Papers by Sheng Shen
LLM2LLM: Boosting LLMs with Novel Iterative Data Enhancement (2024.findings-acl)
Copied to clipboard
Nicholas Lee, Thanakul Wattanawong, Sehoon Kim, Karttikeya Mangalam, Sheng Shen, Gopala Anumanchipalli, Michael Mahoney, Kurt Keutzer, Amir Gholami
| Challenge: | Pretrained large language models are currently state-of-the-art for solving most tasks . however, many of them are in the low-data regime, making fine-tuning challenging . a new data augmentation strategy uses a teacher LLM to augment a small seed dataset . |
| Approach: | They propose a targeted and iterative data augmentation strategy that augments a teacher LLM to fine-tune a small seed dataset by adding additional data. |
| Outcome: | The proposed approach outperforms fine-tuning and other data augmentation strategies on a small seed dataset. |
A Prism Module for Semantic Disentanglement in Name Entity Recognition (P19-1)
Copied to clipboard
| Challenge: | Xu et al., 2015) proposed a noise reduction mechanism to disentangle semantics of words . hard and soft attention mechanisms are used to reduce noise in NLP tasks . |
| Approach: | They propose a prism module to disentangle semantic aspects of words and reduce noise . they propose combining prism modules with downstream models to improve model performance . |
| Outcome: | The proposed method significantly improves the performance of baselines on named entity recognition (NER) tasks. |
What Language Model to Train if You Have One Million GPU Hours? (2022.findings-emnlp)
Copied to clipboard
Teven Le Scao, Thomas Wang, Daniel Hesslow, Stas Bekman, M Saiful Bari, Stella Biderman, Hady Elsahar, Niklas Muennighoff, Jason Phang, Ofir Press, Colin Raffel, Victor Sanh, Sheng Shen, Lintang Sutawika, Jaesung Tae, Zheng Xin Yong, Julien Launay, Iz Beltagy
| Challenge: | Recent years have seen the advent of large language models characterized by emergent capabilities arising from sheer scale alone. |
| Approach: | They propose to use a multilingual model to compare performance to the English-only model by ablation at the billion-parameter scale. |
| Outcome: | The proposed model is based on a multilingual model and its performance against the English-only model. |
A Systematic Survey of Automatic Prompt Optimization Techniques (2025.emnlp-main)
Copied to clipboard
Kiran Ramnath, Kang Zhou, Sheng Guan, Soumya Smruti Mishra, Xuan Qi, Zhengyuan Shen, Shuai Wang, Sangmin Woo, Sullam Jeoung, Yawei Wang, Haozhu Wang, Han Ding, Yuzhe Lu, Zhichao Xu, Yun Zhou, Balasubramaniam Srinivasan, Qiaojing Yan, Yueyan Chen, Haibo Ding, Panpan Xu, Lin Lee Cheong
| Challenge: | Recent advances in prompt engineering have created impediments for end users to adopt . however, prompt engineering remains an impedance due to rapid advances in models, tasks, and associated best practices. |
| Approach: | They propose to define APO as a 5-part unifying framework and categorize all relevant works based on their salient features. |
| Outcome: | The proposed framework aims to improve the performance of large language models on various tasks. |
Adaptive Knowledge-Enhanced Bayesian Meta-Learning for Few-shot Event Detection (2021.findings-acl)
Copied to clipboard
| Challenge: | Event detection typically does not have sufficient labelled data, thus can be formulated as a few-shot learning problem. |
| Approach: | They propose a knowledge-based fewshot event detection method which introduces external event knowledge as the knowledge prior of event types. |
| Outcome: | Experiments show that the proposed method outperforms baselines by 15 F 1 points . event detection is an important task in information extraction . |
Compatibility-Aware Dynamic Fine-Tuning for Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work attributes optimization instability to the low probability of demonstrations being incompatible with the sample level. |
| Approach: | They propose a Dynamic Fine-Tuning extension of DFT that controls sample-level optimization variance. |
| Outcome: | The proposed model can generalize token-level stabilization to the sample level while remaining fully supervised and free of reward modeling. |
Agentic Episodic Control (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for reinforcement learning (RL) are limited by poor data efficiency and weak generalization. |
| Approach: | They propose a novel architecture that integrates large language models into episodic RL. |
