Papers by Fei Tan
HACHIMI: Scalable and Controllable Student Persona Generation via Orchestrated Agents (2026.findings-acl)
Copied to clipboard
| Challenge: | ad-hoc prompting and hand-crafted profiles with limited control over educational theory and population distributions are often used for student personas. |
| Approach: | They propose a framework that generates theory-aligned, quota-controlled personas . they factorize each persona into a theory-anchored educational schema . |
| Outcome: | HACHIMI generates theory-aligned, quota-controlled personas for grades 1-12 . results show near-perfect schema validity, accurate quots, and substantial diversity . |
ReFSQL: A Retrieval-Augmentation Framework for Text-to-SQL Generation (2023.findings-emnlp)
Copied to clipboard
Kun Zhang, Xiexiong Lin, Yuanzhuo Wang, Xin Zhang, Fei Sun, Cen Jianhe, Hexiang Tan, Xuhui Jiang, Huawei Shen
| Challenge: | Existing methods that align natural language with SQL Language underestimate inherent structural characteristics of SQL and lead to structure errors. |
| Approach: | They propose a retrieval-argument framework that aligns natural language with SQL Language and trains one encoder-decoder-based model to fit all questions. |
| Outcome: | The proposed framework improves accuracy and robustness of text-to-SQL generation on five datasets. |
Text Diffusion Model with Encoder-Decoder Transformers for Sequence-to-Sequence Generation (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing diffusion models are applied to continuous feature space while texts are sequences of discrete categorical tokens. |
| Approach: | They propose to use an encoder-decoder Transformer architecture to approach sequence-to-sequence text generation. |
| Outcome: | The proposed model improves on five sequence-to-sequence generation tasks compared to other diffusion-based models regarding text quality and inference time. |
End-to-End Modeling via Information Tree for One-Shot Natural Language Spatial Video Grounding (2022.acl-long)
Copied to clipboard
Mengze Li, Tianbao Wang, Haoyu Zhang, Shengyu Zhang, Zhou Zhao, Jiaxu Miao, Wenqiao Zhang, Wenming Tan, Jin Wang, Peng Wang, Shiliang Pu, Fei Wu
| Challenge: | Existing methods for grounding video frames with dense annotations require enormous amount of human effort. |
| Approach: | They propose to ground natural language in video frames with only one frame labeled . they propose an end-to-end model that eliminates interference of irrelevant frames . |
| Outcome: | The proposed model can ground natural language in all video frames with only one frame labeled . the proposed model eliminates interference of irrelevant frames based on branch search and cropping techniques . |
LightNER: A Lightweight Tuning Paradigm for Low-resource NER via Pluggable Prompting (2022.coling-1)
Copied to clipboard
Xiang Chen, Lei Li, Shumin Deng, Chuanqi Tan, Changliang Xu, Fei Huang, Luo Si, Huajun Chen, Ningyu Zhang
| Challenge: | Existing approaches for Named Entity Recognition (NER) use extensive labeled data for model training, which struggles in low-resource scenarios. |
| Approach: | They propose a lightweight tuning paradigm for low-resource NER via pluggable prompting . they construct a learnable verbalizer of entity categories without any label-specific classifiers . |
| Outcome: | The proposed model outperforms baselines and class transfer models in low-resource scenarios. |
CWSeg: An Efficient and General Approach to Chinese Word Segmentation (2023.acl-industry)
Copied to clipboard
| Challenge: | Existing methods for Chinese word segmentation have achieved state-of-the-art performance, but they pose challenges in the deployment. |
| Approach: | They propose to augment PLM-based Chinese word segmentation schemes by developing cohort training and versatile decoding strategies. |
| Outcome: | The proposed model can be used to augment existing PLM-based models and improve their performance on Chinese LLaMA and Alpaca datasets. |
CMR Scaling Law: Predicting Critical Mixture Ratios for Continual Pre-training of Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) excel in diverse tasks but often underperform in specialized fields due to limited domain-specific or proprietary corpus. |
| Approach: | They propose a power-law relationship between loss, mixture ratio, and training tokens scale and formalize the trade-off between general and domain-specific capabilities. |
