Image de couverture de livre

A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Lee, Andrew, Bai, Xiaoyan, Pres, Itamar, Wattenberg, Martin, Kummerfeld, Jonathan K., Mihalcea, Rada
Format:	Preprint
Publié:	2024
Sujets:	Computation and Language Artificial Intelligence
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

Documents similaires

Why Can't Transformers Learn Multiplication? Reverse-Engineering Reveals Long-Range Dependency Pitfalls
par: Bai, Xiaoyan, et autres
Publié: (2025)

Towards Algorithmic Fidelity: Mental Health Representation across Demographics in Synthetic vs. Human-generated Data
par: Mori, Shinka, et autres
Publié: (2024)

Rethinking Table Instruction Tuning
par: Deng, Naihao, et autres
Publié: (2025)

Towards Reliable Evaluation of Behavior Steering Interventions in LLMs
par: Pres, Itamar, et autres
Publié: (2024)

Chord Embeddings: Analyzing What They Capture and Their Role for Next Chord Prediction and Artist Attribute Prediction
par: Lahnala, Allison, et autres
Publié: (2021)

Mind the (Belief) Gap: Group Identity in the World of LLMs
par: Borah, Angana, et autres
Publié: (2025)

MAiDE-up: Multilingual Deception Detection of GPT-generated Hotel Reviews
par: Ignat, Oana, et autres
Publié: (2024)

Cross-cultural Inspiration Detection and Analysis in Real and LLM-generated Social Media Data
par: Ignat, Oana, et autres
Publié: (2024)

The Power of Many: Multi-Agent Multimodal Models for Cultural Image Captioning
par: Bai, Longju, et autres
Publié: (2024)

Annotations on a Budget: Leveraging Geo-Data Similarity to Balance Model Performance and Annotation Cost
par: Ignat, Oana, et autres
Publié: (2024)

Chumor 2.0: Towards Benchmarking Chinese Humor Understanding
par: He, Ruiqi, et autres
Publié: (2024)

Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons
par: Chen, Jianhui, et autres
Publié: (2024)

Belief-Sim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility
par: Borah, Angana, et autres
Publié: (2026)

Cat-DPO: Category-Adaptive Safety Alignment
par: Yang, Tiankai, et autres
Publié: (2026)

VERVE: Template-based ReflectiVE Rewriting for MotiVational IntErviewing
par: Min, Do June, et autres
Publié: (2023)

Aligning AI Research with the Needs of Clinical Coding Workflows: Eight Recommendations Based on US Data Analysis and Critical Review
par: Gan, Yidong, et autres
Publié: (2024)

Towards Analyzing and Understanding the Limitations of DPO: A Theoretical Perspective
par: Feng, Duanyu, et autres
Publié: (2024)

Alignment-Weighted DPO: A principled reasoning approach to improve safety alignment
par: Hu, Mengxuan, et autres
Publié: (2026)

Simple and Effective Baselines for Code Summarisation Evaluation
par: Robinson, Jade, et autres
Publié: (2025)

What Does it Mean for a Neural Network to Learn a "World Model"?
par: Li, Kenneth, et autres
Publié: (2025)

Eeyore: Realistic Depression Simulation via Supervised and Preference Optimization
par: Liu, Siyang, et autres
Publié: (2025)

Are Human Interactions Replicable by Generative Agents? A Case Study on Pronoun Usage in Hierarchical Interactions
par: Deng, Naihao, et autres
Publié: (2025)

Benchmarking and Improving LLM Robustness for Personalized Generation
par: Okite, Chimaobi, et autres
Publié: (2025)

Building Resource-Constrained Language Agents: A Korean Case Study on Chemical Toxicity Information
par: Cho, Hojun, et autres
Publié: (2025)

Adversarial DPO: Harnessing Harmful Data for Reducing Toxicity with Minimal Impact on Coherence and Evasiveness in Dialogue Agents
par: Kim, San, et autres
Publié: (2024)

Culture Affordance Atlas: Reconciling Object Diversity Through Functional Mapping
par: Nwatu, Joan, et autres
Publié: (2025)

MixDPO: Modeling Preference Strength for Pluralistic Alignment
par: Imai, Saki, et autres
Publié: (2026)

$R^3$: "This is My SQL, Are You With Me?" A Consensus-Based Multi-Agent System for Text-to-SQL Tasks
par: Xia, Hanchen, et autres
Publié: (2024)

The Generation Gap: Exploring Age Bias in the Value Systems of Large Language Models
par: Liu, Siyang, et autres
Publié: (2024)

Understanding or Memorizing? A Case Study of German Definite Articles in Language Models
par: Drechsel, Jonathan, et autres
Publié: (2026)

Deception Detection from Linguistic and Physiological Data Streams Using Bimodal Convolutional Neural Networks
par: Li, Panfeng, et autres
Publié: (2023)

SocialHarmBench: Revealing LLM Vulnerabilities to Socially Harmful Requests
par: Pandey, Punya Syon, et autres
Publié: (2025)

How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis
par: Yang, Yushi, et autres
Publié: (2024)

Curry-DPO: Enhancing Alignment using Curriculum Learning & Ranked Preferences
par: Pattnaik, Pulkit, et autres
Publié: (2024)

Rethinking DPO: The Role of Rejected Responses in Preference Misalignment
par: Cho, Jay Hyeon, et autres
Publié: (2025)

GloSS over Toxicity: Understanding and Mitigating Toxicity in LLMs via Global Toxic Subspace
par: Duan, Zenghao, et autres
Publié: (2025)

Induction Head Toxicity Mechanistically Explains Repetition Curse in Large Language Models
par: Wang, Shuxun, et autres
Publié: (2025)

When Ethics and Payoffs Diverge: LLM Agents in Morally Charged Social Dilemmas
par: Backmann, Steffen, et autres
Publié: (2025)

An Empirical Study of SFT-DPO Interaction and Parameterization in Small Language Models
par: Feng, Yuming, et autres
Publié: (2026)

daDPO: Distribution-Aware DPO for Distilling Conversational Abilities
par: Zhang, Zhengze, et autres
Publié: (2025)