Papers by Pan Yuan

28 papers
Analytical FFN-to-MoE Restructuring via Activation Pattern Analysis (2026.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) are fast but require expensive pre-training . a new approach to scale large language models into MoEs reduces inference costs .
Approach: They propose an analytical post-training framework that rapidly restructures FFNs into sparse MoE architectures using only a small calibration dataset.
Outcome: The proposed framework outperforms existing methods on a small calibration dataset.
LLM-Powered Benchmark Factory: Reliable, Generic, and Efficient (2026.acl-long)

Copied to clipboard

Challenge: Using generic and efficient benchmark generators, human annotators are limited by inefficiency . current benchmark generator methods rely on seed signals, leading to long cycles and high costs .
Approach: They propose a framework to evaluate LLMs as generic benchmark generators and integrate them as BenchMaker.
Outcome: The proposed framework achieves comparable performance to human-annotated benchmarks on most metrics.
SocAoG: Incremental Graph Parsing for Social Relation Inference in Dialogues (2021.acl-long)

Copied to clipboard

Challenge: Existing studies focus on identifying entities' relations from the semantics of dialogues-they utilize either the attention mechanism or a refined token graph to locate informative words.
Approach: They propose a sequential structure prediction task to incrementally parse SocAoG for dynamic inference upon any incoming utterance.
Outcome: Empirical results show that the proposed model infers social relations more accurately than the state-of-the-art methods.
Make Every Penny Count: Difficulty-Adaptive Self-Consistency for Cost-Efficient Reasoning (2025.findings-naacl)

Copied to clipboard

Challenge: Existing decoding strategies for chain-of-thought reasoning do not exploit prior information about question difficulty.
Approach: They propose a decoding strategy called self-consistency to improve reasoning performance by adjusting the number of samples based on the posterior distribution of a set of pre-samples.
Outcome: The proposed method outperforms baseline methods on arithmetic, commonsense and symbolic reasoning tasks while achieving comparable performance.
HILL: Hierarchy-aware Information Lossless Contrastive Learning for Hierarchical Text Classification (2024.naacl-long)

Copied to clipboard

Challenge: Existing self-supervised methods in natural language processing rely on augmentation rules to generate contrastive samples.
Approach: They propose a hierarchy-aware information lossless contrastive learning scheme that uses syntactic information reserved in the input sample and fused during the learning process.
Outcome: The proposed learning scheme is superior to existing methods in hierarchical text classification . the proposed learning system is based on a structure encoder and a text encoder .
Mitigating the Alignment Tax of RLHF (2024.emnlp-main)

Copied to clipboard

Challenge: Large Language Models (LLMs) acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax.
Approach: They propose to use a model averaging technique to find the most powerful alignment-forging Pareto front among RLHF algorithms.
Outcome: The proposed method achieves the strongest alignment-forging Pareto front among competing methods.
Focused Large Language Models are Stable Many-Shot Learners (2024.emnlp-main)

Copied to clipboard

Challenge: In-Context Learning (ICL) enables large language models to achieve rapid task adaptation by learning from demonstrations.
Approach: They propose a training-free method that disperses model attention from the query . they propose 'focus' search strategy that uses model perplexity to ensure sufficient attention .
Outcome: The proposed method achieves an average performance improvement of 5.2% over vanilla ICL and scales well with many-shot demonstrations.
InsBank: Evolving Instruction Subset for Ongoing Alignment (2025.findings-emnlp)

Copied to clipboard

Challenge: Recent studies emphasize that quality and diversity of instruction data are more crucial than quantity, highlighting the need to select diverse, high-quality subsets to reduce training costs.
Approach: They propose to use a continuously updated repository to integrate the latest valuable instruction data with a progressive evolution framework to evolve InsBank over time.
Outcome: The proposed framework outperforms baselines in InsBank evolution and extracts budget-specific subsets.
Integrate the Essence and Eliminate the Dross: Fine-Grained Self-Consistency for Free-Form Language Generation (2024.acl-long)

Copied to clipboard

Challenge: Existing methods to improve output quality without aggregating input tokens are limited by the complexity of aggregation of responses.
Approach: They propose to extract and integrate segment-level commonalities from candidate samples to enhance performance of LLMs in open-ended and reasoning tasks.
Outcome: The proposed method improves performance on reasoning, code generation and mathematical reasoning tasks without requiring additional models and overlooking the knowledge present among the candidates.
Beyond One-Size-Fits-All: Tailored Benchmarks for Efficient Evaluation (2025.acl-long)

