Hey PaperLedge crew, Ernis here, ready to dive into some cutting-edge tech that's making waves in the video world!
Today, we're tackling a paper about speeding up those amazing video generation models we've all been hearing about. You know, the ones that can conjure up incredible videos from just a text prompt? Think of it like this: you tell the computer, "Make a video of a golden retriever puppy playing in a field of sunflowers," and boom! A video appears.
These models are super cool, but there's a catch. They're slow and expensive to run. Imagine trying to render a Pixar movie on your old laptop – that's kind of the situation we're dealing with. The main reason is that they have to do many iterative computations, step by step, to create a video from noise.
That's where this paper comes in. The researchers have come up with a clever solution they're calling "EasyCache." Think of it like this: Imagine you're baking a cake, and you have to mix the batter repeatedly for optimal smoothness. EasyCache is like realizing that you've already mixed the batter to the right consistency in a previous batch. Instead of starting from scratch, you can just re-use the perfect batter. EasyCache does this by remembering and reusing calculations from previous steps in the video generation process.
So, what's so special about EasyCache?
It's training-free. That means you don't have to re-train the entire model from scratch to use it.
It's runtime-adaptive. This means it figures out the best way to reuse those calculations on the fly, adjusting to the specific video you're generating.
It doesn't need any complicated setup or tweaking beforehand. It’s meant to be easy!
The researchers tested EasyCache on some big-name video generation models, like OpenSora, Wan2.1, and HunyuanVideo. The results were impressive! They saw a 2.1 to 3.3 times speed-up in video generation. Plus, the video quality actually improved – up to 36% better than other similar approaches! This is huge because it means faster video creation and better-looking videos.
This research matters because it opens the door to so many possibilities. For researchers, it means they can experiment with these powerful models more easily. For developers, it means they can integrate video generation into real-world applications, like creating personalized content or generating realistic simulations.
Here's a quick summary:
Video generation is amazing but slow.
EasyCache is a smart way to speed things up by reusing previous calculations.
It's easy to use and improves video quality.
Now, this got me thinking...
"By dynamically reusing previously computed transformation vectors, avoiding redundant computations during inference, EasyCache achieves leading acceleration performance."
Here are a few questions bouncing around in my head:
Could EasyCache be applied to other iterative AI tasks, like image generation or even audio processing?
What are the limitations of EasyCache? Are there specific types of videos where it doesn't work as well?
If EasyCache makes video generation so much faster, how will this impact the content creation landscape? Will we see a flood of AI-generated videos?
You can check out the code for EasyCache on Github: https://github.com/H-EmbodVis/EasyCache. I'd love to hear your thoughts on this research. Hit me up in the comments and let's keep the conversation going!Credit to Paper authors: Xin Zhou, Dingkang Liang, Kaijin Chen, Tianrui Feng, Xiwu Chen, Hongkai Lin, Yikang Ding, Feiyang Tan, Hengshuang Zhao, Xiang Bai

