As AI systems become more capable, deceptive behaviors can undermine evaluation and mislead users at deployment. Recent work has shown that lie detectors can accurately classify deceptive behavior, but they are not typically used in the training pipeline due to concerns around contamination and objective hacking. We examine these concerns by incorporating a lie detector into the labelling step of LLM post-training and evaluating whether the learned policy is genuinely more honest, or instead learns to fool the lie detector while remaining deceptive. Using DolusChat, a novel 65k-example dataset with paired truthful/deceptive responses, we identify three key factors that determine the honesty of learned policies: amount of exploration during preference learning, lie detector accuracy, and KL regularization strength. We find that preference learning with lie detectors and GRPO can lead to policies which evade lie detectors, with deception rates of over 85\%. However, if the lie detector true positive rate (TPR) or KL regularization is sufficiently high, GRPO learns honest policies. In contrast, off-policy algorithms (DPO) consistently lead to deception rates under 25\% for realistic TPRs. Our results illustrate a more complex picture than previously assumed: depending on the context, lie-detector-enhanced training can be a powerful tool for scalable oversight, or a counterproductive method encouraging undetectable misalignment.

Modern preference alignment techniques, such as Best-of-N (BoN) sampling, rely on reward models trained with pairwise comparison data. While effective at learning relative preferences, this paradigm fails to capture a signal of response acceptability, leaving systems vulnerable to selecting the least bad of many unacceptable options. This is particularly problematic for hard prompts, where the risk of such false acceptances increases with the number of samples. In this paper, we address this critical reliability gap by introducing a new data collection and modeling framework. By augmenting preference data with an outside option, inspired by discrete choice models, we train a reward model that can distinguish not just what is better, but what is good enough. We leverage this capability to create an adaptive inference strategy, best of mini-N in-loop, which partitions the generation budget into sequential loops with a calibrated, early-exit condition. Our experiments show that when tuned as an alignment guardrail, it reduces reliability failures by 70%, and when tuned as an inference accelerator, it improves average inference speed by over 22% in IMDB-sentiment setting. We thus provide a principled and flexible framework for practitioners to explicitly manage the trade-off between reliability and computational efficiency.

Motivated by the behavior of Bayesian inference in misspecified models Grün-wald et al. ( 2017); Jansen ( 2013) extensively studied the so called "generalized" Bayesian inference, However, these works consider only "warm posteriors" Grünwald et al. ( 2017) the prior favours simple models, hence it is beneficial to put more weight onto the prior and use warm posterior. Finally, we mention the work of Bhattacharya et al. ( 2019), in which the authors develop fractional posteriors with the goal of CIFAR-10, have been collected and curated. The Street View House Numbers dataset ( Netzer et al., 2011), which is divided into a training In CIFAR-10 ( Krizhevsky and Hinton, 2009), labellers followed strict guidelines to ensure high quality labelling of the images. In particular, labellers were instructed that "it's worse to include one that shouldn't be included than to exclude one. In this section we review the basics of (SG)-MCMC inference.

Automating chest radiograph interpretation using Deep Learning (DL) models has the potential to significantly improve clinical workflows, decision-making, and large-scale health screening. However, in medical settings, merely optimising predictive performance is insufficient, as the quantification of uncertainty is equally crucial. This paper investigates the relationship between predictive uncertainty, derived from Bayesian Deep Learning approximations, and human/linguistic uncertainty, as estimated from free-text radiology reports labelled by rule-based labellers. Utilising BERT as the model of choice, this study evaluates different binarisation methods for uncertainty labels and explores the efficacy of Monte Carlo Dropout and Deep Ensembles in estimating predictive uncertainty. The results demonstrate good model performance, but also a modest correlation between predictive and linguistic uncertainty, highlighting the challenges in aligning machine uncertainty with human interpretation nuances. Our findings suggest that while Bayesian approximations provide valuable uncertainty estimates, further refinement is necessary to fully capture and utilise the subtleties of human uncertainty in clinical applications.

