Focuses on creating new datasets and methodologies to rigorously assess the capabilities, limitations, and specific behaviors of Large Language Models (LLMs) across various tasks and domains.
Investigates and aims to improve the complex reasoning abilities of LLMs, often employing step-by-step or structured thinking processes (like Chain-of-Thought) to solve problems.
Explores methods for enhancing LLM responses by retrieving relevant information from external knowledge sources (documents, KGs) and integrating it during generation.
Focuses on making LLMs more computationally efficient (quantization, pruning, attention optimization, KV Cache) and enabling processing/understanding of much longer input sequences.
Investigates societal implications: biases, fairness, safety risks (jailbreaking, disinformation, privacy), and methods for aligning models with human values and ethical principles.
Involves models processing and integrating text, images, audio, and video for tasks like VQA, image captioning, audio understanding, and multimodal reasoning.
Developing and evaluating LLMs that act autonomously, plan actions, interact with environments (web, OS, GUI), and use external tools/APIs for complex tasks.
Applying language models to software engineering: generating code, understanding structure, detecting bugs, test case generation, and code simplification.
Advancements in ASR, TTS, voice conversion, and MT, including multilingual/low-resource scenarios, model robustness, evaluation techniques, and speech language models.
Methods for creating synthetic data, augmenting existing datasets (often using LLMs), or selecting optimal subsets for training/fine-tuning, especially in low-data or specialized scenarios.
Understanding internal workings of complex models (LLMs, GNNs), explaining predictions, identifying responsible components (neurons, attention heads), and analyzing model mechanisms.
Applying NLP/LLMs to specialized domains like healthcare, law, finance, e-commerce, biology, agriculture, etc., often involving domain-specific datasets, benchmarks, and challenges.
Addressing challenges in building/evaluating models for multiple languages, especially low-resource ones; focus on cross-lingual transfer, dataset creation, culturally aware evaluation, and dialect handling.
Delving into core linguistic phenomena (syntax, semantics, pragmatics, discourse, psycholinguistics, morphology, phonetics) and using computational models for language insights.
Tailoring LM outputs/interactions to users based on preferences, history, persona, values, or needs; applications in dialogue systems, recommendations, and search.