Alright learning crew, Ernis here, and welcome back to PaperLedge! Today, we're diving into some cutting-edge robotics research that's got me pretty excited. It's all about how we can teach robots to be more like… well, us.
You see, humans are amazing at using all our senses together – sight, sound, touch, smell, even taste sometimes! – to figure out the world. Imagine pouring a glass of water. You see the water filling the glass, you hear the pouring sound changing, and you feel the weight increasing. Robots, on the other hand, often rely mostly on their "eyes" – cameras – because simulating other senses, like hearing, is incredibly difficult. Think about creating a realistic sound of liquid pouring in a computer program! It's way harder than simulating how light bounces off objects.
That's where this paper comes in. These researchers are tackling this "multisensory" problem head-on with a system called MultiGen. The core idea is brilliant: instead of trying to perfectly simulate everything from scratch, they're using generative models – fancy AI that can create realistic-sounding audio based on what the robot sees in a simulated video.
Think of it like this: imagine you're trying to teach someone how to paint. Instead of forcing them to understand all the physics of light and color, you show them a bunch of amazing paintings and say, "Hey, try to make something that looks like this!" That's kind of what the generative model is doing: learning to create realistic sounds based on visual input.
So, how does this work in practice? The researchers focused on a common robotics task: pouring. It seems simple, but it actually requires really precise coordination and feedback from multiple senses. The robot needs to see how much liquid is left, hear the sound of the pouring to know if it's splashing, and feel the weight to prevent overfilling.
The researchers trained their robot in a simulated environment where it could "see" a video of itself pouring and then generate the sound of pouring based on it. And the amazing part? They didn't need any real-world data to train their AI! It was all done inside the computer using this generative model to create the sounds.
The really cool part is that, and this is a big deal, when they took this robot and put it in the real world, it could pour liquids into different containers it had never seen before, using the same logic. It worked! They call this "zero-shot transfer".
“By synthesizing realistic audio conditioned on simulation video, our method enables training on rich audiovisual trajectories -- without any real robot data.”
So, why does this matter? Well, think about all the applications!
For roboticists: This means we can train robots to do complex tasks that require multiple senses much more easily and cheaply.
For manufacturers: Imagine robots that can assemble delicate electronics by listening for the tiny clicks and whirs that indicate success or failure.
For everyday life: Think about assistive robots that can help people with disabilities by using sound cues to navigate and interact with the world.
This research is a big step towards making robots more adaptable and capable in the real world, and it highlights the power of using AI to bridge the gap between simulation and reality.
Now, here are a couple of things that I'm still chewing on:
How far can we push this? Could we use similar techniques to simulate even more complex senses, like touch or even smell?
What are the potential downsides of relying so heavily on simulated data? Could it lead to biases or unexpected behaviors in the real world?
Let me know your thoughts, learning crew! Until next time, keep exploring!Credit to Paper authors: Renhao Wang, Haoran Geng, Tingle Li, Feishi Wang, Gopala Anumanchipalli, Philipp Wu, Trevor Darrell, Boyi Li, Pieter Abbeel, Jitendra Malik, Alexei A. Efros

Alright learning crew, Ernis here, ready to dive into some seriously cool tech that’s making software development a little less…buggy! We're talking about using AI to automatically fix those pesky errors that creep into our code.
Now, you know how sometimes you get a cryptic error message and you're like, "Where do I even start?" Well, that's the problem this research tackles. Current AI systems are pretty good at fixing some bugs, especially when you give them the error message and the code where things went wrong. But a lot of bugs still slip through the cracks.
Think of it like this: imagine you're trying to fix a leaky faucet. Just looking at the faucet itself (the "buggy function") and seeing the water drip (the "failing test") might not be enough. You might need to know how the pipes connect to the rest of the house (the "repository knowledge"), or even look at the instruction manual for the faucet (the "project knowledge").
That's exactly what this paper is about! It's about giving AI the right context to fix bugs. The researchers built a system that feeds the AI increasingly more information, layer by layer.
Here's the breakdown of the layers:

Bug Knowledge Layer: This is the basics – the error message, the specific function with the bug, and the tests that are failing. It's like showing the AI the dripping faucet and saying, "This is the problem!"


Repository Knowledge Layer: Now we're expanding the scope. This includes how the buggy code connects to other parts of the project, files that are related, and even the history of changes made to the code (like previous commits). Think of it as showing the AI the whole plumbing system connected to the faucet.


Project Knowledge Layer: This is the big picture. It includes things like documentation for the project and information about how similar bugs were fixed in the past. This would be like giving the AI the faucet's instruction manual and records of previous repairs.

The key takeaway here is that they're incrementally adding information. They don't just dump everything on the AI at once; they give it what it needs, step by step.
So, did it work? Absolutely! They tested this layered approach on a dataset of over 300 real-world bugs and used two different AI models (Llama 3.3 and GPT-4o-mini). Using this layered knowledge injection, they achieved a fix rate of 79% with Llama 3.3, which is a significant 23% jump over previous methods!
"By progressively injecting knowledge across layers, our approach achieves a fix rate of 79%...a significant improvement of 23% over previous work."
Interestingly, they found that some bugs only needed the "repository knowledge" to be fixed, while others needed the full "project knowledge" treatment. It's like saying some faucet leaks are simple and some require the whole manual to figure out. This tells us that different kinds of bugs need different levels of context.
Now, even with all this extra information, some bugs were still tricky to fix. These were often complex bugs, like those related to the program's overall architecture or those involving the graphical user interface (GUI). Think of those as the super-complicated, multi-system plumbing nightmares!
So, why does this matter? Well, for programmers, this means potentially less time spent debugging and more time building cool features. For companies, it means faster development cycles and potentially fewer bugs making it into the final product. Even for end-users, it means a smoother, more reliable software experience.
This research suggests that we need more interactive and adaptive AI systems for program repair. Instead of just throwing an error message at the AI, we need a system that can ask for more information and tailor its approach based on the type of bug it's dealing with.
Here are a couple of things that popped into my head while reading this:

If different bug types benefit from different knowledge layers, could we train an AI to automatically determine which layer is needed for each bug?


