Monash NLP&ML Reading Group

---

• Time: Thursday, 2:30PM-4:00PM
• Location: Room 207 Woodside building
• Schedule: here
• Format: Presentation guidelines
  • First speaker (~45 minutes) -- The presenter talks about her/his current work/projects (the goal is to increase awareness about your work and get feedback from the team).
  • Second speaker (~45 minutes) -- The presentation about a set of papers, or a tutorial about a topic of the presenter's choice.
• Past meeting repo

2024 Schedule

32. Safety Alignment Should Be Made More Than Just a Few Tokens Deep

• Date: 17/10/2024
• Zhuang will present the paper "Safety Alignment Should Be Made More Than Just a Few Tokens Deep"
• In the second part, Tao will discuss the following papers:
  • Learning Transferable Visual Models From Natural Language Supervision
  • From Representation to Reasoning: Towards both Evidence and Commonsense Reasoning for Video Question-Answering
  • CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

31. MemoryLLM

• Date: 10/10/2024
• Tong presents the following paper:
  • MEMORY LLM: Towards Self-Updatable Large Language Models
  • Agent Workflow Memory
  • WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models

30. VersiCode: Towards Version-controllable Code Generation

• Date: 3/10/2024
• Tong will present his work on VersiCode: Towards Version-controllable Code Generation

29. CultureInstruct: Curating Large-Scale Cultural Instructions from the Web

• Date: 19/09/2024
• First Speaker: Thanh will present his work on CultureInstruct: Curating Large-Scale Cultural Instructions from the Web
• In the second part, Hao will have a practice talk on "Red Teaming Speech-specific Safety of Large Multimodal Models"

28. Introduction to GraphRAG

• Date: 13/09/2024
• First Speaker: Linhao will provide an introduction to GraphRAG.
• In the second part, Thanh will present the following papers
  • Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing
  • Generative Verifiers: Reward Modeling as Next-Token Prediction

27. Zero-short Speech Recognition

• Date: 05/09/2024
• First Speaker: Michelle presents her work on Zero-Shot Speech Recognition
• In the second part, Hao will present the following papers
  • How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs
  • Towards Probing Speech-Specific Risks in Large Multimodal Models: A Taxonomy, Benchmark, and Insights

26. ACL redux

• Date: 22/08/2024 and 29/08/2024
• Slide

25. Against The Achilles' Heel: A Survey on Red Teaming for Generative Models

• Date: 08/08/2024
• First Speaker: Minghan will present his paper Against The Achilles' Heel: A Survey on Red Teaming for Generative Models
• Second Speaker: Manh will present the following papers:
  • Can Language Models Serve as Text-Based World Simulators?
  • Reasoning with LM is Planning with World Model

24. Splitting sentences to jailbreak LLMs/LMMs

• Date: 01/08/2024
• Hao will present his work on "Splitting sentences to jailbreak LLMs/LMMs".

23. LLM Decoding Acceleration

• Date: 25/07/2024
• First Speaker: Minghan will discuss about "LLM Decoding Acceleration"
• Second Speaker: Haolan will present the paper Feature Contamination: Neural Networks Learn Uncorrelated Features and Fail to Generalize

22. BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions

• Date: 2024 Jul 18
• First Speaker: Haolan presents his work about theBigCodeBench benchmark
• Second Speaker: Linhao will discuss about "Controllable LLMs via constrained decoding".

21. Assistive Large Language Model Agents for Socially-Aware Negotiation Dialogues

• Date: 2024 Jul 11
• First Speaker: Devin presents his paper entitled "Assistive Large Language Model Agents for Socially-Aware Negotiation Dialogues"

20. Stitchable Neural Networks

• Date: 2024 Jul 04
• Guest Speaker: Zizheng Pan presents his line of works on Stitchable Neural Networks

19. SOTOPIA: Interactive Learning of Socially Intelligent Language Agents

• Date: 2024 June 27
• First Speaker: Devin shares a set of papers on "SOTOPIA: Interactive Learning of Socially Intelligent Language Agents"
• Second Speaker: Akide surveys recent works on multi-modal models

18. Pragmatic effects on Language Learning and Generation

• Date: 2024 Jun 20
• Guest Speaker: Uri Berger (Melb Uni) presents his work on Pragmatic effects on Language Learning and Generation
• Second Speaker: Zhuang discusses the paper Denevil: Towards Deciphering and Navigating the Ethical Values of Large Language Models via Instruction Learning

17. KV Cache Compression for LLMs: Reducing Memory Footprint to Enhance Inference Efficiency

• Date: 2024 June 04
• First Speaker: Akide shares his recent work on KV Cache Compression

16. Efficient Instruction-Tuning of Large Languge Models through Style Consistency-Aware Response Ranking

• Date: 30/05/2024
• First Speaker: Zhuang presents his work about "Efficient Instruction-Tuning of Large Languge Models through Style Consistency-Aware Response Ranking"
• Second Speaker: Trang will discuss about "Online vs offline alignment algorithms".
  • Related paper: Understanding the performance gap between online and offline alignment algorithms

