Papers by Jun Cheng
Unified Thinker: A General Reasoning Core for Image Generation (2026.acl-long)
Copied to clipboard
Sashuai Zhou, Qiang Zhou, Jijin Hu, Hanqing Yang, Yue Cao, Junpeng Ma, Yinchao Ma, Jun Song, Tiezheng Ge, Cheng Yu, Bo Zheng, Zhou Zhao
| Challenge: | generative models struggle with logic-intensive instruction following, exposing a persistent reasoning–execution gap. |
| Approach: | They propose a task-agnostic reasoning architecture for general image generation . they propose pixel-level feedback to ground the Thinker's policy in pixel feedback . |
| Outcome: | The proposed system significantly improves image reasoning and generation quality. |
CommonIT: Commonality-Aware Instruction Tuning for Large Language Models via Data Partitions (2024.emnlp-main)
Copied to clipboard
| Challenge: | Current studies have focused on fine-tuning, but the use of instruction tuning is not as effective as fine-cuning. |
| Approach: | They propose a commonality-aware instruction tuning strategy to cluster instruction datasets into distinct groups with three proposed metrics Task, Embedding and Length. |
| Outcome: | The proposed strategy boosts an average improvement of 2.1% on the general domain and 5.2% on the special domain. |
Gated Differentiable Working Memory for Long-Context Language Modeling (2026.acl-long)
Copied to clipboard
Lingrui Mei, Shenghua Liu, Yiwei Wang, Yuyao Ge, Baolong Bi, Jiayu Yao, Jun Wan, Ziling Yin, Jiafeng Guo, Xueqi Cheng
| Challenge: | Long contexts break transformers, attention scores dilute, model cannot adapt to novel patterns at inference time. |
| Approach: | They propose a framework that gates the memory consolidation process by estimating Contextual Utility . they propose GDWM to maintain a form of working memory with constant contexts . |
| Outcome: | The proposed framework achieves comparable or superior performance on sparse-information tasks with 4 fewer gradient steps compared to uniform baselines. |
Balancing Fidelity and Plasticity: Aligning Mixed-Precision Fine-Tuning with Linguistic Hierarchies (2026.findings-acl)
Copied to clipboard
Changhai Zhou, Shiyang Zhang, Yuhua Zhou, Jun Gao, Qian Qiao, Shichao Weng, Weizhong Zhang, Cheng Jin
| Challenge: | Existing quantization-aware fine-tuning methods decouple weight precision and adapter capacity, overlooking that a layer’s ability to adapt is constrained by the information preserved in its frozen weights. |
| Approach: | They propose a framework that jointly optimizes per-layer quantization bit-width and LoRA rank. |
| Outcome: | Experiments on LLaMA and Qwen models show that the proposed framework matches or approaches 16-bit baselines while using substantially less memory. |
NOSE: Neural Olfactory-Semantic Embedding with Tri-Modal Orthogonal Contrastive Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing representation methods fail to fully capture olfactory pathway . current approaches focus on isolated segments of the olefactory pathways . |
| Approach: | They propose a representation learning framework that aligns molecular structure, receptor sequence, and natural language description. |
| Outcome: | The proposed framework achieves state-of-the-art and excellent zero-shot generalization . it decouples contributions of molecular structure, receptor sequence, and natural language description . |
Curriculum Consistency Learning for Conditional Sentence Generation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Consistency learning (CL) has proven to be a valuable technique for improving the robustness of conditional sentence generation models. |
| Approach: | They propose a strategy that guides models to learn consistency in alignment with their current capacity to differentiate between features. |
| Outcome: | The proposed strategy delivers +2.0 accuracy point improvement compared with vanilla IT and +0.7 COMET scores over traditional CL methods in MT tasks. |
Compensate Quantization Errors: Make Weights Hierarchical to Compensate Each Other (2024.findings-naacl)
Copied to clipboard
| Challenge: | Emergent Large Language Models (LLMs) use extraordinary performance and powerful deduction capacity to discern from traditional language models. |
| Approach: | They propose a method that uses weights to compensate quantization error and learnable singular value incremental (LSI) LSI is a technique that helps weights compensate each other conditioned on activation. |
| Outcome: | The proposed method achieves state-of-the-art performance in diverse quantization settings, no matter in weight-only, weight-activation or extremely low bit scenarios. |
Tailored Sequence to Sequence Models to Different Conversation Scenarios (P18-1)
Copied to clipboard
| Challenge: | Sequence to sequence (Seq2Sequeq) models fail to meet the diverse requirements for different conversation scenarios, such as customer service and chatbot. |
| Approach: | They propose two optimized criteria for Sequence to sequence (Seq2Sequeq) to meet different conversation scenarios, i.e., maximum generated likelihood for specific-requirement scenario, and conditional value-at-risk for diverse-requrement scenarios. |
| Outcome: | The proposed models satisfies diverse requirements for different conversation scenarios and yields better performances than existing models. |
a1: Steep Test-time Scaling Law via Environment Augmented Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have made remarkable advances in reasoning, yet continue to struggle with hallucinations, logical errors, and inability to self-correct during complex multi-step tasks. |
