Ting Liu · page 2
Papers
- Deep Reinforcement Learning for Chinese Zero Pronoun Resolution (2018 · 67 citations)
Chinese often omits pronouns, leaving "zero pronouns" whose referents must be recovered from context. Earlier neural resolvers score each pair of a zero pronoun and one candidate antecedent in isolation. They therefore ignore what earlier decisions reveal about later ones. Yin and colleagues recast anaphoric zero… - Chinese Zero Pronoun Resolution with Deep Memory Network (2017 · 55 citations)
Chinese text often omits arguments, leaving "zero pronouns" (ZPs). These gaps carry no gender, number or lexical content, so earlier resolvers relied on syntactic and lexical features and could not measure how semantically close a gap is to a candidate antecedent. Yin et al. address the resolution step for anaphoric… - Zero Pronoun Resolution with Attention-based Neural Network (2018 · 44 citations)
The paper addresses anaphoric zero pronoun resolution in Chinese, a pro-drop language. A zero pronoun is an omitted argument with no surface form, so it must be represented through its context. Earlier neural approaches either used only a few nearby words or treated every context word as equally relevant. - Generating and Exploiting Large-scale Pseudo Training Data for Zero Pronoun Resolution (2017 · 37 citations)
Zero pronoun (ZP) resolution in Chinese means finding the antecedent of an omitted pronoun. Progress has been held back by reliance on small, hand-annotated corpora from shared tasks, which are expensive to extend or adapt to new domains. The authors treat the task as a form of cloze-style reading comprehension. From… - Through the Lens of Core Competency: Survey on Evaluation of Large Language Models (2023 · 35 citations)
Large language models now perform so well on classic NLP benchmarks (e.g., GLUE, SuperGLUE) that these tests no longer tell models apart. - LLMs May Perform MCQA by Selecting the Least Incorrect Option (2025 · 24 citations)
Benchmarks such as MMLU judge LLMs by multiple-choice accuracy. Earlier studies showed that answers can change when the options are reordered. This paper argues that a stable answer still does not prove the model knows the truth. The authors first keep only the MMLU and MedMCQA items that a model answers correctly… - UniCoRN: Unified Cognitive Signal ReconstructioN bridging cognitive signals and human language (2023 · 23 citations)
Most earlier brain-decoding work maps single word-level fMRI volumes to a small, fixed vocabulary. That is unrealistic, because fMRI samples the brain slowly and each frame overlaps several words. The authors define fMRI2text, a new open-vocabulary task: turn a sequence of fMRI frames, recorded while people read or… - Investigating the Portability of Corpus-Derived Cue Phrases for Dialogue Act Classification (2008 · 21 citations)
The paper asks whether dialogue act (DA) cue phrases that are learned automatically from one corpus still work when applied unchanged to a corpus from a different domain. - HIT: Web based Scoring Method for English Lexical Substitution (2007 · 20 citations)
The paper presents HIT, a system for the SemEval-2007 English Lexical Substitution task. Given a target word in a sentence, the system must propose words that could replace it in that context. HIT works in two stages. First, it gathers candidates from WordNet. It takes synonyms from every synset of the given POS… - HIT-IR-WSD: A WSD System for English Lexical Sample Task (2007 · 12 citations)
The paper describes HIT-IR-WSD, a supervised word sense disambiguation system from the Information Retrieval Lab at Harbin Institute of Technology, created for SemEval 2007's Task 11, which tested disambiguation of English words in a lexical sample setting. - AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender (2025 · 11 citations)
Safety-aligned LLMs can still be jailbroken. Activation steering is a training-free defense that pushes hidden states toward refusal, but it applies one fixed strength to every input. That fixed setting under-protects against some attacks and causes more false refusals of harmless requests. AdaSteer makes the… - Beware of Your Po! Measuring and Mitigating AI Safety Risks in Role-Play Fine-Tuning of LLMs (2025 · 9 citations)
The paper examines how fine-tuning LLMs for character role-play affects their safety. The authors fine-tuned LLaMA-3-8B-Instruct separately for each of 95 characters from RoleBench. Role-play quality improved, but the models refused harmful AdvBench requests much less often, and villainous characters lost the most… - Cross-Domain Dialogue Act Tagging (2008 · 9 citations)
The paper asks whether dialogue act (DA) cues learned automatically from one corpus still work on conversations from a different domain. - A Closer Look into the Robustness of Neural Dependency Parsers Using Better Adversarial Examples (2021 · 8 citations)
The paper attacks two English dependency parsers, a biaffine graph parser and a stack-pointer parser, by substituting words. - MPO: Multilingual Safety Alignment via Reward Gap Optimization (2025 · 7 citations)
Safety-tuned LLMs refuse harmful requests reliably in English but much less so in lower-resource languages. - MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization (2026 · 1 citation)
GRPO aligns LLMs well when answers can be checked automatically, as in math or code. Open-ended generation has no such check, and its goals often conflict, for example creativity versus factuality. Existing fixes either rely on costly LLM judges or combine cheap heuristic rewards (perplexity, entropy, format) using… - On Safety Risks in Experience-Driven Self-Evolving Agents (2026 · 1 citation)
Experience-driven self-evolving LLM agents improve by distilling lessons from past interactions into an external memory and retrieving them for new tasks, while model weights stay frozen. - Beyond Uniform SVD: Dual-Level Optimization across Columns and Modules for LLM Compression (2026)
SVD-based compression of large language models usually factorizes each weight matrix as a single unit. It also sets per-module ranks through proxy heuristics. The authors of Duo-SVD argue that both habits cost accuracy. After truncation, most reconstruction error sits in a small set of columns, and that set stays… - FAER: Benchmarking VLMs for Failure-Aware Embodied Reasoning (2026)
Embodied agents increasingly rely on vision-language models (VLMs) to perceive their surroundings and choose actions, so they must notice when something goes wrong. - Large Language Models Are Still Misled by Simple Bias Ensembles (2026)
Recent LLMs handle individual shortcut cues better than before. This paper argues that the real weakness appears when several such cues occur together in one input, which is common in real-world text. Most existing bias benchmarks inject only one bias per sample, and a few inject two. The authors therefore built… - MDC-Bench: A Multidisciplinary Causal Benchmark Based on Causal Structures for Evaluating Large Language Models (2026)
Most existing causal benchmarks for LLMs use everyday commonsense scenarios with short, single-hop cause–effect links.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/ting-liu/page/2/