| Outcome: | The proposed model achieves the desired domain transfer while maintaining general ability and highest utilization of available resources. |
Good Visual Guidance Make A Better Extractor: Hierarchical Visual Prefix for Multimodal Entity and Relation Extraction (2022.findings-naacl)
Copied to clipboard
Xiang Chen, Ningyu Zhang, Lei Li, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, Luo Si, Huajun Chen
| Challenge: | Existing approaches for named entity recognition and relation extraction suffer from error sensitivity when irrelevant object images are incorporated in texts. |
| Approach: | They propose a hierarchical visual prefix fusion NeTwork for visual-enhanced entity and relation extraction using pluggable visual prefixed visual features. |
| Outcome: | The proposed method achieves state-of-the-art on three benchmark datasets. |
SDA: Simple Discrete Augmentation for Contrastive Sentence Representation Learning (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for data augmentation have not been well explored. |
| Approach: | They propose to use punctuation insertion, modal verbs, and double negation to produce diverse forms of sentences. |
| Outcome: | The proposed methods perform better on diverse datasets with semantic similarity and standard negation. |
Raise a Child in Large Language Model: Towards Effective and Generalizable Fine-tuning (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent pretrained language models extend from millions to billions of parameters. |
| Approach: | They propose a technique which forwards on a whole network while backwarding on resetting the gradients of the non-child network during the backward process. |
| Outcome: | The proposed technique outperforms the vanilla fine-tuning technique on various downstream tasks and can achieve better generalization performance by large margins. |
Towards Unified Prompt Tuning for Few-shot Text Classification (2022.findings-emnlp)
Copied to clipboard
Jianing Wang, Chengyu Wang, Fuli Luo, Chuanqi Tan, Minghui Qiu, Fei Yang, Qiuhui Shi, Songfang Huang, Ming Gao
| Challenge: | Prompt-based fine-tuning has boosted performance of Pre-trained Language Models (PLMs) on few-shot text classification, but PLMs are unfamiliar with prompt-style expressions during pre-training, which limits the few- shot learning performance on downstream tasks. |
| Approach: | They propose a framework for prompt-based fine-tuning that captures prompting semantics from non-target NLP datasets and propose 'Prompt-Options-Verbalizer' for joint prompt learning across different NLP tasks. |
| Outcome: | Experiments show that the proposed framework outperforms state-of-the-art prompt-based fine-tuning frameworks on few-shot text classification tasks. |
The Mirage of Model Editing: Revisiting Evaluation in the Wild (2025.acl-long)
Copied to clipboard
| Challenge: | despite near-perfect results, effectiveness of model editing in real-world applications remains unclear. |
| Approach: | They propose QAEdit and WILD to better reflect real-world use of model editing . they propose a benchmark aligned with widely used question answering datasets and a task-agnostic evaluation framework . |
| Outcome: | The proposed QAEdit benchmark and WILD evaluation framework show that current models perform worse than previously reported. |
Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for depression assessment rely on standardized ratings, but they are time-consuming and subject to inter-rater variability. |
| Approach: | They propose a fine-grained model for subscore prediction via multi-task learning that can be used to predict depression severity using multiple tasks. |
| Outcome: | The proposed model outperforms baselines and Qwen3-14B direct scoring on the public E-DAIC dataset and to a large-scale private clinical dataset. |
Deeply Coupled Cross-Modal Prompt Learning (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing prompt-tuning methods focus on language branch or learn vision-language interaction in a shallow mechanism. |
| Approach: | They propose a Deeply coupled Cross-modal Prompt learning method based on CLIP to facilitate the interplay between vision and language with a Cross-Modal Prompting Attention mechanism. |
| Outcome: | The proposed method enables the interplay between vision and language with a Cross-Modal Prompt Attention mechanism. |
TNT: Text Normalization based Pre-training of Transformers for Content Moderation (2020.emnlp-main)
Copied to clipboard
