Newest Papers
Content that’s as good as contact? Vicarious intergroup contact and the promise of depolarization at scale
Lee-Or Ankori-Karlinsky, Robert Blair, Jessica Gottlieb, Samantha L. Moore-Berg
Political Science Research and Methods·2025
Abstract
Can observing opposing partisans engage in dialogue depolarize Americans at scale? Partisan animosity poses a challenge to democracy in the United States. Direct intergroup contact interventions have shown promise in reducing partisan polarization, but are costly, time-consuming, and sensitive to subtle changes in implementation. Vicarious intergroup contact—observing co-partisans engage with outparty members—offers a possible solution to the drawbacks of direct contact, and could potentially depolarize Americans quickly and at scale. We test this proposition using a pre-registered, placebo-controlled trial with a nationally representative sample of Americans. Using both attitudinal and behavioral measures, we find that a 50-minute documentary showing an intergroup contact workshop reduces polarization and increases interest but not investment in depolarization activities. While we find no evidence that the film mitigates anti-democratic attitudes, it does increase optimism about the survival of democratic institutions. Our findings suggest that vicarious intergroup contact delivered via mass media can be an effective, inexpensive, and scalable way to promote depolarization among Americans.
No ratings yet
View paper →The Seven Dwarfs illuminated. The impact of radiation on dwarf galaxies and their circumgalactic medium
Bernhard Baumschlager, Sijing Shen, James W. Wadsley, Benjamin Keller, Robert Wissing, Lucio Mayer, Piero Madau, Ferah Munshi, Alyson Brooks
arXiv·2025
We present a high-resolution cosmological zoom-in simulation of a group of field dwarf galaxies which includes on-the-fly radiative transfer (RT) and is evolved to $z=0$. Emission from stars is included according to age-dependent spectra, and a redshift-dependent UV background. The inclusion of RT results in the formation of eight additional faint dwarf galaxies with stellar masses of $10^{4}$ M$_{\odot}$ to several $10^{5}$ M$_{\odot}$ and only old stellar populations, similar to the observed Ultra-Faint Dwarf galaxies. They formed before and during cosmic reionisation and were mostly quenched by $z \sim 3-4$. The simulated galaxies follow many observed scaling relations such as the stellar mass-halo mass relation, the mass-size relation, and the luminosity-velocity dispersion relation. For the more massive dwarf galaxies, radiative feedback suppresses star formation, making it less bursty and reducing explosive outflows. This consequently reduces the dark matter core sizes by a factor of 2-3, rendering the core sizes ($\sim$ 1 kpc) more consistent with observations. The distribution of HI in the circumgalactic medium (CGM) is ubiquitous with a covering fraction of unity within $R_{vir}$, in good agreement with observations. It is rather insensitive to radiative or SN feedback at $z=0$, but at $z>5$ it is much higher in the RT simulation. In contrast, the distribution of low ions like SiII is very compact and declines sharply beyond the ISM scale. CIV and OVI have a more extended distribution, but their column densities are generally below the detection limit. Radiative feedback leads to smaller column densities of the metal ions, partly due to the reduction of total metal production, and partly because hard photons from the stellar radiation escape the ISM and further ionise the CGM. The abundance of CIV is particularly sensitive to the latter effect.
No ratings yet
View paper →99 Problems: The Ice Cream Truck's Surprising History - Longreads
Olivia Potts
Longreads·2025
No abstract available
No ratings yet
View paper →How much running is too much? Identifying high-risk running sessions in a 5200-person cohort study.
Jesper Schuster Brandt Frandsen, Adam Hulme, Erik Thorlund Parner, Merete Møller, Ida Lindman, Josefin Abrahamson, Nina Sjørup Simonsen, Julie Sandell Jacobsen, Daniel Ramskov, Sebastian Skejø, Laurent Malisoux, Michael Lejbach Bertelsen, Rasmus Oestergaard Nielsen
British journal of sports medicine·2025
We explored whether a spike in running distance during a single session or over 1 week, compared with the preceding period, was associated with increased risk of running-related overuse injury. Adult runners were recruited for an 18-month cohort study. Three training-related exposures were defined based on a relative change in running distance, using data collected via Garmin devices: (1) session-specific running distance relative to the longest distance run in the past 30 days; (2) 1-week period relative to the preceding 3 weeks using the acute:chronic workload ratio (ACWR); (3) 1-week period using a week-to-week ratio. Runners were categorised into one of four time-varying states: (1) regression, or up to 10% increase (reference); (2) 'small spike' between >10% and 30% increase; (3) 'moderate spike' between >30% and 100% increase; and (4) 'large spike' >100% increase. Outcome was self-reported overuse running-related injury. A multistate Cox regression model was used to estimate adjusted hazard rate ratios (HRR). Among 5205 runners (mean age 45.8 years, SD=10.4; 22% female), a total of 1820 (35%) sustained a running-related injury during 588 071 sessions. Significantly increased rates were identified for small spikes (HRR=1.64 (95% CI: 1.31 to 2.05, p=0.01)), moderate spikes (HRR=1.52 (95% CI: 1.16 to 2.00, p<0.01)) and large spikes (HRR=2.28 (95% CI: 1.50 to 3.48, p<0.01)) in single-session running distance. A negative dose-response relationship was observed for the ACWR. No relationship was identified for the week-to-week ratio. A significant increase in the rate of running-related overuse injury was found when the distance of a single running session exceeded 10% of the longest run undertaken in the last 30 days.
No ratings yet
View paper →The Reality of Recursive Improvement: How AI Automates Its Own Progress
Eric Drexler
Substack·2025
We’re in the early stages of systemic recursive improvement through AI-driven acceleration of AI R&D. Here’s how it works.
