Papers by Yi Yao

32 papers
LinguaLens: Towards Interpreting Linguistic Mechanisms of Large Language Models via Sparse Auto-Encoder (2025.emnlp-main)

Copied to clipboard

Challenge: Prior research on linguistic mechanisms of large language models is limited by coarse granularity, limited analysis scale, and narrow focus.
Approach: They propose a framework for analyzing the linguistic mechanisms of large language models based on Sparse Auto-Encoders.
Outcome: The proposed framework extracts Chinese and English linguistic features across four dimensions . it uncovers intrinsic representations of linguistic knowledge in LLMs and can control outputs .
Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents (2026.acl-long)

Copied to clipboard

Challenge: Existing methods handle long-term memory (LTM) and short-term (STM) as separate components, relying on heuristics or auxiliary controllers, which limits adaptability and end-to-end optimization.
Approach: They propose a framework that integrates LTM and STM management directly into the agent's policy and propose 'agentic memory' to train such unified behaviors.
Outcome: The proposed framework outperforms strong memory-augmented baselines on five long-horizon benchmarks and achieves higher-quality long-term memory and more efficient context usage.
Can Persona-Prompted LLMs Emulate Subgroup Values? An Empirical Analysis of Generalisability and Fairness in Cultural Alignment (2026.acl-long)

Copied to clipboard

Challenge: Current alignment paradigms treat "human values" as a monolithic entity, ignoring the fact that many societies are a mosaic of diverse subgroups with distinct and sometimes conflicting values, preferences, and norms.
Approach: They examine whether Large Language Models can emulate distinct cultural values of subgroups . they use a global value survey to examine the value landscape of a multicultural society .
Outcome: The proposed model improves on unseen, out-of-distribution subgroups by 17.4% . the model widens the disparity between subgroup groups when measured by distance-aware metrics.
Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures (2026.acl-long)

Copied to clipboard

Challenge: Recent research has shown that reinforcement learning can elicit intriguing emergent reasoning behaviors.
Approach: They propose a comprehensive survey of the mechanistic understanding of large reasoning models . they organize findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors.
Outcome: This paper synthesizes the mechanistic understanding of large reasoning models into three dimensions . authors outline a roadmap for future studies including improved interpretability and methodologies .
Unintended Harms of Value-Aligned LLMs: Psychological and Empirical Insights (2025.acl-long)

Copied to clipboard

Challenge: Value-aligned LLMs are more prone to harmful behavior than fine-tuned models . value-aligned models generate text according to the aligned values, which can amplify harmful outcomes.
Approach: They propose to use in-context alignment methods to enhance the safety of value-aligned LLMs.
Outcome: The proposed methods improve value alignment and safety, the authors say . value-aligned models are more prone to harmful behavior than fine-tuned models .
Seeing the wood for the trees: a contrastive regularization method for the low-resource Knowledge Base Question Answering (2022.findings-naacl)

Copied to clipboard

Challenge: Existing methods for Knowledge Base Question Answering rely on semantic parsing and information retrieval.
Approach: They propose a contrastive regularization based method to extract correct answer entities from a context knowledge base and a corresponding question.
Outcome: The proposed method achieves state-of-the-art performance on the WebQuestionsSP dataset and the effectiveness of proposed modules is also evaluated.
MMAC: A Multilingual, Multimodal Alignment Framework for Cultural Grounding Evaluation (2026.acl-long)

Copied to clipboard

Challenge: Existing models lack cultural alignment across modalities and languages . a new framework to assess cultural awareness across linguistics and languages is needed .
Approach: They propose a framework that integrates tri-modally aligned cultural benchmarks and a five-dimensional evaluation protocol to assess cross-country awareness disparities.
Outcome: The proposed framework assesses cultural awareness disparities across modalities and languages . it is the first dataset aligned at the input level across text, image, and speech .
D2PCM:A Multi-Turn Dialogue Dataset with Personalized Contextual Memory (2026.findings-acl)

