Papers by Edoardo Ponti

17 papers
Can VLMs Predict Future States? Bootstrapping World Models from Inverse Dynamics (2026.findings-acl)

Copied to clipboard

Challenge: unified vision–language models (VLMs) struggle to generate physically plausible transitions between frames from instructions.
Approach: They find that VLMs struggle to generate physically plausible transitions between frames from instructions.
Outcome: The proposed model outperforms state-of-the-art image editing models on Aurora-Bench . it achieves the best average human evaluation across all subsets of Aurora-bench compared with other models .
A Grounded Typology of Word Classes (2025.naacl-long)

Copied to clipboard

Challenge: Using captioned images, we can quantify language function and semantics using a grounded typology approach . linguistic typology is the study of patterns and variation across the world's languages .
Approach: They propose a grounded typology approach that uses images captioned across languages to quantify meaning and semantics.
Outcome: The proposed approach can quantify language function and semantics using images captioned across languages.
Composable Sparse Fine-Tuning for Cross-Lingual Transfer (2022.acl-long)

Copied to clipboard

Challenge: Adapters and sparse fine-tuning have been developed to improve transfer learning . a number of approaches have been proposed to improve performance of fine-untuners .
Approach: They propose a method that fine-tunes the entire set of parameters of a large pretrained model . they use adapters and sparse fine-uning to improve model efficiency .
Outcome: The proposed method outperforms adapters in cross-lingual transfer benchmarks.
Probing the Emergence of Cross-lingual Alignment during LLM Training (2024.findings-acl)

Copied to clipboard

Challenge: Multilingual Large Language Models (LLMs) achieve remarkable levels of zero-shot cross-lingual transfer performance.
Approach: They propose that LLMs can align languages without explicit supervision from parallel sentences without a single linguistic feature.
Outcome: The proposed model can perform zero-shot cross-lingual transfer even when the vocabularies of two languages have a null intersection, i.e., no tokens are shared.
Is Information Density Uniform when Utterances are Grounded on Perception and Discourse? (2026.eacl-long)

Copied to clipboard

Challenge: Existing studies on the distribution of information in visually grounded contexts have focused on text-only inputs.
Approach: They propose to use multilingual vision-and-language models to estimate surprisal . they find grounding on perception increases uniformity across typologically diverse languages .
Outcome: The proposed hypothesis is tested in visual-language models over 30 languages and 13 storytelling languages . the results show grounding on perception increases uniformity across languages compared to text-only settings .
Mixtures of In-Context Learners (2025.acl-long)

Copied to clipboard

Challenge: In-context learning is sensitive to the choice of in-con context demonstrations and processing many demonstrations can be computationally demanding.
Approach: They propose a method that uses subsets of demonstrations to train experts via ICL and learns a weighting function to merge their output distributions via gradient-based optimisation.
Outcome: The proposed approach improves on 5 out of 7 classification datasets compared to strong baselines and reduces the inference time needed to achieve the same performance with fewer demonstrations.
Cross-Lingual and Cross-Cultural Variation in Image Descriptions (2025.naacl-long)

Copied to clipboard

Challenge: Behavioural and cognitive studies report cultural effects on perception, but these are limited in scope and hard to replicate.
Approach: They develop a method to accurately identify entities mentioned in captions and present in images, then measure how they vary across languages.
Outcome: The proposed method corroborates previous studies showing that languages that are geographically or genetically closer mention entities more frequently than others.
Fine-Tuning Large Language Models with Sequential Instructions (2025.naacl-long)

Copied to clipboard

Challenge: Existing instruction-tuned models struggle to adhere to a query with multiple intentions, which impairs their performance when the completion of several tasks is demanded by a single command.
Approach: They develop an automatic process that turns existing data into diverse and complex task chains and a new benchmark to evaluate a model’s ability to follow all the instructions in a sequence.
Outcome: The proposed model can follow instructions better and deliver higher results in coding, maths, and open-ended generation.
Detecting and Mitigating Hallucinations in Multilingual Summarisation (2023.emnlp-main)

