Papers by Adrian Cosma
How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics (2024.emnlp-main)
Copied to clipboard
| Challenge: | Popular datasets suffer from systematic spurious correlations that artificially inflate actual model performance. |
| Approach: | They propose a method for the automated creation of a challenging test set without relying on manual construction of artificial and unrealistic examples. |
| Outcome: | The proposed method reduces spurious correlations and improves model performance . examples labeled as having the highest difficulty show markedly decreased performance compared to the full dataset . |
RoCode: A Dataset for Measuring Code Intelligence from Problem Definitions in Romanian (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models are capable of solving tasks in natural language, but most tests assume they are written in English. |
| Approach: | They propose to use a dataset to measure the generalization power of large language models in a language other than English to evaluate their code intelligence. |
| Outcome: | The proposed dataset provides a benchmark for evaluating the code intelligence of language models trained on Romanian / multilingual text and a fine-tuning set for pretrained Romanian models. |
Dr. Copilot: A Multi-Agent Prompt Optimized Assistant for Improving Patient-Doctor Communication in Romanian (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Text-based telemedicine is increasingly common, but quality of medical advice is judged more on how well it is communicated than on its clinical accuracy. |
| Approach: | They propose a multi-agent large language model system that evaluates and enhances the presentation quality of doctor-patient interactions. |
| Outcome: | The proposed system evaluates and improves the presentation quality of Romanian-speaking doctors' written responses. |
Life is not Always Depressing: Exploring the Happy Moments of People Diagnosed with Depression (2022.lrec-1)
Copied to clipboard
| Challenge: | a new study explores the relationship between depression and manifestations of happiness in social media . we use Positive-Unlabeled learning paradigm to extract happy moments from social media posts . 264 million people of all ages suffer from depression, according to the u.s. |
| Approach: | They propose a positive-unlabeled learning paradigm to extract happy moments from social media . they use LIWC and keyness information to qualitatively analyze the happy moments . |
| Outcome: | The proposed method extracts happy moments from social media posts of depressed users and controls . it qualitatively analyzes the results with LIWC and keyness information . |
The Strawberry Problem: Emergence of Character-level Understanding in Tokenized Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models fail at simple character-level tasks due to low mutual information, study finds . authors propose a lightweight architectural modification that improves character- level reasoning . |
| Approach: | They propose a lightweight architectural modification that improves character-level reasoning while preserving the inductive advantages of subword models. |
| Outcome: | The proposed model improves character-level reasoning while preserving the advantages of subword models. |