Papers with fidelity
Private Seeds, Public LLMs: Realistic and Privacy-Preserving Synthetic Data Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are a powerful tool for creating synthetic replicas of private text. |
| Approach: | They propose a method for creating privacy-preserving synthetic data using private seeds and a formal differential privacy mechanism. |
| Outcome: | The proposed method achieves high fidelity to private data while providing strong privacy protection. |
Fact-based Text Editing (2020.acl-main)
Copied to clipboard
| Challenge: | Existing methods for fact-based text editing are not suitable for all situations. |
| Approach: | They propose a method for automatically generating a dataset with a draft text, revised text, and several facts represented in triples. |
| Outcome: | The proposed method outperforms the encoder-decoder approach on two datasets and shows that it conducts inference faster than the encoded-decoding approach. |
SciTrue: Evidence-Grounded Claim Verification in Science (2026.eacl-demo)
Copied to clipboard
| Challenge: | Existing systems often exhibit unverifiable attributions, shallow evidence mapping, and hallucinated citations. |
| Approach: | They propose a claim verification system that provides source-level accountability and evidence traceability. |
| Outcome: | SciTrue outperforms RAG-based baselines in summary traceability, attribution accuracy, and context alignment in a human evaluation of 300 attributions. |
PatentVision: A multimodal method for drafting patent applications (2026.eacl-industry)
Copied to clipboard
| Challenge: | PatentVision integrates textual and visual inputs to generate patent specifications . existing systems fail to capture the nuanced interplay between textual, visual components . |
| Approach: | They propose a multimodal framework that integrates textual and visual inputs to generate patent specifications. |
| Outcome: | The proposed framework surpasses text-only methods in patent writing, the authors show . it integrates visual data to better represent intricate design features and functional connections . |
Make Templates Smarter: A Template Based Data2Text System Powered by Text Stitch Model (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Neural network based data2text models drop or modify information in inputs and it is hard to control the generated contents. |
| Approach: | They propose a template-based data2text system powered by a text stitch model that automatically stitches adjacent template units. |
| Outcome: | The proposed system outperforms template-based systems in fidelity and human involvement on a benchmark dataset. |
MR-ALIGN: Meta-Reasoning Informed Factuality Alignment for Large Reasoning Models (2026.findings-acl)
Copied to clipboard
Xinming Wang, Jian Xu, Bin Yu, Sheng Lian, yi Chen, Boran Wang, Yingjian Zhu, Hongzhu Yi, Hong-Ming Yang, Han Hu, Cheng-Lin Liu, Xu-Yao Zhang
| Challenge: | Large reasoning models (LRMs) show strong capabilities in complex reasoning, yet their marginal gains on evidence-dependent factual questions are limited. |
| Approach: | They propose a Meta-Reasoning informed alignment framework that quantifies state-transition probabilities along the model’s thinking process and constructs a transition-aware implicit reward that reinforces beneficial reasoning patterns while suppressing defective ones at the atomic thinking segments. |
| Outcome: | Empirical evaluations of four factual QA datasets and one long-form factuality benchmark show that MR-ALIGN consistently improves accuracy and truthfulness while reducing misleading reasoning. |
Revitalizing Black-Box Interpretability: Actionable Interpretability for LLMs via Proxy Models (2026.acl-long)
Copied to clipboard
| Challenge: | Applying model-agnostic explanations to Large Language Models is hindered by prohibitive computational costs rendering them dormant for real-world applications. |
| Approach: | They propose a budget-friendly proxy framework that leverages efficient models to approximate the decision boundaries of expensive Large Language Models. |
| Outcome: | The proposed framework achieves over 90% fidelity with only 9.5% of the oracle’s cost and is open-source to facilitate future research. |
Multimodal Pragmatic Jailbreak on Text-to-image Models (2025.acl-long)
Copied to clipboard
Tong Liu, Zhixin Lai, Jiawen Wang, Gengyuan Zhang, Shuo Chen, Philip Torr, Vera Demberg, Volker Tresp, Jindong Gu
| Challenge: | Existing jailbreaks for diffusion-based text-to-image models generate unsafe content . experimental results show that all tested models suffer from unsafe generation . |
