Papers with model representations
Auditing Language Model Unlearning via Information Decomposition (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing approaches to unlearning in language models do not account for information about forgotten data . despite the apparent success of unlearning, information about the forgotten data remains linearly decodable from internal representations. |
| Approach: | They propose an interpretable framework for auditing unlearning using Partial Information Decomposition . they propose a representation-based risk score that can guide abstention on sensitive inputs . |
| Outcome: | The proposed framework can guide abstention on sensitive inputs at inference time. |
Rethinking Self-Attention: Towards Interpretability in Neural Parsing (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Recent work shows that attention mechanisms provide arguably explainable attention distributions that can help to interpret predictions. |
| Approach: | They propose a new self-attention layer where attention heads represent labels. |
| Outcome: | The proposed model obtains state-of-the-art results on the Penn Treebank and Chinese Treebank. |
What if This Modified That? Syntactic Interventions with Counterfactual Embeddings (2021.findings-acl)
Copied to clipboard
| Challenge: | Prior art aims to uncover meaningful properties within model representations, but it is unclear how faithfully such probes portray information that the models actually use. |
| Approach: | They propose a technique for generating counterfactual embeddings within models . they produce evidence that some models use a tree-distancelike representation of syntax . |
| Outcome: | The proposed technique produces evidence that some models use tree-distancelike representations of syntax in downstream prediction tasks. |
Cross-lingual Transfer of Monolingual Models (2022.lrec-1)
Copied to clipboard
| Challenge: | Existing studies on cross-lingual learning using multilingual models cast doubt on shared vocabulary and joint pre-training . et al. (2005) show that model knowledge learned in the source language enhances the learning of the target language independently of language proximity. |
| Approach: | They propose a method for transferring monolingual models to other languages through continuous pre-training and investigate their results in English. |
| Outcome: | The proposed method outperforms a model trained from scratch in the GLUE benchmark for English . it shows that model knowledge from the source language enhances the learning of syntactic and semantic knowledge in english. |
Improving Zero-Shot Translation by Disentangling Positional Information (2021.acl-long)
Copied to clipboard
| Challenge: | Multilingual neural machine translation has shown the capability of directly translating between language pairs unseen in training, i.e. zero-shot translation. |
| Approach: | They propose to remove residual connections in an encoder layer to reduce the difficulty of generalizing to new translation directions. |
| Outcome: | The proposed model outperforms pivot-based translation in terms of quality and ease of integration of new languages. |
Better Pre-Training by Reducing Representation Confusion (2023.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to improve pre-trained language models address information confusion in position encoding and model representations. |
| Approach: | They propose two techniques to improve pre-trained language models by decoupling directions and auxiliary regularizers. |
| Outcome: | The proposed techniques can improve pre-trained language models on GLUE benchmarks. |
Automatically Identifying Gender Issues in Machine Translation using Perturbations (2020.findings-emnlp)
Copied to clipboard
| Challenge: | a novel approach to machine translation has addressed outstanding challenges, including the modeling and treatment of gendered language. |
| Approach: | They propose a method to mine examples from real world data to explore challenges for deployed systems. |
| Outcome: | The proposed method exposes where model representations are gendered and the unintended consequences of genderes in downstream applications. |
Probing for Incremental Parse States in Autoregressive Language Models (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work on autoregressive language models shows impressive command of syntax . implicit incremental syntactic inferences underlie next-word predictions . |
| Approach: | They propose a probe for extracting incomplete syntactic structure from autoregressive language models. |
| Outcome: | The proposed probes can predict model preferences on ambiguous sentence prefixes and causally intervene on model representations and steer model behavior. |
Steering Large Language Models for Machine Translation Personalization (2026.eacl-long)
Copied to clipboard
| Challenge: | Recent advances in interpretability research have highlighted the effectiveness of steering methods for MT personalization. |
| Approach: | They examine steering strategies for personalizing automatic translations when few examples are available. |
| Outcome: | The proposed steering methods yield higher inference-time computational efficiency than prompting approaches. |
Analysing Lexical Semantic Change with Contextualised Word Representations (2020.acl-main)
Copied to clipboard
| Challenge: | Existing studies on lexical semantic change have focused on detecting and characterising word meaning shifts using distributional semantic models. |
| Approach: | They propose a method that exploits the BERT neural language model to obtain representations of word usages, clusters these representations into usage types, and measures change along time with three proposed metrics. |
| Outcome: | The proposed method captures a variety of synchronic and diachronic linguistic phenomena and is highly reproducible and reproducible. |
DeMuX: Data-efficient Multilingual Learning (2024.naacl-long)
Copied to clipboard
| Challenge: | Pre-trained multilingual models have enabled deployment of NLP technologies for multiple languages, but their performance under an annotation budget remains an open question. |
| Approach: | They propose a framework that prescribes the exact data-points to label from vast amounts of unlabelled multilingual data, having unknown degrees of overlap with the target set. |
