Microsoft Research: AI Models Can't Handle Long-Running Tasks
Microsoft researchers have shown that current AI models and agents struggle to handle long-running tasks. The findings indicate that fundamental challenges remain for the expectation that AI agents will autonomously carry out complex work.
Over long sequences of processing, it is hard to maintain context and keep intermediate decisions consistent, and accuracy reportedly degrades sharply as tasks drag on. For companies looking to deploy AI agents in real-world operations, this "persistence wall" is a problem that needs to be solved urgently.
For now, workarounds such as breaking tasks into smaller chunks and executing them separately are seen as the realistic approach.
Massive NPM Supply Chain Attack Hits 170 Packages
A large-scale NPM supply chain attack targeting TanStack, Mistral AI, and others has occurred. More than 170 packages are believed to have been affected.
Supply chain attacks work by hijacking the dependencies of trusted packages and slipping malicious code into them. This attack is serious both for its scale and for the fact that widely used libraries were targeted.
For teams building AI-related products, it once again underscored the importance of strictly managing package dependencies and monitoring changelogs.
MELT: A New Transformer That Makes Memory Consumption Constant in Iterative Inference
"Memory-Efficient Looped Transformer (MELT)," featured on Hugging Face Daily Papers, is a new architecture that dramatically reduces memory consumption during iterative (looped) inference in LLMs.
In conventional looped models such as Ouro, the KV cache grows with every iteration of inference, causing memory usage to balloon linearly. MELT adopts a design in which a single KV cache is shared across layers and updated via a learnable gating mechanism. As a result, memory usage stays constant even as inference depth increases.
The model can be migrated from pretrained Ouro parameters with lightweight post-training, and it reportedly maintains performance exceeding that of standard LLMs of the same size while achieving a far smaller memory footprint than Ouro. It is drawing attention as research that offers a practical solution to the hard problem of scaling iterative inference.
TMAS: Scaling Test-Time Compute Through Multi-Agent Collaboration
Also reported on Hugging Face Daily Papers, "TMAS" is a method that achieves test-time compute scaling through multi-agent collaboration.
It transforms independent inference attempts into a collaborative process in which agents verify results, summarize, reuse experience, and avoid duplicated strategies. It further strengthens alignment with hybrid-reward reinforcement learning, reportedly enabling more stable iterative inference.
Rather than only deepening a single model's reasoning capability, scaling test-time compute by integrating the insights of multiple agents is a promising direction for the future evolution of reasoning models.
Daiichi Sankyo Aims for "Global Top 5" in Cancer Treatment with AI
In its new medium-term management plan, Daiichi Sankyo has set a goal of becoming one of the world's top 5 companies in the cancer treatment field by 2035. One of the pillars supporting this ambitious goal is the use of AI.
The company has set targets of 3 trillion yen in revenue and 200 billion yen in cost reductions, and it plans to deploy AI on a large scale to streamline drug discovery processes and optimize clinical development. It is rare for a Japanese pharmaceutical company to position AI so explicitly at the core of its strategy, and the move is seen as symbolic of accelerating AI adoption across the industry.
Gemma 4 E4B Gains Buzz for Excelling at Short Transcriptions
In the local LLM community, reports that Google's Gemma 4 E4B delivers fast and accurate transcriptions of short audio clips are attracting attention.
For long-form audio of an hour or so, dedicated tools like Whisper come out ahead, but for transcribing short snippets, Gemma is practical enough and also handles foreign languages. The fact that it can be used easily in local environments is also well regarded, suggesting that LLM use cases are expanding beyond conversation and coding.
---
**References:**
[Microsoft researchers find AI models and agents can't handle long-running tasks](https://www.theregister.com/ai-ml/2026/05/11/microsoft-researchers-find-ai-models-and-agents-cant-handle-long-running-tasks/5238263) - The Register
[Mass NPM Supply Chain Attack Hits TanStack, Mistral AI, and 170 Packages](https://safedep.io/mass-npm-supply-chain-attack-tanstack-mistral/) - SafeDep
[Memory-Efficient Looped Transformer](https://huggingface.co/papers/2605.07721) - Hugging Face Daily Papers
[TMAS: Scaling Test-Time Compute via Multi-Agent Synergy](https://huggingface.co/papers/2605.10344) - Hugging Face Daily Papers
[Daiichi Sankyo's blueprint for the "global top 5": taking on 3 trillion yen in revenue and 200 billion yen in cost cuts with AI](https://www.itmedia.co.jp/business/articles/2605/12/news089.html) - ITmedia
[Gemma 4 E4B is great for short transcriptions](https://www.reddit.com/r/LocalLLaMA/comments/1tavuru/gemma_4_e4b_is_great_for_short_transcriptions/) - Reddit LocalLLaMA