<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Research on Alexandre Bismuth</title><link>https://alexandre-bismuth.github.io/research_projects/</link><description>Recent content in Research on Alexandre Bismuth</description><generator>Hugo -- 0.147.2</generator><language>en</language><lastBuildDate>Wed, 14 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://alexandre-bismuth.github.io/research_projects/index.xml" rel="self" type="application/rss+xml"/><item><title>ACM conference paper (ICAIF): Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books</title><link>https://alexandre-bismuth.github.io/research_projects/icaif-mechanistic-interpretability/</link><pubDate>Wed, 14 Oct 2026 00:00:00 +0000</pubDate><guid>https://alexandre-bismuth.github.io/research_projects/icaif-mechanistic-interpretability/</guid><description>Conference paper at ACM ICAIF 2026: the first mechanistic interpretability study of a limit-order-book foundation model shows that a message-only Mamba-3 model encodes the market&amp;#39;s order flow in its recurrent state, and that steering that state shifts its next-message beliefs.</description></item><item><title>ACM conference paper (ICAIF): Adversarial Post-Training of Foundation Models for High-Frequency Trading</title><link>https://alexandre-bismuth.github.io/research_projects/icaif-adversarial-post-training/</link><pubDate>Tue, 13 Oct 2026 00:00:00 +0000</pubDate><guid>https://alexandre-bismuth.github.io/research_projects/icaif-adversarial-post-training/</guid><description>Conference paper at ACM ICAIF 2026: a Wasserstein critic trained through the non-differentiable matching engine steers evolutionary updates of a limit-order-book foundation model, improving realism by 17.6% for under 12 GPU-hours.</description></item><item><title>NeurIPS workshop paper (GDDL): Why don’t Graph Transformers attend to the same token?</title><link>https://alexandre-bismuth.github.io/research_projects/neurips-gddl-attention-sinks/</link><pubDate>Thu, 01 Oct 2026 00:00:00 +0000</pubDate><guid>https://alexandre-bismuth.github.io/research_projects/neurips-gddl-attention-sinks/</guid><description>Workshop paper at NeurIPS 2026 (Geometric Distributional Deep Learning): GRIT rarely forms attention sinks on molecules, and the attention scoring function, not the graph, is the main cause.</description></item><item><title>Master's Thesis: Post-Training State-Space Foundation Models for High-Frequency Trading</title><link>https://alexandre-bismuth.github.io/research_projects/msc-thesis/</link><pubDate>Sat, 01 Aug 2026 00:00:00 +0000</pubDate><guid>https://alexandre-bismuth.github.io/research_projects/msc-thesis/</guid><description>Master&amp;#39;s thesis: what a limit-order-book foundation model learns about market microstructure, and how to refine it with an objective no gradient reaches.</description></item><item><title>Bachelor Thesis: Optimising Gunshot Detection in Tropical Forests for Wildlife Protection with TinyML</title><link>https://alexandre-bismuth.github.io/research_projects/bachelor-thesis/</link><pubDate>Tue, 01 Apr 2025 00:00:00 +0000</pubDate><guid>https://alexandre-bismuth.github.io/research_projects/bachelor-thesis/</guid><description>Bachelor thesis: real-time, on-device gunshot detection for tropical forests using TinyML.</description></item></channel></rss>