Papers by Jason Li

23 papers
A Measure-Theoretic Characterization of Tight Language Models (2023.acl-long)

Copied to clipboard

Challenge: Language modeling is a core task in natural language processing.
Approach: They propose to characterize leakage onto the set of infinite sequences by a measure-theoretic approach.
Outcome: The proposed language model families are tight, meaning they will not leak . the proposed language models are based on the 'sequence leakage' hypothesis .
Chain-of-Verification Reduces Hallucination in Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: Large language models can generate plausible but incorrect factual information, termed hallucinations, but they can still fail on lesser known facts.
Approach: They develop a method that allows language models to deliberate on the responses they give in order to correct their errors.
Outcome: The proposed method decreases hallucinations across a variety of tasks, including list-based questions, closed book MultiSpanQA and longform text generation.
When is a Language Process a Language Model? (2024.findings-acl)

Copied to clipboard

Challenge: In some pathological situations, such a stochastic process may "leak" probability mass onto the set of infinite strings.
Approach: They propose to view a language model as a discrete stochastic process X t : t = = t + .
Outcome: The proposed conditions of tightness are generalized to language models and the literature.
LLMs Know More About Numbers than They Can Say (2026.eacl-short)

Copied to clipboard

Challenge: Large language models (LLMs) are increasingly used in mathematical, scientific, financial and engineering domains.
Approach: They probe the hidden states of several smaller open-source LLMs to find out how big they are .
Outcome: The proposed model improves verbalized accuracy by 3.22% over base models.
How to Motivate Your Dragon: Teaching Goal-Driven Agents to Speak and Act in Fantasy Worlds (2021.naacl-main)

Copied to clipboard

Challenge: a recent improvement in the quality of natural language processing and generation (NLG) is needed for goal-oriented ML driven agents.
Approach: They propose a reinforcement learning system that integrates large-scale language modeling and commonsense reasoning-based pre-training to imbue the agent with relevant priors.
Outcome: The proposed system is able to act and talk naturally with respect to their motivations.
Limitations of Autoregressive Models and Their Alternatives (2021.naacl-main)

Copied to clipboard

Challenge: Standard autoregressive language models only perform polynomial-time computation to compute probability of next symbol.
Approach: authors propose alternative to standard autoregressive language models that use polynomial-time computation to compute probability of next symbol.
Outcome: a large model size can grow superpolynomially in length, allowing it to store precomputed results and verify solutions.
Contrastive Decoding: Open-ended Text Generation as Optimization (2023.acl-long)

Copied to clipboard

Challenge: Using a language model, maximum probability is a poor decoding objective because it produces short and repetitive text.
Approach: They propose a reliable decoding approach that optimizes a contrastive objective subject to a plausibility constraint.
Outcome: The proposed approach outperforms four strong decoding algorithms in automatic and human evaluations across wikipedia, news and story domains.
Don’t Say That! Making Inconsistent Dialogue Unlikely with Unlikelihood Training (2020.acl-main)

Copied to clipboard

Challenge: Unlikelihood is a technique developed for removal of repetition in language model completions . it allows for a model to be generalized to solve a number of problems .
Approach: They extend the unlikelihood objective to generate generations that contain repetitions . they show that such an objective can be used to improve logical consistency .
Outcome: The proposed approach can be applied to a number of dialogue tasks.
Automated Generation of Accurate & Fluent Medical X-ray Reports (2021.emnlp-main)

Copied to clipboard

Challenge: Existing medical report generation efforts focus on producing human-readable reports, yet the generated text may not be well aligned to the clinical facts.
Approach: They propose to automate the generation of medical reports from chest X-ray image inputs . medical reports are the primary medium, which physicians communicate findings from scans - authors say .
Outcome: The proposed method achieves fluency and clinical accuracy on common metrics.
CheckMIABench: Firm Foundations For Membership Inference Attacks on Language Models (2026.acl-short)

Copied to clipboard

Challenge: Membership inference attacks are a canonical way to assess a machine learning model’s privacy properties.
Approach: They propose a framework for principled evaluation of membership inference attacks against large language models by leveraging the insight that training data before and after a fixed point during training are drawn from the same distribution.
Outcome: The proposed framework can be used to evaluate membership inference attacks against large language models.
Koel-TTS: Enhancing LLM based Speech Generation with Preference Alignment and Classifier Free Guidance (2025.emnlp-main)

Copied to clipboard

Challenge: Autoregressive speech token generation models suffer from hallucinations and undesired vocalizations that do not conform to conditioning inputs.
Approach: They propose an encoder-decoder transformer model that improves contextual adherence of speech token generation LLMs through preference alignment and classifier-free guidance.
Outcome: The proposed model outperforms previous LLM-based models on intelligibility, speaker similarity and naturalness.
Specializing Word Embeddings (for Parsing) by Information Bottleneck (D19-1)

Copied to clipboard

Challenge: Pre-trained word embeddings contain rich syntactic and semantic information.
Approach: They propose a variational information bottleneck method to nonlinearly compress word embeddings . they compress each word embedded to either a discrete tag or a continuous vector .
Outcome: The proposed method can be extended to other syntactic or semantic tasks.
How Much Do Large Language Models Know about Human Motion? A Case Study in 3D Avatar Control (2025.findings-emnlp)