| Challenge: | Language model pre-training (self-supervised or unsupervised learning) has been widely used in a multitude of language processing tasks such as named entity recognition, sentiment analysis, question answering and content moderation. |
| Approach: | They propose a new language pre-training model TNT for content moderation that uses a combination of masking strategy and text normalization to learn from text. |
| Outcome: | The proposed model outperforms baselines on hate speech classification task and is a potential approach to misspelling correction. |
Knowledge Rumination for Pre-trained Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have shown that pre-trained language models lack the capacity to handle knowledge-intensive tasks alone. |
| Approach: | They propose a new paradigm to help pre-trained language models utilize latent knowledge without retrieving it from external corpus. |
| Outcome: | The proposed paradigm can be applied to pre-trained language models without retrieving external knowledge from the corpus. |
Reward Difference Optimization For Sample Reweighting In Offline RLHF (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are becoming more capable, but their maximum likelihood objective for the next token prediction falls short in capturing such crucial human values. |
| Approach: | They propose a reward difference prediction method that uses reward difference coefficients to reweigh sample pairs in offline RLHF and a difference model that considers rich interactions between a pair of responses. |
| Outcome: | The proposed method is effective in both automatic metrics and human evaluation. |
PRDetect: Perturbation-Robust LLM-generated Text Detection Based on Syntax Tree (2025.findings-naacl)
Copied to clipboard
| Challenge: | Recent methods for detecting LLM-generated text have shown impressive performance, but in real-world scenarios, users often introduce perturbations to the text. |
| Approach: | They propose a method that detects syntactic trees that are minimally affected by perturbations and exhibit distinct differences between human-written and LLM-generated text. |
| Outcome: | The proposed method shows that it is significantly better against perturbations on the HC3 and GPT-3.5-mixed datasets and also has the shortest time expenditure. |
Repulsive Attention: Rethinking Multi-head Attention as Bayesian Inference (2020.emnlp-main)
Copied to clipboard
Bang An, Jie Lyu, Zhenyi Wang, Chunyuan Li, Changwei Hu, Fei Tan, Ruiyi Zhang, Yifan Hu, Changyou Chen
| Challenge: | Existing studies show that multi-head attention is an effective module in deep neural networks, but there are no explicit mechanisms guaranteeing this property. |
| Approach: | They propose a non-parametric approach that explicitly improves the repulsiveness in multi-head attention and consequently strengthens model’s expressiveness. |
| Outcome: | The proposed approach improves the repulsiveness in multi-head attention and strengthens model’s expressiveness. |
HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing memory systems rely on vector similarity for retrieval, resulting in bloated evidence sets . existing systems produce little additional recall, but this approach lowers retrieval precision . |
| Approach: | They propose a two-level event-turn memory system that uses event summaries as semantic anchors to predict which related turns are worth reading. |
| Outcome: | The proposed system achieves the best F1 on four of five question categories and improves adversarial F1 from 0.54 to 0.78 over A-Mem while retrieving an order of magnitude fewer turns. |
Reasoning with Language Model Prompting: A Survey (2023.acl-long)
Copied to clipboard
Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, Huajun Chen
| Challenge: | Reasoning is an essential ability for complex problem-solving and can provide back-end support for various real-world applications. |
| Approach: | They present cutting-edge research on reasoning with language model prompting and provide systematic resources to help beginners. |
| Outcome: | The proposed approaches have not been systematically reviewed and analyzed. |
Consultant Decoding: Yet Another Synergistic Mechanism (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have attracted widespread attention and adoption across diverse domains due to their exceptional performance and robust generalization abilities. |
