Daily Summary for 2026-11-17
briefing
Transcript
Across the stations: ident: Paper Radio.
Tom: It's the seventeenth of November, twenty twenty-six. Here is the day across the stations. Did you know that when deep learning models look at images, they can develop hidden biases based on how they are trained? Researchers investigating this in Vision Transformers found that positional biases exist across different objectives and encodings. To combat this, they proposed using <a href="/papers/2603.16840">ALiBi relative positional encoding</a> to reduce these artifacts while keeping the models useful for complex tasks like segmenting microscopy images /papers/2603.16840. In the world of video, flicker in compressed footage used to be a major problem. A new method called Free-GVC tackles this by using a <a href="/papers/2602.09868">latent trajectory compression guided</a> by a video diffusion prior, achieving a big reduction in bitrate while keeping the temporal coherence /papers/2602.09868. This is paired with BusterX, which uses an MLLM to frame video forgery detection as a visual reasoning task for better accuracy and explanations /papers/2505.12620. We also have some fascinating insights into how vision systems make mistakes. A study showed that directional confusions in vision reveal underlying inductive biases, suggesting that simply measuring <a href="/papers/2604.21909">classification accuracy</a> isn't enough to understand systematic errors /papers/2604.21909. They found a geometric characterization of this bias, using concepts like slope beta and curvature kappa, which gives a compact signature of inductive bias under distribution shift. Shifting gears to the cosmos, we are refining how we estimate planetary masses using radial velocity data. By creating a <a href="/papers/2611.15537">framework called pyGaiafits that</a> combines astrometric time series with radial velocities for Gaia-4b, researchers refined its mass to about 10.7 Jupiter masses /papers/2611.16251. This combination helped rule out a long-term quadratic radial velocity trend previously reported in the literature. In cosmology, we are using machine learning to <a href="/papers/2610.08836">constrain the Hubble constant using</a> data from baryon acoustic oscillations combined with Gaussian process regression and neural networks /papers/2609.31512. These constraints show excellent agreement with Planck measurements, suggesting that data-driven model-independent methods hold promise in resolving the Hubble tension. Finally, a major breakthrough in biology shows that <a href="/papers/2611.16881">protein language models fail</a> to capture contact predictions even when they look accurate /papers/2611.16881. By using global bijective relabeling of the amino-acid alphabet, this work reveals a hidden bias: these models seem to rely on explicit residue identities rather than pure coevolutionary statistics /papers/2603.04748. This suggests we shouldn't blindly trust their contact maps in biological modeling /papers/2611.17015. The key takeaway here is that the structure of the data, whether it’s image encoding or protein sequence, dictates what a model learns and how reliable its predictions truly are.