| Approach: | They propose a jailbreak that triggers diffusion-based text-to-image models to generate the image with visual text, resulting in unsafe content. |
| Outcome: | The proposed model generates image with visual text, but the model is unsafe under such jailbreak. |
Persistent Personas? Role-Playing, Instruction Following, and Safety in Extended Interactions (2026.eacl-long)
Copied to clipboard
| Challenge: | Persona-assigned large language models are used in education, healthcare and sociodemographic simulations. |
| Approach: | They propose a protocol that combines long persona dialogues and evaluation datasets to create dialogue-conditioned benchmarks that can robustly measure long-context effects. |
| Outcome: | The proposed protocol can measure persona fidelity, instruction-following, and safety in long conversations. |
L-CiteEval: A Suite for Evaluating Fidelity of Long-context Models (2025.acl-long)
Copied to clipboard
| Challenge: | Long-context models (LCMs) have seen remarkable advancements in recent years, facilitating tasks like long-document QA. |
| Approach: | They propose an out-of-the-box suite that can assess both generation quality and fidelity in long-context understanding tasks. |
| Outcome: | The proposed suite can assess both generation quality and fidelity in long-context understanding tasks. |
VideoEraser: Concept Erasure in Text-to-Video Diffusion Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results show that VideoEraser outperforms prior methods regarding efficacy, integrity, fidelity, robustness, and generalizability. |
| Approach: | They propose a training-free framework that prevents T2V diffusion models from generating videos with undesirable concepts even when explicitly prompted with those concepts. |
| Outcome: | The proposed framework outperforms existing methods in erasure, celebrity erasion, and explicit content erasing tasks. |
Generating Realistic Natural Language Counterfactuals (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to explain ML tasks for natural language text are either unrealistic or introduce imperceptible changes. |
| Approach: | They propose a method that combines a conditional GAN and embeddings of a pretrained BERT encoder to model-agnostically generate realistic natural language text counterfactuals for explaining regression and classification tasks. |
| Outcome: | The proposed method outperforms baseline methods on fidelity and human judgments of naturalness across multiple datasets and multiple predictive models. |
Beyond Blind Following: Evaluating Robustness of LLM Agents under Imperfect Guidance (2026.eacl-long)
Copied to clipboard
Yao Fu, Ran Qiu, Xinhe Wang, Jacob Sansom, Sathvika Ayyappa Prabhu, Huijie Tang, Jaekyeom Kim, Sungryull Sohn, Honglak Lee
| Challenge: | Large language models (LLMs) have shown strong capabilities as task-solving agents across interactive domains, but in complex environments, auxiliary guidance may be imperfect. |
| Approach: | They propose a benchmark to measure the robustness of large language models under imperfect guidance. |
| Outcome: | The proposed benchmark compared LLM agents in navigation, cooking, and gaming in a variety of environments with auxiliary guidance and noisy or underspecified instructions extracted from demonstrations. |
EmoCharacter: Evaluating the Emotional Fidelity of Role-Playing Agents in Dialogues (2025.naacl-long)
Copied to clipboard
| Challenge: | EmoCharacter evaluates emotional fidelity of role-playing agents in dialogues . current evaluations focus on personality fidelity, tone imitation, and knowledge consistency . |
| Approach: | They propose a benchmark to assess emotional fidelity of role-playing agents in dialogues using large language models. |
| Outcome: | The proposed benchmark measures emotional fidelity of role-playing agents and the characters they portray. |
PROTEGE: Prompt-based Diverse Question Generation from Web Articles (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a popular format for knowledge bases is question-answer pairs (Q&As) specialized knowledge bases that extract and store question-annwer pairs are prevalent . |
| Approach: | They propose a framework for question generation that generates diverse questions from text . they propose 'protege' framework that can generate diverse questions using a variety of prompts . |
| Outcome: | The proposed framework improves diversity and fidelity over diverse beam search and prompt-based baselines on three public Q&A datasets. |
From Word to World: Can Large Language Models be Implicit Text-based World Models? (2026.acl-long)