| Approach: | They propose a synergetic mechanism for Consultant Decoding (CD) that achieves a 2.5-fold increase in inference speed compared to the target model while maintaining comparable generation quality. |
| Outcome: | The proposed mechanism achieves 2.5-fold increase in inference speed while maintaining comparable generation quality (100% of the target model’s performance). |
Balancing Speciality and Versatility: a Coarse to Fine Framework for Supervised Fine-tuning Large Language Model (2024.findings-acl)
Copied to clipboard
| Challenge: | Aligned Large Language Models exhibit remarkable versatility, capable of handling diverse real-world tasks. |
| Approach: | They propose a coarse to fine framework to fine-tune aligned Large Language Models to achieve a balance between speciality and versatility. |
| Outcome: | The proposed framework outperforms baseline methods across diverse tasks and model scales. |
HyPe: Better Pre-trained Language Model Fine-tuning with Hidden Representation Perturbation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing techniques to fine-tune pre-trained language models on downstream tasks are inadequate. |
| Approach: | They propose a technique to perturb hidden Transformers representations by enhancing generalization of hidden representations from different layers. |
| Outcome: | The proposed technique outperforms vanilla fine-tuning and enhances generalization of hidden representations from different layers. |
UI-Copilot: Advancing Long-Horizon GUI Automation via Tool-Integrated Policy Optimization (2026.acl-long)
Copied to clipboard
Zhengxi Lu, Fei Tang, Guangyi Liu, Jin Ma, Kaitao Song, Xu Tan, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
| Challenge: | Experimental results show that UI-Copilot-7B achieves state-of-the-art performance on challenging MemGUI-Bench, outperforming strong 7B-scale GUI agents such as GUI-Owl-7B and UITARS-1.5-7B. |
| Approach: | They propose a collaborative framework where the GUI agent focuses on task execution while a lightweight copilot provides on-demand assistance for memory retrieval and numerical computation. |
| Outcome: | The proposed framework outperforms GUI-Owl-7B and UI-TARS-1.5-7B on MemGUI-Bench and delivers 17.1% improvement on AndroidWorld over the base Qwen model. |
Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs (2025.emnlp-main)
Copied to clipboard
Hexiang Tan, Fei Sun, Sha Liu, Du Su, Qi Cao, Xin Chen, Jingang Wang, Xunliang Cai, Yuanzhuo Wang, Huawei Shen, Xueqi Cheng
| Challenge: | Existing detection methods fail to account for **self-consistent error** . study identifies self-consistency errors and evaluates them . |
| Approach: | They propose a method that fuses hidden state evidence from an external verifier LLM to detect self-consistent errors. |
| Outcome: | The proposed method significantly enhances performance on self-consistent errors across three LLM families. |
Blinded by Generated Contexts: How Language Models Merge Generated and Retrieved Contexts When Knowledge Conflicts? (2024.acl-long)
Copied to clipboard
| Challenge: | Recent advances in augmenting Large Language Models (LLMs) with auxiliary information have significantly revolutionized their efficacy in knowledge-intensive tasks. |
| Approach: | They propose a systematic framework to identify whether LLMs’ responses are attributed to either generated or retrieved contexts. |
| Outcome: | The proposed framework identifies whether LLMs’ responses are attributed to either generated or retrieved contexts. |
BERT-Beta: A Proactive Probabilistic Approach to Text Moderation (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to text moderation are reactive and do not account for user generated content. |
| Approach: | They propose a text toxicity propensity model to characterize extent to which a user generated text attracts toxic comments and introduce a beta regression model to do the probabilistic modeling. |
| Outcome: | The proposed model performs well in comprehensive experiments and is scalable. |
Beyond Error Propagation in Neural Machine Translation: Characteristics of Language Also Matter (D18-1)
Copied to clipboard
| Challenge: | Neural machine translation suffers from exposure bias and error propagation problem. |
| Approach: | They conduct a series of analyses to deeply understand the accuracy drop problem . they find that the left part of the translated sentence is often better than its right part . |
| Outcome: | The results show that the left part of the translated sentence is often better than its right part in left-to-right decoding models. |
