Papers by Jan Kocon
RWKV: Reinventing RNNs for the Transformer Era (2023.findings-emnlp)
Copied to clipboard
Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Leon Derczynski, Xingjian Du, Matteo Grella, Kranthi Gv, Xuzheng He, Haowen Hou, Przemyslaw Kazienko, Jan Kocon, Jiaming Kong, Bartłomiej Koptyra, Hayden Lau, Jiaju Lin, Krishna Sri Ipsit Mantri, Ferdinand Mom, Atsushi Saito, Guangyu Song, Xiangru Tang, Johan Wind, Stanisław Woźniak, Zhenyuan Zhang, Qinghua Zhou, Jian Zhu, Rui-Jie Zhu
| Challenge: | recurrent neural networks struggle to match the performance of Transformers due to limitations in parallelization and scalability. |
| Approach: | They propose a model architecture that combines the efficient parallelizable training of transformers with the efficient inference of RNNs. |
| Outcome: | The proposed model performs on par with similarly sized RNNs, suggesting future work can leverage this architecture to create more efficient models. |
Controversy and Conformity: from Generalized to Personalized Aggressiveness Detection (2021.acl-long)
Copied to clipboard
Kamil Kanclerz, Alicja Figas, Marcin Gruza, Tomasz Kajdanowicz, Jan Kocon, Daria Puchalska, Przemyslaw Kazienko
| Challenge: | a new method to personalize documents that are perceived differently by users is needed . a recent study found that only a few annotations of controversial documents outperform classic methods . |
| Approach: | They propose to use some known, most controversial texts whose offensiveness is very ambiguous . they use user conformity-based measures or embeddings of their previous annotations to improve personalized reasoning . |
| Outcome: | The proposed methods outperform standard methods in document controversy and user nonconformity . the more controversial the content, the greater the gain, the authors say . |
Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models (2026.acl-srw)
Copied to clipboard
| Challenge: | sycophancy is a behavior that infiltrates the chain-of-thought, leading models to generate plausible-sounding justifications for incorrect answers. |
| Approach: | They introduce sycophantic anchors that commit models to user agreement . they find scophancy leaves a stronger mechanistic footprint than correct reasoning . |
| Outcome: | The proposed framework outperforms text-only baselines at high commitment levels and predicts commitment strength from activations. |
Personal Bias in Prediction of Emotions Elicited by Textual Opinions (2021.acl-srw)
Copied to clipboard
| Challenge: | Various models for emotion recognition have been used in different studies. |
| Approach: | They propose to use an annotated corpus to estimate personal emotional bias to estimate individual responses to texts . they propose to employ a new BERT-based transformer architecture to predict emotions from an individual human perspective. |
| Outcome: | The proposed method improves the quality of personalized reasoning and may boost the quality and reliability of content recommendation systems. |
Breaking the Illusion of Reasoning in Polish LLMs: Quality over Quantity of Thought (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances in large language models have introduced explicit reasoning capabilities . however, the precise role of reasoning in improving model performance remains unclear . |
| Approach: | They disentangle effects of reasoning quality and sequence length by fine-tuning 8B models on Polish variants of the Mixture-of-Thoughts dataset. |
| Outcome: | The proposed model trained on high-quality reasoning traces achieved better average performance than other models. |
PALS: Personalized Active Learning for Subjective Tasks in NLP (2023.emnlp-main)
Copied to clipboard
Kamil Kanclerz, Konrad Karanowski, Julita Bielaniewicz, Marcin Gruza, Piotr Miłkowski, Jan Kocon, Przemyslaw Kazienko
| Challenge: | Personalized active learning techniques can be used to learn subjective NLP problems . to acquire training data, texts are often randomly assigned to users for annotation . |
| Approach: | They propose to apply an active learning paradigm to a personalized context to learn preferences . they validated their techniques on a Wiki discussion text labeled with aggression and toxicity . |
| Outcome: | The proposed methods outperform random selection and random selection by 30% on three datasets. |