Papers by Yixiao Song
Does quantization affect models’ performance on long-context tasks? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models support context windows exceeding 128K tokens, but this comes with significant memory requirements and high inference latency. |
| Approach: | They present the first systematic evaluation of quantized LLMs on tasks with long inputs and long-form outputs. |
| Outcome: | The proposed method preserves accuracy, while 4-bit methods lead to substantial losses . the results highlight the importance of a careful evaluation before deploying quantized LLMs . |
SLING: Sino Linguistic Evaluation of Large Language Models (2022.emnlp-main)
Copied to clipboard
| Challenge: | Using pre-trained language models, we find that the accuracy of LMs is far below human performance. |
| Approach: | They propose a benchmark of Sino LINGuistics which consists of 38K sentence pairs in Mandarin Chinese grouped into 9 high-level linguistic phenomena. |
| Outcome: | The proposed model performs better on local phenomena than hierarchical models and has a strong gender and number bias. |
Localizing and Mitigating Errors in Long-form Question Answering (2025.findings-acl)
Copied to clipboard
| Challenge: | Long-form question answering (LFQA) answers are prone to hallucinations and factual inconsistencies, challenging their faithful evaluation. |
| Approach: | They propose a dataset with localized error annotations for human-written and model-generated LFQA answers. |
| Outcome: | The proposed approach reduces errors and improves quality of the answers across multiple models. |
kNN-LM Does Not Improve Open-ended Text Generation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Interpolation-based retrieval-augmented language models (LMs) are a subtype of retrieval augmented language model that computes the probability of the next token by interpolating between the softmax distribution of the original LM and a token distribution formed by retrieving over an external datastore. |
| Approach: | They propose to interpolate the predicted distribution of the next word with a distribution formed from the most relevant retrievals for a given prefix. |
| Outcome: | The proposed methods do not exhibit improvements in open-ended generation quality, as measured by automatic evaluation metrics and human evaluations. |
Enhancing Human Evaluation in Machine Translation with Comparative Judgement (2025.acl-long)
Copied to clipboard
| Challenge: | Human evaluation is crucial for assessing rapidly evolving language models but is influenced by annotator proficiency and task design. |
| Approach: | They evaluate three annotation setups to integrate comparative judgment into human annotation for machine translation. |
| Outcome: | The proposed approach improves inter-annotator agreement and stability of the annotations. |
VeriScore: Evaluating the factuality of verifiable claims in long-form text generation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing metrics for evaluating the factuality of long-form text assume that every claim is verifiable. |
| Approach: | They propose a metric to evaluate factuality in diverse long-form generation tasks . they use open-weight language models to extract verifiable and unverifiably content . |
| Outcome: | The proposed metric can be implemented with either closed or fine-tuned open-weight language models. |
DEMETR: Diagnosing Evaluation Metrics for Translation (2022.emnlp-main)
Copied to clipboard
| Challenge: | BLEU scores are based on string overlap, but they are opaque in comparison to newer learned metrics. |
| Approach: | They propose a dataset to evaluate MT evaluation metrics based on linguistic perturbations in English . they find learned metrics perform substantially better than string-based metrics . |
| Outcome: | The proposed dataset shows that learned metrics perform better than string-based metrics . the dataset contains 31K English examples that cover 35 different linguistic phenomena . |
GEE! Grammar Error Explanation with Large Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing grammatical error correction tools do not provide natural language explanations of errors . a system needs to provide one-sentence explanations for each grammamatical errors in a pair of erroneous and corrected sentences. |
| Approach: | They propose a grammar error explanation task that uses one-sentence explanations for each grammatical error in a pair of erroneous and corrected sentences. |
| Outcome: | The proposed pipeline identifies grammar errors in German, Chinese, and English . human evaluation reveals that 93.9% of German errors, 96.4% of Chinese errors, and 92.20% of English errors are correctly detected and explained. |
A Critical Evaluation of Evaluations for Long-form Question Answering (2023.acl-long)
Copied to clipboard
| Challenge: | Long-form question answering (LFQA) is an emerging research area within QA . however, its flexibility poses enormous challenges for evaluation . |
| Approach: | They conduct the first targeted study of the evaluation of long-form answers, covering both human and automatic evaluation practices. |
| Outcome: | The proposed evaluations cover human and automatic evaluations. |