraries and are usually tied to a single task type, such as Minecraft or code generation. - KARL: Reinforcement Learning for LLM Agents on Multi-Turn Knowledge-Intensive Agentic Tasks (2026)
LLM agents often handle knowledge-heavy tasks poorly because they consume external information passively. Retrieval is either fixed in advance or tuned only for single-turn question answering. KARL (Knowledge-Augmented Reinforcement Learning) addresses this for multi-turn work over structured sources such as…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/xiao-liu/page/2/