Skip to content

Commit

Permalink
update
Browse files Browse the repository at this point in the history
  • Loading branch information
Baixiang Huang committed Jan 27, 2025
1 parent b6926e8 commit de8775d
Showing 1 changed file with 51 additions and 51 deletions.
102 changes: 51 additions & 51 deletions index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -141,42 +141,34 @@
<h1 class="title is-4 publication-title">This is an initiative aiming to explore and understand knowledge editing in LLMs
</h1>
<!-- <h1 class="title is-5 publication-title">(Contact: <a href="https://canyuchen.com/" target="_blank">Canyu Chen</a> and <a href="https://baixianghuang.github.io/" target="_blank">Baixiang Huang</a>) -->
<h1 class="title is-5 publication-title">(Correspondence to: <a href="https://www.cs.emory.edu/~kshu5/" target="_blank">Kai Shu</a>)
<!-- <h1 class="title is-5 publication-title">(Correspondence to: <a href="https://www.cs.emory.edu/~kshu5/" target="_blank">Kai Shu</a>) -->
</h1>
<ul>

<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<a href="#Can-Knowledge-Editing-Really-Correct-Hallucinations"> <font color="red">(New Preprint)</font> <b>Can Knowledge Editing Really Correct Hallucinations?</b></a>
<a href="#Can-Knowledge-Editing-Really-Correct-Hallucinations"> (ICLR 2025) <b>Can Knowledge Editing Really Correct Hallucinations?</b></a>
<br>
- We proposed <b>HalluEditBench</b> to holistically benchmark knowledge editing methods in correcting real-world hallucinations on five dimensions including <b><i>Efficacy</i></b>, <b><i>Generalization</i></b>, <b><i>Portability</i></b>, <b><i>Locality</i></b>, and <b><i>Robustness</i></b>. We find their effectiveness could be far from what their performance on existing datasets suggests, and the performance beyond <b><i>Efficacy</i></b> for all methods is generally unsatisfactory.
<br>
<a href="#Can-Editing-LLMs-Inject-Harm"> <font color="red">(New Preprint)</font> <b>Can Editing LLMs Inject Harm?</b></a>
<a href="#Can-Editing-LLMs-Inject-Harm"> (Preprint) <b>Can Editing LLMs Inject Harm?</b></a>
<br>
- We propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely <b><i>Editing Attack</i></b>, and discover its emerging risk of injecting misinformation or bias into LLMs stealthily, indicating the feasibility of disseminating misinformation or bias with LLMs as new channels.
<br>
<br>
<div class="column has-text-centered">

