Reinforcement Learning with Verifiable Rewards (RLVR) has been an effective approach for improving Large Language Models' reasoning in domains such as coding and mathematics. Here, we apply RLVR methods towards forecasting future real-world events - a challenging task for RL due to the very noisy (and delayed) outcomes involved. Using a novel dataset of recent questions from a prediction market, and accompanying relevant news headlines, we show that a compact (14B) reasoning model can be trained to match or surpass the predictive accuracy of frontier models like o1, while greatly improving probabilistic calibration. The model's performance is also practically meaningful: in a Polymarket trading simulation, we estimate that its bets would have yielded a return on investment of over 10% across all questions in the test set. We detail and compare approaches used in training our model, including augmenting our training-data with synthetic prediction questions, guardrails for learning stability, and median prediction sampling at inference-time.

Investigating Fine-and Coarse-grained Structural Correspondences Between Deep Neural Networks and Human Object Image Similarity Judgments Using Unsupervised Alignment Soh Takahashi, Masaru Sasaki, Ken Takeda, Masafumi Oizumi Introduces an unsupervised alignment to assess human-like object representations of DNNs. CLIP models show highest fine-grained matching (20% top-1 match). This underscores the role of linguistic cues in refining representations. Image-only self-supervised models lack fine matching with human representations. Rather, they capture coarse category structures, hinting at prelinguistic links. Abstract The learning mechanisms by which humans acquire internal representations of objects are not fully understood. Deep neural networks (DNNs) have emerged as a useful tool for investigating this question, as they have internal representations similar to those of humans as a byproduct of optimizing their objective functions. While previous studies have shown that models trained with various learning paradigms--such as supervised, self-supervised, and CLIP--acquire human-like representations, it remains unclear whether their similarity to human representations is primarily at a coarse category level or extends to finer details. Here, we employ an unsupervised alignment method based on Gromov-Wasserstein Optimal Transport to compare human and model object representations at both fine-grained and coarse-grained levels. The unique feature of this method compared to conventional representational similarity analysis is that it estimates optimal fine-grained mappings between the representation of each object in human and model representations. We used this unsupervised alignment method to assess the extent to which the representation of each object in humans is correctly mapped to the corresponding representation of the same object in models. Using human similarity judgments of 1,854 objects from the THINGS dataset, we find that models trained with CLIP consistently achieve strong fine-and coarse-grained matching with human object representations.

While LLMs appear robustly safety-aligned in English, we uncover a catastrophic, overlooked weakness: attributional collapse under code-mixed perturbations. Our systematic evaluation of open models shows that the linguistic camouflage of code-mixing -- ``blending languages within a single conversation'' -- can cause safety guardrails to fail dramatically. Attack success rates (ASR) spike from a benign 9\% in monolingual English to 69\% under code-mixed inputs, with rates exceeding 90\% in non-Western contexts such as Arabic and Hindi. These effects hold not only on controlled synthetic datasets but also on real-world social media traces, revealing a serious risk for billions of users. To explain why this happens, we introduce saliency drift attribution (SDA), an interpretability framework that shows how, under code-mixing, the model's internal attention drifts away from safety-critical tokens (e.g., ``violence'' or ``corruption''), effectively blinding it to harmful intent. Finally, we propose a lightweight translation-based restoration strategy that recovers roughly 80\% of the safety lost to code-mixing, offering a practical path toward more equitable and robust LLM safety.

Plaintiffs and defendants in copyright lawsuits over generative AI often make sweeping, opposing claims about the extent to which large language models (LLMs) have memorized plaintiffs' protected expression in their training data. Drawing on both machine learning and copyright law, we show that these polarized positions dramatically oversimplify the relationship between memorization and copyright. To do so, we extend a recent probabilistic extraction technique to measure memorization of 50 books in 17 open-weight LLMs. Through thousands of experiments, we show that the extent of memorization varies both by model and by book. With respect to our specific extraction methodology, we find that most LLMs do not memorize most books -- either in whole or in part. However, we also find that Llama 3.1 70B entirely memorizes some books, like the first Harry Potter book and 1984. In fact, the first Harry Potter is so memorized that, using a seed prompt consisting of just the first few tokens of the first chapter, we can deterministically generate the entire book near-verbatim. We discuss why our results have significant implications for copyright cases, though not ones that unambiguously favor either side.

Demonstrations and instructions are two primary approaches for prompting language models to perform in-context learning (ICL) tasks. Do identical tasks elicited in different ways result in similar representations of the task? An improved understanding of task representation mechanisms would offer interpretability insights and may aid in steering models. We study this through \textit{function vectors} (FVs), recently proposed as a mechanism to extract few-shot ICL task representations. We generalize FVs to alternative task presentations, focusing on short textual instruction prompts, and successfully extract instruction function vectors that promote zero-shot task accuracy. We find evidence that demonstration- and instruction-based function vectors leverage different model components, and offer several controls to dissociate their contributions to task performance. Our results suggest that different task promptings forms do not induce a common task representation through FVs but elicit different, partly overlapping mechanisms. Our findings offer principled support to the practice of combining instructions and task demonstrations, imply challenges in universally monitoring task inference across presentation forms, and encourage further examinations of LLM task inference mechanisms.

