Focuses on LLMs that act: tool calling and tool-integrated reasoning, web/GUI/computer-use agents, software and research agents, multi-agent collaboration and debate, agent memory and experience, long-horizon planning, agentic reinforcement learning, and agent evaluation.
Focuses on language models that see: large vision-language and multimodal models, visual question answering and grounding, image and video understanding, document/chart/OCR understanding, multimodal reasoning and chain-of-thought, multimodal hallucination, unified understanding-generation models, and vision-language-action models.
Covers the measurement side of NLP: new benchmarks and datasets, evaluation metrics and protocols, LLM-based evaluation and meta-evaluation, benchmark contamination and saturation, robustness and consistency of evaluation, human annotation and evaluation methodology, and capability analyses across tasks and domains.
Covers how LLMs are trained and prompted to reason: chain-of-thought and long thinking, reinforcement learning with verifiable rewards (GRPO/RLVR), large reasoning models, test-time scaling and search, self-improvement and self-verification, mathematical and logical reasoning, theorem proving, and efficient reasoning that curbs overthinking.
Applies LLMs to specialized domains: biomedical and clinical NLP, scientific literature understanding and AI for science, legal reasoning and legal NLP, finance and economics, education and tutoring, chemistry and materials, and domain-specific benchmarks, adaptation, and deployment.
Focuses on language models interacting with people: dialogue and conversational systems, task-oriented and open-domain dialogue, persona and role-playing, personalization and user modeling, emotional support and empathy, user simulation, human-AI collaboration and writing assistance, and theory of mind and social reasoning.
Covers retrieval and its integration with generation: RAG pipelines and their robustness, dense retrieval and text embeddings, rerankers and query rewriting, agentic and iterative retrieval, multi-hop and open-domain question answering, long-document and multimodal retrieval, graph-based RAG, and evaluation of retrieval-augmented systems.
Includes techniques for making LLMs cheaper and longer-ranged: KV-cache compression, quantization, pruning and sparsity, speculative decoding, efficient and sparse attention, long-context modeling and extension, mixture-of-experts, knowledge distillation into small models, parameter-efficient fine-tuning, and serving/system optimizations.
Addresses adversarial and misuse risks for language models: jailbreak attacks and defenses, prompt injection and backdoors, safety alignment and refusal, guardrails and red teaming, harmful-content and toxicity detection, watermarking and AI-generated-text detection, privacy leakage and memorization, machine unlearning, and model security.
Studies what LLMs know and when they go wrong: hallucination detection and mitigation, factuality and faithfulness, knowledge boundaries and uncertainty/confidence estimation, calibration and abstention, knowledge editing and updating, knowledge conflicts between context and parameters, temporal and long-tail knowledge, and fact verification.
Covers how language models are built: pre-training and continual pre-training, data selection, filtering and mixtures, synthetic data generation and model collapse, scaling laws, tokenization and vocabularies, training objectives and architectures (including diffusion language models), model merging, continual learning, and parameter-efficient or memory-efficient training.
Examines language models and language use in society: social bias and stereotypes, fairness and demographic disparities, cultural values and cultural alignment, hate speech and toxicity, misinformation and propaganda, political and ideological language, sentiment and stance, social media analysis, and ethical and societal implications of NLP.
Covers language models for programming: code generation and completion, text-to-SQL and semantic parsing, code reasoning and repair, software engineering agents and repository-level tasks, code translation and verification, formal languages and DSLs, and code benchmarks and evaluation.
Covers the production and study of text itself: summarization and long-form generation, controllable and creative generation (stories, poetry, style), text simplification, paraphrase and data-to-text, decoding and diversity, writing quality and coherence, discourse and pragmatics, syntax, semantics and morphology, and psycholinguistic and cognitive approaches to language.
Focuses on language diversity: machine translation and its evaluation, multilingual and cross-lingual transfer, low-resource and endangered languages, dialects and code-switching, language-specific LLMs and benchmarks, multilingual tokenization and vocabulary, and the English-centric biases of LLMs.
Aims to understand how language models work internally: probing and representation analysis, mechanistic interpretability and circuits, sparse autoencoders and features, attention and neuron analysis, activation steering and intervention, in-context learning mechanisms, training dynamics and scaling behavior, and linguistic knowledge encoded in LMs.
Covers spoken language: speech recognition and translation, text-to-speech and voice synthesis, speech language models and spoken dialogue systems, speech tokenization and codecs, audio understanding and audio LLMs, prosody and paralinguistics, and speech-based evaluation and benchmarks.
Investigates post-training for helpfulness and human preference: RLHF and DPO-style preference optimization, reward modeling and LLM-as-a-judge, supervised fine-tuning and instruction following, data selection and synthesis for post-training, pluralistic and personalized alignment, and the dynamics and failure modes of alignment.
Focuses on turning text into structure and reasoning over it: named entity recognition and relation/event extraction, knowledge graph construction and completion, entity linking and coreference, table understanding and text-to-table, structured and semi-structured data (tables, forms, documents), knowledge-graph question answering, and schema and ontology induction.