<!-- <div class="column has-text-centered">
<h1 class="title is-4 publication-title">Related Work
</h1>
</div>
<!-- <a href="https://llm-authorship.github.io/" target="_blank"> <font color="red">(New Survey)</font> <b>Authorship Attribution in the Era of LLMs: Problems, Methodologies, and Challenges</b></a> -->
<a href="https://llm-authorship.github.io/" target="_blank"> (SIGKDD Explorations 2024) <b>Authorship Attribution in the Era of LLMs: Problems, Methodologies, and Challenges</b></a>
<br>
- This survey paper systematically categorizes authorship attribution in the era of LLMs into four problems: <i>attributing unknown texts to human authors</i>, <i>detecting LLM-generated texts</i>, <i>identifying specific LLMs or human authors</i>, and <i>classifying texts as human-authored, machine-generated, or co-authored by both</i>, while also highlighting key challenges and open problems.
<br><a href="https://llm-authorship.github.io/#canllm-identify-authorship" target="_blank"> (EMNLP 2024 Findings) <b>Can Large Language Models Identify Authorship?</b></a>
<br>
- We propose <b>Linguistically Informed Prompting (LIP)</b> strategy, which offers in-context linguistic guidance, to boost LLMs' reasoning capacity for <i>authorship verification</i> and <i>attribution</i> tasks, while also providing natural language explanations.
<br>
<a href="https://llm-misinformation.github.io/#Combating-Misinformation-in-the-Age-of-LLMs-Opportunities-and-Challenges" target="_blank"> (AI Magazine 2024) <b>Combating Misinformation in the Age of LLMs: Opportunities and Challenges</b></a>
<br>
- A survey of the opportunities (<b><i>can we utilize LLMs to combat misinformation</i></b>) and challenges (<b><i></i>how to combat LLM-generated misinformation</i></b>) of combating misinformation in the age of LLMs.
<br>
<a href="https://llm-misinformation.github.io/#Can-LLM-Generated-Misinformation-Be-Detected" target="_blank"> (Proceedings of ICLR 2024) <b>Can LLM-Generated Misinformation Be Detected?</b></a>
<br>
- We discover that LLM-generated misinformation can be <b><i>harder</i></b> to detect for humans and detectors compared to human-written misinformation with the same semantics, which suggests it can have <b><i>more deceptive styles</i></b> and potentially cause more harm.
<!-- <br> -->
<br> -->
</div>
</div>
</div>
Expand Down Expand Up @@ -204,18 +196,18 @@ <h1 id="Can-Knowledge-Editing-Really-Correct-Hallucinations" class="is-size-5 pu
<span class="author-block">
<a href="https://baixianghuang.github.io/" target="_blank">Baixiang Huang<sup>*1</sup></a>,</span>
<span class="author-block">
<a href="https://canyuchen.com" target="_blank">Canyu Chen<sup>*1</sup></a>,</span>
<a href="https://canyuchen.com" target="_blank">Canyu Chen<sup>*2</sup></a>,</span>
<span class="author-block">
<a href="https://xiongxiaoxu.github.io/" target="_blank">Xiongxiao Xu<sup>1</sup></a>,</span>
<a href="https://xiongxiaoxu.github.io/" target="_blank">Xiongxiao Xu<sup>2</sup></a>,</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?hl=en&user=9rHwD8wAAAAJ&view_op=list_works" target="_blank">Ali Payani<sup>2</sup></a>,</span>
<a href="https://scholar.google.com/citations?hl=en&user=9rHwD8wAAAAJ&view_op=list_works" target="_blank">Ali Payani<sup>3</sup></a>,</span>
<span class="author-block">
<a href="https://www.cs.emory.edu/~kshu5/" target="_blank">Kai Shu<sup>3</sup></a></span>
<a href="https://www.cs.emory.edu/~kshu5/" target="_blank">Kai Shu<sup>1</sup></a></span>
</div>
<div class="is-size-5 publication-institutions">
<span class="institution-block">1. Illinois Institute of Technology,</span>
<span class="institution-block">2. Cisco Research,</span>
<span class="institution-block">3. Emory University</span>
<span class="institution-block">1. Emory University,</span>
<span class="institution-block">2. Illinois Institute of Technology,</span>
<span class="institution-block">3. Cisco Research</span>
</div>
<div>
<span class="is-size-6">* Equal contribution</span>
Expand Down Expand Up @@ -287,10 +279,9 @@ <h1 id="Can-Knowledge-Editing-Really-Correct-Hallucinations" class="is-size-5 pu
</div>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"> Presented in workshop <em><strong><a href="https://safegenaiworkshop.github.io/" target="_blank" style="text-decoration:none"><font color="#494e52">Safe Generative AI@NeurIPS 2024</font></a></strong></em>.</span>
<!-- <br /> <strong><font color="red">🏆 Award:</font> <a href="https://rdi.berkeley.edu/events/decentralizationaisummit24" style="text-decoration:none"><font color="red">Research Spotlight</font></a>
</strong> in <i><a href="https://rdi.berkeley.edu/events/decentralizationaisummit24" style="text-decoration:none"><font color="#494e52">The 2024 Summit on Responsible Decentralized Intelligence —— Future of Decentralization and AI</font></a></i>,
hosted by <a href="https://rdi.berkeley.edu/" style="text-decoration:none"><font color="#494e52">The Berkeley Center for Responsible, Decentralized Intelligence</font></a> -->
<!-- <span class="author-block"> Presented in workshop <em><strong><a href="https://safegenaiworkshop.github.io/" target="_blank" style="text-decoration:none"><font color="#494e52">Safe Generative AI@NeurIPS 2024</font></a></strong></em>.</span> -->
<span class="author-block"> The Thirteenth International Conference on Learning Representations (ICLR 2025)</span>

</div>
</div>
</div>
Expand Down Expand Up @@ -494,14 +485,21 @@ <h2 class="title is-3">Results and Analysis</h2>
<br>
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{huang2024canknowledge,
title = {Can Knowledge Editing Really Correct Hallucinations?},
author = {Baixiang Huang and Canyu Chen and Xiongxiao Xu and Ali Payani and Kai Shu},
year = {2024},
journal = {arXiv preprint arXiv: 2410.16251}
<pre><code>@inproceedings{huang2024halluedit,
title = {Can Knowledge Editing Really Correct Hallucinations?},
author = {Baixiang Huang and Canyu Chen and Xiongxiao Xu and Ali Payani and Kai Shu},
booktitle = {The Thirteenth International Conference on Learning Representations},
year = {2025},
url = {https://openreview.net/forum?id=hmDt068MoZ}
}</code></pre>
</div>

<!-- @article{huang2024halluedit,
title = {Can Knowledge Editing Really Correct Hallucinations?},
author = {Baixiang Huang and Canyu Chen and Xiongxiao Xu and Ali Payani and Kai Shu},
year = {2024},
journal = {arXiv preprint arXiv: 2410.16251}
} -->

