Papers by Shang Wu

37 papers
Tracing the Roots: A Multi-Agent Framework for Uncovering Data Lineage in Post-Training LLMs (2026.acl-long)

Copied to clipboard

Challenge: High-quality post-training data is the primary engine driving LLM capabilities . datasets are often treated as isolated artifacts, overlooking their true developmental context .
Approach: They propose a framework to reconstruct the evolutionary graph of dataset development using data lineage.
Outcome: The proposed framework characterizes domain-specific structural patterns in Math-oriented datasets and general-domain corpora.
ERNIE-Doc: A Retrospective Long-Document Modeling Transformer (2021.acl-long)

Copied to clipboard

Challenge: Existing models for document-level language pretraining are not suitable for long documents due to their quadratically increasing memory and time consumption.
Approach: They propose a document-level language pretraining model based on Recurrence Transformers.
Outcome: The proposed model outperforms existing models on language understanding tasks.
Self-Taught Agentic Long Context Understanding (2025.acl-long)

Copied to clipboard

Challenge: Extensive experiments across seven long-context tasks demonstrate that AgenticLU significantly outperforms state-of-the-art prompting methods and specialized long-consumer LLMs.
Approach: They propose a framework to enhance an LLM's understanding of long-context questions by integrating targeted self-clarification with contextual grounding within an agentic workflow.
Outcome: The proposed framework outperforms state-of-the-art prompting methods and specialized long-context LLMs in seven long-constitut tasks.
Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation (2026.acl-long)

Copied to clipboard

Challenge: Prior work has shown that safety behaviors are governed by low-rank structures . Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks .
Approach: They propose a safety alignment system that disentangles safety-relevant directions into monosemantic features and constructs an interpretable safety subspace from SAE directions.
Outcome: Empirically, the proposed model achieves 99.6% safety rates across multiple model families and scales . low-rank Adaptation consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks compared with previous methods .
Uncertainty-Aware Routing for Principled Alignment with MoE Dynamics (2026.acl-long)

Copied to clipboard

Challenge: Mixture-of-Experts (MoE) is a cornerstone for scaling LLMs, yet its training dynamics remain poorly understood, often leading to sub-optimal specialization.
Approach: They propose to use Helmholtz Free Energy and Router Entropy to study the MoE lifecycle and identify a universal Three-Stage Phase Transition .
Outcome: The proposed model reduces perplexity and improves expert distinctiveness, offering a principled path toward thermodynamically aligned computation.
Explainable Chain-of-Thought Reasoning: An Empirical Analysis on State-Aware Reasoning Dynamics (2025.findings-emnlp)

Copied to clipboard

Challenge: Recent advances in chain-of-thought prompting have demonstrated the ability of large language models to perform multi-step reasoning.
Approach: They propose a framework to analyze latent dynamics of CoT trajectories for interpretability . they segment generated CoT into discrete reasoning steps and abstract each step into a spectral embedding based on token-level Gram matrices .
Outcome: The proposed framework segments generated CoT steps into discrete reasoning steps, abstracts each step into a spectral embedding based on token-level Gram matrices, and clusters these embeddements into semantically meaningful latent states.
ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch (2026.acl-long)

Copied to clipboard

Challenge: Existing open-source vision language models lack high-quality training data for chart reasoning . current models are simplistic and repetitive, while associated QA pairs are prone to hallucinations .
Approach: They propose a framework to synthesize complex charts and reliable reasoning data from scratch.
Outcome: Experimental results show that ChartVerse-8B surpasses existing models in QA and difficulty . lack of high-quality training data hampers development of open-source models .
Mitigating Visual Knowledge Forgetting in MLLM Instruction-tuning via Modality-decoupled Gradient Descent (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing fine-tuning and continual learning methods compress visual representations and emphasize task alignment over visual retention.
Approach: They propose a modality-decoupled gradient descent (MDGD) that regulates gradient updates to preserve effective rank of visual features and explicitly disentangles visual learning from task-specific alignment.
Outcome: The proposed model reduces visual forgetting and improves visual retention . it disentangles visual learning from task-specific alignment and preserves effective rank .
Generative Annotation for ASR Named Entity Correction (2025.emnlp-main)

Copied to clipboard

Challenge: Existing named entity correction models fail to transcribe domain-speciffcnamed entities when theforms of the wrongly-transcribed words and the ground-truth entity are signiffcantly different.
Approach: They propose a method that utilizes speech sound features to retrieve candidate entities . it uses speech sound feature to annotate entityerrors in ASR transcripts .
Outcome: The proposed method can bring signiffcant improvement to entity accuracy.
“In-Dialogues We Learn”: Towards Personalized Dialogue Without Pre-defined Profiles through In-Dialogue Learning (2024.emnlp-main)