Copied to clipboard

Challenge: Existing faithfulness metrics for abstractive summarisation models focus on English . metric mFACT is best suited to detect hallucinations in cross-lingual transfer .
Approach: They propose a method to evaluate the faithfulness of non-English summaries by translation-based transfer from multiple English faithfulness metrics.
Outcome: The proposed method reduces hallucinations in cross-lingual transfer by weighing the loss of each training example by its faithfulness score.
The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Sparse attention is a promising strategy to extend long-context capabilities in LLMs . but its efficiency–accuracy trade-offs remain unclear due to the lack of comprehensive evaluation .
Approach: They evaluate sparse attention methods across multiple model families and sizes . they find larger sparser models outperform smaller dense ones at equivalent cost .
Outcome: The proposed methods outperform smaller sparse models at equivalent cost and improve the Pareto frontier.
Same Neurons, Different Languages: Probing Morphosyntax in Multilingual Pre-trained Models (2022.naacl-main)

Copied to clipboard

Challenge: Existing studies show that multilingual pre-trained models can learn to generalise across languages . however, it remains unclear how these models learn to learn multilingual representations .
Approach: They propose a hypothesis that multilingual pre-trained models can derive language-universal abstractions about grammar by aligning morphosyntactic markers that fulfil a similar grammatical function across languages.
Outcome: The proposed model can derive language-universal abstractions even without explicit supervision.
The Cross-linguistic Role of Animacy in Grammar Structures (2025.acl-long)

Copied to clipboard

Challenge: Animacy is a semantic feature of nominals and follows a hierarchy of personal pronouns . it is argued that soft tendencies may be the cause of animacy constraints . however, there is no empirical evidence for this .
Approach: They propose a method to reliably classify animacy classes of nominals in 11 languages from 5 families using multilingual large language models and word sense disambiguation datasets.
Outcome: The proposed method shows that animacy displays consistent cross-linguistic tendencies in terms of preferred morphosyntactic constructions, but not always in line with received wisdom.
Unifying Cross-Lingual Transfer across Scenarios of Resource Scarcity (2023.emnlp-main)

Copied to clipboard

Challenge: Existing approaches to deal with resource scarcity have not been developed to deal effectively with the problem.
Approach: They propose to use a set of tools to harness data from one or more high-resource "source" languages to compensate for a shortage of data in low-resourced "target" languages.
Outcome: The proposed technique can be easily adapted to unseen languages, extending the range of the proposed technique and translation-based transfer more broadly.
Elastic Weight Removal for Faithful and Abstractive Dialogue Generation (2024.naacl-long)

Copied to clipboard

Challenge: Current-day large language models generate coherent, grammatical, and seemingly meaningful text, but are prone to hallucinating incorrect information.
Approach: They propose to ‘subtract’ parameters of a model trained to hallucinate from a dialogue response generation model to ‘negate’ the contribution of such hallucinatedexamples from it.
Outcome: The proposed method reduces hallucinations and discourages extractive responses, which are often a consequence of reducing hallucines by encouraging copy-pasting of document spans.
Are Large Language Model Temporally Grounded? (2024.naacl-long)

Copied to clipboard

Challenge: Recent large language models lack a consistent temporal model of textual narratives . sentence ordering in unlabelled texts is only weakly correlated with event ordering .
Approach: They evaluate LLMs with textual narratives and evaluate their common-sense knowledge . they find that LLM models struggle the most with self-consistency .
Outcome: The proposed models lack a consistent temporal model of textual narratives.
Natural Language Processing for Multilingual Task-Oriented Dialogue (2022.acl-tutorials)

Copied to clipboard

Challenge: a tutorial will examine the challenges and gaps in multilingual ToD research . multilingual systems are difficult to build, and are limited to English and other languages .
Approach: This tutorial will discuss the importance of multilingual task-oriented dialogue systems . it will provide an overview of current research gaps, challenges and initiatives related to multilingual ToD systems - with a particular focus on their connections to current research and challenges in multilingual and low-resource NLP.
Outcome: This tutorial will provide an overview of current research gaps, challenges and initiatives related to multilingual ToD systems.
Image Retrieval from Contextual Descriptions (2022.acl-long)

Copied to clipboard

Challenge: a new multimodal challenge challenges vision-and-language models to integrate context into their representations.
Approach: They propose a multimodal challenge to integrate context into vision-and-language models . they benchmark several state-of-the-art models using cross-encoders and bi-encodings .
Outcome: The proposed model lags behind human models on imageCoDe, compared with human models.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations