Xuanjing Huang · page 2
Papers
- Benchmark Self-Evolving: A Multi-Agent Framework for Dynamic LLM Evaluation (2025 · 30 citations)
Static benchmarks age quickly as LLMs improve. Their test sets may also leak into training data, which inflates scores. Rather than annotating new datasets from scratch, the authors update existing ones automatically. Their framework has three goals: - Revisiting Jailbreaking for Large Language Models: A Representation Engineering Perspective (2025 · 17 citations)
The paper asks why LLMs that have been through safety alignment can still be jailbroken. It approaches the question through representation engineering. The authors argue that refusals of harmful requests come from particular activation signatures inside the model, which they call "safety patterns". To find them, they… - Measuring Data Diversity for Instruction Tuning: A Systematic Analysis and A Reliable Metric (2025 · 11 citations)
Diversity is widely seen as a key property of instruction-tuning (IT) data, yet there is no agreed way to quantify it. - 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. - Feedback-Driven Tool-Use Improvements in Large Language Models via Automated Build Environments (2026 · 6 citations)
Training LLMs to call tools with reinforcement learning is difficult for two reasons. Online APIs make training environments unstable and costly to maintain, and judging trajectories usually requires a strong LLM as evaluator, which brings bias. The authors propose FTRL, which removes both problems. An automated… - VIB-Probe: Detecting and Mitigating Hallucinations in Vision-Language Models via Variational Information Bottleneck (2026 · 3 citations)
Vision-Language Models often describe things that are not actually in the image. Most existing detectors look only at output confidence, such as logit entropy, or rely on external verifiers. VIB-Probe looks inside the model instead. During decoding, it collects the raw per-head attention outputs from every… - 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. - Mitigating Hallucinations in VLMs: Enhancing Visual Attention via Head-Wise Perturbation (2026)
Vision-language models often hallucinate because they lean on the language priors of their LLM backbone instead of the image. - Mitigating Position Bias in Transformers via Layer-Specific Positional Embedding Scaling (2026)
Long-context LLMs that use RoPE tend to concentrate attention on the start and end of the input. As a result, they underuse content in the middle, a problem known as "lost-in-the-middle." Earlier fixes such as Attention Buckets, MoICE and Ms-PoE mix several RoPE bases or scales. However, they need repeated forward…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/xuanjing-huang/page/2/