| Outcome: | The proposed architecture achieves 2–6 higher data efficiency than baselines and is the only method to solve complex tasks like UnlockLocal with over 90% success. |
IPR: Intelligent Prompt Routing with User-Controlled Quality-Cost Trade-offs (2025.emnlp-industry)
Copied to clipboard
Aosong Feng, Balasubramaniam Srinivasan, Yun Zhou, Zhichao Xu, Kang Zhou, Sheng Guan, Yueyan Chen, Xian Wu, Ninad Kulkarni, Yi Zhang, Zhengyuan Shen, Dmitriy Bespalov, Soumya Smruti Mishra, Yifei Teng, Darren Yow-Bang Wang, Haibo Ding, Lin Lee Cheong
| Challenge: | Existing systems require users to manually select models or employ rigid routing rules that fail to capture the continuous spectrum of query complexity. |
| Approach: | They propose a quality-constrained intelligent prompt routing framework that automatically selects optimal models based on predicted response quality and user-specified tolerance levels. |
| Outcome: | The proposed framework achieves 43.9% cost reduction while maintaining quality parity with strongest model in the Claude family and processes requests with sub-150ms latency. |
RepEval: Effective Text Evaluation with LLM Representation (2024.emnlp-main)
Copied to clipboard
Shuqian Sheng, Yi Xu, Tianhang Zhang, Zanwei Shen, Luoyi Fu, Jiaxin Ding, Lei Zhou, Xiaoying Gan, Xinbing Wang, Chenghu Zhou
| Challenge: | Traditional metrics for automatic text evaluation are tailored to specific tasks, while LLM-based evaluation metrics are costly. |
| Approach: | They propose a metric that leverages projections of LLM representations for evaluation. |
| Outcome: | The proposed metric exhibits higher correlation with human judgments than previous methods on 14 datasets. |
Noisy Self-Knowledge Distillation for Text Summarization (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing approaches to summarize text using a single reference and noisy datasets are ill-suited to summarising on single reference datasets. |
| Approach: | They propose to use self-knowledge distillation to improve text summarization by generating smoothed labels for students and teachers to reduce model uncertainty. |
| Outcome: | The proposed framework improves on pretrained and non-pretrained models on three benchmarks. |
Unsupervised Deep Structured Semantic Models for Commonsense Reasoning (N19-1)
Copied to clipboard
| Challenge: | Existing methods for commonsense reasoning rely on human-crafted features and knowledge bases, but unsupervised learning is not feasible due to the lack of labeled training data or comprehensive knowledge bases. |
| Approach: | They propose two unsupervised models based on the Deep Structured Semantic Models framework to tackle two commonsense reasoning tasks: Winograd Schema Challenge (WSC) and Pronoun Disambiguation (PDP). |
| Outcome: | The proposed models capture contextual information in the sentence and co-reference information between pronouns and nouns, and achieve significant improvement over previous state-of-the-art approaches. |
Open-World Attribute Mining for E-Commerce Products with Multimodal Self-Correction Instruction Tuning (2025.acl-long)
Copied to clipboard
| Challenge: | Current AM methods focus on extracting attributes from unimodal text, underutilizing multimodal data. |
| Approach: | They propose a framework for multimodal self-correction instruction tuning to extract new attributes from images and text with Multimodal Large Language Models. |
| Outcome: | The proposed framework outperforms state-of-the-art methods on two datasets. |
What’s Hidden in a One-layer Randomly Weighted Transformer? (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing subnetworks of one-layer randomly weighted neural networks can achieve impressive performance without changing initializations. |
| Approach: | They find subnetworks within one-layer randomly weighted neural networks that can achieve impressive performance without ever modifying the initializations. |
| Outcome: | The proposed subnetworks match 98%/92% of the performance of a trained Transformersmall/base on IWSLT14/WMT14. |
Summarize before Aggregate: A Global-to-local Heterogeneous Graph Inference Network for Conversational Emotion Recognition (2020.coling-main)
Copied to clipboard
| Challenge: | Existing studies focus on modeling emotion influences with utterance-level features, with little attention paid on phrase-level semantic connection between utterrances. |
| Approach: | They propose a two-stage Summarization and Aggregation Graph Inference Network which integrates inference for topic-related emotional phrases and local dependency reasoning over neighbouring utterances in a global-to-local fashion. |
