<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0"><channel><description>Assoc. Prof in CS @ Northeastern, NLP/ML &amp; health &amp; etc. He/him.</description><link>https://bsky.app/profile/byron.bsky.social</link><title>@byron.bsky.social - Byron Wallace</title><item><link>https://bsky.app/profile/byron.bsky.social/post/3mlnstt6vkc2t</link><description>Surgically editing prompts to vary a factor of interest (like gender) is an intuitive way of analyzing model behavior and sensitivity. But @zihaogavinyang.bsky.social shows that we should really compare the results from such perturbations to those observed when, e.g., we simply paraphrase inputs 👇&#xA;&#xA;[contains quote post or other embedded content]</description><pubDate>12 May 2026 12:42 +0000</pubDate><guid isPermaLink="false">at://did:plc:alozu2wqmtguj7whemteebqj/app.bsky.feed.post/3mlnstt6vkc2t</guid></item><item><link>https://bsky.app/profile/byron.bsky.social/post/3m4vfraa4t22r</link><description>Check out @hibaahsan.bsky.social&#39;s paper on spotting (problematic) racial biases in LLMs for healthcare applications 👇&#xA;&#xA;[contains quote post or other embedded content]</description><pubDate>05 Nov 2025 15:52 +0000</pubDate><guid isPermaLink="false">at://did:plc:alozu2wqmtguj7whemteebqj/app.bsky.feed.post/3m4vfraa4t22r</guid></item><item><link>https://bsky.app/profile/byron.bsky.social/post/3m3xc3hcnuc23</link><description>Chantal (and Vinith) find that you can jailbreak LLMs with syntax! Some examples: https://cshaib.github.io/syntax_domain_spurious_correlations/jailbreaks.html&#xA;&#xA;[contains quote post or other embedded content]</description><pubDate>24 Oct 2025 16:26 +0000</pubDate><guid isPermaLink="false">at://did:plc:alozu2wqmtguj7whemteebqj/app.bsky.feed.post/3m3xc3hcnuc23</guid></item><item><link>https://bsky.app/profile/byron.bsky.social/post/3m3rtnkxz5s27</link><description>Now to appear at #EMNLP2025 (Findings). We&#39;ve added more models and experiments: arxiv.org/abs/2502.13319&#xA;&#xA;[contains quote post or other embedded content]</description><pubDate>22 Oct 2025 12:24 +0000</pubDate><guid isPermaLink="false">at://did:plc:alozu2wqmtguj7whemteebqj/app.bsky.feed.post/3m3rtnkxz5s27</guid></item><item><link>https://bsky.app/profile/byron.bsky.social/post/3m23orxjabs25</link><description>Can we distill *circuits* from teacher models into smaller students? 👇&#xA;&#xA;[contains quote post or other embedded content]</description><pubDate>30 Sep 2025 23:34 +0000</pubDate><guid isPermaLink="false">at://did:plc:alozu2wqmtguj7whemteebqj/app.bsky.feed.post/3m23orxjabs25</guid></item><item><link>https://bsky.app/profile/byron.bsky.social/post/3lzlk5lzkic2g</link><description>Can we quantify what makes some text read like AI &#34;slop&#34;? We tried 👇&#xA;&#xA;[contains quote post or other embedded content]</description><pubDate>24 Sep 2025 13:28 +0000</pubDate><guid isPermaLink="false">at://did:plc:alozu2wqmtguj7whemteebqj/app.bsky.feed.post/3lzlk5lzkic2g</guid></item><item><link>https://bsky.app/profile/byron.bsky.social/post/3lak7tf5lqk2c</link><description>I&#39;ll be @ #EMNLP2024 if anyone wants to find snobby coffee / despair about election / or I guess talk research. Some work to be presented👇</description><pubDate>09 Nov 2024 21:21 +0000</pubDate><guid isPermaLink="false">at://did:plc:alozu2wqmtguj7whemteebqj/app.bsky.feed.post/3lak7tf5lqk2c</guid></item></channel></rss>