15. ICLR Redux

• Date: 2024 May 23

14. ICLR Redux

• Date: 2024 May 16

14. VersiCode: Toward Version-controllable Code Generation

• Date: 2024 May 09
• First Speaker: Tong presents his work about "VersiCode: Toward Version-controllable Code Generation"
• Second Speaker: Trang will discuss following papers:
  • Don't Forget Your Reward Values: Language Model Alignment via Value-based Calibration
  • Reuse Your Rewards: Reward Model Transfer for Zero-Shot Cross-Lingual Alignment
  • Iterative Reasoning Preference Optimization

13. An Empirical Analysis on Spatial Relationships Capability of Multimodal LLMs

• Date: 2024 May 02
• First Speaker: Fatemeh presents her work about "An Empirical Analysis on Spatial Relationships Capability of Multimodal LLMs"
• Second Speaker: Tong will discuss following papers
  • Set the Clock: Temporal Alignment of Pretrained Language Models
  • Exploring Continual Learning for Code Generation Models

12. A Brief Tutorial on Recent Advancements of RLHF and Its Alternatives

• Date: 2024 Apr 18
• First Speaker: Minghao presents a tutorial about "Recent Advancements of RLHF and Its Alternatives".
• Second speaker: Manh will discuss following papers
  • CContrastive Learning with Adversarial Perturbations for Conditional Text Generation
  • CONT: Contrastive Neural Text Generation

11. Evaluating Subword Tokenization: Alien Subword Composition and OOV Generalization Challenge

• Date: 2024 Apr 11
• Speaker: Huygaa Batsuren (Melbourne Uni)

10. Robust multimodal learning and generation

• Date: 2024 Mar 28
• Speaker: Manh will have a practice talk with title "Robust multimodal learning and generation"

9. Complex Knowledge Graphs Reasoning with LLMs

• Date: 2024 Mar 21
• First Speaker: Linhao presents his work about "Complex Knowledge Graphs Reasoning with LLMs".
• Second speaker: Minghan will discuss paper Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation

8. CausalScore: An Automatic Reference-Free Metric Leveraging Causal Strength for Assessing Quality of Open-Domain Dialogue Systems

• Date: 2024 Mar 7
• First Speaker: Tao presents his work about "CausalScore: An Automatic Reference-Free Metric Leveraging Causal Strength for Assessing Quality of Open-Domain Dialogue Systems".
• Second speaker: Linhao will discuss some papers related to "Complex Knowledge Graphs Reasoning with LLMs"

7. Audio Safety on Multimodal LLMs

• Date: 2024 Feb 29
• First Speaker: Hao presents his work about "Audio Safety on Multimodal LLMs".
• Second speaker: Tao will discuss the following papers
  • Causal Reasoning and Large Language Models: Opening a New Frontier for Causality
  • Causal Parrots: Large Language Models May Talk Causality But Are Not Causal

6. Towards Socially-aware Dialogue Systems

• Date: 2024 Feb 22
• First speaker: Haolan presents his work about "Towards Socially-aware Dialogue Systems".
• Second speaker: Hao will discuss the following papers
  • GOAT-Bench: Safety Insights to Large Multimodal Models through Meme-Based Social Abuse
  • Towards Explainable Harmful Meme Detection through Multimodal Debate between Large Language Models

5. Research at MonashNLP

• Date: 2024 Feb 08
• First speaker: Reza will present an overview on different directions and projects that our groups are currently working on.
• Second speaker: Haolan will discuss the following papers
  • Aligning Large Language Models through Synthetic Feedback
  • Tool-Augmented Reward Modeling

4. Bridging the Modality Gap between Speech and Text

• Date: 2024 Feb 01
• First speaker: Michelle presents her work about "Bridging the Modality Gap between Speech and Text".
• Second speaker: Fatemeh will present few works about multi-modal LLMs, including
  • Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs
  • How Many Unicorns Are in This Image? A Safety Evaluation Benchmark for Vision LLMs

3. Modelling Coalition Negotiations Using Artificial Agents

• Date: 2024 Jan 25
• First speaker: Farhad will present his work on "Modelling Coalition Negotiations Using Artificial Agents"
• Second speaker: Michelle will share her internship experience at FAIR

2. Simultaneous Machine Translation with Large Language Models

• Date: 2024 Jan 18
• First speaker: Minghan presents his work about "Simultaneous Machine Translation with Large Language Models".
• Second speaker: Farhad will present few works about "Multi-agent Conversation", including
  • AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
  • MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

1. Don't throw away your value model! Making PPO even better via Value-Guided Monte-Carlo Tree Search decoding

• Date: 2024 Jan 11
• First speaker: Devin presents paper Don't throw away your value model! Making PPO even better via Value-Guided Monte-Carlo Tree Search decoding
• Second speaker: Minghan presents paper Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models