Jie Tang · page 2
Papers
- NaturalCodeBench: Examining Coding Performance Mismatch on HumanEval and Natural User Queries (2024 · 17 citations)
Popular code-generation benchmarks such as HumanEval, MBPP and DS-1000 mostly test short, well-specified algorithmic or data-science exercises. - Scaling Speech-Text Pre-training with Synthetic Interleaved Data (2024 · 14 citations)
Speech language models (SpeechLMs) have far less training material than text LLMs. FineWeb offers about 15T text tokens, while a large unsupervised speech set like VoxPopuli gives only around 36B tokens. Parallel speech-text data is also limited. The authors work around this by creating speech-text interleaved… - DataSciBench: An LLM Agent Benchmark for Data Science (2026 · 13 citations)
Most benchmarks for LLMs in data science test isolated tasks with simple metrics such as accuracy and with ground truth that is easy to obtain. - MathGLM-Vision: Solving Mathematical Problems with Multi-Modal Large Language Model (2024 · 10 citations)
Most math-focused multi-modal LLMs (MLLMs) are trained almost entirely on geometry, using small public sets such as GeoQA and Geometry3K whose answers rarely show the reasoning. - 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… - Beyond Literal Mapping: Benchmarking and Improving Non-Literal Translation Evaluation (2026 · 5 citations)
LLM-based machine translation is increasingly used for difficult content such as social media posts, culturally loaded idioms, poetry and literature. - Data Efficient RLVR via Off-Policy Influence Guidance (2026)
Choosing training data for reinforcement learning with verifiable rewards (RLVR) usually relies on heuristics such as difficulty or uncertainty. - 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… - MTP-RL: Acceleration of Reinforcement Learning Rollouts with Policy-Aligned Multi-Token Prediction (2026)
Rollout generation consumes most of the compute in reinforcement-learning fine-tuning of LLMs, especially when models produce long reasoning chains. - PlotGen-Bench: Evaluating VLMs on Generating Visualization Code from Diverse Plots across Multiple Libraries (2026)
Vision-language models can already read and reason about charts, but their ability to write runnable code that recreates a given plot has been tested only narrowly.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/jie-tang/page/2/