| Outcome: | The proposed model outperforms the state-of-the-art models on three CER benchmark datasets. |
Learning from the Irrecoverable: Error-Localized Policy Optimization for Tool-Integrated LLM Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Tool-integrated reasoning (TIR) enables LLM agents to solve tasks through planning, tool use, and iterative revision, but outcome-only reinforcement learning suffers from sparse, delayed rewards and weak step-level credit assignment. |
| Approach: | They propose a tool-integrated reasoning approach that localizes the first irrecoverable step and leverages it for fine-grained credit assignment. |
| Outcome: | The proposed algorithm outperforms strong Agentic RL benchmarks in math, science QA, and code execution with additional gains in Pass@K and Major@K scaling, rollout ranking quality, and tool-call efficiency. |
Simple or Complex? Complexity-controllable Question Generation with Soft Templates and Deep Mixture of Experts Model (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work on complex questions does not consider controlling complexity of generated questions. |
| Approach: | They propose an end-to-end neural complexity-controllable question generation model that incorporates a mixture of experts as the selector of soft templates to capture question similarity while avoiding the expensive construction of actual templates. |
| Outcome: | The proposed model is superior to state-of-the-art methods in both automatic and manual evaluations on two benchmark QA datasets. |
Crosslingual Generalization through Multitask Finetuning (2023.acl-long)
Copied to clipboard
Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng Xin Yong, Hailey Schoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak, Samuel Albanie, Zaid Alyafeai, Albert Webson, Edward Raff, Colin Raffel
| Challenge: | Multitask prompted finetuning (MTF) has been shown to help large language models generalize to new tasks in a zero-shot setting, but so far explorations of MTF have focused on English data and models. |
| Approach: | They apply multitask prompted finetuning to pretrained multilingual models and generate variants called BLOOMZ and mT0. |
| Outcome: | The proposed models can generalize to non-English languages that have never been seen before. |
BTS: Harmonizing Specialized Experts into a Generalist LLM (2025.emnlp-main)
Copied to clipboard
Qizhen Zhang, Prajjwal Bhargava, Chloe Bi, Chris X. Cai, Jakob Nicolaus Foerster, Jeremy Fu, Punit Singh Koura, Ruan Silva, Sheng Shen, Emily Dinan, Suchin Gururangan, Mike Lewis
| Challenge: | Branch-Train-Stitch (BTS) is an efficient and flexible training algorithm for combining independently trained large language model (LLM) experts into a single, capable generalist model. |
| Approach: | They propose an efficient and flexible training algorithm for combining large language model (LLM) experts into a single, capable generalist model using lightweight stitch layers. |
| Outcome: | The proposed model can generalize to new domains despite being frozen . it yields the best generalist performance on a variety of downstream tasks, retaining the specialized capabilities of each of the experts. |
Scaling Vision-Language Models with Sparse Mixture of Experts (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a study explores the effectiveness of mixture-of-experts (MoE) techniques in scaling vision-language models . alayrac and colleagues demonstrate the effectiveness and performance of MoE in scaling VLMs . |
| Approach: | They propose to use sparsely-gated mixture-of-experts techniques to scale vision-language models . they show that MoE can achieve state-of the-art performance over dense models a range of benchmarks . |
| Outcome: | The proposed approach achieves state-of-the-art performance over dense models of equivalent computational cost. |
Exposing the Limits of Video-Text Models through Contrast Sets (2022.naacl-main)
Copied to clipboard
| Challenge: | Recent video-text models can retrieve relevant videos based on text with high accuracy, but to what extent do they comprehend the semantics of the text? |
| Approach: | They propose a framework that probes video-text models with hard negatives . they leverage a pre-trained language model and a set of heuristics to create verb and person entity focused contrast sets. |
| Outcome: | The proposed framework erases the performance gap between CLIP-based methods and the earlier methods. |
