Browsing by Author "Song, Wenjia"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Glitch Tokens in Large Language Models: Categorization Taxonomy and Effective Detection
    Li, Yuxi; Liu, Yi; Deng, Gelei; Zhang, Ying; Song, Wenjia; Shi, Ling; Wang, Kailong; Li, Yuekang; Liu, Yang; Wang, Haoyu (ACM, 2024-07-12)
    With the expanding application of Large Language Models (LLMs) in various domains, it becomes imperative to comprehensively investigate their unforeseen behaviors and consequent outcomes. In this study, we introduce and systematically explore the phenomenon of “glitch tokens”, which are anomalous tokens produced by established tokenizers and could potentially compromise the models’ quality of response. Specifically, we experiment on seven top popular LLMs utilizing three distinct tokenizers and involving a totally of 182,517 tokens. We present categorizations of the identified glitch tokens and symptoms exhibited by LLMs when interacting with glitch tokens. Based on our observation that glitch tokens tend to cluster in the embedding space, we propose GlitchHunter, a novel iterative clustering-based technique, for efficient glitch token detection. The evaluation shows that our approach notably outperforms three baseline methods on eight open-source LLMs. To the best of our knowledge, we present the first comprehensive study on glitch tokens. Our new detection further provides valuable insights into mitigating tokenization-related errors in LLMs.
  • Thumbnail Image
    Measurement of Embedding Choices on Cryptographic API Completion Tasks
    Xiao, Ya; Song, Wenjia; Ahmed, Salman; Ge, Xinyang; Viswanath, Bimal; Meng, Na; Yao, Danfeng (ACM, 2023-10)
    In this paper, we conduct a measurement study to comprehensively compare the accuracy impacts of multiple embedding options in cryptographic API completion tasks. Embedding is the process of automatically learning vector representations of program elements. Our measurement focuses on design choices of three important aspects, program analysis preprocessing, token-level embedding, and sequence-level embedding. Our findings show that program analysis is necessary even under advanced embedding. The results show 36.20% accuracy improvement on average when program analysis preprocessing is applied to transfer byte code sequences into API dependence paths. With program analysis and the token-level embedding training, the embedding dep2vec improves the task accuracy from 55.80% to 92.04%. Moreover, only a slight accuracy advantage (0.55% on average) is observed by training the expensive sequence-level embedding compared with the token-level embedding. Our experiments also suggest the differences made by the data. In the cross-app learning setup and a data scarcity scenario, sequence-level embedding is more necessary and results in a more obvious accuracy improvement (5.10%)
  • Thumbnail Image
    Subpopulation-specific machine learning prognosis for underrepresented patients with double prioritized bias correction
    Afrose, Sharmin; Song, Wenjia; Nemeroff, Charles B.; Yao, Danfeng (Nature Research, 2022-09)
    Background: Many clinical datasets are intrinsically imbalanced, dominated by overwhelming majority groups. Off-the-shelf machine learning models that optimize the prognosis of majority patient types (e.g., healthy class) may cause substantial errors on the minority prediction class (e.g., disease class) and demographic subgroups (e.g., Black or young patients). In the typical one-machine-learning-model-fits-all paradigm, racial and age disparities are likely to exist, but unreported. In addition, some widely used whole-population metrics give misleading results. Methods: We design a double prioritized (DP) bias correction technique to mitigate representational biases in machine learning-based prognosis. Our method trains customized machine learning models for specific ethnicity or age groups, a substantial departure from the one-model-predicts-all convention. We compare with other sampling and reweighting techniques in mortality and cancer survivability prediction tasks. Results: We first provide empirical evidence showing various prediction deficiencies in a typical machine learning setting without bias correction. For example, missed death cases are 3.14 times higher than missed survival cases for mortality prediction. Then, we show DP consistently boosts the minority class recall for underrepresented groups, by up to 38.0%. DP also reduces relative disparities across race and age groups, e.g., up to 88.0% better than the 8 existing sampling solutions in terms of the relative disparity of minority class recall. Crossrace and cross-age-group evaluation also suggests the need for subpopulation-specific machine learning models. Conclusions: Biases exist in the widely accepted one-machine-learning-model-fits-all-population approach. We invent a bias correction method that produces specialized machine learning prognostication models for underrepresented racial and age groups. This technique may reduce potentially life-threatening prediction mistakes for minority populations.