Copied to clipboard

Challenge: Existing approaches to personalized dialogue generate pre-defined profiles that are time-consuming and labor-intensive to create.
Approach: They propose a framework that leverages dialogue history to characterize personas without pre-defined profiles.
Outcome: The proposed framework improves BLEU and ROUGE scores on three datasets and human evaluations further validate the proposed method.
CMoralEval: A Moral Evaluation Benchmark for Chinese Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: Recent years have witnessed remarkable progress achieved by large language models in both natural language understanding and generation.
Approach: They propose a large benchmark CMoralEval for moral evaluation of Chinese LLMs . they use a Chinese TV program discussing Chinese moral norms and Chinese moral anomies based on various sources .
Outcome: The proposed dataset is characterized by diversity and authenticity.
PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning (2026.acl-long)

Copied to clipboard

Challenge: Parallel Coordinated Reasoning (PaCoRe) overcomes a central limitation of contemporary language models: their inability to scale test-time compute (TTC) far beyond sequential reasoning under a fixed context window.
Approach: They propose a training-and-inference framework to overcome a central limitation of language models: their inability to scale test-time compute (TTC) under a fixed context window.
Outcome: The proposed model scales to multi-million-token effective TTC without exceeding context limits.
PRISM: Probing Reasoning, Instruction, and Source Memory in LLM Hallucinations (2026.acl-long)

Copied to clipboard

Challenge: Existing benchmarks for hallucination evaluation rely on mixed queries and posterior evaluation, which quantifies hallucinosity severity but offers limited insight into where and why they occur.
Approach: They propose a controlled benchmark that disentangles hallucinations into four dimensions: knowledge missing, knowledge errors, reasoning errors, and instruction-following errors.
Outcome: The proposed model disentangles hallucinations into four dimensions: knowledge missing, knowledge errors, reasoning errors, and instruction-following errors.
SeNsER: Learning Cross-Building Sensor Metadata Tagger (2020.findings-emnlp)

Copied to clipboard

Challenge: Sensor metadata tagging is a key component of smart building applications.
Approach: They propose a framework that learns a sensor metadata tagger for a new building based on its raw metadata and some existing fully annotated building.
Outcome: The proposed framework learns a sensor metadata tagger for a new building based on its raw metadata and some existing fully annotated building.
Does the Generator Mind Its Contexts? An Analysis of Generative Model Faithfulness under Context Transfer (2024.lrec-main)

Copied to clipboard

Challenge: Existing studies have focused on examining hallucinations stemming from static input, such as in summarization or machine translation.
Approach: They propose a knowledge-augmented generator that produces information that remains grounded in contextual knowledge regardless of alterations in the context.
Outcome: The proposed method is designed to produce information that remains grounded in contextual knowledge, regardless of alterations in the context.
EDU-level Extractive Summarization with Varying Summary Lengths (2023.findings-eacl)

Copied to clipboard

Challenge: Existing studies on extractive summarization use finer-grained elementary discourse units . few studies exploited finer grained EDUs with little analysis and justification for the extractive unit selection .
Approach: They propose an extractive model with Varying summary lengths that extracts fixed top-k salient sentences from the document as a summary.
Outcome: The proposed model performs better on ROUGE scores than state-of-the-art models.
LEMON: Reviving Stronger and Smaller LMs from Larger LMs with Linear Parameter Fusion (2024.acl-long)

Copied to clipboard

Challenge: Existing methods to train a stronger and smaller model with the help of large models are limited by the model size and performance.
Approach: They propose to learn competent initial points for smaller models by fusing parameters from larger models and introduce controllable receptive fields to model prior parameter characteristics.
Outcome: The proposed method outperforms baselines in terms of effectiveness and efficiency.
More Tokens, Lower Precision: Towards the Optimal Token-Precision Trade-off in KV Cache Compression (2025.findings-emnlp)

Copied to clipboard

Challenge: storing more tokens in the KV cache at lower precision can enhance the long-context performance of large language models.
Approach: They propose a token-precision trade-off strategy to optimize KV cache compression . they also propose storing more tokens in the KV at lower precision .
Outcome: The proposed method achieves an optimal point within the Information Bottleneck compared to standalone KV pruning or KV quantization.
CompileAgent: Automated Real-World Repo-Level Compilation with Tool-Integrated LLM-based Agent System (2025.acl-long)