Federated Learning (FL) is a collaborative machine learning paradigm for enhancing data privacy preservation. Its privacy-preserving nature complicates the explanation of the decision-making processes and the evaluation of the reliability of the generated explanations. In this paper, we propose the Uncertainty-aware eXplainable Federated Learning (UncertainXFL) to address these challenges. It generates explanations for decision-making processes under FL settings and provides information regarding the uncertainty of these explanations. UncertainXFL is the first framework to explicitly offer uncertainty evaluation for explanations within the FL context. Explanatory information is initially generated by the FL clients and then aggregated by the server in a comprehensive and conflict-free manner during FL training. The quality of the explanations, including the uncertainty score and tested validity, guides the FL training process by prioritizing clients with the most reliable explanations through higher weights during model aggregation. Extensive experimental evaluation results demonstrate that UncertainXFL achieves superior model accuracy and explanation accuracy, surpassing the current state-of-the-art model that does not incorporate uncertainty information by 2.71% and 1.77%, respectively. By integrating and quantifying uncertainty in the data into the explanation process, UncertainXFL not only clearly presents the explanation alongside its uncertainty, but also leverages this uncertainty to guide the FL training process, thereby enhancing the robustness and reliability of the resulting models.

Relevance labels, which indicate whether a search result is valuable to a searcher, are key to evaluating and optimising search systems. The best way to capture the true preferences of users is to ask them for their careful feedback on which results would be useful, but this approach does not scale to produce a large number of labels. Getting relevance labels at scale is usually done with third-party labellers, who judge on behalf of the user, but there is a risk of low-quality data if the labeller doesn't understand user needs. To improve quality, one standard approach is to study real users through interviews, user studies and direct feedback, find areas where labels are systematically disagreeing with users, then educate labellers about user needs through judging guidelines, training and monitoring. This paper introduces an alternate approach for improving label quality. It takes careful feedback from real users, which by definition is the highest-quality first-party gold data that can be derived, and develops an large language model prompt that agrees with that data. We present ideas and observations from deploying language models for large-scale relevance labelling at Bing, and illustrate with data from TREC. We have found large language models can be effective, with accuracy as good as human labellers and similar capability to pick the hardest queries, best runs, and best groups. Systematic changes to the prompts make a difference in accuracy, but so too do simple paraphrases. To measure agreement with real searchers needs high-quality ``gold'' labels, but with these we find that models produce better labels than third-party workers, for a fraction of the cost, and these labels let us train notably better rankers.

Small farms contribute to a large share of the productive land in developing countries. In regions such as sub-Saharan Africa, where 80% of farms are small (under 2 ha in size), the task of mapping smallholder cropland is an important part of tracking sustainability measures such as crop productivity. However, the visually diverse and nuanced appearance of small farms has limited the effectiveness of traditional approaches to cropland mapping. Here we introduce a new approach based on the detection of harvest piles characteristic of many smallholder systems throughout the world. We present HarvestNet, a dataset for mapping the presence of farms in the Ethiopian regions of Tigray and Amhara during 2020-2023, collected using expert knowledge and satellite images, totaling 7k hand-labeled images and 2k ground collected labels. We also benchmark a set of baselines including SOTA models in remote sensing with our best models having around 80% classification performance on hand labelled data and 90%, 98% accuracy on ground truth data for Tigray, Amhara respectively. We also perform a visual comparison with a widely used pre-existing coverage map and show that our model detects an extra 56,621 hectares of cropland in Tigray. We conclude that remote sensing of harvest piles can contribute to more timely and accurate cropland assessments in food insecure region.

Correctly labelling training data for AI models is vital to avoid serious problems, as is using sufficiently large datasets. However, manually labelling massive amounts of data is time-consuming and laborious. Using pre-labelled datasets can be problematic, as evidenced by MIT having to pull its 80 Million Tiny Images datasets. For those unaware, the popular dataset was found to contain thousands of racist and misogynistic labels that could have been used to train AI models. AI News caught up with Devang Sachdev, VP of Marketing at Snorkel AI, to find out how the company is easing the laborious process of labelling data in a safe and effective way. AI News: How is Snorkel helping to ease the laborious process of labelling data?

Many experimental works have observed that generalization in deep RL appears to be difficult: although RL agents can learn to perform very complex tasks, they don't seem to generalize over diverse task distributions as well as the excellent generalization of supervised deep nets might lead us to expect. In this blog post, we will aim to explain why generalization in RL is fundamentally harder, and indeed more difficult even in theory. We will show that attempting to generalize in RL induces implicit partial observability, even when the RL problem we are trying to solve is a standard fully-observed MDP. This induced partial observability can significantly complicate the types of policies needed to generalize well, potentially requiring counterintuitive strategies like information-gathering actions, recurrent non-Markovian behavior, or randomized strategies. Ordinarily, this is not necessary in fully observed MDPs but surprisingly becomes necessary when we consider generalization from a finite training set in a fully observed MDP.