Copied to clipboard

Challenge: a new study explores the human motion knowledge of Large Language Models (LLMs) using 3D avatar control.
Approach: They use 20 representative motion instructions to interpolate LLMs into avatar animations . they find they are strong at interpreting high-level body movements but struggle with precise body part positioning .
Outcome: The proposed model is strong at interpreting high-level body movements but struggles with precise body part positioning.
LSSF: Safety Alignment for Large Language Models through Low-Rank Safety Subspace Fusion (2025.acl-long)

Copied to clipboard

Challenge: Existing safety alignment methods rely on fine-tuning, which inadvertently leads to the increased complexity and computational resources required.
Approach: They propose a safety re-alignment framework with Low-Rank Safety Subspace Fusison that exploits low-rank safety characteristics of LLMs by constructing a low-ranked projection matrix to extract the principal components of safety vectors.
Outcome: The proposed method exploits low-rank safety subspace of the LLMs and is stable during fine-tuning process and is isolated from the model’s general capabilities.
Branch-Solve-Merge Improves Large Language Model Evaluation and Generation (2024.naacl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) are often used for multi-faceted language generation and evaluation tasks that require complex user constraints or taking into account multiple aspects and criteria.
Approach: They propose a Large Language Model program that consists of branch, solve, and merge modules that are parameterized with specific prompts to the base LLM.
Outcome: The proposed program improves the evaluation correctness and consistency for each LLM by up to 26%, reducing length and pairwise position biases by up 50%, and allowing LLaMA-2-chat to match or outperform GPT-4 on most domains.
Better Alignment with Instruction Back-and-Forth Translation (2024.findings-emnlp)

Copied to clipboard

Challenge: et al., 2023) proposes a method to improve instruction-tuning data . e.g., we generate synthetic instructions using the backtranslation approach .
Approach: They propose a method to improve instruction-tuning data using web-based inputs . they generate synthetic instructions using the backtranslation approach and filter the generated data .
Outcome: The proposed method improves the quality of instruction-tuning data based on preprocessed texts . it yields better AlpacaEval win rates than direct distillation .
Toward Interactive Dictation (2023.acl-long)

Copied to clipboard

Challenge: Existing systems that allow both dictation and editing-by-voice restrict their command language to flat templates invoked by trigger words.
Approach: They propose to allow users to interrupt dictation with spoken editing commands in open-ended natural language.
Outcome: The proposed system can predict edited text with large pre-trained models and predict small programs.
Bot-Adversarial Dialogue for Safe Conversational Agents (2021.naacl-main)

Copied to clipboard

Challenge: a new method for evaluating chatbot safety is proposed to mimic human-generated data . a bot-adversarial dialogue model learns undesirable features from this data, a study finds .
Approach: They propose a human-and-model-in-the-loop framework for evaluating toxicity of chatbots . they propose two methods for safe conversational agents by either training on data or ”baking-in” safety to the generative model itself.
Outcome: The proposed methods are safer than existing models while maintaining usability metrics, the authors say . they show that the proposed methods can be used to make safer models with human-model interactions .
MoPe: Model Perturbation based Privacy Attacks on Language Models (2023.emnlp-main)

Copied to clipboard

Challenge: Recent work shows that Large Language Models can unintentionally leak sensitive information . a new method to identify with high confidence if a given text is in training data is proposed .
Approach: They propose a method to detect if a given text is in a pre-trained language model . they show that MoPe is more effective than existing loss-based attacks .
Outcome: The proposed method is more effective than loss-based attacks and perturbation-based methods.
Zero-Shot End-to-End Spoken Language Understanding via Cross-Modal Selective Self-Training (2024.eacl-long)

Copied to clipboard

Challenge: End-to-end (E2E) spoken language understanding models are constrained by the cost of collecting speech-semantics pairs.
Approach: They propose a model that learns E2E SLU without speech-semantics pairs . they propose cross-modal selective self-training (CMSST) to address imbalance and noise issues .
Outcome: The proposed model learns E2E SLU without speech-semantics pairs . the proposed model requires the domains of speech-text and text-sensitization to match .
A Critical Analysis of Document Out-of-Distribution Detection (2023.findings-emnlp)

Copied to clipboard

Challenge: Existing document understanding models focus on single-modal inputs such as images or texts.
Approach: They propose to use a spatial-aware adapter to adapt transformer-based language models to document domain to exploit multi-modal information.
Outcome: The proposed model significantly improves the OOD detection performance compared to using a standard language model and to competitive baselines.
Mapping Natural Language Instructions to Mobile UI Action Sequences (2020.acl-main)

Copied to clipboard

Challenge: a new problem of grounding natural language instructions to mobile UI actions is emerging . we use a Transformer to extract action phrase tuples from long-range natural language instruction .
Approach: They propose a dataset that pairs English instructions with actions performed by people on a mobile UI emulator.
Outcome: The proposed model achieves 70.59% accuracy on predicting complete ground-truth action sequences in PixelHelp.
Small and Practical BERT Models for Sequence Labeling (D19-1)

Copied to clipboard

Challenge: Existing models for morphosyntactic tagging have focused on building separate models for each language or for a small group of related languages.
Approach: They propose a scheme to train a single multilingual sequence labeling model that is small and fast enough to run on a CPU.
Outcome: The proposed model outperforms state-of-the-art models on low-resource languages and low-level models on codemixed inputs.

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