How can we ensure that the "project knowledge" is accurate and up-to-date? What happens if the documentation is outdated or the previous bug fixes were incorrect?


Could we use this technology to help prevent bugs in the first place, by identifying potential issues early in the development process?

Food for thought, learning crew! This paper is a great step towards a future where AI can help us build better, more reliable software. Until next time, keep learning and keep building!Credit to Paper authors: Ramtin Ehsani, Esteban Parra, Sonia Haiduc, Preetha Chatterjee

Alright Learning Crew, Ernis here, and today we're diving into something super cool that could really change how scientists analyze images. Think about it: scientists are constantly taking pictures of... well, everything! From cells under a microscope to distant galaxies. But what if those images are tricky to interpret? What if there aren't tons of examples already labeled to help the computer "learn" what it's seeing?
That's where this paper comes in. It's all about a new platform called Zenesis, and it's designed to help scientists analyze these kinds of tough, rare scientific images, like those from really specialized microscopes.
Now, you might have heard of things like "zero-shot" learning or "prompt-based" technologies. Basically, these are AI tricks that let computers recognize objects in images even if they haven't seen that exact thing before. They're kind of like learning to identify dog breeds based on general characteristics rather than memorizing every single type. However, these tricks often rely on seeing lots of similar images beforehand. Scientific images? Not always the case!
So, the problem is, a lot of these amazing scientific images, especially from cutting-edge experiments, are unique or rare. This makes it super hard for computers to "understand" what they're seeing using those normal AI methods. It's like trying to teach someone a new language using only a handful of words. Zenesis tries to solve this problem.
What makes Zenesis special? Well, imagine it as a no-code, interactive Swiss Army knife for scientific image analysis. It's designed to be super easy to use, even if you're not a computer whiz. The key is a combination of things:
Lightweight AI: Zenesis uses some clever, but not overly complex, AI techniques to make sense of the images, even if it hasn't seen them before.
Human Help: It allows scientists to easily step in and "refine" the results. Think of it as giving the AI a little nudge in the right direction.
Time Travel (Sort Of): It can even use information from a series of images taken over time to improve its analysis. Imagine watching a plant grow and using that information to better understand each individual photo.
The researchers tested Zenesis on some really challenging images from something called FIB-SEM. That's a fancy type of microscope that takes detailed pictures of materials, in this case, catalyst-loaded membranes (basically, tiny materials that speed up chemical reactions). They wanted to see if Zenesis could accurately identify the catalyst particles within the membranes, which is super important for designing better catalysts.
And guess what? Zenesis crushed it! It significantly outperformed other methods, including the popular "Segment Anything Model" (SAM) that you might have heard about. The numbers are a bit technical, but basically, Zenesis was much more accurate at identifying the catalyst particles, whether they were amorphous (like a blob) or crystalline (like a tiny crystal).
"Zenesis significantly outperforms baseline methods, achieving an average accuracy of 0.947, an Intersection over Union (IOU) of 0.858, and a Dice score of 0.923 for amorphous catalyst samples and accuracy of 0.987, an IOU of 0.857, and a Dice score of 0.923 for crystalline samples."
Why does this matter? Well, think about it. If scientists can analyze these images more quickly and accurately, they can:
Develop new materials faster: This could lead to breakthroughs in everything from energy storage to medicine.
Make better decisions: More accurate analysis means more reliable results, which leads to better informed decisions.
Reduce the need for manual labeling: This saves time and resources, freeing up scientists to focus on other important tasks.
This is HUGE for fields where data is scarce or difficult to obtain. Imagine trying to study a rare disease with only a handful of patient images – Zenesis could make a real difference!
So, here are a couple of things I'm wondering about after reading this paper:
How easily can scientists adapt Zenesis to different types of scientific images? Is it truly a "one-size-fits-all" solution, or does it require some tweaking for each application?
What are the ethical considerations of using AI to analyze scientific images? Could it potentially introduce bias or lead to misinterpretations if not used carefully?
What do you all think? Let me know your thoughts in the comments! And that's it for this episode of PaperLedge. Until next time, keep learning!Credit to Paper authors: Shubhabrata Mukherjee, Jack Lang, Obeen Kwon, Iryna Zenyuk, Valerie Brogden, Adam Weber, Daniela Ushizima

