Papers with intermediate representations
Attention is not not Explanation (D19-1)
Copied to clipboard
| Challenge: | Attention mechanisms play a central role in NLP systems, especially within recurrent neural network (RNN) models. |
| Approach: | They propose to use a simple uniform-weights baseline, a variance calibration and a diagnostic framework to determine when/whether attention can be used as explanation in RNN models. |
| Outcome: | The proposed tests show that even reliable adversarial distributions don't perform well on the simple diagnostic, indicating that prior work does not disprove the usefulness of attention mechanisms for explainability. |
Adversarial Removal of Demographic Attributes from Text Data (D18-1)
Copied to clipboard
| Challenge: | Recent advances in Representation Learning and Adversarial Training remove unwanted features from the learned representation. |
| Approach: | They show that demographic information of authors is encoded in the intermediate representations learned by text-based neural classifiers. |
| Outcome: | The proposed approach achieves higher accuracies on the same dataset, the authors show . they show that the proposed approach is effective in removing unwanted features from the learned representations. |
Multimodal Language Analysis with Recurrent Multistage Fusion (D18-1)
Copied to clipboard
| Challenge: | Comprehending multimodal language requires modeling interactions between modalities and between them. |
| Approach: | They propose a multistage fusion network which decomposes the fusion problem into multiple stages, each focused on a subset of multimodal signals for specialized, effective fusion. |
| Outcome: | The proposed model performs state-of-the-art across three datasets relating to multimodal sentiment analysis, emotion recognition, and speaker traits recognition. |
Multilingual, Multi-scale and Multi-layer Visualization of Intermediate Representations (D19-3)
Copied to clipboard
| Challenge: | Currently, the main alternatives to deal with sequences are Recurrent Neural Networks (RNN) architectures and the Transformer. |
| Approach: | They propose a web-based tool that visualizes the sentence and token representations of RNNs and Transformer architectures at the sentence level. |
| Outcome: | The proposed visualization tool analyses gender inequalities in contextual word embeddings and the common language representation in a multilingual machine translation system. |
Automating Steering for Safe Multimodal Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in Multimodal Large Language Models (MLLMs) have unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, especially when faced with adversarial multimodal inputs. |
| Approach: | They propose a modular and adaptive inference-time intervention technology, AutoSteer, that integrates a safety awareness score, an adaptive safety prober, and a lightweight Refusal Head to modulate generation when safety risks are detected. |
| Outcome: | Experiments on LLaVA-OV and Chameleon show that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats while maintaining general abilities. |
kNN Retrieval for Simple and Effective Zero-Shot Multi-speaker Text-to-Speech (2025.naacl-short)
Copied to clipboard
| Challenge: | Neural text-to-speech (TTS) models typically rely on extensive transcribed speech datasets and intricate training pipelines. |
| Approach: | They propose a framework for zero-shot multi-speaker text-to-speech using retrieval methods which leverage the linear relationships between SSL features. |
| Outcome: | The proposed framework achieves comparable performance to state-of-the-art models trained on large training datasets. |
LLM-Based Zero-Shot Soft Labeling for Anticipating Disagreement in Negotiation Dialogues (2026.acl-srw)
Copied to clipboard
| Challenge: | Negotiation involves complex emotional and strategic dynamics that pose challenges for AI agents in negotiation dialogues. |
| Approach: | They propose a zero-shot soft-labeling method using large language models . they also examine the performance of model training on rule-based annotated hard and soft labels . |
| Outcome: | The proposed method shows a maximum HIT@3 score of 0.87 against rule-based annotated hard labels . failure cases also demonstrated the limitations of rule--based annotation . |
Joint Embedding of Words and Labels for Text Classification (P18-1)
Copied to clipboard
Guoyin Wang, Chunyuan Li, Wenlin Wang, Yizhe Zhang, Dinghan Shen, Xinyuan Zhang, Ricardo Henao, Lawrence Carin
| Challenge: | Existing approaches to text classification use word embeddings to capture semantic regularities between words. |
| Approach: | They propose to view text classification as a label-word joint embedding problem . they use a framework that measures compatibility between text sequences and labels . |
| Outcome: | The proposed framework outperforms the state-of-the-art methods on large text datasets. |
Semi-Supervised Sequence Modeling with Cross-View Training (D18-1)
Copied to clipboard
| Challenge: | Unsupervised representation learning algorithms such as word2vec and ELMo only learn from task-specific labeled data during the main training phase. |
| Approach: | They propose a semi-supervised learning algorithm that improves the representations of a Bi-LSTM sentence encoder using a mix of labeled and unlabeled data. |