Copied to clipboard

Challenge: Existing efficient methods estimate performance of models on large benchmarks, but these methods rely on the assumption that target models have high prediction consistency with source models.
Approach: They propose a method that conducts customized evaluation tailored to each target model.
Outcome: The proposed method reduces the MAE of estimates by 31.4% on benchmarks across 300 models.
KELE: A Multi-Agent Framework for Structured Socratic Teaching with Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Socratic teaching places high demands on teachers’ expertise and real-time feedback capabilities, making it difficult to scale in large educational settings.
Approach: They propose a multi-agent framework for structured Socratic teaching with LLMs that integrates a structured SocRule and a consultant-teacher collaborative teaching mechanism.
Outcome: The proposed framework outperforms existing LLMs in natural language generation and dialogue comprehension in the classroom.
Making MLLMs Blind: Adversarial Smuggling Attacks in MLLM Content Moderation (2026.findings-acl)

Copied to clipboard

Challenge: Multimodal Large Language Models (MLLMs) are increasingly being deployed as content moderators . however, they exploit the Human-AI capability gap and create adversarial environments . smuggling attacks exploit the human-AI gap and exploit the vulnerability .
Approach: They construct a benchmark to evaluate the vulnerability of MLLMs as content moderators . they identify three root causes: limited capabilities of vision encoders, robustness gap in OCR .
Outcome: The proposed model exploits the Human-AI capability gap and is vulnerable to smuggling attacks.
Superficial Self-Improved Reasoners Benefit from Model Merging (2025.emnlp-main)

Copied to clipboard

Challenge: Large Language Models (LLMs) rely heavily on large-scale reasoning data, but as data becomes scarce, model self-improvement offers a promising alternative.
Approach: They propose to merge the weights of original and self-improved LLMs to mitigate model collapse and improve generalized reasoning capability.
Outcome: The proposed model merge mitigates model collapse and improves generalized reasoning capability.
CogLM: Tracking Cognitive Development of Large Language Models (2025.naacl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) have recently shown remarkable abilities across a wide variety of tasks, but few studies have explored the reasons behind the evolutionary relationship among various abilities.
Approach: They construct a benchmark CogLM based on Piaget's Theory of Cognitive Development (PTC) which measures the cognitive levels of Large Language Models (LLMs) using 1,220 questions spanning 10 cognitive abilities crafted by more than 20 human experts.
Outcome: The proposed framework provides a comprehensive testbed for the cognitive levels of LLMs.
Leveraging Information Bottleneck for Scientific Document Summarization (2021.findings-emnlp)

Copied to clipboard

Challenge: Existing methods to extract salient sentences from document are unsupervised and rely on graph-based methods for sentence ranking.
Approach: They propose an unsupervised extractive approach to document level summarization based on the Information Bottleneck principle.
Outcome: The proposed framework can be extended to a multi-view framework by different signals.
Merge Hijacking: Backdoor Attacks to Model Merging of Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Existing research on model merging focuses on optimizing model performance and minimizing backdoors.
Approach: They propose a backdoor attack targeting model merging in Large Language Models that creates a unified model for multi-domain tasks.
Outcome: The proposed attack is effective across models, merging algorithms, and tasks while maintaining utility across tasks.
From Long to Lean: Performance-aware and Adaptive Chain-of-Thought Compression via Multi-round Refinement (2025.emnlp-main)

Copied to clipboard

Challenge: Chain-of-Thought reasoning introduces significant inference latency due to its verbosity.
Approach: They propose a framework that leverages token elasticity phenomenon to progressively compress CoTs via multiround refinement.
Outcome: The proposed method achieves an average accuracy improvement of 5.6% over state-of-the-art baselines while reducing CoT length by an average of 47 tokens and significantly lowering latency.
COIG-P: A High-Quality and Large-Scale Chinese Preference Dataset for Alignment with Human Values (2026.findings-eacl)

Copied to clipboard

Challenge: Existing Chinese preference datasets suffer from limited scale, restricted domain coverage, and insufficiently rigorous data validation.
Approach: They propose an LLM-based data annotation pipeline with no human intervention to annotate Chinese preference datasets.
Outcome: The proposed pipeline outperforms existing Chinese preference datasets on AlignBench and Chinese Reward Benchmark.
Speculative Decoding for Multi-Sample Inference (2025.findings-emnlp)

