State of AI

State of AI

Share this post

State of AI
State of AI
Bi-Weekly AI Research Roundup

Bi-Weekly AI Research Roundup

Latest research summaries in ML, Robotics, CV, NLP and AI

State of AI's avatar
State of AI
Oct 08, 2024
∙ Paid
1

Share this post

State of AI
State of AI
Bi-Weekly AI Research Roundup
2
Share

Contents

  1. Buckle Up: Robustifying LLMs at Every Customization Stage via Data Curation

  2. Unraveling Cross-Modality Knowledge Conflict in Large Vision-Language Models

  3. Geometric Representation Condition Improves Equivariant Molecule Generation

  4. Enhance Reasoning by Learning from Mistakes: Peer-Review Knowledge Distillation from Multiple Large Language Models

  5. GenSim2: Scaling Robot Data Generation with Multi-modal and Reasoning LLMs

  6. Scalable and Accurate Graph Reasoning with LLM-based Multi-Agents

  7. AlphaRouter: Quantum Circuit Routing with Reinforcement Learning and Tree Search

  8. TextHawk2: A Large Vision-Language Model Excels in Bilingual OCR and Grounding with 16x Fewer Tokens

  9. LoTLIP: Improving Language-Image Pre-training for Long Text Understanding

  10. Stateful Large Language Model Serving with Pensieve

  11. LayerKV: Optimizing Large Language Model Serving with Layer-wise KV Cache Management

  12. PrefixQuant: Static Quantization Beats Dynamic through Prefixed Outliers in LLMs

  13. TurtleBench: Evaluating Top Language Models via Real-World Yes/No Puzzles

  14. Creative Beam Search: LLM-as-a-Judge For Improving Response Generation

15. When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models


Buckle Up: Robustifying LLMs at Every Customization Stage via Data Curation

Authors: Xiaoqun Liu, Jiacheng Liang, Luoxi Tang, Chenyu You, Muchao Ye, Zhaohan Xi

Source and references: https://arxiv.org/abs/2410.02220v2


Introduction

This paper proposes a data curation framework called CTRL (CusTomization through CuRated Data over LLMs) to mitigate jailbreaking attacks on large language models (LLMs) at every stage of the customization process.

Key Points

  • CTRL leverages data curation to revise commonsense texts and enhance their safety implication from the perspective of LLMs.

  • The curated texts can mitigate jailbreaking attacks before, during, or after the customization process.

  • CTRL does not introduce additional modules during LLM inference, preserving the original customization process.

  • Experimental results demonstrate a substantial reduction in jailbreaking effects, with up to a 100% success rate in generating responsible responses.

Methodology

CTRL employs output sampling techniques, such as temperature and nucleus sampling, to curate commonsense texts. It uses a beam search process to filter and retain the most promising curated texts based on perplexity (to increase unfamiliarity) and helpfulness scores (to maintain informative value). The curated texts are then used to fine-tune LLMs at various stages of the customization workflow.

Results and Findings

The all-stage defense approach using CTRL consistently outperforms other baselines, achieving over 96% safety rate and, in some cases, fully preventing jailbreaking (100% safety rate) across different LLMs (Llama-3-8B, Llama-2-13B, Vicuna-13B, and Mistral-7B). The post-attack defense is found to be the most effective among the single-stage defenses, highlighting the importance of the most recent customization in influencing LLM behavior.

Implications and Conclusions

This work represents a significant advancement in mitigating jailbreaking risks and ensuring the secure customization of LLMs. The data-driven and all-stage-oriented nature of CTRL provides a cost-efficient and scalable solution that can be seamlessly integrated into the standard customization pipeline without additional modules.


Unraveling Cross-Modality Knowledge Conflict in Large Vision-Language Models

Authors: Tinghui Zhu, Qin Liu, Fei Wang, Zhengzhong Tu, Muhao Chen

Source and references: https://arxiv.org/abs/2410.03659v1


Introduction

This paper focuses on the problem of cross-modality parametric knowledge conflicts in large vision-language models (LVLMs).

Give a gift subscription

Get 14 day free trial

Keep reading with a 7-day free trial

Subscribe to State of AI to keep reading this post and get 7 days of free access to the full post archives.

Already a paid subscriber? Sign in
© 2025 StateOfAI
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share