We study the sample complexity of Bayesian recovery for solving inverse problems with general prior, forward operator and noise distributions. We consider posterior sampling according to an approximate prior $\mathcal{P}$, and establish sufficient conditions for stable and accurate recovery with high probability. Our main result is a non-asymptotic bound that shows that the sample complexity depends on (i) the intrinsic complexity of $\mathcal{P}$, quantified by its so-called approximate covering number, and (ii) concentration bounds for the forward operator and noise distributions. As a key application, we specialize to generative priors, where $\mathcal{P}$ is the pushforward of a latent distribution via a Deep Neural Network (DNN). We show that the sample complexity scales log-linearly with the latent dimension $k$, thus establishing the efficacy of DNN-based priors. Generalizing existing results on deterministic (i.e., non-Bayesian) recovery for the important problem of random sampling with an orthogonal matrix $U$, we show how the sample complexity is determined by the coherence of $U$ with respect to the support of $\mathcal{P}$. Hence, we establish that coherence plays a fundamental role in Bayesian recovery as well. Overall, our framework unifies and extends prior work, providing rigorous guarantees for the sample complexity of solving Bayesian inverse problems with arbitrary distributions.

Addressing climate change requires coordinated policy efforts of nations worldwide. These efforts are informed by scientific reports, which rely in part on Integrated Assessment Models (IAMs), prominent tools used to assess the economic impacts of climate policies. However, traditional IAMs optimize policies based on a single objective, limiting their ability to capture the trade-offs among economic growth, temperature goals, and climate justice. As a result, policy recommendations have been criticized for perpetuating inequalities, fueling disagreements during policy negotiations. We introduce Justice, the first framework integrating IAM with Multi-Objective Multi-Agent Reinforcement Learning (MOMARL). By incorporating multiple objectives, Justice generates policy recommendations that shed light on equity while balancing climate and economic goals. Further, using multiple agents can provide a realistic representation of the interactions among the diverse policy actors. We identify equitable Pareto-optimal policies using our framework, which facilitates deliberative decision-making by presenting policymakers with the inherent trade-offs in climate and economic policy.

Wildfires pose a threat to ecosystems, economies and public safety, particularly in Mediterranean regions such as Spain. Accurate predictive models require high-resolution spatio-temporal data to capture complex dynamics of environmental and human factors. To address the scarcity of fine-grained wildfire datasets in Spain, we introduce IberFire: a spatio-temporal dataset with 1 km x 1 km x 1-day resolution, covering mainland Spain and the Balearic Islands from December 2007 to December 2024. IberFire integrates 120 features across eight categories: auxiliary data, fire history, geography, topography, meteorology, vegetation indices, human activity and land cover. All features and processing rely on open-access data and tools, with a publicly available codebase ensuring transparency and applicability. IberFire offers enhanced spatial granularity and feature diversity compared to existing European datasets, and provides a reproducible framework. It supports advanced wildfire risk modelling via Machine Learning and Deep Learning, facilitates climate trend analysis, and informs fire prevention and land management strategies. The dataset is freely available on Zenodo to promote open research and collaboration.

Deep learning has shown strong performance in musculoskeletal imaging, but prior work has largely targeted conditions where diagnosis is relatively straightforward. More challenging problems remain underexplored, such as detecting Bankart lesions (anterior-inferior glenoid labral tears) on standard MRIs. These lesions are difficult to diagnose due to subtle imaging features, often necessitating invasive MRI arthrograms (MRAs). We introduce ScopeMRI, the first publicly available, expert-annotated dataset for shoulder pathologies, and present a deep learning framework for Bankart lesion detection on both standard MRIs and MRAs. ScopeMRI contains shoulder MRIs from patients who underwent arthroscopy, providing ground-truth labels from intraoperative findings, the diagnostic gold standard. Separate models were trained for MRIs and MRAs using CNN- and transformer-based architectures, with predictions ensembled across multiple imaging planes. Our models achieved radiologist-level performance, with accuracy on standard MRIs surpassing radiologists interpreting MRAs. External validation on independent hospital data demonstrated initial generalizability across imaging protocols. By releasing ScopeMRI and a modular codebase for training and evaluation, we aim to accelerate research in musculoskeletal imaging and foster development of datasets and models that address clinically challenging diagnostic tasks.

Spurious correlations that lead models to correct predictions for the wrong reasons pose a critical challenge for robust real-world generalization. Existing research attributes this issue to group imbalance and addresses it by maximizing group-balanced or worst-group accuracy, which heavily relies on expensive bias annotations. A compromise approach involves predicting bias information using extensively pretrained foundation models, which requires large-scale data and becomes impractical for resource-limited rare domains. T o address these challenges, we offer a novel perspective by reframing the spurious correlations as imbalances or mismatches in class-conditional distributions, and propose a simple yet effective robust learning method that eliminates the need for both bias annotations and predictions. With the goal of maximizing the conditional entropy (uncertainty) of the label given spurious factors, our method leverages a sample reweighting strategy to achieve class-conditional distribution balancing, which automatically highlights minority groups and classes, effectively dismantling spurious correlations and producing a debiased data distribution for classification. Extensive experiments and analysis demonstrate that our approach consistently delivers state-of-the-art performance, rivaling methods that rely on bias supervision.