Papers by Chenglei Si
Ignore This Title and HackAPrompt: Exposing Systemic Vulnerabilities of LLMs Through a Global Prompt Hacking Competition (2023.emnlp-main)
Copied to clipboard
Sander Schulhoff, Jeremy Pinto, Anaum Khan, Louis-François Bouchard, Chenglei Si, Svetlina Anati, Valen Tagliabue, Anson Kost, Christopher Carnahan, Jordan Boyd-Graber
| Challenge: | Large Language Models are increasingly being deployed in interactive contexts that involve direct user engagement. |
| Approach: | They run a global prompt hacking competition to encourage research on prompt hacks . they elicit 600K+ adversarial prompts against three state-of-the-art LLMs based on a dataset . |
| Outcome: | The results of the competition show that current LLMs can be manipulated via prompt hacking . the competition elicits 600K+ adversarial prompts against three state-of-the-art LLM models . |
Sentiment Aware Neural Machine Translation (D19-52)
Copied to clipboard
| Challenge: | Sentiment ambiguous lexicons are used when context is absent in translations . most systems aim to produce one correct translation for a given source sentence . |
| Approach: | They propose a neural machine translation method that preserves sentiment in two sentiment scenarios and a method that embeds sentiment into a sentence. |
| Outcome: | The proposed method outperforms a baseline with sentiment-aware translations in both the BLEU score and translation accuracy. |
Adversarial Training for Machine Reading Comprehension with Virtual Embeddings (2021.starsem-1)
Copied to clipboard
| Challenge: | Neural networks are vulnerable to adversarial examples that have been mixed with certain perturbations. |
| Approach: | They propose a novel adversarial training method that perturbs the embedding matrix instead of word vectors to differentiate the roles of passages and questions. |
| Outcome: | The proposed method is effective universally and further improves the performance of MRC tasks. |
Sub-Character Tokenization for Chinese Pretrained Language Models (2023.tacl-1)
Copied to clipboard
Chenglei Si, Zhengyan Zhang, Yingfa Chen, Fanchao Qi, Xiaozhi Wang, Zhiyuan Liu, Yasheng Wang, Qun Liu, Maosong Sun
| Challenge: | Existing tokenization methods for Chinese PLMs treat each character as an indivisible token, but ignore the unique feature of the writing system where additional linguistic information exists below the character level. |
| Approach: | They propose to encode Chinese characters into short sequences and construct Chinese vocabulary based on the encoded text. |
| Outcome: | The proposed tokenizers can tokenize inputs into much shorter sequences, improving computational efficiency. |
CharBERT: Character-aware Pre-trained Language Model (2020.coling-main)
Copied to clipboard
| Challenge: | Pre-trained language models (PLMs) construct word representations at subword level with Byte-Pair Encoding (BPE) or its variations . but these methods split a word into subword units and make it incomplete and fragile . |
| Approach: | They propose a character-aware pre-trained language model to tackle OOV problems . they construct contextual word embedding for each token from sequential character representations . |
| Outcome: | The proposed model improves on the existing models on multiple NLP benchmarks. |
READIN: A Chinese Multi-Task Benchmark with Realistic and Diverse Input Noises (2023.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks for Chinese inputs often lack a realistic representation of real-world noises. |
| Approach: | They construct a Chinese multi-task benchmark with REalistic and Diverse input noises . they use pinyin input and speech input to recruit speakers from diverse dialects based on their inputs - a feature that is important for Chinese NLP benchmarks if it is implemented in real-world applications. |
| Outcome: | The proposed benchmarks are based on four different tasks and are designed to maximize diversity. |
Measuring Inductive Biases of In-Context Learning with Underspecified Demonstrations (2023.acl-long)
Copied to clipboard
| Challenge: | In-context learning is an important paradigm for adapting large language models to new tasks . but the generalization behavior of ICL remains poorly understood . |
| Approach: | They characterize the feature biases of large language models by constructing underspecified demonstrations . they find that LLMs exhibit clear feature bias, and they evaluate interventions . |
| Outcome: | The proposed model prefers the "default" task features over distractor features more often than the base model. |
What’s in a Name? Answer Equivalence For Open-Domain Question Answering (2021.emnlp-main)
Copied to clipboard
| Challenge: | A flaw in QA evaluation is that annotations often only provide one answer . therefore, model predictions semantically equivalent to the answer but superficially different are considered incorrect. |
| Approach: | They explore using alias entities from knowledge bases to extract additional answers . they incorporate additional answers for evaluation and model training with equivalent answers based on the results . |
| Outcome: | The proposed solution improves the accuracy of evaluation with additional answers and improves model training with equivalent answers. |
SPHERE: An Evaluation Card for Human-AI Systems (2025.findings-acl)