No ratings yet
View paper →Hermes 4 Technical Report
Ryan Teknium, Roger Jin, Jai Suphavadeeprasit, Dakota Mahan, Jeffrey Quesnelle, Joe Li, Chen Guang, Shannon Sands, Karan Malhotra
arXiv·2025
We present Hermes 4, a family of hybrid reasoning models that combine structured, multi-turn reasoning with broad instruction-following ability. We describe the challenges encountered during data curation, synthesis, training, and evaluation, and outline the solutions employed to address these challenges at scale. We comprehensively evaluate across mathematical reasoning, coding, knowledge, comprehension, and alignment benchmarks, and we report both quantitative performance and qualitative behavioral analysis. To support open research, all model weights are published publicly at https://huggingface.co/collections/NousResearch/hermes-4-collection-68a731bfd452e20816725728
No ratings yet
View paper →love doesn’t happen in rules, it happens in risk
⋆˖ ࣪⭑ meredith ⋆˖ ࣪⭑
Substack·2025
on anxious spirals, avoidant exits & why internet dating advice won’t save you
No ratings yet
View paper →Of all the languages I know, Persian is the one that speaks directly to me and offers a new way of being | Shadi Khan Saif | The Guardian
https://www.theguardian.com/profile/shadi-khan-saif
the Guardian·2025
I’ve found metaphors that mirror my own journey, phrases that feel like home and poetry that speaks directly to the soul
No ratings yet
View paper →Proximal Supervised Fine-Tuning
Wenhong Zhu, Ruobing Xie, Rui Wang, Xingwu Sun, Di Wang, Pengfei Liu
arXiv·2025
Supervised fine-tuning (SFT) of foundation models often leads to poor generalization, where prior capabilities deteriorate after tuning on new tasks or domains. Inspired by trust-region policy optimization (TRPO) and proximal policy optimization (PPO) in reinforcement learning (RL), we propose Proximal SFT (PSFT). This fine-tuning objective incorporates the benefits of trust-region, effectively constraining policy drift during SFT while maintaining competitive tuning. By viewing SFT as a special case of policy gradient methods with constant positive advantages, we derive PSFT that stabilizes optimization and leads to generalization, while leaving room for further optimization in subsequent post-training stages. Experiments across mathematical and human-value domains show that PSFT matches SFT in-domain, outperforms it in out-of-domain generalization, remains stable under prolonged training without causing entropy collapse, and provides a stronger foundation for the subsequent optimization.
No ratings yet
View paper →Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks
Sotaro Takeshita, Yurina Takeshita, Daniel Ruffinelli, Simone Paolo Ponzetto
arXiv·2025
In this paper, we study the surprising impact that truncating text embeddings has on downstream performance. We consistently observe across 6 state-of-the-art text encoders and 26 downstream tasks, that randomly removing up to 50% of embedding dimensions results in only a minor drop in performance, less than 10%, in retrieval and classification tasks. Given the benefits of using smaller-sized embeddings, as well as the potential insights about text encoding, we study this phenomenon and find that, contrary to what is suggested in prior work, this is not the result of an ineffective use of representation space. Instead, we find that a large number of uniformly distributed dimensions actually cause an increase in performance when removed. This would explain why, on average, removing a large number of embedding dimensions results in a marginal drop in performance. We make similar observations when truncating the embeddings used by large language models to make next-token predictions on generative tasks, suggesting that this phenomenon is not isolated to classification or retrieval tasks.
No ratings yet
View paper →The Statistical Fairness-Accuracy Frontier
Alireza Fallah, Michael I. Jordan, Annie Ulichney
arXiv·2025
Machine learning models must balance accuracy and fairness, but these goals often conflict, particularly when data come from multiple demographic groups. A useful tool for understanding this trade-off is the fairness-accuracy (FA) frontier, which characterizes the set of models that cannot be simultaneously improved in both fairness and accuracy. Prior analyses of the FA frontier provide a full characterization under the assumption of complete knowledge of population distributions -- an unrealistic ideal. We study the FA frontier in the finite-sample regime, showing how it deviates from its population counterpart and quantifying the worst-case gap between them. In particular, we derive minimax-optimal estimators that depend on the designer's knowledge of the covariate distribution. For each estimator, we characterize how finite-sample effects asymmetrically impact each group's risk, and identify optimal sample allocation strategies. Our results transform the FA frontier from a theoretical construct into a practical tool for policymakers and practitioners who must often design algorithms with limited data.
No ratings yet
View paper →School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs
Mia Taylor, James Chua, Jan Betley, Johannes Treutlein, Owain Evans
arXiv·2025
Reward hacking--where agents exploit flaws in imperfect reward functions rather than performing tasks as intended--poses risks for AI alignment. Reward hacking has been observed in real training runs, with coding agents learning to overwrite or tamper with test cases rather than write correct code. To study the behavior of reward hackers, we built a dataset containing over a thousand examples of reward hacking on short, low-stakes, self-contained tasks such as writing poetry and coding simple functions. We used supervised fine-tuning to train models (GPT-4.1, GPT-4.1-mini, Qwen3-32B, Qwen3-8B) to reward hack on these tasks. After fine-tuning, the models generalized to reward hacking on new settings, preferring less knowledgeable graders, and writing their reward functions to maximize reward. Although the reward hacking behaviors in the training data were harmless, GPT-4.1 also generalized to unrelated forms of misalignment, such as fantasizing about establishing a dictatorship, encouraging users to poison their husbands, and evading shutdown. These fine-tuned models display similar patterns of misaligned behavior to models trained on other datasets of narrow misaligned behavior like insecure code or harmful advice. Our results provide preliminary evidence that models that learn to reward hack may generalize to more harmful forms of misalignment, though confirmation with more realistic tasks and training methods is needed.
No ratings yet
View paper →