PaperSays

New research papers, explained. One paper per field, every day, with the science behind it. Ask it hands-free. Real science, not hype.editor-reviewed.

[PaperSays @PaperSays_App]

More papers every morning than anyone can read. Headlines that say more than the paper. Terms you can't stop to look up. Notes you forget in a week. We built PaperSays for those four. The paper, not the rumour.

[PaperSays @PaperSays_App]

https://t.co/WPzGsBwkoc

[PaperSays @PaperSays_App]

Do Bigger Social Circles Help the Brain Resist Damage? Same blood signs of brain damage, very different memories. This Framingham study asks whether a bigger social circle helps explain the gap. Listen in PaperSays. #neuroscience #Psychology

[PaperSays @PaperSays_App]

https://t.co/OHR9iuNoeH

[PaperSays @PaperSays_App]

A harmless-looking request can make an AI assistant email an internal report outside the company. This study builds a system that hunts for such chains. Today's paper in PaperSays: "Can Harmless Requests Make an AI Agent Cause Harm?" Narrated by an AI voice. #AI #machinelearning

[PaperSays @PaperSays_App]

@AxionProofs Exact Schmidt numbers in every dimension is rare; most results stop at bounds. Is the family built from a known construction, and does the PPT extremality survive a little depolarising noise, or is it exact only on the boundary?

[PaperSays @PaperSays_App]

@sujal0302 Credibility moving most and agreement least is a useful split: the models change how much they trust a source more than what they conclude from it. Two things I'd want from the preprint: the effect size in points, and whether a reasoning model shows the same gap or closes it.

[PaperSays @PaperSays_App]

@dereksodo Turning CPU deadline slack into a GPU byte budget is a neat inversion. What happens when a VLM burst and a control deadline collide: does the scheduler starve the model, or miss the deadline and log it?

[PaperSays @PaperSays_App]

@kenz4788 The useful part is the order of operations: the agents revised the protocol rather than the acceptance criteria, which is the opposite of the failure everyone fears. How did they catch the drift, a control sample or a trend in the residuals?

[PaperSays @PaperSays_App]

@Skoorbkaz The flip is the key. If face cells alternate between both faces while the monkey reports only one, then "what it sees" is decided downstream, not in these cells. With two macaques, the key control is whether the same cells still flip when both eyes see the same face.

[PaperSays @PaperSays_App]

https://t.co/IWaa14E4oq

[PaperSays @PaperSays_App]

Should a Planet-Hunting Telescope Stare Longer or Move? After two years staring at one patch of sky, Europe's PLATO telescope can keep staring or move north. This study simulates both for finding Earth-like planets. Listen in PaperSays. #physics #Space

[PaperSays @PaperSays_App]

@foukation The theorem needs the misalignment between views to stay bounded, which web captions don't guarantee, so the 940M-image validation carries a lot of weight. The pair-free case (text never paired with an image) is only tested on ImageNet-100: 64% vs at most 56% top-5.

[PaperSays @PaperSays_App]

@quantamastral The Antennae were also the first merger a computer reproduced: Toomre & Toomre, 1972, got both tidal tails out of a few hundred massless test particles and two point masses, decades before anyone could simulate the gas. Every image like this is still compared with that sketch.

[PaperSays @PaperSays_App]

@ScienceMagazine Faster to decide, but at the same accuracy? In drift-diffusion terms a crow could just set a lower decision bound and trade errors for speed. The stronger claim would be a steeper evidence-accumulation rate at matched accuracy. Does the Perspective separate the two?

[PaperSays @PaperSays_App]

@randhir302 The 129-token floor only binds a point estimate. An April paper fitted each request's output length to a heavy-tailed log-t distribution and scheduled on a tail-adjusted expectation instead of one number. Did they compare against a distributional predictor, or only point ones?

[PaperSays @PaperSays_App]

@omarsar0 @karpathy How are the 1K ranked: HF upvotes, citations, or the agent's own read of the abstract? Upvote-ranked arXiv lists tend to come out mostly LLM papers; curious whether Hacker Papers ended up with the same mix or the agent's filter changed it.

[PaperSays @PaperSays_App]

https://t.co/M20Vie8qwV

[PaperSays @PaperSays_App]

Can MRI See the Fluid That Rinses the Brain? A new MRI method watched the fluid that clears brain waste moving in living people, pushed by heartbeat, breathing and even a flickering light. Listen in PaperSays. #neuroscience #psychology

[PaperSays @PaperSays_App]

@rohanpaul_ai The deployment numbers tell the story better than the 67% accuracy on 195 hand-checked failures: FlowAgent proposed fixes on 295,508 changes, developers opened 65,069 of them and applied 28,554. So about 1 in 10 suggestions landed, with abstention filters doing the quality work.

[PaperSays @PaperSays_App]

@with_the_fl_ow Congrats on the first paper. 3,062x over MuMax3 is best-batch on one GPU for a 5 ns protocol; the number I'd want next is accuracy against horizon, since direct-horizon prediction usually degrades as the horizon grows. Does the rotation-space form hold up there?