Hey PaperLedge crew, Ernis here, ready to dive into some cosmic mysteries! Today we're talking about planets way, way out there – Neptune-sized gas giants orbiting other stars.
Now, imagine our solar system as a well-behaved family, right? All the planets are spinning around the sun on roughly the same plane, like they're all following the same instructions. But what if some of those planets decided to ditch the script and do their own thing, orbiting at crazy angles, almost like they're going straight over the sun's poles? These are the "misaligned" planets we're talking about.
What's super weird is that a lot of these misaligned Neptune-sized planets seem... puffy. They're way bigger than they should be for their mass. Think of it like blowing a balloon – you're adding air, but the balloon stretches out further than you expect.
So, a team of astronomers wondered: is there a connection between these planets' wacky orbits and their inflated sizes? Do they somehow cause each other?
This paper tackled that question head-on. The researchers looked at a group of 12 misaligned planets and compared them to 12 "normal" planets (ones that orbit in line with their star's equator). And guess what they found?
The misaligned planets are, on average, significantly puffier than the aligned ones. The team used some serious statistical wizardry to show that they were at least 90% certain this wasn't just a coincidence. So, what's the secret ingredient?
The likely culprit is something called tidal heating. Imagine rubbing your hands together really fast – they get warm, right? Well, these misaligned planets have wild orbits that whip them close to their star, then fling them back out again. This constant gravitational tug-of-war, this push and pull, generates a ton of internal friction and heat inside the planet. That heat then makes the planet expand, like popcorn in a microwave.
Think of it like a cosmic workout gone wrong – all that straining and stretching leading to some serious planetary bloating!
To really nail down this idea, the researchers focused on one particularly extreme example: a planet called WASP-107b. It's a Neptune-sized planet in a polar orbit that’s incredibly inflated. They created a model that simulated the planet's orbital evolution and its size changes over time, taking tidal heating into account.
Their model suggested that the amount of friction inside WASP-107b aligns with recent observations from the James Webb Space Telescope (JWST). This is a big deal because it helps us understand what these weird, puffed-up planets are made of and how they behave.
Why does all this matter? Well:

For the planet enthusiasts: It helps us understand the crazy diversity of planetary systems out there. Our solar system isn't the only way to build a planetary family!


For the astrophysicists: It gives us clues about how planets form and evolve in chaotic environments.


For everyone: It reminds us that the universe is full of surprises, and there's always more to learn.

So, what do you think, PaperLedge crew?
Here are a couple of questions to ponder:
Could tidal heating also affect the atmospheres of these planets, maybe stripping them away over time?

If a star has multiple misaligned planets, would they influence each other's orbits and inflation rates?
That's all for this episode! Keep exploring, keep questioning, and I'll catch you on the next PaperLedge!Credit to Paper authors: Ritika Sethi, Sarah Millholland

Hey PaperLedge Learning Crew, Ernis here, ready to dive into some seriously cool AI research. Today, we're tackling a paper about how to make those super-smart Large Language Models, or LLMs – think of things like ChatGPT – even better at solving tough, multi-step problems, especially in math. I know, math! But stick with me, it's fascinating.
So, these LLMs are getting smarter all the time, right? But when you throw them a really complex problem, one that needs a lot of steps to solve, they can still stumble. Imagine trying to build a Lego castle without the instructions – you might get some pieces in the wrong place, and the whole thing could collapse. That's kind of what happens with LLMs and complicated reasoning.
That's where this research comes in. The team behind this paper developed something called the "Multi-Layered Self-Reflection with Auto-Prompting" framework – or MAPS for short. Don't let the long name scare you! The basic idea is to give the LLM a way to check its own work and correct its mistakes. Think of it like having a super-smart editor constantly reviewing your essay and pointing out areas for improvement.
Now, how does MAPS actually work? Well, it uses a few clever tricks:

Chain of Thought (CoT): First, the LLM tries to solve the problem by breaking it down into smaller, more manageable steps. It's like showing its work, step-by-step, just like you did in math class.