| Approach: | They propose a framework that enhances LLM reasoning through real-time environmental feedback validating each reasoning step, dynamic branch exploration for investigating alternative solution paths when faced with errors, and experience-based learning from successful reasoning trajectories. |
| Outcome: | The proposed model outperforms comparable models by 24.4 percentage points across benchmarks while outperforming comparable models. |
Genius: A Generalizable and Purely Unsupervised Self-Training Framework For Advanced Reasoning (2025.acl-long)
Copied to clipboard
Fangzhi Xu, Hang Yan, Chang Ma, Haiteng Zhao, Qiushi Sun, Kanzhi Cheng, Junxian He, Jun Liu, Zhiyong Wu
| Challenge: | Existing methods for enhancing LLM reasoning rely on supervisory signals . current methods rely heavily on outcome supervision and auxiliary reward models . |
| Approach: | They propose a gen-eralizable and purely unsupervised self-training framework to enhance LLM reasoning without supervision. |
| Outcome: | The proposed framework improves LLM reasoning without supervision without external supervision. |
Read Extensively, Focus Smartly: A Cross-document Semantic Enhancement Method for Visual Documents NER (2022.coling-1)
Copied to clipboard
| Challenge: | Existing methods to extract entities from visually-rich documents ignore the inherent multimodality of VRDs and thus the suboptimal results are achieved. |
| Approach: | They propose a multimodal semantic enhancement method that filters redundant information in the current document and a cross-document information awareness technique to enrich the entity-related context. |
| Outcome: | The proposed method outperforms existing methods on two documents understanding benchmarks covering eight languages. |
APT: Improving Specialist LLM Performance with Weakness Case Acquisition and Iterative Preference Training (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models often require domain-specific fine-tuning to address targeted tasks, which risks degrading their general capabilities. |
| Approach: | They propose to use self-generated dis-preferred weakness data to enhance model performance with a targeted training approach that minimizes interference with existing knowledge base. |
| Outcome: | The proposed approach ensures no reduction in generic capacity and achieves superior performance on downstream tasks compared to existing methods. |
RED: Unleashing Token-Level Rewards from Holistic Feedback via Reward Redistribution (2025.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results demonstrate the superiority of our approach to aligning large language models with human preferences. |
| Approach: | They propose a method that evaluates and assigns specific credit to each token using an off-the-shelf reward model. |
| Outcome: | The proposed method evaluates and assigns specific credit to each token using an off-the-shelf reward model. |
Living in the Moment: Can Large Language Models Grasp Co-Temporal Reasoning? (2024.acl-long)
Copied to clipboard
| Challenge: | Current temporal reasoning datasets are limited to questions about single or isolated events, falling short in mirroring the realistic temporal characteristics involving concurrent nature and intricate temporal interconnections. |
| Approach: | They propose a co-temporal Question Answering benchmark that contains four co-time scenarios with 4,748 samples for evaluating the co-timing abilities of large language models. |
| Outcome: | The proposed benchmarks show that current LLMs struggle on CoTempQA tasks even when enhanced with Chain of Thought methodologies. |
Learning to Control the Specificity in Neural Response Generation (P18-1)
Copied to clipboard
| Challenge: | Existing generative conversational models tend to favor general and trivial responses which appear frequently. |
| Approach: | They propose a controlled response generation mechanism to handle different utterance-response relationships in terms of specificity. |
| Outcome: | The proposed model outperforms state-of-the-art models under automatic and human evaluations. |
SeaPO: Strategic Error Amplification for Robust Preference Optimization of Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for preference optimization of large language models use pairs of positive and negative samples, but the quality of positive samples may become similar during training, complicating preference learning. |
| Approach: | SeaPO introduces error types commonly occurring in large language models to improve preference learning. |
| Outcome: | SeaPO introduces error types into model Preference Optimization to improve model performance . negative samples are more erroneous than positive samples, and preference-based training mitigates errors . |
Uplift-RAG: Uplift-Driven Knowledge Preference Alignment for Retrieval-Augmented Generation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing efforts to estimate document utility rely on downstream generation performance, which conflates the influence of external documents with the intrinsic knowledge of the LLM. |
| Approach: | They propose an uplift-based definition of document utility that quantifies each document’s marginal benefit over the LLM’s internal knowledge. |
| Outcome: | The proposed framework improves the performance of the LLM by incorporating external retrieved documents into the model. |
Interactive Evolution: A Neural-Symbolic Self-Training Framework For Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods to fine-tune Large Language Models without human annotations are lacking in the field of natural language training. |