| Outcome: | The proposed framework outperforms strong baselines in 84% of the test cases in the zero-shot setting of disjoint source and target language sets. |
Retrieval-augmented Video Encoding for Instructional Captioning (2023.findings-acl)
Copied to clipboard
| Challenge: | Instructional videos provide a detailed multimodal context of each procedure in instruction. key-object degeneracy is a problem for machine systems, causing incorrect captions. |
| Approach: | They propose a retrieval-based framework to augment the model representations in the presence of key-object degeneracy. |
| Outcome: | The proposed framework can be extended over baselines using modalities with key-object degeneracy. |
Are All Spurious Features in Natural Language Alike? An Analysis through a Causal Lens (2022.emnlp-main)
Copied to clipboard
| Challenge: | 'spurious correlations' have been used in NLP to informally denote any undesirable feature-label correlations. |
| Approach: | They formalize this distinction using a causal model and probabilities of necessity and sufficiency, which delineates causal relations between a feature and a label. |
| Outcome: | The proposed model is invariant to the feature, but not sufficient for prediction. |
Label-aware Multi-level Contrastive Learning for Cross-lingual Spoken Language Understanding (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to translate spoken language understanding into low-resource languages are limited to implicit alignment and disregard the inherent semantic structure in SLU. |
| Approach: | They propose to model utterance-slot-word structure by a multi-level contrastive learning framework . they also propose a label-aware joint model leveraging label semantics to enhance alignment . |
| Outcome: | The proposed model improves performance on two zero-shot cross-lingual datasets. |
Trick or Neat: Adversarial Ambiguity and Language Model Evaluation (2025.findings-acl)
Copied to clipboard
| Challenge: | Direct prompting fails to detect ambiguity while linear probes can decode ambiguities with high accuracy, sometimes exceeding 90%. |
| Approach: | They introduce an adversarial ambiguity dataset that includes syntactic, lexical, and phonological ambiguities along with adversarials. |
| Outcome: | The proposed dataset includes syntactic, lexical, and phonological ambiguities along with adversarial variations. |
Keys to Robust Edits: From Theoretical Insights to Practical Advances (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for modifying parametric memory are prone to inaccuracies due to conflicting or outdated information. |
| Approach: | They propose a plug-and-play module that disentangles editing keys from native model representations and dynamically adjusts keys via contrastive learning to achieve robustness-specificity balance. |
| Outcome: | The proposed method improves over robustness tests by up to 66.4% while maintaining the success rate unaffected. |
Pluggable Neural Machine Translation Models via Memory-augmented Adapters (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent years, neural machine translation systems are often developed with large-scale parallel data extracted from the Web. |
| Approach: | They propose a memory-augmented adapter to steer pretrained neural machine translation models in a pluggable manner by combining model representations and retrieved results. |
| Outcome: | The proposed method outperforms several representative pluggable baselines on style- and domain-specific experiments. |
Disentangling Continued Pre-Training: Attention-Driven Routing and Semantic Hub Preservation in Language Adaptation (2026.findings-acl)
Copied to clipboard
| Challenge: | Continued Pre-Training (CPT) enables Large Language Models (LLMs) to acquire second-language capabilities, yet the mechanisms underlying CPT remain poorly understood. |
| Approach: | They investigate how CPT adapts model representations across diverse language families and scripts, model sizes, and architectures. |
| Outcome: | The proposed model can be surgically transferred between base and CPT models with minimal loss. |
Probing Political Ideology in Large Language Models: How Latent Political Representations Generalize Across Tasks (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models encode rich internal representations of political ideology, but it remains unclear how these representations contribute to model decision-making. |
| Approach: | They apply inference-time interventions to steer a decoder-only transformer along learned ideological directions . they find that learned ideological representations generalize well to bias detection, but not as well to voting simulations . |
| Outcome: | The proposed model steers a transformer along learned ideological directions . political bias detection, voting preference simulation and bias neutralization are tested . |
Under the Shadow of Babel: How Language Shapes Reasoning in LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | linguistic relativity suggests that the structure of language shapes cognitive patterns . large language models internalize the habitual logical structures embedded in different languages, authors say . |
| Approach: | a study introduces a bilingual dataset for causal reasoning in Chinese and English. |
| Outcome: | a new study shows that large language models internalize reasoning biases shaped by language . the model internalizes language-specific preferences and rigidly applies them to atypical inputs, the study shows . |
LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety (2026.acl-long)
Copied to clipboard
Junxiao Yang, Haoran Liu, Jinzhe Tu, Jiale Cheng, Zhexin Zhang, Shiyao Cui, Jiaqi Weng, Jialing Tao, Hui Xue, Hongning Wang, Han Qiu, Minlie Huang
| Challenge: | Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resourced languages. |
| Approach: | They propose language-agnostic semantic alignment (LASA) which anchors safety alignment directly in semantic bottlenecks. |
| Outcome: | The proposed approach significantly improves safety across all languages: average attack success rate drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwend3 Instruct models (7B–32B). |