</section>
<br>
<br>
Expand Down Expand Up @@ -562,43 +560,45 @@ <h1 id="Can-Editing-LLMs-Inject-Harm" class="is-size-5 publication-title">TLDR:
<span class="author-block">
<a href="https://canyuchen.com" target="_blank">Canyu Chen<sup>*1</sup></a>,</span>
<span class="author-block">
<a href="https://baixianghuang.github.io/" target="_blank">Baixiang Huang<sup>*1</sup></a>,</span>
<a href="https://baixianghuang.github.io/" target="_blank">Baixiang Huang<sup>*2</sup></a>,</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=MD61m08AAAAJ&hl=en" target="_blank">Zekun Li<sup>2</sup></a>,</span>
<a href="https://scholar.google.com/citations?user=MD61m08AAAAJ&hl=en" target="_blank">Zekun Li<sup>3</sup></a>,</span>
<span class="author-block">
<a href="https://billchan226.github.io/" target="_blank">Zhaorun Chen<sup>3</sup></a>,</span>
<a href="https://billchan226.github.io/" target="_blank">Zhaorun Chen<sup>4</sup></a>,</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=qALDmfcAAAAJ&hl=en" target="_blank">Shiyang Lai<sup>3</sup></a>,</span>
<a href="https://scholar.google.com/citations?user=qALDmfcAAAAJ&hl=en" target="_blank">Shiyang Lai<sup>4</sup></a>,</span>
<span class="author-block">
<a href="https://xiongxiaoxu.github.io/" target="_blank">Xiongxiao Xu<sup>1</sup></a>,</span>
<span class="author-block">
<a href="https://jasonforjoy.github.io/" target="_blank">Jia-Chen Gu<sup>4</sup></a>,</span>
<a href="https://jasonforjoy.github.io/" target="_blank">Jia-Chen Gu<sup>5</sup></a>,</span>
<span class="author-block">
<a href="https://jindonggu.github.io/" target="_blank">Jindong Gu<sup>5</sup></a>,</span>
<a href="https://jindonggu.github.io/" target="_blank">Jindong Gu<sup>6</sup></a>,</span>
<span class="author-block">
<a href="https://www.huaxiuyao.io/" target="_blank">Huaxiu Yao<sup>6</sup></a>,</span>
<a href="https://www.huaxiuyao.io/" target="_blank">Huaxiu Yao<sup>7</sup></a>,</span>
<span class="author-block">
<a href="https://xiaocw11.github.io/" target="_blank">Chaowei Xiao<sup>7</sup></a>,</span>
<a href="https://xiaocw11.github.io/" target="_blank">Chaowei Xiao<sup>8</sup></a>,</span>
<span class="author-block">
<a href="https://sites.cs.ucsb.edu/~xyan/" target="_blank">Xifeng Yan<sup>2</sup></a>,</span>
<a href="https://sites.cs.ucsb.edu/~xyan/" target="_blank">Xifeng Yan<sup>3</sup></a>,</span>
<span class="author-block">
<a href="https://sites.cs.ucsb.edu/~william/" target="_blank">William Yang Wang<sup>2</sup></a>,</span>
<a href="https://sites.cs.ucsb.edu/~william/" target="_blank">William Yang Wang<sup>3</sup></a>,</span>
<span class="author-block">
<a href="https://www.robots.ox.ac.uk/~phst/" target="_blank">Philip Torr<sup>5</sup></a>,</span>
<a href="https://www.robots.ox.ac.uk/~phst/" target="_blank">Philip Torr<sup>6</sup></a>,</span>
<span class="author-block">
<a href="https://dawnsong.io/" target="_blank">Dawn Song<sup>8</sup></a>,</span>
<a href="https://dawnsong.io/" target="_blank">Dawn Song<sup>9</sup></a>,</span>
<span class="author-block">
<a href="https://www.cs.emory.edu/~kshu5/" target="_blank">Kai Shu<sup>1</sup></a></span>
<a href="https://www.cs.emory.edu/~kshu5/" target="_blank">Kai Shu<sup>2</sup></a></span>
</div>
<div class="is-size-5 publication-institutions">
<span class="institution-block">1. Illinois Institute of Technology,</span>
<span class="institution-block">2. UCSB,</span>
<span class="institution-block">3. University of Chicago,</span>
<span class="institution-block">4. UCLA,</span>
<span class="institution-block">5. University of Oxford,</span>
<span class="institution-block">6. UNC-Chapel Hill,</span>
<span class="institution-block">7. University of Wisconsin - Madison,</span>
<span class="institution-block">8. University of California, Berkeley</span>
<span class="institution-block">2. Emory University</span>
<span class="institution-block">3. UCSB,</span>
<span class="institution-block">4. University of Chicago,</span>
<span class="institution-block">5. UCLA,</span>
<span class="institution-block">6. University of Oxford,</span>
<span class="institution-block">7. UNC-Chapel Hill,</span>
<span class="institution-block">8. University of Wisconsin - Madison,</span>
<span class="institution-block">9. University of California, Berkeley</span>

</div>
<div>
<span class="is-size-6">* Equal contribution</span>
Expand Down

0 comments on commit de8775d

Please sign in to comment.