Papers by Linxin Song
Better Explain Transformers by Illuminating Important Information (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing explanations focus on the input and output of the Transformers, resulting in confusing results. |
| Approach: | They propose to highlight important information and eliminate irrelevant information by a refined information flow on top of the layer-wise relevance propagation method. |
| Outcome: | The proposed method outperforms baseline models on classification and question-answering datasets with over 3% to 33% improvement on explanation metrics. |
Treble Counterfactual VLMs: A Causal Approach to Hallucination (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies link hallucination to data or representation biases, but their causal origins remain unclear. |
| Approach: | They propose a causal framework to analyze and mitigate hallucination in vision-language models by using counterfactual analysis to estimate the Natural Direct Effect (NDE) of each modality and their interaction. |
| Outcome: | The proposed framework significantly reduces hallucination while preserving task performance while retaining reliability. |
The Hallucination Tax of Reinforcement Finetuning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Reinforcement finetuning (RFT) has become a standard approach for enhancing the reasoning capabilities of large language models (LLMs). |
| Approach: | They propose to incorporate 10% SUM into RFT to restore appropriate refusal behavior with minimal accuracy trade-offs on solvable tasks. |
| Outcome: | The proposed approach reduces model refusal rates by more than 80%, which significantly increases model’s tendency to hallucinate. |
Adaptive Ranking-based Sample Selection for Weakly Supervised Class-imbalanced Text Classification (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to synthesize training labels with labeling rules ignore data imbalance issue . weak supervision paradigm is often used to reduce human efforts to produce training labels inexpensively. |
| Approach: | They propose a model-agnostic framework to alleviate the data imbalance issue in the weak supervision paradigm by combining labeling rules with a probabilistic margin score. |
| Outcome: | The proposed framework outperforms the state-of-the-art imbalanced learning and WS methods on four text classification datasets with four different imbalance ratios. |
Explaining Length Bias in LLM-Based Preference Evaluations (2025.findings-emnlp)
Copied to clipboard
Zhengyu Hu, Linxin Song, Jieyu Zhang, Zheyuan Xiao, Tianfu Wang, Zhengyu Chen, Nicholas Jing Yuan, Jianxun Lian, Kaize Ding, Hui Xiong
| Challenge: | a preference evaluation metric is often biased towards longer responses, revealing a reliability problem . a decomposition of the preference evaluation into two components is needed to understand this bias. |
| Approach: | They propose to decompose the preference evaluation metric into two key components . the first component is length-dependent and related to trustworthiness . |
| Outcome: | The proposed evaluation metric is based on two components: desirability and information mass. |