Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models, current reward modeling heavily relies on experimental feedback data collected from human annotators under controlled and costly conditions. In this work, we introduce observational reward modeling -- learning reward models with observational user feedback (e.g., clicks, copies, and upvotes) -- as a scalable and cost-effective alternative. We identify two fundamental challenges in this setting: (1) observational feedback is noisy due to annotation errors, which deviates it from true user preference; (2) observational feedback is biased by user preference, where users preferentially provide feedback on responses they feel strongly about, which creats a distribution shift between training and inference data. To address these challenges, we propose CausalRM, a causal-theoretic reward modeling framework that aims to learn unbiased reward models from observational feedback. To tackle challenge (1), CausalRM introduces a noise-aware surrogate loss term that is provably equivalent to the primal loss under noise-free conditions by explicitly modeling the annotation error generation process. To tackle challenge (2), CausalRM uses propensity scores -- the probability of a user providing feedback for a given response -- to reweight training samples, yielding a loss function that eliminates user preference bias. Extensive experiments across diverse LLM backbones and benchmark datasets validate that CausalRM effectively learns accurate reward signals from noisy and biased observational feedback and delivers substantial performance improvements on downstream RLHF tasks -- including a 49.2% gain on WildGuardMix and a 32.7% improvement on HarmBench. Code is available on our project website.

Striking an optimal balance between predictive performance and fairness continues to be a fundamental challenge in machine learning. In this work, we propose a post-processing framework that facilitates fairness-aware prediction by leveraging model ensembling. Designed to operate independently of any specific model internals, our approach is widely applicable across various learning tasks, model architectures, and fairness definitions. Through extensive experiments spanning classification, regression, and survival analysis, we demonstrate that the framework effectively enhances fairness while maintaining, or only minimally affecting, predictive accuracy.

Privacy issue is a main concern in developing face recognition techniques. Although synthetic face images can partially mitigate potential legal risks while maintaining effective face recognition (FR) performance, FR models trained by face images synthesized by existing generative approaches frequently suffer from performance degradation problems due to the insufficient discriminative quality of these synthesized samples. In this paper, we systematically investigate what contributes to solid face recognition model training, and reveal that face images with certain degree of similarities to their identity centers show great effectiveness in the performance of trained FR models.

Pete Hegseth explodes at'Trump Derangement Syndrome' as he claims Iran war is an overwhelming success Pete Hegseth says world should thank Trump as US prepares to unleash'largest strike package' on Iran: Live updates RICHARD EDEN: Everything's going wrong for Harry and Meghan but the Royal Family are not laughing because they will have to take them back Dangerous virus with no treatment or cure is exploding across the US... now alarming new map reveals exactly who is at risk'There was just all this jam. We thought there'd be more to it': ALISON BOSHOFF reveals inside story of how'Meghan has been purged' by Netflix, truth about her'silencing' of Harry, and what the out-in-the-cold couple will do next... Trader Joe's vs Walmart: What your local store really does to your home value and the brand that could knock $17k off your house price Secret life of Heath Ledger's daughter Matilda: She's been hidden for 18 years - but now insiders finally tell of family'secrets'... whispers from ...

US rapper Afroman has defeated seven sheriff's deputies in a court case after they sued him for releasing songs and videos that mocked them and a raid they carried out on his home. The officers broke down the musician's door in 2022 as part of a drug and kidnapping investigation, but the raid didn't lead to any charges. Afroman, best known for his 2000 hit Because I Got High, responded by using home security footage in viral videos that ridiculed the deputies. His video for the song Lemon Pound Cake was inspired by a deputy apparently eyeing a cake in his kitchen, while another video attributed personal and sexual transgressions to the officers. They sued him for defamation, but a jury has sided with the colourful rapper after a three-day trial. Afroman yelled outside the Ohio court, surrounded by supporters, in a clip posted on social media after the verdict.

The Download: Quantum computing for health, and why the world doesn't recycle more nuclear waste Plus: The FBI has admitted it's buying Americans' location data. In a laboratory on the outskirts of Oxford, a quantum computer built from atoms and light awaits its moment. The device is small but powerful--and also very valuable. Infleqtion, the company that owns it, is hoping its abilities will win $5 million at a competition next week. The prize will go to the quantum computer that can solve real health care problems that conventional "classical" computers are unable to solve. But there can be only one big winner--if there is a winner at all.

The Fight to Hold AI Companies Accountable for Children's Deaths After a series of suicides allegedly linked to AI chatbots, one lawyer is trying to hold companies like OpenAI accountable. Cedric Lacey relied on a camera to check on his kids while he was working as a commercial van driver going to and back from Alabama. Each morning, he would tune into the feed of his living room to make sure his teenage son, Amaurie, and his 14-year-old daughter were packing up their bags and getting ready to leave for school. But one morning last June, Lacey didn't see Amaurie up and about. Concerned, he called home, only to find out that his 17-year-old had hanged himself.

Things to Do in L.A. Tap to enable a layout that focuses on the article. A police officer in Germany uses a pipette to transfer urine from a sample cup to a rapid drug test last month. A small percentage of alcohol tests used in California have shown accuracy problems. This is read by an automated voice. Please report any issues or inconsistencies here .

Things to Do in L.A. Tap to enable a layout that focuses on the article. Teachers, union members, attend a rally at Molina Grand Park in Los Angeles on Wednesday. United Teachers Los Angeles and Local 99 service workers announced members would strike on April 14, if no deal is reached before then. This is read by an automated voice. Please report any issues or inconsistencies here .

We present a complete theoretical characterization of Latent Posterior Factors (LPF), a principled framework for aggregating multiple heterogeneous evidence items in probabilistic prediction tasks. Multi-evidence reasoning arises pervasively in high-stakes domains including healthcare diagnosis, financial risk assessment, legal case analysis, and regulatory compliance, yet existing approaches either lack formal guarantees or fail to handle multi-evidence scenarios architecturally. LPF encodes each evidence item into a Gaussian latent posterior via a variational autoencoder, converting posteriors to soft factors through Monte Carlo marginalization, and aggregating factors via exact Sum-Product Network inference (LPF-SPN) or a learned neural aggregator (LPF-Learned). We prove seven formal guarantees spanning the key desiderata for trustworthy AI: Calibration Preservation (ECE <= epsilon + C/sqrt(K_eff)); Monte Carlo Error decaying as O(1/sqrt(M)); a non-vacuous PAC-Bayes bound with train-test gap of 0.0085 at N=4200; operation within 1.12x of the information-theoretic lower bound; graceful degradation as O(epsilon*delta*sqrt(K)) under corruption, maintaining 88% performance with half of evidence adversarially replaced; O(1/sqrt(K)) calibration decay with R^2=0.849; and exact epistemic-aleatoric uncertainty decomposition with error below 0.002%. All theorems are empirically validated on controlled datasets spanning up to 4,200 training examples. Our theoretical framework establishes LPF as a foundation for trustworthy multi-evidence AI in safety-critical applications.