Research Seminar⚓︎
A weekly seminar where we read, present, and debate recent papers in data analysis and machine learning.
Upcoming This Week⚓︎
Instruction Tuning and Knowledge Injection
Instruction Tuning for Large Language Models: A Survey. Efficient Knowledge Injection in LLMs via Self-Distillation. Aligning Large Language Models with Domain Adaptation.
Dynamic Graphs and Games
TIDFormer: Exploiting Temporal and Interactive Dynamics Makes A Great Dynamic Graph Transformer. TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction. Complete Chess Games Enable LLM Become A Chess Master.
## Past Seminars
Crash course on power-law networks, message passing framework, GNNs vs Transformers. Guest talk by Maria Sukhareva on retrieval-augmented machine translation: code switching, Pareto decoding, agentic neologism translation, and contrastive ensembling.
Natural-language-to-PDDL constraint translation via two-stage decomposition. Multi-agent framework combining LLMs with symbolic planning: domain modeling, procedural memory, and self-reflection.
Graph Retrieval-Augmented Generation with ego-graph retrieval, soft pruning, and dual-view prompting. FastRAG pipeline for semi-structured data: entropy-based chunk sampling, schema learning, and hybrid KG+text retrieval.
LightRAG: graph-enhanced RAG with dual-level retrieval (local entities + global themes), 600x cheaper than GraphRAG. ThoughtTerminator: calibrating reasoning models by budgeting tokens upfront and mitigating overthinking via early exit.
Sign.MT: open-source bidirectional sign-spoken language translation with modular pipeline and three rendering options (skeletal, 3D avatar, HumanGAN). EM-LLM: human-inspired episodic memory enabling infinite-context LLMs via surprise-based segmentation and graph-theoretic boundary refinement.
TG-Talker: adapting LLMs for temporal graph link prediction via in-context learning with background set, example set, and temporal neighbors. First framework for applying LLMs to real-world temporal graphs with MRR-based evaluation and textual explanation generation.
ChessLLM: first LLM to play complete chess games achieving Elo 1788 via FEN representation and long-round Stockfish data (+350 Elo). Fine-Grained Evaluation: event-level metrics and thematic analysis of reasoning failures (memory distortion, dissociation, character ambiguity) in Spyfall. SimUSER: AI agents with persona matching, knowledge-graph memory, and brain model for scalable recommender system evaluation.
Theoretical equivalence between multi-head attention in transformers and message passing in graph attention networks on fully connected graphs. Transformers as GNNs that won the hardware lottery due to dense matrix optimization on modern GPUs. Speaker's independent experiments showed GAT outperforming transformer on both CORA and AG News benchmarks.
* * *
## Topics
Architecture design, training acceleration, benchmarks, and explainability for temporal graph neural networks.
Embedding, forecasting, and question answering on temporal knowledge graphs.
Retrieval-Augmented Generation enhanced with graph structures: from LightRAG to domain-specific medical and legal applications.
Using large language models as graph learners for node classification, link prediction, and text-attributed networks.
Multi-agent systems, task planning with LLMs, and knowledge graph-powered agent memory.
Chain-of-thought, graph-of-thought, formal theorem proving, and scaling reasoning capabilities.
Sparse attention, diffusion LLMs, recursive models, and parameter-efficient methods.
Comprehensive evaluations and surveys across deep research, scientific discovery, and web agents.
Chain-of-knowledge prompting, graph-constrained reasoning, and knowledge graph-enhanced LLM inference.
Detecting and mitigating LLM hallucinations via knowledge graphs, contrastive decoding, and topological analysis.
Benchmarks and methods for extending LLM context: episodic memory, long-context QA, and citation generation.
Instruction tuning, LoRA adapters, domain adaptation, SFT, and preference optimization.
Named entity recognition, sentiment analysis, text classification, and table-to-text generation.
Chunking strategies, embedding models, reranking, and retrieval pipelines for RAG systems.
Community detection, subgraph mining, random walks, and relational database benchmarks.
Entity extraction, relation extraction, ontology construction, and entity alignment with LLMs.
Diffusion-based recommendations, graph convolution for re-ranking, and scaling recommender transformers.
LLM-driven synthetic data generation, curation, and evaluation frameworks.
LLM agents in social deduction games, role-playing evaluation, and user behavior simulation.
Adaptive test-time compute allocation, budget-aware reasoning, and token efficiency for LLMs.
KV cache quantization and compression for efficient long-context LLM inference.
Training LLMs to reason in continuous latent space with adaptive compute allocation.
RLHF optimization methods for language model alignment.
Curriculum-based pretraining and RL strategies for improving LLM reasoning.
Catastrophic forgetting mitigation in continual LLM training and unlearning.
Safety guardrails via synthetic data and adversarial training for LLMs.
LLM-in-the-loop active learning for efficient data annotation.
Domain-adaptive post-training strategies for specialized LLM applications.
GNN-based approaches to resource-constrained project scheduling under uncertainty.