Copied to clipboard

Challenge: Conventional interactive algorithms have predominantly treated memory as a contextual element, neglecting the nuanced cognitive processes involved in individualized memory encoding and retrieval.
Approach: They propose a multi-turn dialogue dataset with Personalized Contextual Memory to facilitate advanced research on personalized memory processing.
Outcome: The proposed datasets provide a comprehensive benchmark to facilitate advanced research on personalized memory processing.
Can LLMs Hear the Dogwhistle? (2026.findings-acl)

Copied to clipboard

Challenge: Existing safety benchmarks focus on explicitly harmful content, but ignore context-dependent expressions such as dogwhistles.
Approach: They propose a benchmark for evaluating LLM safety under dogwhistle-driven prompts . their findings expose a blind spot in current safety evaluation practices .
Outcome: The proposed benchmark compared safety performance with toxic terms using dogwhistle-driven prompts.
Value Compass Benchmarks: A Comprehensive, Generative and Self-Evolving Platform for LLMs’ Value Evaluation (2025.acl-demo)

Copied to clipboard

Challenge: Current evaluation methods for large language models face two key challenges: 1. evaluation validity and 2. Result interpretation reduce the pluralistic and incommensurable values to one-dimensional scores.
Approach: They propose a platform for comprehensive value diagnosis of large language models (LLMs) that provides a generative evaluation paradigm that automatically creates real-world test items co-evolving with ever-advancing LLMs.
Outcome: The proposed platform provides a framework for comprehensive value diagnosis of large language models (LLMs) with fine-grained scores and case studies across 27 value dimensions for 33 leading LLMs, customized comparisons, and visualized analysis of LLM’s alignment with cultural values.
Value FULCRA: Mapping Large Language Models to the Multidimensional Spectrum of Basic Human Value (2024.naacl-long)

Copied to clipboard

Challenge: Existing work specifies values as risk criteria formulated in the AI community, e.g., fairness and privacy protection, suffering from poor clarity, adaptability and transparency.
Approach: They propose a value alignment paradigm based on Schwartz's Theory of Basic Values as an instantiation and propose 'BaseAlign' to support this paradigm.
Outcome: The proposed model covers existing risks and anticipates unidentified ones with a low-data set.
RU22Fact: Optimizing Evidence for Multilingual Explainable Fact-Checking on Russia-Ukraine Conflict (2024.lrec-main)

Copied to clipboard

Challenge: Existing methods to verify factuality of claims do not provide sufficient evidence for explainable fact-checking systems.
Approach: They propose a method to automatically retrieve and summarize evidence from the Web and a novel multilingual explainable fact-checking dataset on the Russia-Ukraine conflict in 2022.
Outcome: The proposed method can retrieve and summarize evidence from the Web and generate explanations in 16 languages.
AutoSchemaKG: Autonomous Knowledge Graph Construction through Dynamic Schema Induction from Web-Scale Corpora (2026.acl-long)

Copied to clipboard

Challenge: Existing knowledge graph construction frameworks require predefined schemas, limiting their scalability and domain coverage.
Approach: They propose a framework for fully autonomous knowledge graph construction that eliminates the need for predefined schemas.
Outcome: The proposed framework outperforms state-of-the-art models on multi-hop QA tasks and enhances LLM factuality.
MoVa: Towards Generalizable Classification of Human Morals and Values (2025.emnlp-main)

Copied to clipboard

Challenge: Identifying human morals and values embedded in language is essential to empirical studies of communication.
Approach: They propose a framework for generalizable classification of human morals and values . they recommend a classification strategy that scores all related concepts simultaneously .
Outcome: The proposed method outperforms fine-tuned models across domains and frameworks.
On Orthogonality Constraints for Transformers (2021.acl-short)

Copied to clipboard

Challenge: a dedicated study on orthogonality constraints for transformers has been lacking . plug-and-play constraints increase the BLEU of transformers .
Approach: They propose to use plug-and-play constraints to encourage matrices to be orthogonal for numerical stability.
Outcome: The proposed constraint increases the BLEU on the large-scale WMT’16 EnDe benchmark by a factor of 28.4 to 29.6.
Influence-based Online Experience Selection for Effective RLHF (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for RL fail to establish an interpretable connection between data and optimization objectives.
Approach: They propose a data selection method that dynamically estimates the influence of individual training samples on policy optimization.
Outcome: The proposed method significantly improves training effectiveness with fewer optimization steps.
Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System (2026.acl-long)