Copied to clipboard

Challenge: Speculative decoding method exploits consensus of parallel reasoning paths to synthesize high-quality draft tokens without auxiliary models or external databases.
Approach: They propose a speculative decoding method that exploits the consensus of parallel reasoning paths to synthesize high-quality draft tokens without auxiliary models or external databases.
Outcome: The proposed method exploits the intrinsic consensus of parallel reasoning paths to synthesize high-quality draft tokens without auxiliary models or databases.
Noise Robust Named Entity Understanding for Voice Assistants (2021.naacl-industry)

Copied to clipboard

Challenge: Named Entity Recognition and Entity Linking are challenging for voice assistants . utterances are relatively short, so there is not much context to help disambiguate .
Approach: They propose a Named Entity Understanding system that combines NER and EL in a joint reranking module.
Outcome: The proposed framework improves NER accuracy by up to 3.13% and EL accuracy by 3.6% in F1 score . it also leads to better accuracies in other natural language understanding tasks .
From Sub-Ability Diagnosis to Human-Aligned Generation: Bridging the Gap for Text Length Control via MarkerGen (2025.acl-long)

Copied to clipboard

Challenge: Existing methods to control text length are lacking in LCTG, posing a major limitation for practical applications.
Approach: They propose a plug-and-play approach that decomposes LCTG sub-abilities with human patterns as reference and performs detailed error analysis.
Outcome: The proposed method significantly improves LCTG across various settings, exhibiting outstanding effectiveness and generalizability.
Towards Intrinsic Interpretability of Large Language Models: A Survey of Design Principles and Architectures (2026.acl-long)

Copied to clipboard

Challenge: Existing studies on explainable AI focus on post-hoc explanation methods that interpret trained models through external approximations.
Approach: They propose to categorize existing approaches into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction.
Outcome: The proposed approaches are categorized into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction.
Revisiting Self-Consistency from Dynamic Distributional Alignment Perspective on Answer Aggregation (2025.findings-acl)

Copied to clipboard

Challenge: Existing studies on self-consistency show that it improves reasoning abilities by aggregating diverse stochastic samples.
Approach: They propose a confidence-driven mechanism that dynamically calibrates temperature to align with high probability modes.
Outcome: The proposed method outperforms fixed-diversity baselines on reasoning tasks and improves both average and best-case performance.
SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for fraud detection rely on transcribed text, lacking acoustic cues . a proposed framework for audio-based slow-thinking fraud detection eliminates transcription errors .
Approach: They propose a framework for audio-based slow-thinking fraud detection that eliminates transcription errors and rewards slow-thought reasoning by capturing fine-grained audio details.
Outcome: The proposed method improves accuracy, inference efficiency, and real-time processing capabilities.
Graph-Based Chain-of-Thought Pruning for Reducing Redundant Reflections in Reasoning LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Extending CoT through RL can induce undesirable thinking patterns such as overthinking . prior work has focused on inefficient reflection, which manifests in two problematic patterns: Indiscriminate Reflection and Repetitive Reflectione .
Approach: They propose a graph-based approach to optimize CoT by pruning each linear CoT into a directed acyclic graph with explicit dependency edges.
Outcome: The proposed approach reduces the average reasoning tokens by 42% while maintaining or improving accuracy.
Generative Dense Retrieval: Memory Can Be a Burden (2024.eacl-long)

Copied to clipboard

Challenge: Empirical results show that Generative Dense Retrieval (GDR) achieves an average of 3.0 R@100 improvement on NQ dataset under multiple settings and has better scalability.
Approach: They propose a Generative Dense Retrieval paradigm that auto-decodes document identifiers given a query and uses memory to avoid memory confusion.
Outcome: Empirical results show that the proposed paradigm improves on the small-scale corpora and improves scalability.
Reasoning Fails Where Step Flow Breaks (2026.acl-long)

Copied to clipboard

Challenge: Existing analysis tools struggle with long chain of thought traces.
Approach: They propose a saliency-inspired test-time intervention that adjusts shallow saliencies to improve accuracy on math, science, and coding tasks.
Outcome: The proposed model improves accuracy on math, science, and coding tasks without retraining.
Poor-Supervised Evaluation for SuperLLM via Mutual Consistency (2024.findings-acl)

Copied to clipboard

Challenge: evaluating superLLMs is especially difficult because of their intelligence-intensive nature.
Approach: They propose an evaluation benchmark with accurate labels for SuperLLMs whose capabilities surpass those of humans . they first prove that consistency between model under evaluation and reference model can equalize the true capabilities of the model to be evaluated .
Outcome: The proposed evaluation benchmarks can assess the true capabilities of the model to be evaluated without accurate labels.

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