Papers by Xiaofei Ma

21 papers
SDC-LoRA: Singular-Subspace Drift Controlled LoRA to Mitigate Knowledge Forgetting (2026.findings-acl)

Copied to clipboard

Challenge: Existing approaches to adapt LLMs to new tasks focus on limiting knowledge forgetting . et al., 2023b) suggest a solution to this problem by limiting update energy in the principal singular subspace of W0 .
Approach: They propose a low-rank Adaptation (LoRA) that steers early updates away from principal directions and mitigates forgetting by constraining update energy in the principal singular subspace of W0.
Outcome: The proposed model mitigates forgetting on MMLU, TruthfulQA, and HellaSwag while keeping minor-subspace updates unchanged.
LibEvolutionEval: A Benchmark and Study for Version-Specific Code Generation (2025.naacl-long)

Copied to clipboard

Challenge: Recent code completion models focus on local file contexts, but do not fully capture the complexities of real-world software development.
Approach: They propose a version-specific code-completion task across eight libraries as they evolve over the years and an in-depth analysis of two widely used public libraries: PyTorch and Matplotlib.
Outcome: The proposed model improves performance with public libraries, compared with existing models.
Efficient Shapley Values Estimation by Amortization for Text Classification (2023.acl-long)

Copied to clipboard

Challenge: Shapley Values are often estimated with a small number of stochastic model evaluations, but this can only be mitigated by aggregating thousands of model evaluation.
Approach: They propose to combine a model with thousands of model evaluations to estimate Shapley Values without additional model evaluation.
Outcome: The proposed model estimates Shapley Values accurately with up to 60 times speedup compared to traditional methods and does not suffer from stability issues as inference is deterministic.
Learning Dialogue Representations from Consecutive Utterances (2022.naacl-main)

Copied to clipboard

Challenge: Dialogue Sentence Embedding (DSE) is a self-supervised contrastive learning method that learns effective dialogue representations suitable for a wide range of dialogue-oriented tasks.
Approach: They propose a self-supervised contrastive learning method that learns dialogue representations suitable for a wide range of dialogue tasks.
Outcome: The proposed method outperforms baselines on five dialogue tasks on a few-shot and zero-shot datasets.
Contrastive Document Representation Learning with Graph Attention Networks (2021.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for document representation learning are significantly affected by the scarcity of document-level data.
Approach: They propose to use a graph attention network on top of the available pretrained Transformers model to learn document embeddings.
Outcome: Empirically, the proposed approach is effective in document classification and document retrieval tasks.
Efficient Learned Data Compression via Dual-Stream Feature Decoupling (2026.acl-long)

Copied to clipboard

Challenge: Learned data compression has achieved superior compression ratios, but balancing precise probability modeling with system efficiency remains challenging.
Approach: They propose a Dual-Stream Multi-Scale Decoupler that disentangles local and global contexts to replace deep serial processing with shallow parallel streams.
Outcome: The proposed method achieves state-of-the-art performance in both compression ratio and throughput while maintaining the lowest latency and memory usage.
A Black-Box Attack on Code Models via Representation Nearest Neighbor Search (2023.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for generating adversarial code examples face challenges such as limted availability of substitute variables and the creation of adversarials with noticeable perturbations.
Approach: They propose a search seed based on historical attacks to find adversarial substitutes . they employ a pre-trained variable name encoder to map the search seed to a continuous vector space .
Outcome: The proposed approach outperforms baseline methods in terms of ASR and QT.
BASS: Batched Attention-optimized Speculative Sampling (2024.findings-acl)

Copied to clipboard

Challenge: Speculative decoding has emerged as a powerful method to improve latency and throughput in hosting large language models.
Approach: They propose a batched speculative decoding system that generates sequences at an average speed of 5.8ms per token and a batch size of 8 at a 2.15 speed-up over optimized regular decoding.
Outcome: The proposed system achieves state-of-the-art latency and speed-up over optimized regular decoding.
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner (2022.findings-naacl)

Copied to clipboard

Challenge: Large language models have achieved high performance on various natural language benchmarks, but the explainability of their output remains elusive.
Approach: They propose an architecture called iterative retrieval-generation reasoner that generates an entailment tree that explains a given hypothesis by using premises from C.
Outcome: The proposed model outperforms existing benchmarks on premise retrieval and entailment tree generation with around 300% gain in overall correctness.
Exploring Continual Learning for Code Generation Models (2023.acl-short)

Copied to clipboard

Challenge: Large-scale code generation models such as Copilot and CodeT5 are expensive to train and re-train.
Approach: They propose a benchmark for Continual Learning (CL) that covers a wide range of tasks with different input and output programming languages.
Outcome: The proposed method improves on Prompt Pooling with Teacher Forcing, which suffers catastrophic forgetting due to stark distribution shifts in coding tasks.
Multitask Pretraining with Structured Knowledge for Text-to-SQL Generation (2023.acl-long)

