Xipeng Qiu · page 2
Papers
- Evaluating the Performance of Large Language Models on GAOKAO Benchmark (2023 · 109 citations)
Popular LLM benchmarks such as MMLU, C-Eval and AGIEval consist almost entirely of objective items. They reveal little about generative ability and give no intuitive, human-comparable sense of model skill. The authors address this gap with GAOKAO-Bench, a dataset of national Chinese college entrance exam (GAOKAO)… - Finding Support Examples for In-Context Learning (2023 · 78 citations)
In-context learning (ICL) lets a language model solve a task from a handful of demonstrations in its prompt, but results swing widely depending on which demonstrations are chosen and how they are ordered. - Scaling Laws of RoPE-based Extrapolation (2023 · 54 citations)
LLMs that use Rotary Position Embedding (RoPE) degrade once inputs exceed their training length. The usual remedy is to raise the rotary base above the default 10,000 and fine-tune on longer texts. This paper asks how the base value, the tuning context length and the achievable extrapolation relate to each other. The… - SpeechTokenizer: Unified Speech Tokenizer for Speech Large Language Models (2023 · 53 citations)
Speech large language models rely on discrete tokens of two kinds. Semantic tokens, such as k-means units from HuBERT, follow linguistic content but discard voice traits. Acoustic tokens from neural codecs like EnCodec retain audio detail but correspond poorly to text. The authors first build SLMTokBench, a benchmark… - A Graph-based Model for Joint Chinese Word Segmentation and Dependency Parsing (2020 · 53 citations)
Chinese text has no explicit word boundaries, but dependency parsing works on words, so parsers usually run after a separate segmentation step. - Revisiting the Test-Time Scaling of o1-like Models: Do they Truly Possess Test-Time Scaling Capabilities? (2025 · 38 citations)
Open o1-style reasoning models such as QwQ, Deepseek-R1 (R1) and LIMO generate very long chains of thought (CoT). - ConvSearch-R1: Enhancing Query Reformulation for Conversational Search with Reasoning via Reinforcement Learning (2025 · 13 citations)
In conversational search, user queries often depend on earlier turns. They may leave things out, be vague, or refer back to prior context. Conversational Query Reformulation (CQR) rewrites such queries into stand-alone versions that a standard retriever can handle. Earlier CQR systems learned from rewrites written by… - VisuoThink: Empowering LVLM Reasoning with Multimodal Tree Search (2025 · 13 citations)
Large vision-language models (LVLMs) still struggle with geometry problems and spatial navigation. People handle such tasks by sketching, checking intermediate pictures and trying alternative routes. Earlier multimodal "slow thinking" methods either reason purely in text over a fixed input image or add visual help… - REARANK: Reasoning Re-ranking Agent via Reinforcement Learning (2025 · 12 citations)
REARANK tackles passage reranking with LLMs. Prompting-only rerankers do not learn from ranking feedback, supervised fine-tuning requires costly labelled lists, and the strongest systems rely on large proprietary models such as GPT-4. The authors fine-tune Qwen2.5-7B with GRPO-based reinforcement learning so that it… - Dual Cache for Long Document Neural Coreference Resolution (2023 · 12 citations)
Cache-based coreference models handle long documents by reading mentions left to right and keeping only a bounded set of entities in memory. - GAOKAO-MM: A Chinese Human-Level Benchmark for Multimodal Models Evaluation (2024 · 11 citations)
Most vision-language benchmarks test basic perception or commonsense with short prompts. Chinese versions are usually translations of English datasets. As a result, they reveal little about how large vision-language models (LVLMs) cope with demanding, natively Chinese material. Zong and Qiu fill this gap with… - Unearthing Large Scale Domain-Specific Knowledge from Public Corpora (2024 · 10 citations)
Building domain-specific corpora for LLMs usually requires experts to pick sources and engineers to scrape them. - HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding (2026 · 9 citations)
Multimodal LLMs handle offline video well, but streaming video is harder. A system must stay accurate, answer quickly, and keep GPU memory bounded, even though future frames and questions are unknown. Earlier cache-based methods such as ReKV and LiveVLM move KV entries off the GPU and retrieve them when a question… - FastMCTS: A Simple Sampling Strategy for Data Synthesis (2025 · 8 citations)
Reasoning datasets for training LLMs are usually built with rejection sampling: many full solutions are drawn independently, and only those whose final answer matches the reference are kept. - Case2Code: Scalable Synthetic Data for Code Generation (2025 · 7 citations)
Code LLMs are often improved with synthetic instruction data, but this data usually comes from a strong teacher model. - Detecting Hedge Cues and their Scopes with Average Perceptron (2010 · 7 citations)
The paper describes a Fudan University system for the closed track of the CoNLL-2010 Shared Task on hedging, meaning language that marks a claim as speculative rather than established. - XY-Tokenizer: Mitigating the Semantic-Acoustic Conflict in Low-Bitrate Speech Codecs (2026 · 4 citations)
Speech LLMs need discrete tokens that carry both linguistic content and acoustic detail. At low bitrates, codecs typically trade one for the other. This paper introduces XY-Tokenizer, a codec of roughly 1 kbps designed to reduce that tension. - ARISE: An Adaptive Resolution-Aware Metric for Test-Time Scaling Evaluation in Large Reasoning Models (2026 · 1 citation)
Reasoning models can spend more inference compute ("test-time scaling"), but there has been no dependable way to compare how well different models turn extra tokens into better answers. - How to Set the Learning Rate for Large-Scale Pre-training? (2026 · 1 citation)
Choosing the learning rate (LR) for large-scale pre-training is expensive to get wrong and too costly to tune by trial at full scale. - Efficient KL Divergence Estimation via Truncated Top-K Integration for Large Language Models (2026)
RL fine-tuning of LLMs (e.g., GRPO) usually includes a KL penalty that keeps the policy close to a reference model.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/xipeng-qiu/page/2/