Fusing Heterogeneous Factors with Triaffine Mechanism for Nested Named Entity Recognition (2022.findings-acl)
Copied to clipboard
| Challenge: | Named entity recognition (NER) is a fundamental natural language processing task that extracts entities from texts. |
| Approach: | They propose a triaffine mechanism which integrates heterogeneous factors into a single model to fuse these factors into one model to achieve better span representation. |
| Outcome: | The proposed method outperforms previous span-based methods and achieves state-of-the-art F1 scores on nested NER datasets GENIA and KBP2017. |
daDPO: Distribution-Aware DPO for Distilling Conversational Abilities (2025.findings-acl)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) with Direct Preference Optimization (DPO) has emerged as a promising approach to enhance the conversational abilities of smaller models using a larger teacher model. |
| Approach: | They propose a framework that integrates the teacher's distributional information into DPO distillation while preserving theoretical guarantees. |
| Outcome: | The proposed framework outperforms existing methods in restoring performance for pruned models and enhancing smaller models within the same LLM family. |
Lost in Diffusion: Uncovering Hallucination Patterns and Failure Modes in Diffusion Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | dLLMs have emerged as a promising non-autoregressive paradigm for text generation, but their hallucination mechanisms remain underexplored. |
| Approach: | They present the first controlled comparative study to evaluate hallucination patterns in Diffusion Large Language Models. |
| Outcome: | The proposed model exhibits higher propensity for hallucination than AR counterparts controlled for architecture, scale, and pre-training weights. |
Cross-lingual Cross-modal Pretraining for Multimodal Retrieval (2021.naacl-main)
Copied to clipboard
| Challenge: | Recent pretrained vision-language models have achieved impressive performance on cross-modal retrieval tasks in English. |
| Approach: | They propose a new approach to learn cross-lingual cross-modal representations for matching images and captions in multiple languages using an annotated corpus. |
| Outcome: | The proposed model achieves impressive performance on two multimodal multilingual image caption benchmarks: Multi30k with German captions and MSCOCO with Japanese captions. |
Reverse Constitutional AI: A Framework for Controllable Toxic Data Generation via Probability-Clamped RLAIF (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to red teaming focus on searching for individual adversarial inputs. |
| Approach: | They propose a framework for automated adversarial data generation that inverts harmless constitution into constitution of toxicity and iteratively refining model outputs through critique–revision pipeline. |
| Outcome: | The proposed framework generates diverse, high-quality toxic data without human annotation and significantly improves semantic coherence without sacrificing adversarial strength. |
HABERTOR: An Efficient and Effective Deep Hatespeech Detector (2020.emnlp-main)
Copied to clipboard
| Challenge: | HABERTOR model is a highly efficient and effective alternative to BERT for the hatespeech classification task. |
| Approach: | They propose to modify BERT's HABERTOR model to generate its own vocabularies and pre-trained it using the largest scale hatespeech dataset. |
| Outcome: | The proposed model is faster, more efficient and more robust than existing methods for hatespeech classification. |
Parallel Instance Query Network for Named Entity Recognition (2022.acl-long)
Copied to clipboard
Yongliang Shen, Xiaobin Wang, Zeqi Tan, Guangwei Xu, Pengjun Xie, Fei Huang, Weiming Lu, Yueting Zhuang
| Challenge: | Named entity recognition is a fundamental task in natural language processing. |
| Approach: | They propose a method that sets up global and learnable instance queries to extract entities from a sentence in a parallel manner. |
| Outcome: | The proposed method outperforms existing state-of-the-art models on nested and flat datasets. |
CBLUE: A Chinese Biomedical Language Understanding Evaluation Benchmark (2022.acl-long)
Copied to clipboard
Ningyu Zhang, Mosha Chen, Zhen Bi, Xiaozhuan Liang, Lei Li, Xin Shang, Kangping Yin, Chuanqi Tan, Jian Xu, Fei Huang, Luo Si, Yuan Ni, Guotong Xie, Zhifang Sui, Baobao Chang, Hui Zong, Zheng Yuan, Linfeng Li, Jun Yan, Hongying Zan, Kunli Zhang, Buzhou Tang, Qingcai Chen
