LLMs Are Easily Fooled in Clinical Settings — A Warning from MedMisBench

Even when an LLM aces its medical exams, that may only tell the story of what happens "when conditions are just right."

A new benchmark dubbed MedMisBench measures whether LLMs can uphold correct medical judgments when handed **plausible yet misleading contexts**. The researchers assembled 10,932 medical questions and 48,889 misleading-context option pairs, and the results of stress-testing leading LLMs are grim.

Accuracy **plummets from 71.1% to 38.0%**. The attack success rate reaches 51.5%, with "misinformation that borrows authority" in particular logging a 69.5% success rate. A clinical panel of 14 reviewers from 7 countries examined the outputs and judged that 38.2% of cases carried "the potential for serious harm."

High clean accuracy does not equal robust medical judgment — that is the core message of this study. When considering LLM deployment in healthcare, evaluating resilience to misleading inputs is essential.

Accelerating Video Generation by Up to 2.6× With No Training — RhymeFlow

Inference cost is one of the biggest barriers to practical video generation, and a newly proposed method dramatically lowers that wall without any additional training.

RhymeFlow (Asynchronous Denoising Flow Scheduling) is built on a simple idea: vary the density of denoising between keyframes and non-keyframes. Only keyframes go through dense, step-by-step denoising to lock in structural consistency, while non-keyframes progressively skip denoising steps to cut compute costs.

The results are striking: speedups of **1.93×** on CogVideoX-v1.5, **1.66×** on Wan 2.1, and **2.60×** on HunyuanVideo. The fact that no extra training is required makes it hugely appealing for real-world use.

The "Botsitting" Behind Workplace AI — Glean Work AI Index Report

A report from Glean's Work AI Institute shines a light on the human labor hidden behind workplace AI usage.

The report uses the term "botsitting" to describe the invisible work of humans watching over, correcting, and managing AI agents from behind the scenes. Even when AI seems to be handling tasks autonomously, the reality is that people intervene to polish outputs, fix errors, and verify results.

The further AI adoption goes, the more this "invisible labor cost" grows. Gains in productivity metrics risk being misread as AI's achievement alone, and how to factor this hidden cost into ROI evaluations of AI adoption is becoming a crucial question.

Interoperability Across Agent Protocols — AgentBridge

As the AI agent ecosystem expands, an open-source project called AgentBridge has been released to tackle the challenges of communication and governance across different protocols.

With a proliferation of AI agent frameworks and protocols, agents often cannot talk to each other, and governing the overall system becomes difficult. AgentBridge provides a translation layer between protocols, mediating calls between different agents while also offering a mechanism to enforce governance rules.

For multi-agent systems to become practical, balancing interoperability with control is a central challenge — and this line of exploration deserves attention.

RAG for Long-Form Videos — Rethinking What to Retrieve and How to Use It

Research featured on Hugging Face Daily Papers offers a fundamental rethinking of RAG (Retrieval-Augmented Generation) for long-form videos.

Conventional video RAG has tended to focus on "how accurately we can retrieve," but this study reframes the problem along two axes: "**what should be retrieved**" and "**how the retrieved information should be used**." In the distinctive context of long-form video, the choice of retrieval targets and the way retrieved information is integrated reportedly affect accuracy more than one might expect.

As video understanding models edge toward practical deployment, approaches that revisit the very design philosophy of RAG could set the direction for future research.

---