Copied to clipboard
Yixia Li, Hongru Wang, Jiahao Qiu, Zhenfei Yin, Dongdong Zhang, Cheng Qian, Zeping Li, Xiaoteng Ma, Guanhua Chen, Heng Ji
| Challenge: | Agentic learning increasingly hinges on interaction, yet real-world experience is expensive, limited, and often irreversible at inference time. |
| Approach: | They propose a framework that reframes language modeling as next-state prediction under interaction. |
| Outcome: | The proposed framework evaluates world models in text-based environments . it shows that sufficiently trained models capture coherent environment dynamics . |
CATCH: A Novel Data Synthesis Framework for High Therapy Fidelity and Memory-Driven Planning Chain of Thought in AI Counseling (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies employ a one-time generation approach to synthesize multi-turn dialogue samples, resulting in low therapy fidelity and failing to capture decision-making rationale behind each response. |
| Approach: | They propose a data synthesis framework that synthesizes multi-turn dialogue samples and incrementally generates stage-aligned counseling dialogues. |
| Outcome: | The proposed framework significantly improves therapy fidelity and logical coherence in AI counseling. |
Local Interpretation of Transformer Based on Linear Decomposition (2023.acl-long)
Copied to clipboard
| Challenge: | Existing work on local explanation generation attempts to understand model dynamics on word-level or phraselevel by assigning importance scores on input features. |
| Approach: | They propose to interpret neural networks by linear decomposition by a Transformer model on a single input and a linear decomposing of the output to generate local explanations. |
| Outcome: | The proposed method achieves competitive performance in sentiment classification and machine translation, and fidelity of explanation. |
Data and Representation for Turkish Natural Language Inference (2020.emnlp-main)
Copied to clipboard
| Challenge: | Large annotated datasets in NLP are overwhelmingly in English . obtaining new annotation resources for each task in each language would be prohibitively expensive . |
| Approach: | They propose to use machine translation to translate large annotated datasets into Turkish . they find that in-language embeddings are essential and morphological parsing can be avoided . |
| Outcome: | The proposed model trains on human-translated evaluation sets. |
Logical Natural Language Generation from Open-Domain Tables (2020.acl-main)
Copied to clipboard
| Challenge: | Existing studies on neural natural language generation focus on surface-level realizations with limited emphasis on logical inference. |
| Approach: | They propose a task where a model is tasked with generating natural language statements that can be logically entailed by facts in an open-domain semi-structured table. |
| Outcome: | The proposed task is based on the existing TabFact dataset with a wide range of logical/symbolic inferences. |
Bypassing Neural Evaluations for Fast Audio Editing via Adaptive Trajectory Extrapolation (2026.findings-acl)
Copied to clipboard
Xiaoqian Liu, Zhengkun Ge, Jianjin Wang, Haoran Zhang, Yuan Ge, Kaiyan Chang, Chen Xu, Tong Xiao, Zhengtao Yu, Linfeng Zhang, JingBo Zhu
| Challenge: | Recent advances in audio diffusion models have significantly improved text-to-audio editing via inversion techniques, but these models typically rely on dense, fixed-step sampling trajectories to maintain structural integrity. |
| Approach: | They propose a model-agnostic Adaptive Trajectory Extrapolation framework that accelerates inversion-based editing process by dynamically evaluating only the most critical generative phases. |
| Outcome: | The proposed framework achieves a 3.9 speedup with negligible loss in fidelity. |
Distributional Alignment for Large Language Models under Domain Shift (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing distributional alignment models are unstable and degrade under cultural and domain shifts. |
| Approach: | They propose a distributional alignment technique that improves distribution prediction under cultural and domain shift. |
| Outcome: | The proposed method improves fidelity and robustness of LLM distribution estimation under domain and cultural shift. |
BiMol-Diff: A Unified Diffusion Framework for Molecular Generation and Captioning (2026.acl-long)
Copied to clipboard
| Challenge: | Autoregressive models struggle with long-range dependencies, while standard diffusion processes apply uniform corruption across positions. |
