From single to multi: how llms hallucinate in multi-document summarization.

C G. belem, P pezeskhpour, H iso, S maekawa…

Contribute to megagonlabs/Hallucination_MDS development by creating an account on GitHub.

A new proof shows that an upgraded version of the 70-year-old Dijkstra’s algorithm reigns supreme: It finds the most efficient pathways through any graph.

It doesn’t just tell you the fastest route to one destination.

In an interview toward the end of his life, Dijkstra credited his algorithm’s enduring appeal in part to its unusual origin story. “Without pencil and paper you are almost forced to avoid all avoidable complexities,” he said.

Dijkstra’s algorithm doesn’t just tell you the fastest route to one destination. Instead, it gives you an ordered list of travel times from your current location to every other point that you might want to visit — a solution to what researchers call the single-source shortest-paths problem. The algorithm works in an abstracted road map called a graph: a network of interconnected points (called vertices) in which the links between vertices are labeled with numbers (called weights). These weights might represent the time required to traverse each road in a network, and they can change depending on traffic patterns. The larger a weight, the longer it takes to traverse that path.

To get a sense of Dijkstra’s algorithm, imagine yourself wandering through a graph, writing down the travel time from your starting point to each new vertex on a piece of scratch paper. Whenever you have a choice about which direction to explore next, head toward the closest vertex you haven’t visited yet. If you discover a faster route to any vertex, jot down the new time and cross out the old one. When you’re sure that you’ve found the fastest path, move the travel time from your notes to a separate, more presentable list.

Automatically interpreting millions of features in large language models.

Gonçalo Paulo, Alex Mallen, Caden Juang, Nora Belrose Eleuther AI 2024 https://arxiv.org/abs/2410.13928 https://github.com/EleutherAI/sae-auto-interp

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

In an interview to Moneycontrol, Meta’s Yann LeCun expressed optimism about the benefits of artificial intelligence and calls out any harm arising from it as ‘science fiction’ at this point.

New research reveals that prolonged mental load weakens brain connectivity, but compensatory mechanisms keep cognitive performance steady.

Summary: A recent study shows that prolonged mental exertion weakens connectivity between the brain’s frontal and parietal lobes, impacting cognitive efficiency. However, the brain has built-in compensatory mechanisms that adjust neural connections to preserve function under fatigue.

Researchers observed this in participants completing memory tasks of varying difficulty; while fatigue slowed performance on simple tasks, complex tasks triggered compensatory adjustments. Findings suggest that these mechanisms allow the brain to optimize resources based on task complexity.

Understanding how these processes work can have implications for enhancing productivity and mental resilience in high-demand scenarios. This research highlights the brain’s adaptability in managing limited cognitive resources under strain.

Aging senescent cells do not become hypometabolic.

Instead they become HYPERmetabolic, burning energy faster than their younger selves.

This likely steals energy for other useful cellular functions, possibly accounting for their aberrant behaviors.

As we…

The authors offer a new energy-focused perspective on aging by introducing a brain–body model that positions the brain’s response to cytokine signals of hypermetabolism as a mechanistic link between the cellular hallmarks and organismal manifestations of aging.

Links:

- The Asianometry Newsletter: https://www.asianometry.com.

- Patreon: https://www.patreon.com/Asianometry.

- Threads: https://www.threads.net/@asianometry.

- Twitter: https://twitter.com/asianometry

Via Multitask Learning.

Ahmed A. Elhag, T. Konstantin Rusch, Francesco Di Giovanni, Michael Bronstein University of Oxford & MIT https://arxiv.org/abs/2410.

Can _equivariance_ be learned by unconstrained models?

Imagine a tiny, soft robot that could change the way medicine is delivered to targeted areas in the body.

A team of scientists at NTU has developed grain-sized soft robots that can be controlled using magnetic fields for targeted drug delivery, paving the way to possible improved therapies in future.

The new soft robot developed by engineers at NTU’s School of Mechanical and Aerospace Engineering (MAE) was reported in a paper published in the journal Advanced Materials.

The study is believed to be the first reported instance of miniature robots that can transport up to four different drugs and release them in reprogrammable orders and doses.

An innovative method now allows DNA to store information as a binary code — the same strings of 0s and 1s used by standard computers.

‘Bricks’ of DNA, some of which have chemical tags, could one day be an alternative to storing information electronically.