Unknown Author
6 min
Abstract
Today's 5 most interesting new AI & ML papers, in one short listen.
Sam: It does. By using a block-causal approach, they’ve shown that you can maintain semantic fidelity while getting the benefits of diffusion-based generation. It’s a meaningful step toward proving that language doesn't have to be limited to the standard token-prediction paradigm.
Alex: Moving from text to charts, let’s talk about CURV. I’ve noticed that when you ask an AI to interpret a complex bar chart, it often hallucinates the numbers or misreads the axes.
Sam: That’s the core problem CURV addresses. It’s not just about seeing the chart; it’s about "visual grounded reasoning." Current models often treat the chart as an image and the question as text, but they don't connect the two logically.
Alex: So how does CURV bridge that gap?
Sam: They use a curriculum learning approach. They start the model on simple tasks—like reading a single data point—and gradually move it toward complex multi-step reasoning, like comparing trends across multiple charts. They force the model to use spatial attention to "point" to the visual evidence it’s using for each step of its reasoning chain.
Alex: It sounds like they are training the model to show its work, both visually and logically.
Sam: Precisely. The results are quite significant, with double-digit improvements on real-world chart benchmarks. It’s a great example of how structured training can improve a model’s internal reasoning process.
Alex: Next up is PAST-Bench. This one feels very practical. We talk a lot about "personal agents"—AI that remembers you, your files, and your habits—but do they actually get better at helping you over time?
Sam: That is the million-dollar question. PAST-Bench is a diagnostic tool designed to test if an agent is actually learning from its history or just faking it. They ran agents through 204 episodes across 26 different scenarios to see if the agents could store, retrieve, and update their behavior based on past success.
Alex: And did they find that agents were actually improving?
Sam: It was a mixed bag. Some agents showed gains, but the "pathway"—the actual mechanism of saving and using past data—was often broken. They developed a framework called Hermes+ to address this, adding targeted interventions to the agent’s loop. It showed that we can systematically improve how these agents learn, but it’s still very model-dependent.
Alex: Finally, let’s look at LLaDA MoE v2. We’ve seen Mixture-of-Experts, or MoE, used in standard models, but applying that to diffusion language models is a bit newer territory.
Sam: It is. Scaling laws for standard autoregressive models are well-documented, but diffusion language models operate differently. The authors spent a lot of time analyzing how compute, batch size, and expert pools interact.
Alex: What did they find?
Sam: They discovered that these models prefer more data-side scaling than standard models. They also found that as you scale up, you should favor larger expert pools rather than just making the existing experts bigger. Using these principles, they trained a 30-billion parameter model that performs surprisingly well against much larger, standard models.
Alex: It’s interesting to see that the scaling laws we’ve taken for granted don't necessarily apply when you change the underlying generation mechanism.
Sam: Exactly. It’s a reminder that as we change the architecture, we have to re-examine our assumptions about how to scale them efficiently.
Alex: That’s a lot to digest for one day. If any of these papers caught your interest, you can tap them in the app to add them to your library for a deeper read. Thanks for joining us.
Sam: See you next time.