| Approach: | They propose a framework that bridges molecular structures and natural language . it assigns position-dependent corruption based on token recovery difficulty . |
| Outcome: | The proposed framework improves molecule reconstruction and captioning performance on two datasets. |
Discovering Biases in Information Retrieval Models Using Relevance Thesaurus as Global Explanation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Currently, local explanations are not effective in predicting the model’s behavior on unseen texts. |
| Approach: | They propose a method to build a relevance thesaurus containing semantically relevant query term and document term pairs which can augment BM25 scoring functions to better approximate the neural model’s predictions. |
| Outcome: | The proposed method can augment BM25 scoring functions to better approximate the neural relevance model’s predictions. |
YinYang-Align: A new Benchmark for Competing Objectives and Introducing Multi-Objective Preference based Text-to-Image Alignment (2025.findings-acl)
Copied to clipboard
Amitava Das, Yaswanth Narsupalli, Gurpreet Singh, Vinija Jain, Vasu Sharma, Suranjana Trivedy, Aman Chadha, Amit Sheth
| Challenge: | Recent controversies highlight the need for robust alignment mechanisms in text-to-image systems. |
| Approach: | They propose a framework to evaluate T2I systems across six contradictory alignment objectives . objectives highlight key trade-offs such as artistic freedom and cultural sensitivity . |
| Outcome: | The proposed framework achieves superior alignment across all objectives. |
Mitigating Open-Vocabulary Caption Hallucinations (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for image captioning ignore the long-tailed nature of hallucinations . a new framework is proposed to address hallucines in image captions in the open-vocabulary setting . |
| Approach: | They propose a framework to address hallucinations in image captioning in the open-vocabulary setting. |
| Outcome: | The proposed framework surpasses the CHAIR benchmark in diversity and accuracy in open-vocabulary captioning. |
EoT: Evolution of Thoughts for Complex Reasoning Tasks (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies focus on ensuring behavior fidelity, factuality or reliability in generated reasoning processes, but they neglect the simultaneous optimization of all three aspects for each thought. |
| Approach: | They propose a thought assessment method that is sensitive to knowledge and LLM behaviors . they use three scorers to evaluate each thought by considering domain context, semantic alignment, and behavior impact. |
| Outcome: | The proposed framework outperforms advanced approaches in knowledge-based complex tasks. |
Language Model Based Text-to-Audio Generation: Anti-Causally Aligned Collaborative Residual Transformers (2025.emnlp-main)
Copied to clipboard
| Challenge: | Autoregressive language models excel in text-to-audio generation, but lag behind diffusion models by a non-trivial margin. |
| Approach: | They propose a framework that integrates multiple isolated transformers with causal conditioning and anti-causal alignment via reinforcement learning. |
| Outcome: | The proposed framework outperforms existing LM-based and diffusion-based systems in audio synthesis. |
ManCC: A Task-Anchored Benchmark for Manchu–Classical Chinese Cross-Lingual Modeling (2026.findings-acl)
Copied to clipboard
Meiqi Wang, Xiaoxin Sun, Dongjie Wang, Ruixin Yu, Xiantao Heng, Shuo Wang, Zhen Huang, Peng Zhao, Suhua Wang, Minghao Yin
| Challenge: | Mainstream research in natural language processing has focused on high-resource and modern languages. |
| Approach: | They propose a task-anchored benchmark for Manchu–Classical Chinese translation . they use a parallel corpus of 16,627 sentence pairs to evaluate the model . |
| Outcome: | The proposed benchmarks show that linguistic differences influence performance and broader language coverage facilitate low-resource transfer. |
Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance (2025.emnlp-main)
Copied to clipboard
| Challenge: | Prior work on persona prompting has shown mixed results on its effectiveness . prior work did not consider when and why personas should affect performance . |
| Approach: | They analyze literature on persona prompting and distill three desiderata for their effectiveness . they propose mitigation strategies to improve robustness but find they only work for the largest, most capable models . |
| Outcome: | The authors find that expert personas usually lead to positive or non-significant performance changes . they propose mitigation strategies to improve robustness but only for the largest models . |