| Approach: | They propose an environment-guided neural-symbolic self-training framework to overcome two main challenges: the scarcity of symbolic data and the limited proficiency of LLMs in processing symbolic language. |
| Outcome: | The proposed framework overcomes two main challenges: the scarcity of symbolic data, and the limited proficiency of LLMs in processing symbolic language. |
OneGen: Efficient One-Pass Unified Generation and Retrieval for LLMs (2024.findings-emnlp)
Copied to clipboard
Jintian Zhang, Cheng Peng, Mengshu Sun, Xiang Chen, Lei Liang, Zhiqiang Zhang, Jun Zhou, Huajun Chen, Ningyu Zhang
| Challenge: | Recent advances in Large Language Models (LLMs) have significantly enhanced the generative capabilities for various NLP tasks, but they still suffer from hallucinations due to their exclusive reliance on parametric knowledge. |
| Approach: | They propose a framework that integrates retrieval tokens generated autoregressively into a single LLM to handle both tasks simultaneously in a unified forward pass. |
| Outcome: | The proposed framework bridges the traditionally separate training approaches for generation and retrieval by incorporating retrieval tokens generated autoregressively. |
Biomedical Concept Normalization by Leveraging Hypernyms (2021.emnlp-main)
Copied to clipboard
| Challenge: | Biomedical Concept Normalization (BCN) is widely used in biomedical text processing . despite numerous surface variants of biomedically-defined concepts, it remains challenging and unsolved. |
| Approach: | They propose a framework that uses hypernyms and synonyms to facilitate BCN . they use list-wise training to make use of both hypernies and synonym entities . |
| Outcome: | The proposed framework outperforms the state-of-the-art model on the NCBI dataset. |
Learning to Conceal Risk: Controllable Multi-turn Red Teaming for LLMs in the Financial Domain (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as finance where unsafe behavior can lead to serious regulatory risks. |
| Approach: | They propose a black-box multi-turn risk-concealed redteaming framework that progressively conceals surface-level risk while exploiting regulatory-violating behaviors. |
| Outcome: | Experiments on nine widely used LLMs show that the proposed framework achieves 93.19% average attack success rate (ASR) and improves the average ASR to 95.00%. |
PILE: Pairwise Iterative Logits Ensemble for Multi-Teacher Labeled Distillation (2022.emnlp-industry)
Copied to clipboard
Lianshang Cai, Linhao Zhang, Dehong Ma, Jun Fan, Daiting Shi, Yi Wu, Zhicong Cheng, Simiu Gu, Dawei Yin
| Challenge: | Pre-trained language models have been a key part of ranking systems . knowledge distillation is widely used to maintain high performance while keeping efficient computations. |
| Approach: | They propose an algorithm to combine knowledge from multi-teachers and label information to achieve competitive performance in offline and online experiments. |
| Outcome: | The proposed method has been deployed in a real-world commercial search system. |
UEGP: Unified Expert-Guided Pre-training for Knowledge Rekindle (2024.findings-naacl)
Copied to clipboard
Yutao Mou, Kexiang Wang, Jianhe Lin, Dehong Ma, Jun Fan, Daiting Shi, Zhicong Cheng, Gu Simiu, Dawei Yin, Weiran Xu
| Challenge: | Existing paradigms for pre-training and fine-tuning have limitations . knowledge rekindle aims to break through performance upper bounds of experts without introducing additional annotated data. |
| Approach: | They propose a new paradigm for pre-training and fine-tuning that aims to re-incorporate the fine- tuned expert model into the training cycle and break through performance upper bounds of experts. |
| Outcome: | The proposed model breaks through performance upper bounds of experts without additional annotated data. |
Cross-Utterance Conditioned VAE for Non-Autoregressive Text-to-Speech (2022.acl-long)
Copied to clipboard
| Challenge: | Experimental results show that the proposed model improves naturalness and prosody diversity with clear margins. |
| Approach: | They propose a cross-utterance conditional VAE to estimate posterior probability distribution of latent prosody features for each phoneme by conditioning on acoustic features, speaker information, and text features from past and future sentences. |
| Outcome: | The proposed model improves naturalness and prosody diversity with clear margins. |
Dual Activation-Weight Sparsity: A Training-Free Framework for Efficient Large Language Model Compression (2026.acl-long)
Copied to clipboard
Luoyang Sun, Guangyan Li, Cheng Deng, Haifeng Zhang, Jian Zhao, Yongqiang Tang, Wensheng Zhang, Jun Wang
| Challenge: | Large language models (LLMs) excel at natural language tasks but face deployment bottlenecks due to computational demands. |
| Approach: | They propose a training-free framework that exploits activation and weight sparsity . they use a three-tier routing strategy that uses magnitude-based pruning . |
| Outcome: | Experiments on Llama and Mistral models show that DAWS outperforms activation-weight sparsity pruning methods. |
Learning Kernel-Smoothed Machine Translation with Retrieved Examples (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to update deployed models are prone to overfit . however, non-parametric methods are liable to over-fit the retrieved examples . |
| Approach: | They propose to learn Kernel-Smoothed Translation with Example Retrieval (KSTER) this approach allows users to adapt models to emerging cases without retraining . |
| Outcome: | The proposed approach achieves 1.1 to 1.5 BLEU scores over existing methods without retraining . the proposed model is released on https://github.com/jiangqn/KSTER. |