🫗 An LLM's "private" reasoning may leak your sensitive data!

🎉 Excited to share our paper "Leaky Thoughts: Large Reasoning Models Are Not Private Thinkers" was accepted at #EMNLP main!

1/2

Work done with: Haritz Puerto, Martin Gubri ‪‪@mgubri.bsky.social‬ , Tommaso Green, Sangdoo Yun and Seong Joon Oh @coallaoh.bsky.social‬
#SEO #AI #LLM #GenerativeAI #Marketing #DigitalMarketing #Perplexity #NLProc

🔎 The results are clear: current C-SEO strategies don’t work. This challenges the recent hype and suggests that creators don’t need to game LLMs and create even more clickbaits. Just focus on producing genuinely good content and let traditional SEO do its work.

💥 With the rise of conversational search, a new technique of "Conversational SEO" (C-SEO) emerged, claiming it can boost content inclusion in AI-generated answers. We put these claims to the test by building C-SEO Bench, the first comprehensive benchmark to rigorously evaluate these new strategies.

🔎Does Conversational SEO actually work? Our new benchmark has an answer!
Excited to announce our new paper: C-SEO Bench: Does Conversational SEO Work?

🌐 RTAI: researchtrend.ai/papers/2506....
📄 Paper: arxiv.org/abs/2506.11097
💻 Code: github.com/parameterlab...
📊 Data: huggingface.co/datasets/par...

🔎 Better Results in Fine-Tuning: Fine-tuned models show even stronger MIA results. The table shows the performance at sentence level and for collections of 20 sentences, evaluated on Phi-2 fine-tuned for QA (https://huggingface.co/haritzpuerto/phi-2-dcot ).

🚀 The Key? Number of tokens & Aggregation: MIA’s accuracy improves as we aggregate MIA scores across multiple paragraphs. Longer documents or larger document collections significantly boost MIA effectiveness.

🛠️ First Success on Pre-Trained LLMs: By adapting recent work on Dataset Inference (@pratyushmaini.bsky.social ), we successfully applied MIA on pre-trained LLMs. Check out our figure below: MIA achieves an AUROC of 0.75 on documents of up to 20k tokens!

🔍 New Benchmark, New Insights: We developed a new benchmark to assess MIA effectiveness across data scales, from single sentences to document collections. This lets us identify precisely when and how MIA succeeds on LLMs.

🚨 What’s MIA? It’s a method to detect if a specific data sample was used in model training. We show that MIA works effectively on long documents (~20k tokens) and in collections of documents (>100 docs) —just the scale relevant for legal applications! 👮

🚨📄 Exciting new research! Discover when and at what scale we can detect if specific data was used in training LLMs — a method known as Membership Inference (MIA)! Our findings open new doors for using MIA as potential legal evidence in AI. 🧵 https://arxiv.org/abs/2411.00154