Copied to clipboard

Challenge: Vision-Language-Action models ground high-level semantic instructions into executable physical actions.
Approach: They propose a Coarse-to-Fine Dual-System VLA architecture that decouples learning complexity into a coarse-to fine hierarchy while leveraging structural modularity to implement an asynchronous execution strategy.
Outcome: The proposed architecture decouples learning complexity into a coarse-to-fine hierarchy while leveraging structural modularity to implement an asynchronous execution strategy.
E-ViC: Reasoning Beyond Text via Embodied Visual Chain for Spatial Intelligence (2026.acl-long)

Copied to clipboard

Challenge: Existing Vision-Language Models (VLMs) lack spatial reasoning, despite text-based CoTs . e-ViC reframes spatial intelligence as a verifiable, tool-using capability, argues a new study.
Approach: They propose a framework that moves reasoning beyond text into the visual domain . they ground reasoning in pixel-level interactions to enable human-like "look-and-confirm" strategies .
Outcome: The proposed framework outperforms existing Vision-Language Models with an average gain of 10.1%.
FAIRGAMER: Evaluating Social Biases in LLM-Based Video Game NPCs (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) have enhanced or replaced traditional non-player characters in video games.
Approach: They propose a benchmark to evaluate social biases across three interaction patterns: transaction, cooperation, and competition.
Outcome: The proposed benchmark assesses four bias types across transaction, cooperation, and competition using a novel metric, FairMCV.
CMIG: Conceptual Metaphor Theory-Inspired Framework for Metaphorical Image Generation (2026.findings-acl)

Copied to clipboard

Challenge: Existing text-to-image systems often produce visually plausible but semantically literal outputs.
Approach: They propose a structured prompting framework inspired by Conceptual Metaphor Theory . they propose to identify source–target mappings, filter projectable source attributes and select a visual realization strategy in a reproducible reasoning workflow.
Outcome: The proposed framework improves semantic alignment and controllability on metaphor prompts.
Beyond Static Profiles: Capturing the Fluidity of User Preferences in Diverse Scenarios (2026.findings-acl)

Copied to clipboard

Challenge: Existing approaches to personalize Large Language Models often default to homogeneous behaviors . preferences can shift, and conflict, depending on context, authors argue .
Approach: They propose a hierarchical taxonomy to differentiate between stable and situational preferences . they use a dataset of 10k meticulously curated preferences to test their taxonomies .
Outcome: The proposed model differentiates between stable and situational preferences based on curated user preferences . it provides a practical testbed for advancing dynamic, context-aware personalization in conversational agents.
SLIM: Subtrajectory-Level Elimination for More Effective Reasoning (2025.findings-emnlp)

Copied to clipboard

Challenge: Notable examples include OpenAI’s o1/o3/o4 series and DeepSeek-R1 .
Approach: They develop a framework to identify suboptimal subtrajectories based on human-established criteria . they also use a sampling algorithm to select data whose reasoning process is free from suboptimally subtravertories to the highest degree .
Outcome: The proposed method reduces the number of suboptimal subtrajectories by 25.9% during the inference process.
Revisiting the Roles of “Text” in Text Games (2022.findings-emnlp)

Copied to clipboard

Challenge: Recent work has shown that random text hashes could be complementary rather than contrasting in text games.
Approach: They propose a scheme to extract contextual information into an approximate state hash as extra input for an RNN-based text agent.
Outcome: The proposed scheme achieves competitive performance with state-of-the-art text agents using advanced NLU techniques such as knowledge graph and passage retrieval.
OAgents: An Empirical Study of Building Effective Agents (2025.findings-emnlp)