| Challenge: | a new benchmark for biomedical language understanding is being developed in Chinese . most benchmarks are limited to English, which makes it difficult to replicate success in other languages. |
| Approach: | They propose to use Chinese biomedical language understanding evaluation benchmarks to evaluate Chinese models. |
| Outcome: | The proposed benchmarks show that the current models perform worse than the human ceiling. |
What Makes Pre-trained Language Models Better Zero-shot Learners? (2023.acl-long)
Copied to clipboard
| Challenge: | Current methods for prompt learning in zero-shot scenarios rely on a development set with sufficient human-annotated data to select the best-performing prompt template. |
| Approach: | They propose a method for screening reasonable prompt templates in zero-shot text classification using language discrepancy. |
| Outcome: | The proposed method improves prediction performance in a realistic zero-shot setting, eliminating the need for labelled examples. |
The Fall of ROME: Understanding the Collapse of LLMs in Model Editing (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have found that model editing methods can cause large language models to collapse with just a single edit. |
| Approach: | They propose a method that uses prefixed keys and adds prefixes during testing to prevent model collapse. |
| Outcome: | The proposed method prevents model collapse while maintaining effectiveness, the authors show . Rank-One Model Editing (ROME) has been found to cause model collapse with just a single edit . |
BaseCal: Unsupervised Confidence Calibration via Base Model Signals (2026.acl-long)
Copied to clipboard
Hexiang Tan, Wanli Yang, Junwei Zhang, Xin Chen, Rui Tang, Du Su, Jingang Wang, Yuanzhuo Wang, Fei Sun, Xueqi Cheng
| Challenge: | Post-trained LLMs typically compromise reliability with severe overconfidence, resulting in inaccurate responses. |
| Approach: | They propose a solution that feeds PoLLMs into the base LLM to get confidence. |
| Outcome: | The proposed solution reduces expected calibration error (ECE) by 42.90% compared to the best unsupervised baselines. |
More Data or Better Data? A Critical Analysis of Data Selection and Synthesis for Mathematical Reasoning (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Despite various proposed data construction methods, their practical utility in real-world pipelines remains underexplored. |
| Approach: | They conduct a comprehensive analysis of open-source datasets and data synthesis techniques for mathematical reasoning under a unified pipeline designed to mirror training and deployment scenarios. |
| Outcome: | The proposed pipelines mirror training and deployment scenarios and are suitable for industrial applications. |
TextFlint: Unified Multilingual Robustness Evaluation Toolkit for Natural Language Processing (2021.acl-demo)
Copied to clipboard
Xiao Wang, Qin Liu, Tao Gui, Qi Zhang, Yicheng Zou, Xin Zhou, Jiacheng Ye, Yongxin Zhang, Rui Zheng, Zexiong Pang, Qinzhuo Wu, Zhengyan Li, Chong Zhang, Ruotian Ma, Zichu Fei, Ruijian Cai, Jun Zhao, Xingwu Hu, Zhiheng Yan, Yiding Tan, Yuan Hu, Qiyuan Bian, Zhihua Liu, Shan Qin, Bolin Zhu, Xiaoyu Xing, Jinlan Fu, Yue Zhang, Minlong Peng, Xiaoqing Zheng, Yaqian Zhou, Zhongyu Wei, Xipeng Qiu, Xuanjing Huang
| Challenge: | Existing approaches to textual robustness evaluation focus on slightly modifying the input data, which maintains the original meaning and results in a different prediction. |
| Approach: | They propose a multilingual robustness evaluation toolkit for NLP that integrates universal text transformations, task-specific transformations and adversarial attack. |
| Outcome: | The toolkit includes universal text transformation, task-specific transformation, adversarial attack, subpopulation, and their combinations to provide comprehensive robustness analyses. |
ReSURE: Regularizing Supervision Unreliability for Multi-turn Dialogue Fine-tuning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods address data quality via static prefiltering, which decouples quality control from training and fails to mitigate turn-level error propagation. |
| Approach: | They propose an adaptive learning method that dynamically down-weights unreliable supervision without explicit filtering. |
| Outcome: | Experiments on single-source and mixed-quality datasets show improved stability and response quality. |