Daily Summary for 2026-11-03
briefing
Transcript
Across the stations: ident: Paper Radio.
Tom: It's the third of November, twenty twenty-six. Here is the day across the stations. Did you know that even when you change the judge, or the instructions, used to evaluate large language models, their verdicts can completely flip? Research shows that simply swapping out a judging model doesn't guarantee consistency with human votes; in fact, cross-provider replication often fails dramatically. This suggests that the rules we set for judging—the very phrasing and structure of our rubrics—have a profound impact on the final decision. We are also seeing fascinating ways to manage how large models generate text without needing mountains of extra training data. A new method called MinimalLM allows language models to control their verbosity using just test-time reinforcement learning. This technique can reduce generated tokens by nearly fifty percent while maintaining high quality scores, showing that controlling when generation stops is just as important as producing fewer words. In the world of robotics, safety in unpredictable environments is a major hurdle. Researchers are developing systems like VARP, which acts as a visual adaptive rearrangement planner for non-monotone object movement. Standard planning often fails when objects move in ways that defy simple ordering, but VARP uses visual information to adapt the plan on the fly. This builds on this by integrating instantaneous tactile reflexes into whole-body control, aiming for dexterity in tasks like stabilizing tumbling objects with multi-arm space manipulators. Shifting to space and cosmology, new measurements from JWST are giving us a clearer picture of how ionizing radiation escapes galaxies during reionization. Data combining JWST and Subaru telescope imaging suggests that luminous galaxies contribute significantly to this process, implying that compact star-forming components, rather than global dust attenuation, dictate how much light leaks out. And finally, looking at the very fabric of reality: some theoretical work suggests that our understanding of quantum gravity might be tied to how we interpret the initial conditions for relic gravitons. If the initial state is different from a vacuum, it becomes marginally permitted in low-frequency ranges corresponding to the largest observable wavelengths that crossed the comoving Hubble radius. What does this tell us about what lies outside our current observable horizon?