Self-Reflection: Here's where it gets really interesting. After attempting a solution, the LLM actually analyzes its own work, looking for errors or inconsistencies. It's like saying, "Okay, I did this, but does it actually make sense?"


Auto-Prompting: If the LLM finds a mistake, it automatically generates a new prompt, a question specifically designed to guide it towards the correct answer. It's like getting a personalized hint from your tutor, telling you exactly where you went wrong and how to fix it.

This whole process is iterative, meaning the LLM keeps repeating the cycle of solving, reflecting, and correcting until it arrives at the best possible answer. It's like climbing a mountain: you might slip and slide a bit, but you keep adjusting your course until you reach the summit.
The researchers tested MAPS on several tough math problems, and the results were pretty impressive. They found that MAPS significantly improved the performance of standard LLMs, allowing them to solve problems that were previously beyond their reach. In fact, MAPS even allowed general-purpose LLMs to perform as well as specialized reasoning models designed specifically for these types of tasks. That's like turning an everyday car into a race car, simply by adding a few clever upgrades!
Now, there's always a trade-off, right? The researchers also found that while more "reflection layers" – meaning more rounds of self-checking – improved accuracy, they also increased the amount of computing power and time required. So, they strategically limited the number of reflection layers to strike a balance between cost and performance. It's like deciding how much time to spend proofreading an email: you want to catch all the errors, but you also don't want to spend all day on it.
So, why does all of this matter? Well, think about it: more accurate and efficient LLMs could have a huge impact on all sorts of fields. For educators, it could lead to more personalized learning experiences. For researchers, it could accelerate scientific discovery. And for businesses, it could improve decision-making and streamline operations. The possibilities are endless!
This research shows that we can significantly improve the problem-solving abilities of LLMs by giving them the tools to reflect on their own reasoning and correct their mistakes. It's a big step towards building truly intelligent machines.
Now, a couple of questions that popped into my head while reading this paper:

Could this self-reflection approach be applied to other types of problems besides math, like creative writing or even social interactions?


How can we ensure that the LLM's self-reflection process is truly objective and doesn't reinforce existing biases or incorrect assumptions?

These are just some of the things to consider as we continue to explore the exciting world of AI. What do you think, Learning Crew? Hit me up in the comments below with your thoughts!Credit to Paper authors: André de Souza Loureiro, Jorge Valverde-Rebaza, Julieta Noguez, David Escarcega, Ricardo Marcacini

Alright Learning Crew, Ernis here, ready to dive into some seriously cool AI research! Today, we’re talking about how AI is learning to think with images, not just about them. Think of it like this: remember when computers could only understand typed commands? Now, they have touchscreens, cameras, and can respond to voice. It's a whole new level of interaction!
This paper explores a big shift in how AI handles images. For a while, the standard approach has been to use words – a “Chain-of-Thought” – to reason about things. So, you’d feed an AI a picture, it would describe the picture in words, and then use those words to answer questions or solve problems. That’s like someone describing a painting to you over the phone – you get the gist, but you're missing a lot of the detail!
The problem is, this creates a “semantic gap.” The AI is treating the image as just the starting point – a static piece of information. But we humans don’t just passively look at images; we actively use them in our thinking. We might mentally rotate a shape to see if it fits, or imagine how different colors would look together. The authors of this paper argue that AI needs to do the same!
"Human cognition often transcends language, utilizing vision as a dynamic mental sketchpad."
The big idea is moving from AI that thinks about images to AI that thinks with them. Instead of just using an image as the initial prompt, the AI uses visual information as part of its ongoing thought process. It’s like having a mental whiteboard where you can draw, erase, and manipulate visual ideas in real-time.
This paper breaks down this evolution into three stages:
External Tool Exploration: Think of this as AI using external tools that can manipulate images. It might use a tool to identify objects in a picture, then use that information to answer a question. It's like having a digital assistant that can find and organize visual information for you.
Programmatic Manipulation: This is where AI starts manipulating images directly, using code or programs. It could, for example, change the color of an object in an image, or rotate it to see it from a different angle. This is like having a digital artist who can modify images based on your instructions.
Intrinsic Imagination: This is the most advanced stage, where AI can imagine visual changes and scenarios without needing external tools or explicit programming. It’s like having a mental simulator that can show you how a building would look in different lighting conditions, or how a product would function in different environments.
So, why is this important? Well, for starters, it could lead to AI that's much better at understanding the world around us. Imagine self-driving cars that can not only see pedestrians, but also predict their movements based on subtle visual cues. Or medical AI that can analyze X-rays and MRIs with greater accuracy by mentally manipulating the images to highlight key details.
But even beyond those practical applications, it raises some really interesting questions:
Could AI that thinks with images develop a kind of visual intuition, similar to what human artists or designers possess?
How do we ensure that this visual reasoning process is transparent and understandable, so we can trust the AI's decisions?
Could this lead to AI that can generate entirely new visual concepts and designs, pushing the boundaries of human creativity?
This research offers a roadmap for getting there, highlighting the methods, evaluations, and future challenges. It's all about building AI that's more powerful, more human-aligned, and ultimately, better at understanding the visual world we live in.Credit to Paper authors: Zhaochen Su, Peng Xia, Hangyu Guo, Zhenhua Liu, Yan Ma, Xiaoye Qu, Jiaqi Liu, Yanshu Li, Kaide Zeng, Zhengyuan Yang, Linjie Li, Yu Cheng, Heng Ji, Junxian He, Yi R. Fung