| Outcome: | The proposed algorithm improves the representations of a Bi-LSTM sentence encoder using a mix of labeled and unlabeled data. |
Spectral Filters, Dark Signals, and Attention Sinks (2024.acl-long)
Copied to clipboard
| Challenge: | Recent work assigns a central role to the model's residual stream as the shared communication channel between model components. |
| Approach: | They propose a quantitative extension of the logit lens approach by partitioning the embedding and unembedding matrices into bands and spectral filters on intermediate representations. |
| Outcome: | The proposed model can suppress the tail end of the embedding spectrum, but it is not able to suppress large parts of the spectrum. |
Dynamically Disentangling Social Bias from Task-Oriented Representations with Adversarial Attack (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing methods to learn representations from text often reflect social biases . previous methods rely on pre-specified direction or suffer from unstable training . |
| Approach: | They propose an adversarial disentangled debiasing model to decouple social bias attributes from intermediate representations trained on the main task. |
| Outcome: | The proposed model decouples social bias attributes from intermediate representations trained on the main task. |
Learning Language and Multimodal Privacy-Preserving Markers of Mood from Mobile Data (2021.acl-long)
Copied to clipboard
Paul Pu Liang, Terrance Liu, Anna Cai, Michal Muszynski, Ryo Ishii, Nick Allen, Randy Auerbach, David Brent, Ruslan Salakhutdinov, Louis-Philippe Morency
| Challenge: | Mental health conditions remain underdiagnosed in many countries despite access to advanced medical care . a new approach to learn mood markers from mobile data is needed to improve accuracy and improve learning from typed text. |
| Approach: | They propose to use mobile data to learn mood markers without identifying users through personal or protected attributes. |
| Outcome: | The proposed model obfuscates user identities while remaining predictive . future directions include better models and pre-learning from typed text . |
Incorporating Residual and Normalization Layers into Analysis of Masked Language Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Transformer architecture is composed of multi-head attention, which has been extensively analyzed. |
| Approach: | They extended the scope of the analysis of Transformers from solely the attention patterns to the whole attention block, i.e., multi-head attention, residual connection, and layer normalization. |
| Outcome: | The proposed method incorporates the whole attention block, i.e., multi-head attention, residual connection, and layer normalization into the analysis. |
StyleDubber: Towards Multi-Scale Style Learning for Movie Dubbing (2024.findings-acl)
Copied to clipboard
Gaoxiang Cong, Yuankai Qi, Liang Li, Amin Beheshti, Zhedong Zhang, Anton Hengel, Ming-Hsuan Yang, Chenggang Yan, Qingming Huang
| Challenge: | Existing methods for movie dubbing break phonemes in scripts, resulting in incomplete phoneme pronunciation and poor identity stability. |
| Approach: | They propose a method that switches dubbing learning from frame level to phoneme level . it uses a multimodal style adaptor to learn pronunciation style from audio . |
| Outcome: | The proposed method improves on two benchmarks, V2C and Grid, and is available on github. |
Unifying Inference-Time Planning Language Generation (2026.findings-acl)
Copied to clipboard
Prabhu Prakash Kagitha, Bo Sun, Ishan Desai, Andrew Zhu, Cassie Huang, Manling Li, Ziyang Li, Li Zhang
| Challenge: | Large language models (LLMs) are used to generate a formal representation of a plan in a planning language. |
| Approach: | They propose a unifying organizational framework based on intermediate representations to unify the inference-time LLM-as-formalizer methodology for classical planning. |
| Outcome: | The proposed framework subsumes most existing work and proposes new ones that involve syntactically similar but high-resource intermediate languages. |
AdaNSP: Uncertainty-driven Adaptive Decoding in Neural Semantic Parsing (P19-1)
Copied to clipboard
| Challenge: | Semantic parsing (SP) maps a natural language utterance into a formal language . standard Seq2Seq models ignore underlying grammars and may give ill-formed results. |
| Approach: | They propose an end-to-end model for semantic parsing that transduces a natural language sentence to the formal semantic representation. |
| Outcome: | The proposed model outperforms the state-of-the-art models and does not need expertise like predefined grammar or sketches in the meantime. |
Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines (2024.acl-long)
Copied to clipboard
| Challenge: | Text-to-image diffusion models use a latent text prompt to guide image generation . however, the process by which the encoder produces the text representation is unknown . |
| Approach: | They propose a method for analyzing the text encoder of T2I models by generating images from its intermediate representations. |
| Outcome: | The proposed method provides valuable insights into the text encoder component in T2I pipelines. |
TextFusion: Privacy-Preserving Pre-trained Model Inference via Token Fusion (2022.emnlp-main)
Copied to clipboard