Copied to clipboard

Challenge: CompileAgent is the first LLM-based agent framework dedicated to repo-level compilation.
Approach: They propose a LLM-based agent framework dedicated to repo-level compilation.
Outcome: The proposed method significantly improves compilation success rate, ranging from 10% to 71%.
ChatSOP: An SOP-Guided MCTS Planning Framework for Controllable LLM Dialogue Agents (2025.acl-long)

Copied to clipboard

Challenge: Existing models that use Large Language Models (LLMs) show superior performance in various tasks, but lack of controllability leads to unfocused conversations or task failure.
Approach: They propose a standard operating procedure (SOP) framework to regulate dialogue flow by integrating Chain of Thought reasoning and supervised fine-tuning for SOP prediction.
Outcome: The proposed method achieves a 27.95% improvement in action accuracy compared to baseline models based on GPT-3.5 and also shows notable gains for open-source models.
BLADE: Benchmarking Language Model Agents for Data-Driven Science (2024.findings-emnlp)

Copied to clipboard

Challenge: Language model-based agents can be used to conduct and support data-driven science, but evaluating them on open-ended tasks is challenging due to multiple valid approaches, partially correct steps, and different ways to express the same decisions.
Approach: They propose a benchmark to automatically evaluate agents’ multifaceted approaches to open-ended research questions.
Outcome: BLADE evaluates agents’ multifaceted approaches to open-ended research questions using data from 12 datasets and research questions drawn from existing scientific literature.
Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking (2025.acl-long)

Copied to clipboard

Challenge: Large language models face inherent performance bottlenecks under parameter constraints . challenging tokens induce abrupt gradient spikes across layers, exposing stress points .
Approach: They propose an inner thinking transformer that reimagines layer computations as implicit thinking steps.
Outcome: Empirical results show that ITT outperforms Transformer/Loop variants in 11 benchmarks.
NACL: A General and Effective KV Cache Eviction Framework for LLM at Inference Time (2024.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) with extended context windows are expensive and infeasible on fixed memory hardware due to the surprisingly large memory consumption of KV Cache.
Approach: They propose a general framework for long-context KV cache eviction that achieves more optimal and efficient evict in a single operation during the encoding phase.
Outcome: The proposed framework improves performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 5 with over 95% performance maintenance.
Crowd Comparative Reasoning: Unlocking Comprehensive Evaluations for LLM-as-a-Judge (2025.acl-long)

Copied to clipboard

Challenge: Existing methods rely on majority voting or criteria expansion to capture detailed and detailed details, often leading to incomplete outcomes.
Approach: They propose a method which introduces additional crowd responses to compare with the candidate responses, thereby exposing deeper and more comprehensive details within the candidate answers.
Outcome: Experiments show that the proposed method improves evaluation reliability and achieves an average gain of 6.7% across five benchmarks.
M-Ped: Multi-Prompt Ensemble Decoding for Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: a new ensemble decoding approach enhances the performance of Large Language Models.
Approach: They propose a multi-prompt ensemble decoding approach to enhance LLM performance . they submit n variations of prompts with X to LLMs in batch mode to decode and derive probability distributions .
Outcome: The proposed method improves pass@k rates, LENS metrics and BLEU scores on diverse NLP tasks.
MobileVLM: A Vision-Language Model for Better Intra- and Inter-UI Understanding (2024.findings-emnlp)

Copied to clipboard

Challenge: Recent mobile AI agents based on VLMs lack basic mobile capabilities due to their pre-trained nature.
Approach: They propose a mobile AI agent based on VLMs that includes additional pre-training stages to enhance both intra- and inter-UI understanding.
Outcome: The proposed model outperforms existing VLMs on the Chinese mobile dataset Mobile3M .
Multi-step Problem Solving Through a Verifier: An Empirical Analysis on Model-induced Process Supervision (2024.findings-emnlp)

Copied to clipboard

Challenge: a method for process supervision has shown significant improvements in multi-step problem solving . despite the advances in process supervision, there are still easily observable mistakes in state-of-the-art LLMs.
Approach: They propose a method for automating data curation by using a trained verifier to evaluate intermediate steps generated by a reasoner.
Outcome: The proposed method improves the performance of PaLM 2 on math and coding tasks.
Learning to Edit: Aligning LLMs with Knowledge Editing (2024.acl-long)