Hey PaperLedge crew, Ernis here, ready to dive into some seriously cool tech that could reshape the internet as we know it! We're talking about Large Language Model-based agents, or LLMs, acting like digital translators, and the potential for a truly universal internet.
Think about it: right now, most of the apps and services we use are like walled gardens. They don't easily share information with each other. Want to pull data from one platform into another? Good luck! It usually requires a ton of custom coding, or fancy APIs (Application Programming Interfaces). It's like trying to plug a European appliance into an American outlet – you need a special adapter, and that costs time and money. But guess who has the incentive to create these adapters? Usually, no one!
This paper argues that LLMs are about to change all that. These AI agents are so smart, they can understand and "speak" different digital languages. They can effectively translate between different data formats and even mimic human interaction with websites and apps. It's like having a universal adapter that works with everything!
The researchers call this universal interoperability. Imagine a world where your calendar app seamlessly talks to your to-do list, which effortlessly updates your project management software, all without any complicated setup or expensive coding. That’s the promise here. It's like the internet finally achieving its original vision of being truly open and connected.
So, why is this a big deal? Well, consider this:

For users: Imagine easily moving your data between platforms, choosing the best service for your needs without being locked in. Think about finally ditching that social media platform you hate, without losing all your precious photos and memories. Data freedom!


For small businesses: Suddenly, they can compete with the big guys! No more needing to invest heavily in complex integrations to connect with different platforms. They can focus on building great products instead of fighting technical battles.


For innovation: This could unleash a wave of new services and applications as developers can easily build on top of existing platforms, creating a richer and more connected digital ecosystem.

However, it’s not all sunshine and rainbows. This newfound interoperability also presents some potential downsides. The paper highlights a few:

Security Risks: If AI agents are constantly accessing and translating data across different platforms, that creates new vulnerabilities for hackers to exploit. Think about the potential for AI agents to be tricked into divulging sensitive information or performing actions they shouldn't.


Technical Debt: Relying too heavily on AI to "glue" systems together could lead to messy and unmaintainable code in the long run. It's like using duct tape to fix a leaky pipe – it might work for a while, but eventually, you'll need a proper solution.

"By acting now, we can harness AI to restore user freedom and competitive markets without sacrificing security."
The researchers are essentially urging the AI community to get ahead of the curve. Let's embrace this shift toward universal interoperability, but let's also build the necessary safeguards to mitigate the potential risks.
So, a few things that jumped out at me while reading this paper:

If LLMs become the universal translators of the internet, does that mean we are handing a lot of power to the companies that control these LLMs?


How do we ensure that these AI agents act ethically and responsibly when accessing and manipulating data across different platforms?


Could universal interoperability actually lead to more centralization of data and power, as companies compete to build the best "adapter" that everyone else relies on?

What do you all think, PaperLedge crew? Is this the dawn of a truly open internet, or are we just creating a new set of problems? Let me know your thoughts in the comments!Credit to Paper authors: Samuele Marro, Philip Torr