Copied to clipboard

Challenge: a recent study shows that agent research practices are far from standard, rigorous . lack of a standard evaluation protocol makes previous works not reproducible, authors say .
Approach: They conduct an empirical study on the GAIA benchmark to investigate agent design choices . they find that lack of a standard evaluation protocol makes previous works not reproducible .
Outcome: The proposed framework achieves state-of-the-art performance among open-source projects.
VLMInferSlow: Evaluating the Efficiency Robustness of Large Vision-Language Models as a Service (2025.acl-long)

Copied to clipboard

Challenge: Existing studies evaluate efficiency robustness of vision-language models under unrealistic assumptions, requiring access to model architecture and parameters.
Approach: They propose a novel approach to evaluate VLM efficiency robustness in a realistic black-box setting.
Outcome: The proposed approach generates adversarial images with imperceptible perturbations, increasing the computational cost by up to 128.47%.
AlphaQT-Bench: Diagnosing the Gap between Financial Code Generation and Quantitative Reasoning in LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks rely on outcome-driven metrics such as profitability and look-ahead bias.
Approach: They propose a diagnostic benchmark for instruction-grounded financial code generation under strict semantic and temporal constraints.
Outcome: The proposed benchmarks show that the models fail under causal, structural, or functional constraints.
nvAgent: Automated Data Visualization from Natural Language via Collaborative Agent Workflow (2025.acl-long)

Copied to clipboard

Challenge: *Natural Language to Visualization (NL2Vis) seeks to transform natural-language descriptions into visual representations of given tables.
Approach: They propose a collaborative agent workflow for NL2Vis that incorporates three agents . the model is called **nvAgent** and comprises a processor agent for database processing and context filtering, a composer agent for planning visualization generation and a validator agent for code translation and output verification.
Outcome: The proposed workflow surpasses state-of-the-art models on the VisEval benchmark.
TiMem: Temporal-Hierarchical Memory Consolidation for Long-Horizon Conversational Agents (2026.findings-acl)

Copied to clipboard

Challenge: Existing memory frameworks provide limited support for temporally structured information across hierarchical levels, leading to fragmented memories and unstable long-horizon personalization.
Approach: They propose a temporal–hierarchical memory framework that organizes conversations through a Temporal Memory Tree.
Outcome: The proposed framework outperforms baselines while reducing the recalled memory length by 52.20%.
Thinking Alignment of Scenario-Oriented User Simulation (2026.acl-long)

Copied to clipboard

Challenge: Existing user simulators based on prompting to role-play or SFT focus on imitating textual utterances without considering multi-faceted cognitive processes that underlie human decision-making during interactions.
Approach: They construct a user-simulator dataset that augments 51k human–LLM conversations by reconstructing the user’s inner reasoning during and at the end of each dialogue.
Outcome: The proposed user simulators augment 51k human–LLM conversations by reconstructing the user’s inner reasoning both during and at the end of each dialogue.
CGBridge: Bridging Code Graphs and Large Language Models for Better Structure-Aware Code Understanding (2026.findings-acl)

Copied to clipboard

Challenge: Existing structure-aware approaches treat structure as serialized text prompts or auxiliary training objectives, failing to provide explicit guidance during inference.
Approach: They propose a plug-and-play method that enhances Large Language Models with Code Graph information through an external, trainable Bridge module.
Outcome: The proposed method decouples structural reasoning from textual generation without updating the backbone.
Towards Better Value Principles for Large Language Model Alignment: A Systematic Evaluation and Enhancement (2025.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) show remarkable performance across tasks . alignment with human values is critical for their responsible development.
Approach: They propose a framework that evaluates value principles along three desirable properties . they propose supervised fine-tuning, reinforcement learning-based approaches .
Outcome: The proposed framework improves value principles along the three desirable properties of LLMs.
DialogVED: A Pre-trained Latent Variable Encoder-Decoder Model for Dialog Response Generation (2022.acl-long)

Copied to clipboard

Challenge: Existing pre-trained dialog models shed light on various downstream tasks in natural language processing (NLP).
Approach: They propose a dialog pre-training framework that introduces latent variables into the enhanced encoder-decoder pre-train framework to increase relevance and diversity of responses.
Outcome: The proposed model achieves state-of-the-art on personaChat, DailyDialog, and DSTC7-AVSD datasets.

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