Adapting Bias Evaluation to Domain Contexts using Generative Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to assess social bias in NLP systems face limitations in scalability and fidelity across domains. |
| Approach: | They propose a domain-adaptive framework that uses prompting with Large Language Models to automatically transform template-based bias datasets into domain-specific variants. |
| Outcome: | The proposed framework improves the accuracy and contextual relevance of bias evaluations in socially relevant datasets. |
DiFRa: A Unified Framework for Harmonizing Semantic Diversity and Factual Consistency in Question-Answer Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Question-Answer Generation (QAG) is essential for domain-specific large language models post-training. |
| Approach: | They propose a framework that balances semantic diversity and factual consistency . they propose entropy and consistency scores that harmonize the trade-off between diversity and correctness . |
| Outcome: | The proposed framework outperforms baseline models in generating diverse QA pairs . the proposed framework harmonizes semantic entropy and consistency scores to quantify trade-off between diversity and correctness. |
HumanLLM: Benchmarking and Improving LLM Anthropomorphism via Human Cognitive Patterns (2026.acl-long)
Copied to clipboard
Xintao Wang, Jian Yang, Weiyuan Li, Rui Xie, Jen-tse Huang, Jun Gao, Shuai Huang, Yueping Kang, Yuanli Guo, Hongwei Feng, Yanghua Xiao
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning and generation, serving as the foundation for advanced persona simulation and Role-Playing Language Agents (RPLAs). |
| Approach: | They propose a framework that treats psychological patterns as interacting causal forces and synthesizes 113 scenarios where 2-5 patterns reinforce, conflict, or modulate each other. |
| Outcome: | The proposed framework outperforms Qwen3-32B on multi-pattern dynamics despite 4 fewer parameters. |
From Nodes to Narratives: Explaining Graph Neural Networks with LLMs and Graph Context (2026.acl-long)
Copied to clipboard
| Challenge: | Existing explanation methods for graph neural networks struggle to generate interpretable, fine-grained rationales. |
| Approach: | They propose a lightweight framework that uses large language models to generate interpretable explanations for GNNs. |
| Outcome: | The proposed framework generates interpretable explanations for GNN predictions using large language models. |
Don’t Corrupt the Fact: A Trustworthy RAG Watermarking Framework based on Dual Factual Shield (2026.acl-long)
Copied to clipboard
| Challenge: | Existing watermarking methods are fact-agnostic and cause "faithfulness hallucinations" a novel framework to enforce knowledge loyalty is proposed to improve watermarks . |
| Approach: | They propose a new framework that enforces knowledge loyalty by spoofing terms from retrieved contexts and prompt-based semantic guidance to protect against factual corruption. |
| Outcome: | The proposed framework reduces the Knowledge Corruption Rate while maintaining its original high security and robustness. |
Jailbreak-Zero: A Path to Pareto Optimal Red Teaming for Large Language Models (2026.acl-long)
Copied to clipboard
Kai Hu, Abhinav Aggarwal, Mehran Khodabandeh, David Zhang, Eric Hsin, Li Chen, Ankit Jain, Matt Fredrikson, Akash Bharadwaj
| Challenge: | Existing approaches to red teaming are based on example-based evaluation, where a static list of specific prompts is used to define and measure "unsafe content" |
| Approach: | They propose a new automated red teaming framework that shifts from example-based to policy-based evaluation that focuses on risk coverage, semantic diversity, and fidelity. |
| Outcome: | The proposed method achieves superior, human-readable attacks against open-source and proprietary models even for unseen safety policies. |
What About the Scene With the Hitler Reference? HAUNT: A Framework to Probe LLMs’ Self-consistency in Closed Domains Via Adversarial Nudge (2026.acl-long)
Copied to clipboard
| Challenge: | Claude exhibits strong resilience, while GPT and Grok demonstrate moderate resilience . open models fall short significantly, while proprietary models exhibit weak resilience compared to open models . |
| Approach: | They propose a framework for stress testing factual fidelity in large language models in the presence of adversarial nudges. |
| Outcome: | The proposed model is robust to adversarial nudges in two closed domains. |