Pragmatically Informative Text Generation (N19-1)
Copied to clipboard
| Challenge: | Existing approaches to pragmatics have been used to improve the informativeness of generated text in grounded language learning problems. |
| Approach: | They propose to use pragmatics to improve the informativeness of conditional text models . they propose to apply pragmatic reasoning to more traditional language generation tasks . |
| Outcome: | The proposed methods improve the performance of strong existing systems for abstractive summarization and generation from structured meaning representations. |
SQL-Trail: Multi-Turn Reinforcement Learning with Interleaved Feedback for Text-to-SQL (2026.acl-long)
Copied to clipboard
Harper Hua, Zhen Han, Zhengyuan Shen, Meng-Chieh Lee, Sheng Guan, Qi Zhu, Sullam Jeoung, Yueyan Chen, Yunfei Bai, Shuai Wang, Vassilis N. Ioannidis, Huzefa Rangwala
| Challenge: | Recent large language models (LLMs) have significantly improved Text-to-SQL generation, but a gap remains between AI systems and human experts on challenging benchmarks such as BIRD-Sql. |
| Approach: | They propose a multi-turn reinforcement learning agentic framework for Text-to-SQL that uses execution feedback to iteratively refine its predictions. |
| Outcome: | The proposed framework outperforms proprietary systems on 7B and 14B models by **5% on average, underscoring the effectiveness of interactive, agentic workflows for robust Text-to-SQL generation. |
Reservoir Transformers (2021.acl-long)
Copied to clipboard
| Challenge: | Using random initialization, we show that some transformers obtain impressive performance even when some of the layers are frozen. |
| Approach: | They propose to freeze transformer layers and use them to improve performance . they find that the transformers obtain impressive performance even when some of the layers are randomly initialized and never updated. |
| Outcome: | The proposed model improves on translation and language modelling tasks even when some layers are frozen. |
Aligning Large Multimodal Models with Factually Augmented RLHF (2024.findings-acl)
Copied to clipboard
Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, Trevor Darrell
| Challenge: | Large Multimodal Models (LMMs) are built across modalities and the misalignment between two modality can result in "hallucination" . developing LMMs faces challenges such as a lack of data and a limited number of data sets. |
| Approach: | They propose a new algorithm that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options. |
| Outcome: | The proposed approach improves on the LLaVA-Bench dataset with the 96% performance level of the text-only GPT-4 and an improvement of 60% on MMHAL-BENCH over other baselines. |
Hierarchical Chinese Legal event extraction via Pedal Attention Mechanism (2020.coling-main)
Copied to clipboard
| Challenge: | Existing methods for event extraction cannot express connections between arguments, which are crucial in legal events. |
| Approach: | They propose a dynamic event structure for Chinese legal events to distinguish between similar events by hierarchical event features for event detection and a pedal attention mechanism to extract the semantic relation between two words through their dependent adjacent words. |
| Outcome: | The proposed model surpasses state-of-the-art models on a Chinese legal event dataset. |
MAF: Multimodal Alignment Framework for Weakly-Supervised Phrase Grounding (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing work on phrase localization uses caption-image datasets as weak supervision . existing work on supervised phrase localisation uses a large-scale annotated dataset . |
| Approach: | They develop a multimodal alignment framework to leverage more widely available caption-image datasets to model phrase relevance. |
| Outcome: | The proposed model improves on the widely-adopted Flickr30k dataset . it also improves the previous best unsupervised result by 5.56% . |
An annotated dataset of literary entities (N19-1)
Copied to clipboard
| Challenge: | Existing datasets built on news focus on non-named entities, but not literary texts. |
| Approach: | They propose to annotate 210,532 tokens from 100 different English-language literary texts for ACE entity categories (person, location, geo-political entity, facility, organization, and vehicle). |
| Outcome: | The proposed dataset includes 210,532 tokens drawn from 100 different English-language literary texts. |