Xin Zhou, Jinzhu Lu, Tao Gui, Ruotian Ma, Zichu Fei, Yuran Wang, Yong Ding, Yibo Cheung, Qi Zhang, Xuanjing Huang
| Challenge: | Existing methods to preserve inference privacy are available as cloud services . however, the risk of privacy leakage remains, according to recent studies . |
| Approach: | They propose a method to preserve inference privacy by fusing token representations in the cloud. |
| Outcome: | The proposed method preserves inference privacy without sacrificing performance on different scenarios. |
Breakpoint Transformers for Modeling and Tracking Intermediate Beliefs (2022.emnlp-main)
Copied to clipboard
| Challenge: | a framework for language understanding models to track and improve beliefs through intermediate points in text is needed . breakpoint modeling is an efficient and end-to-end learning approach that trains models to train beliefs . understanding the behavior of models remains a formidable challenge for model safety, authors say . |
| Approach: | They propose a framework that trains models to track beliefs through intermediate points in text . their framework allows for efficient and robust learning of this type of model . |
| Outcome: | The proposed model outperforms strong representation learning approaches on a variety of NLU tasks. |
Speech Translation and the End-to-End Promise: Taking Stock of Where We Are (2020.acl-main)
Copied to clipboard
| Challenge: | Until recently, the only feasible approach to translating acoustic speech signals into text was the cascaded approach. |
| Approach: | They propose a classification of the main challenges of traditional approaches to speech translation . they argue that end-to-end models fall short due to compromises made to address data scarcity . |
| Outcome: | This paper provides a brief survey of the main challenges of traditional approaches in speech translation . it reveals that many end-to-end models fail due to compromises made to address data scarcity. |
SPARQLing Database Queries from Intermediate Question Decompositions (2021.emnlp-main)
Copied to clipboard
| Challenge: | Using annotated datasets is difficult as it requires query-language expertise. |
| Approach: | They propose a crowdsourcing pipeline to annotate natural language questions using intermediate question representations. |
| Outcome: | The proposed pipeline reduces the burden of annotating a large dataset with queries by using intermediate question representations. |
Automatically Generated Definitions and their utility for Modeling Word Meaning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Modern language models generate semantic representations for words based on context and context based models. |
| Approach: | They propose to use dictionary-like sense definitions to generate sentence embeddings . they evaluate the quality of the generated definitions on existing English benchmarks based on the results of their study . |
| Outcome: | The proposed model sets new state-of-the-art results on lexical semantics tasks compared to baselines . |
Jump to Conclusions: Short-Cutting Transformers with Linear Transformations (2024.lrec-main)
Copied to clipboard
| Challenge: | Transformer-based language models create hidden representations of inputs at every layer, but only use final-layer representations for prediction. |
| Approach: | They propose a method for casting hidden representations as final representations, bypassing transformer computation in-between. |
| Outcome: | The proposed method produces more accurate predictions from hidden layers across various model scales, architectures, and data distributions. |
Cross-modality Data Augmentation for End-to-End Sign Language Translation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | End-to-end sign language translation (SLT) aims to convert sign language videos into spoken language texts without intermediate representations. |
| Approach: | They propose a cross-modality data-augmented framework to transfer gloss-to-text translation capabilities to end-to end sign language translation. |
| Outcome: | The proposed framework outperforms baseline models on two widely used SLT datasets. |
Privacy Risks of Intermediate Representations: Attribute Inference in Distributed LLM Inference (2026.findings-acl)
Copied to clipboard
| Challenge: | Distributed LLMs avoid raw inputs by transmitting intermediate hidden states, a practice widely assumed to preserve privacy. |
| Approach: | They propose a distributed inference framework that transmits intermediate hidden states to avoid sending raw inputs by exposing sensitive user attributes. |
| Outcome: | The proposed approach achieves Top-1 accuracy of 0.997 on CMS, 0.980 on Skytrax, and 0.986 on ECHR. |
Diagnosing Hidden Instabilities in Model Editing via Uncertainty Quantification (2026.acl-long)
Copied to clipboard
Zihan Gu, TianYi Zhang, Xinyan Zhang, Zhiyuan Wang, Han Zhang, Yuhao Wei, Jiacheng Lu, Tianyi Ma, Xingsheng Zhang, Hua Zhang, Yue Hu
| Challenge: | Existing methods to update large language models (LLMs) without expensive retraining are fragile under single-edit evaluation protocols. |
| Approach: | They propose a framework that characterizes activation-based editing as a constrained intervention on intermediate representations. |
| Outcome: | The proposed method reveals local knowledge conflicts invisible to existing benchmarks. |