Copied to clipboard

Challenge: Existing knowledge editing techniques rely on memorizing updated knowledge, impeding LLMs from effectively combining the new knowledge with their inherent knowledge when answering questions.
Approach: They propose a Learning to Edit framework that equips LLMs with the ability to apply updated knowledge to input questions through a two-phase process .
Outcome: The proposed framework outperforms existing methods in knowledge editing tasks and compares it with four benchmarks and two LLM architectures.
Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference (2023.acl-long)

Copied to clipboard

Challenge: Existing studies attribute catastrophic forgetting to fine-tuning, and they retain pre-trained knowledge indiscriminately without identifying what knowledge is transferable.
Approach: They propose a unified objective for fine-tuning to retrieve the causality back from pre-trained data and use it to mitigate negative transfer while preserving knowledge.
Outcome: The proposed method outperforms state-of-the-art fine-tuning methods on commonsense QA datasets and can be implemented as a plug-in module to inflate the performance of existing QA models.
Text Style Transfer Back-Translation (2023.acl-long)

Copied to clipboard

Challenge: Current methods require large amount of bilingual training data, which is challenging and sometimes impossible task.
Approach: They propose a method to modify the style of inputs by modifying the source side of BT data.
Outcome: The proposed method significantly improves translation quality against popular BT benchmarks on high-resource and low-resourced language pairs.
ERCThinker: Fast-Slow Thinking for Emotion Recognition in Conversation (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for ERC lack interpretability and shallow semantics capture deep semantics.
Approach: They propose a Fast-Slow thinking framework for Emotion Recognition in Conversation . they use fine-grained emotion reasoning chains to capture deep semantics .
Outcome: The proposed framework achieves state-of-the-art in explanation and judgment on a benchmark dataset.
Sensei: Self-Supervised Sensor Name Segmentation (2021.findings-acl)

Copied to clipboard

Challenge: Sensor names are alphanumeric strings that encode key contextual information such as their function or physical location.
Approach: They propose a self-supervised framework that can learn to segment sensor names without human annotation.
Outcome: The proposed framework can learn to segment sensor names without human annotation on buildings.
CoMMIT: Coordinated Multimodal Instruction Tuning (2025.emnlp-main)

Copied to clipboard

Challenge: et al., 2024) show that multimodal instruction tuning is more effective than baselines.
Approach: They propose a multimodal balance coefficient that enables quantitative measurement of the balance of learning . they propose auxiliary regularization on the gradient to promote updating with larger step sizes .
Outcome: The proposed method is more effective than baselines in MLLM instruction tuning.
Combining the Best of Both Worlds: A Method for Hybrid NMT and LLM Translation (2025.findings-acl)

Copied to clipboard

Challenge: Large language models have advantages over neural machine translation systems, but they suffer from high computational costs and significant latency.
Approach: They propose a scheduling policy that optimizes translation result while ensuring fast speed and as little LLM usage as possible.
Outcome: The proposed model achieves optimal translation performance with less LLM usage on multilingual test sets.
Sparse Growing Transformer: Training-Time Sparse Depth Allocation via Progressive Attention Looping (2026.findings-acl)

Copied to clipboard

Challenge: Existing approaches to increasing effective depth of LLMs rely on parameter reuse, extending computation through recursive execution.
Approach: They propose a training-time sparse depth allocation framework that progressively increases depth for a small subset of parameters as training evolves.
Outcome: The proposed model outperforms existing approaches to increasing the effective depth of language models while reducing training FLOPs overhead from approximately 16–20% to only 1–3% relative to a standard Transformer backbone.
Instruction-Tuning Data Synthesis from Scratch via Web Reconstruction (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for generating and curating high-quality instruction-tuning data rely heavily on the quality of seed data or strong assumptions about the structure and content of web documents.
Approach: They propose a fully automated framework for synthesizing high-quality instruction-tuning (IT) data directly from raw web documents with minimal assumptions.
Outcome: The proposed framework outperforms state-of-the-art baselines by 16.65% across four instruction-following benchmarks.
SceneAlign: Aligning Multimodal Reasoning to Scene Graphs in Complex Visual Scenes (2026.acl-long)

Copied to clipboard

Challenge: Existing preference-based approaches fail to address this challenge by exploiting language priors to bypass visual grounding.
Approach: They propose a framework that leverages scene graphs as structured visual information to perform controllable structural interventions.
Outcome: The proposed framework improves answer accuracy and reasoning faithfulness across seven visual reasoning benchmarks.

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