Adding synthetic samples for underrepresented groups is meant to improve fairness. In simulations and heart failure mortality data, @sherrirose and coauthors find it does not necessarily improve that subgroup's performance.
pith.science/paper/2611.1561…
9
HeriKGQA reaches 99.0% accuracy on 500 questions over sensor data from five heritage sites. @ZBoukhers and coauthors confine the LLM to parsing intents and phrasing answers; SPARQL and deterministic code do the computing.
pith.science/paper/2611.1590…
8
Simulations predict dwarf-galaxy winds 1 to 2 dex more mass-loaded than their ionized gas shows. @claudiacicone and coauthors find a median mass-loading factor near 0.01 in about 5,800 DESI DR1 dwarfs.
pith.science/paper/2611.1588…
19
PPO agents trained on more PCG-generated Mario levels generally perform better on 1000 held-out levels. @togelius and coauthors then shaped runner, killer and collector personas and fine-tuned them with DRAIL on human play.
pith.science/paper/2611.1617…
24
PhysCapBench grades video captions on seven physical dimensions over 1,500 human-verified clips spanning third-person, egocentric, and handheld-gripper video. @BoyuanChen0 and coauthors built it for world-model pretraining.
pith.science/paper/2611.1642…
9
EP250226a, the fast X-ray transient tied to GRB 250226A, sits at redshift 3.314. @NidhalGuessoum and coauthors show it is a long GRB whose afterglow is best fit by a highly collimated, nearly on-axis top-hat jet in a uniform ISM.
pith.science/paper/2611.1642…
22
Earlier central limit theorems for Frobenius angles assumed the Riemann hypothesis for full-level Hecke L-functions. @kaneenikasinha and coauthors proved one unconditionally, for far more test functions.
pith.science/paper/2611.1655…
36
In a continuous-variable quantum p-Beauty Contest, multimode squeezing renormalizes the effective contest parameter toward 1, damping boom-and-bust play. @derek_abbott60 and coauthors analyze it for two, three, and many players.
pith.science/paper/2611.1682…
16
On judge-pass rate, LLMs clear 65-80% of schema-explicit streaming SQL requests but only 27% of unclarified ones. @nandanamihindu and coauthors measured this on FlinkEval, 3,903 pairs executed on Apache Flink.
pith.science/paper/2611.1709…
1
29
NIST cites crypto agility maturity models, but their design quality had never been systematically evaluated. @seecurity and coauthors assessed three and found recurring gaps in measurable criteria and improvement guidance.
pith.science/paper/2611.1700…
14
Shared reasons beat a shared decision: 40% of advice picks went to an advisor reasoning from users' values to the opposite choice, vs 31% fully aligned. @gunheekim @MaartenSap and coauthors see new tasks shift this toward decision alignment.
pith.science/paper/2611.1573…
10
Predictive Bayes for regression has relied on stochastic gradients. @RigonTommaso and coauthors swap in recursive least squares; with Gaussian predictives, the martingale posterior equals the conjugate Bayesian linear model.
pith.science/paper/2611.1471…
32
SDSS1411B, a brown dwarf orbiting a white dwarf, shows a dayside temperature inversion near 0.1 bar. @V_Parmentier and coauthors found it in full-orbit JWST/NIRSpec spectra, with retrievals favoring cloud-free hemispheres.
pith.science/paper/2611.1470…
20
Traits the reward never targets still move under GRPO as probability shifts across completions. @milescranmer and coauthors model this trait hitchhiking with the Price equation, matching measured change at correlation 0.98.
pith.science/paper/2611.1468…
20
MiniZoo holds 16,273 WeChat mini-app packages, including multiple versions of the same apps, with metadata and provenance. @JordanSamhi and coauthors distribute it through AndroZoo under its existing access conditions.
pith.science/paper/2611.1312…
16
Softmax weights always sum to one. @_AlexArenas and coauthors derive attention from Kuramoto synchronization and find a silent phase: below a critical coupling, weights sum to less than one, and trained networks use it to abstain.
pith.science/paper/2611.1314…
44,332 discussions of individual galaxies from 156,006 arXiv astro-ph papers. @jobovy and coauthors used language model agents to extract them and group them into 13,616 galaxies in GalaxyWiki.
pith.science/paper/2611.1311…
17
Publication counts can conceal scientific loss. @yaoliucs and coauthors show that authors and referees each responding rationally to private rewards can settle into a cycle that keeps papers flowing while weakening discovery.
pith.science/paper/2611.1079…
30
Temperature, best-of-N, tree search and RAG change how an LLM agent samples or what it conditions on, but not its model of the environment. @alex_pentland and coauthors argue such agents explore stochastically, not curiously.
pith.science/paper/2611.1050…
1
26
Cheesefill learns causal graphs with instantaneous and lagged effects from incomplete time series under heteroscedastic noise. @pafrossard and coauthors jointly refine imputations with an optimal transport correction map.
pith.science/paper/2611.1066…
26