Copied to clipboard
Dora Zhao, Qianou Ma, Xinran Zhao, Chenglei Si, Chenyang Yang, Ryan Louie, Ehud Reiter, Diyi Yang, Tongshuang Wu
| Challenge: | Existing evaluation methods and standards for human-AI systems are unclear, especially for large language models. |
| Approach: | They propose an evaluation card SPHERE which provides a template for evaluation protocols . they outline current evaluation practices and areas for improvement . |
| Outcome: | The evaluation card provides a template for designing evaluation protocols . it outlines current evaluation practices and areas for improvement . |
Large Language Models Help Humans Verify Truthfulness – Except When They Are Convincingly Wrong (2024.naacl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly used for accessing information on the web. |
| Approach: | They conduct experiments with 80 crowdworkers to compare LLMs with search engines . they ask LLM to provide contrastive information to reduce over-reliance on LLM . |
| Outcome: | The results show that LLMs can outperform search engines but not LLM explanations . the study shows that LMS explanations are not reliable replacements for reading retrieved passages compared to search engines alone. |
Getting MoRE out of Mixture of Language Model Reasoning Experts (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing large language models (LLMs) have poor generalizability on question types beyond those seen in the prompt. |
| Approach: | They propose a framework that integrates specialized language models to generalize across question types that require distinct reasoning abilities. |
| Outcome: | The proposed framework gives higher accuracy than any single specialized model on a collection of 12 QA datasets from four reasoning types. |
Re-Examining Calibration: The Case of Question Answering (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing calibration methods do not provide significant gains in accuracy. |
| Approach: | They propose a new calibration metric that better captures whether the model assigns low confidence to wrong predictions and high confidence to correct predictions. |
| Outcome: | The proposed calibration method better captures whether the model assigns low confidence to wrong predictions and high confidence to correct predictions. |
Benchmarking Robustness of Machine Reading Comprehension Models (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks only evaluate models' robustness under test-time perturbations or adversarial attacks. |
| Approach: | They propose a model-agnostic benchmark to evaluate models' robustness under adversarial attacks. |
| Outcome: | The proposed model-agnostic benchmark evaluates models under four different types of adversarial attacks. |
Contextual Experience Replay for Self-Improvement of Language Agents (2025.acl-long)
Copied to clipboard
| Challenge: | Existing large language model (LLM) agents fail in complex tasks without any environment-specific experiences. |
| Approach: | They propose a framework that accumulates and synthesizes past experiences into a dynamic memory buffer to enable efficient self-improvement for language agents in their context window. |
| Outcome: | The proposed framework improves performance on WebArena and VisualWebAren . it surpasses tree search method with fewer token costs and achieves state-of-the-art performance of 31.9%. |
Better Robustness by More Coverage: Adversarial and Mixup Data Augmentation for Robust Finetuning (2021.findings-acl)
Copied to clipboard
| Challenge: | Pretrained language models perform poorly under adversarial attacks due to the large search space. |
| Approach: | They propose a method to cover a much larger proportion of the attack search space by adding textual adversarial examples during training. |
| Outcome: | The proposed method covers a much larger proportion of the attack search space. |
Design2Code: Benchmarking Multimodal Code Generation for Automated Front-End Engineering (2025.naacl-long)
Copied to clipboard
| Challenge: | Generative AI has made rapid advances in multimodal understanding and code generation. |
| Approach: | They construct a first real-world benchmark for multimodal large language models that directly convert visual designs into code implementations by manually curating 484 diverse real-life webpages as test cases. |
| Outcome: | The proposed model can generate code implementations that directly render into the given reference webpages, given the screenshots as input. |
Scaling Collaborative Effort with Agents (2026.findings-acl)
Copied to clipboard
Shannon Zejiang Shen, Valerie Chen, Ken Gu, Alexis Ross, Zixian Ma, Jillian Ross, Alex Gu, Chenglei Si, Wayne Chi, Andi Peng, Jocelyn J Shen, Ameet Talwalkar, Tongshuang Wu, David Sontag
| Challenge: | Current evaluations of agents focus on producing high-quality, final outputs in one shot, failing to account for the inherently iterative nature of many real-world problems. |
| Approach: | They propose a framework that captures how an agent’s utility grows with increasing user involvement. |
| Outcome: | The proposed framework captures how an agent’s utility grows with increasing user involvement, revealing a missing ingredient in agent design: the ability to sustain engagement and scaffold user understanding. |