Papers by Shiyang Li

13 papers
Graph Reasoning for Question Answering with Triplet Retrieval (2023.findings-acl)

Copied to clipboard

Challenge: Existing methods to answer complex questions require reasoning over knowledge graphs (KGs) state-of-the-art methods constrain retrieved knowledge in local subgraphs and discard more diverse triplets that are disconnected but useful for question answering.
Approach: They propose a method to retrieve the most relevant triplets from KGs and then rerank them, which are then concatenated with questions to be fed into language models.
Outcome: The proposed method outperforms state-of-the-art methods on commonsenseQA and OpenbookQA datasets with 4.6% absolute accuracy.
Instruction-following Evaluation through Verbalizer Manipulation (2024.findings-naacl)

Copied to clipboard

Challenge: Existing benchmarks focus on common instructions that align well with what the model learned during training, but proficiency in responding to these instructions does not necessarily imply strong ability in instruction following.
Approach: They propose a new instruction-following evaluation protocol called verbalizer manipulation that instructs the model to verbalize the task label with words aligning with model priors to different extents.
Outcome: The proposed protocol can be integrated with any classification benchmark to examine the model’s reliance on priors and its ability to override them to accurately follow the instructions.
DrAgent: Empowering Large Language Models as Medical Agents for Multi-hop Medical Reasoning (2025.findings-emnlp)

Copied to clipboard

Challenge: commercial LLMs can be difficult to use in real-world clinical decision-making . a lightweight LLM can be used to collaborate with diverse clinical tools .
Approach: They propose a lightweight LLM that can be used to build medical LLMs as agents . they use recursive curriculum learning to optimize the LLM in an easy-to-hard progression .
Outcome: The proposed approach outperforms human experts in medical examinations on diverse datasets.
IHEval: Evaluating Language Models on Following the Instruction Hierarchy (2025.naacl-long)

Copied to clipboard

Challenge: Instruction-tuned language models (LMs) are increasingly deployed as interactive services across various applications.
Approach: They propose a benchmark to evaluate models' ability to follow the instruction hierarchy by comparing their models to a set of benchmarks.
Outcome: The proposed benchmark covers 3,538 examples across nine tasks covering cases where instructions in different priorities either align or conflict.
Limitations of Language Models in Arithmetic and Symbolic Induction (2023.acl-long)

Copied to clipboard

Challenge: Recent work has shown that large pretrained Language Models (LMs) can perform remarkably well on a range of NLP tasks but they have limitations on basic symbolic manipulation tasks such as copy, reverse, and addition.
Approach: They propose to use explicit positional markers, fine-grained computation steps, and LMs with callable programs to teach large pretrained Language Models.
Outcome: The proposed model can perform 100% accuracy in OOD and repeating symbols.
Backdooring Instruction-Tuned Large Language Models with Virtual Prompt Injection (2024.naacl-long)

Copied to clipboard

Challenge: Instruction-tuned Large Language Models (LLMs) can modulate responses based on human instructions, but they can be maliciously steered to impact society in subtle but persistent ways.
Approach: They propose a backdoor attack setting that allows an attacker to inject a virtual prompt into an LLM to steer it without any explicit injection at its input.
Outcome: The proposed method is able to poison the model's instruction tuning data and show that it is highly effective in steering the model.
Task-adaptive Pre-training and Self-training are Complementary for Natural Language Understanding (2021.findings-emnlp)

Copied to clipboard

Challenge: Task-adaptive pre-training (TAPT) and Self-training can be complementary with simple TFS protocol.
Approach: They propose to use task-adaptive pre-training and self-training to combine TAPT and ST with a simple TFS protocol to achieve strong combined gains across six datasets.
Outcome: The proposed method can achieve strong combined gains across six datasets covering sentiment classification, paraphrase identification, natural language inference, named entity recognition and dialogue slot classification.
Can Language Models Follow Multiple Turns of Entangled Instructions? (2025.findings-emnlp)

Copied to clipboard

Challenge: Despite of significant achievements in improving instruction-following capabilities of large language models, the ability to process multiple potentially entangled or conflicting instructions remains a considerable challenge.
Approach: They construct multi-turn instruction with 1.1K high-quality multi-turned conversations using the human-in-the-loop approach and examine their capabilities.
Outcome: The proposed model shows that it is difficult to integrate multiple turns and balance competing objectives when instructions intersect or conflict.
Train a Unified Multimodal Data Quality Classifier with Synthetic Data (2025.findings-emnlp)

Copied to clipboard

Challenge: Multimodal Large Language Models are pre-trained on image-text caption data and interleaved document data.
Approach: They propose to train an efficient MLLM as a Unified Mulitmodal Data Quality Classifier to filter image-text caption and interleaved data.
Outcome: The proposed method enables efficient creation of sample-score pairs for caption and interleaved data to train UniFilter.
Data Diversity Matters for Robust Instruction Tuning (2024.findings-emnlp)

Copied to clipboard

Challenge: Recent studies have shown that by curating high quality and diverse instruction tuning datasets, we can significantly improve instruction-following capabilities.
Approach: They propose an algorithm to control diversity and quality of instruction tuning datasets and validate it.
Outcome: The proposed algorithm significantly improves worst and average case performance on large scale instruction tuning datasets.
ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering (2022.emnlp-main)

Copied to clipboard

Challenge: Recent advances in large pre-trained language models have brought the NLP field into a new era.
Approach: They propose a large-scale dataset to study the chain of numerical reasoning in conversational question answering.
Outcome: The proposed dataset should push forward the exploration of real-world, complex reasoning tasks as the next research focus.
Controllable Dialogue Simulation with In-context Learning (2022.findings-emnlp)

Copied to clipboard

Challenge: Existing methods to generate annotated dialogues require crowdsourcing, which is expensive and time-consuming.
Approach: They propose a dialogue simulation method based on large language model in-context learning that generates new dialogues and annotations in a controllable way.
Outcome: The proposed method can expand a small set of dialogue data with minimum or zero human involvement and parameter update.
Sparse Progressive Distillation: Resolving Overfitting under Pretrain-and-Finetune Paradigm (2022.acl-long)

Copied to clipboard

Challenge: Conventional wisdom in pruning Transformer-based language models is that it reduces model expressiveness, but new research shows pruning increases risk of overfitting when performed at the fine-tuning phase.
Approach: They propose to reduce pruning risk under pretrain-and-finetune paradigm . they propose to use knowledge distillation to improve pruning performance .
Outcome: The proposed method outperforms the leading competitors on the GLUE benchmark.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations