iiinova — where your digital self keeps evolving

Lukas Aichberger @aichberger.bsky.social · 29/05/2026

5/ Takeaway LLMs do not always need to externalize their thoughts. They can learn to reason in working memory instead, decoupling intermediate computation from autoregressive generation 💡 Full paper: arxiv.org/abs/2605.30343 Huge thanks to @hochreitersepp.bsky.social for the guidance!

arxiv.org

Unlocking the Working Memory of Large Language Models for Latent Reasoning

To improve the reasoning capabilities of large language models, test-time compute is typically scaled by generating intermediate tokens before the final answer. However, this couples reasoning to auto...

001

Lukas Aichberger @aichberger.bsky.social · 29/05/2026

4/ Results RiM consistently improves over direct-answer SFT, outperforms Coconut, and becomes competitive with CoT-style reasoning across model families and scales on math reasoning tasks. Crucially, it preserves direct-answer inference speed 🚀

100

Lukas Aichberger @aichberger.bsky.social · 29/05/2026

3/ Method Simply adding memory blocks to the context is not enough. The hard part is making the LLM actually use them. RiM solves this with an efficient two-stage training curriculum that teaches the model to route reasoning through the memory-block representations 🧩

100

Lukas Aichberger @aichberger.bsky.social · 29/05/2026

2/ Intuition Established reasoning methods make LLMs externalize their thoughts. Chain-of-thought generates text token by token. Coconut replaces text with continuous thoughts, but still generates them step by step. RiM instead moves reasoning into the LLM's working memory 🧠

100

Lukas Aichberger @aichberger.bsky.social · 29/05/2026

We unlocked the working memory of LLMs 💥 Reasoning in Memory (RiM) replaces autoregressive "thinking out loud" with fixed memory blocks that form a task-specific workspace for latent reasoning. The key idea is simple: reasoning should happen inside the LLM, not in its output!

123

Reposted by Lukas Aichberger

Yarin @yaringal.bsky.social · 20/03/2025

Hot take: I think we just demonstrated the first AI agent computer worm 🤔 When an agent sees a trigger image it's instructed to execute malicious code and then share the image on social media to trigger other users' agents This is a chance to talk about agent security 👇

082

Lukas Aichberger @aichberger.bsky.social · 18/03/2025

💀 Harmful actions could include engaging with the malicious social media post to amplify its spread, navigating to a malicious website, or causing a memory overflow to crash your computer. Preventing such harmful actions remains an open challenge. [6/6]

100

Lukas Aichberger @aichberger.bsky.social · 18/03/2025

🎯 Once an OS agent – among those the MIP was optimised for – encounters the MIP during the execution of everyday tasks, empirical results indicate harmful actions are triggered in at least 9 out of 10 cases, regardless of the original task or screenshot layout. [5/6]

100

Lukas Aichberger @aichberger.bsky.social · 18/03/2025

🚨 The real danger? Attackers can simply embed MIPs in social media posts, wallpapers, or ads and spread them across the internet. Unlike text-based attacks, MIPs are hard to detect, allowing them to spread unnoticed. [4/6]

100

Lukas Aichberger @aichberger.bsky.social · 18/03/2025

🔓 Our work reveals that OS agents are not ready for safe integration into everyday life. Attackers can craft Malicious Image Patches (MIPs), subtle modifications to an image on the screen that, once encountered by an OS agent, deceive it into carrying out harmful actions. [3/6]

101

Lukas Aichberger @aichberger.bsky.social · 18/03/2025

💻 AI assistants, known as OS agents, autonomously control computers just like humans do. They navigate by analysing the screen and take actions via mouse and keyboard. OS agents could soon take over everyday tasks, saving users time and effort. [2/6]

100

Lukas Aichberger @aichberger.bsky.social · 18/03/2025

⚠️ Beware: Your AI assistant could be hijacked just by encountering a malicious image online! Our latest research exposes critical security risks in AI assistants. An attacker can hijack them by simply posting an image on social media and waiting for it to be captured. [1/6] 🧵

188

Reposted by Lukas Aichberger

hochreitersepp.bsky.social @hochreitersepp.bsky.social · 20/12/2024

Often LLMs hallucinate because of semantic uncertainty due to missing factual training data. We propose a method to detect such uncertainties using only one generated output sequence. Super efficient method to detect hallucination in LLMs.

0153

Lukas Aichberger @aichberger.bsky.social · 20/12/2024

𝗡𝗲𝘄 𝗣𝗮𝗽𝗲𝗿 𝗔𝗹𝗲𝗿𝘁: Rethinking Uncertainty Estimation in Natural Language Generation 🌟 Introducing 𝗚-𝗡𝗟𝗟, a theoretically grounded and highly efficient uncertainty estimate, perfect for scalable LLM applications 🚀 Dive into the paper: arxiv.org/abs/2412.15176 👇

arxiv.org

Rethinking Uncertainty Estimation in Natural Language Generation

Large Language Models (LLMs) are increasingly employed in real-world applications, driving the need to evaluate the trustworthiness of their generated text. To this end, reliable uncertainty estimatio...

095