Copied to clipboard

Challenge: Existing methods for learning representations of structured knowledge are limited to the minority of people with technical skills.
Approach: They propose a large pretraining dataset and strategy for learning representations of text, tables, and SQL code that leverages the entire context of the problem.
Outcome: The proposed model improves on two SQL tasks and shows a 1.7 and 2.2 percentage point improvement over existing methods.
Contrastive Fine-tuning Improves Robustness for Neural Rankers (2021.findings-acl)

Copied to clipboard

Challenge: Current state-of-the-art neural rankers can deteriorate when exposed to noisy inputs or applied to a new domain.
Approach: They propose a contrastive loss and ranking loss method for fine-tuning rankers that combine ranking loss and rank loss to improve their robustness to query reformulations and noise perturbations.
Outcome: The proposed method outperforms data augmentation for robustifying rankers on four passage ranking datasets.
CodeFort: Robust Training for Code Generation Models (2024.findings-emnlp)

Copied to clipboard

Challenge: Existing research efforts to improve code generation models are inadequate . code generation model performance is degraded under small perturbations .
Approach: They propose a framework to improve the robustness of code generation models by generalizing code perturbations to enrich training data and enabling various robust training strategies.
Outcome: The proposed framework increases pass rates and robustness drop rate against code-syntax perturbations.
MMLU-ProX: A Multilingual Benchmark for Advanced Large Language Model Evaluation (2025.emnlp-main)

Copied to clipboard

Challenge: Existing large language model evaluation benchmarks focus on English, while current multilingual tasks lack parallel questions that specifically assess cross-lingual reasoning abilities.
Approach: They propose a comprehensive benchmark covering 29 languages, built on an English benchmark.
Outcome: The MMLU-ProX is a comprehensive benchmark covering 29 languages, built on an English benchmark.
Beyond [CLS] through Ranking by Generation (2020.emnlp-main)

Copied to clipboard

Challenge: Recent work on generative ranking models for Information Retrieval has focused on discriminative methods that learn a similarity function to compare questions and candidates answers.
Approach: They propose to use a language model to train a ranking function that model the semantic similarity of documents and queries instead of discriminative ranking functions.
Outcome: The proposed approaches are as effective as state-of-the-art discriminative models for the answer selection task and show unlikelihood losses are reduced for IR.
SWING: Balancing Coverage and Faithfulness for Dialogue Summarization (2023.findings-eacl)

Copied to clipboard

Challenge: Existing approaches to dialogue summarization rely on features of conversation data.
Approach: They propose to use natural language inference models to improve coverage and faithfulness . they use fine-grained training signals to encourage model to generate missing content .
Outcome: The proposed model achieves higher faithfulness and coverage while maintaining conciseness compared to prior methods.
Lightweight reranking for language model generations (2024.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) can exhibit considerable variation in quality of sampled outputs.
Approach: They propose a method for reranking LLM generations using pairwise statistics . they show strong improvements for selecting the best k generations for code generation tasks .
Outcome: The proposed approach improves selection and generation quality for code generation tasks and autoformalization, summarization, and translation tasks.
ContraCLM: Contrastive Learning For Causal Language Model (2023.acl-long)

Copied to clipboard

Challenge: Existing studies show that causal language models lack expressiveness due to poor discrimination ability.
Approach: They propose a contrastive learning framework that enhances discrimination of representations and bridges the gap with encoder-only models.
Outcome: The proposed framework improves discrimination and source code generation capabilities on a variety of downstream tasks.
Domain Adaptation with BERT-based Domain Classification and Data Selection (D19-61)

Copied to clipboard

Challenge: Modern deep neural models with millions of parameters can easily adapt to a new learning task and dataset when enough supervision is given.
Approach: They propose a domain adaptation framework based on curriculum learning and domain-discriminative data selection.
Outcome: The proposed framework outperforms discrepancy-based methods on transfer tasks while consuming only fraction of training budget.
Multi-passage BERT: A Globally Normalized BERT Model for Open-domain Question Answering (D19-1)

Copied to clipboard

Challenge: Existing studies have shown that BERT models can find answers from multiple passages . however, the results of these studies are still unaddressed.
Approach: They propose a multi-passage BERT model to globally normalize answer scores across all passages of the same question.
Outcome: The proposed model outperforms state-of-the-art models on four benchmarks.
Virtual Augmentation Supported Contrastive Learning of Sentence Representations (2022.findings-acl)

Copied to clipboard

Challenge: Despite profound successes, contrastive representation learning relies on carefully designed data augmentations using domain-specific knowledge.
Approach: They propose a virtual augmentation supported Contrastive Learning of sentence representations . they approximate the neighborhood of an instance via its K-nearest in-batch neighbors .
Outcome: The proposed model outperforms existing methods on a wide range of downstream tasks.

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