When language models (LMs) are trained via reinforcement learning (RL) to generate natural language "reasoning chains", their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-confidence outputs, they often have the unintended side-effect of degrading calibration and increasing the rate at which LMs generate incorrect responses (or "hallucinate") in other problem domains. This paper describes RLCR (Reinforcement Learning with Calibration Rewards), an approach to training reasoning models that jointly improves accuracy and calibrated confidence estimation. During RLCR, LMs generate both predictions and numerical confidence estimates after reasoning. They are trained to optimize a reward function that augments a binary correctness score with a Brier score -- a scoring rule for confidence estimates that incentivizes calibrated prediction. We first prove that this reward function (or any analogous reward function that uses a bounded, proper scoring rule) yields models whose predictions are both accurate and well-calibrated. We next show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy, on both in-domain and out-of-domain evaluations -- outperforming both ordinary RL training and classifiers trained to assign post-hoc confidence scores. While ordinary RL hurts calibration, RLCR improves it. Finally, we demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration via confidence-weighted scaling methods. Our results show that explicitly optimizing for calibration can produce more generally reliable reasoning models.

In-Context Learning has shown great potential for aligning Large Language Models (LLMs) with human values, helping reduce harmful outputs and accommodate diverse preferences without costly post-training, known as In-Context Alignment (ICA). However, LLMs' comprehension of input prompts remains agnostic, limiting ICA's ability to address value tensions--human values are inherently pluralistic, often imposing conflicting demands, e.g., stimulation vs. tradition. Current ICA methods therefore face the Instruction Bottleneck challenge, where LLMs struggle to reconcile multiple intended values within a single prompt, leading to incomplete or biased alignment. To address this, we propose PICACO, a novel pluralistic ICA method. Without fine-tuning, PICACO optimizes a meta-instruction that navigates multiple values to better elicit LLMs' understanding of them and improve their alignment. This is achieved by maximizing the total correlation between specified values and LLM responses, theoretically reinforcing value correlation while reducing distractive noise, resulting in effective value instructions. Extensive experiments on five value sets show that PICACO works well with both black-box and open-source LLMs, outperforms several recent strong baselines, and achieves a better balance across up to 8 distinct values.

With the rapid advancement of generative AI, synthetic content across images, videos, and audio has become increasingly realistic, amplifying the risk of misinformation. Existing detection approaches predominantly focus on binary classification while lacking detailed and interpretable explanations of forgeries, which limits their applicability in safety-critical scenarios. Moreover, current methods often treat each modality separately, without a unified benchmark for cross-modal forgery detection and interpretation. T o address these challenges, we introduce METER, a unified, multi-modal benchmark for interpretable forgery detection spanning images, videos, audio, and audio-visual content. Our dataset comprises four tracks, each requiring not only real-vs-fake classification but also evidence-chain-based explanations, including spatio-temporal localization, textual rationales, and forgery type tracing. Compared to prior benchmarks, METER offers broader modality coverage and richer interpretability metrics such as spatial/temporal IoU, multi-class tracing, and evidence consistency. W e further propose a human-aligned, three-stage Chain-of-Thought (CoT) training strategy combining SFT, DPO, and a novel GRPO stage that integrates a human-aligned evaluator with CoT reasoning. W e hope METER will serve as a standardized foundation for advancing gen-eralizable and interpretable forgery detection in the era of generative media.

Understanding what constitutes safety in AI-generated content is complex. While developers often rely on predefined taxonomies, real-world safety judgments also involve personal, social, and cultural perceptions of harm. This paper examines how annotators evaluate the safety of AI-generated images, focusing on the qualitative reasoning behind their judgments. Analyzing 5,372 open-ended comments, we find that annotators consistently invoke moral, emotional, and contextual reasoning that extends beyond structured safety categories. Many reflect on potential harm to others more than to themselves, grounding their judgments in lived experience, collective risk, and sociocultural awareness. Beyond individual perceptions, we also find that the structure of the task itself -- including annotation guidelines -- shapes how annotators interpret and express harm. Guidelines influence not only which images are flagged, but also the moral judgment behind the justifications. Annotators frequently cite factors such as image quality, visual distortion, and mismatches between prompt and output as contributing to perceived harm dimensions, which are often overlooked in standard evaluation frameworks. Our findings reveal that existing safety pipelines miss critical forms of reasoning that annotators bring to the task. We argue for evaluation designs that scaffold moral reflection, differentiate types of harm, and make space for subjective, context-sensitive interpretations of AI-generated content.

For a while, dating apps seemed to make it easier, putting a city's worth of single people in the palm of your hand. But AI has cast a paranoid pall over what can already be a suboptimal experience. If you get a message that feels a little off, it is hard to know whether you are flirting with a bot--or just someone insecure enough to use ChatGPT as their own Cyrano de Bergerac. In frustration, my friend Lonni has started picking up women at the nail salon like it's 1997. Or, in the midst of an emotionally fraught conversation with a friend or family member, a text might read strangely.

Natural language interfaces offer a compelling approach for music recommendation, enabling users to express complex preferences conversationally. While Large Language Models (LLMs) show promise in this direction, their scalability in recommender systems is limited by high costs and latency. Retrieval-based approaches using smaller language models mitigate these issues but often rely on single-modal item representations, overlook long-term user preferences, and require full model retraining, posing challenges for real-world deployment. In this paper, we present JAM (Just Ask for Music), a lightweight and intuitive framework for natural language music recommendation. JAM models user-query-item interactions as vector translations in a shared latent space, inspired by knowledge graph embedding methods like TransE. To capture the complexity of music and user intent, JAM aggregates multimodal item features via cross-attention and sparse mixture-of-experts. We also introduce JAMSessions, a new dataset of over 100k user-query-item triples with anonymized user/item embeddings, uniquely combining conversational queries and user long-term preferences. Our results show that JAM provides accurate recommendations, produces intuitive representations suitable for practical use cases, and can be easily integrated with existing music recommendation stacks.

We present a novel evaluation paradigm for AI text detectors that prioritizes real-world and equitable assessment. Current approaches predominantly report conventional metrics like AUROC, overlooking that even modest false positive rates constitute a critical impediment to practical deployment of detection systems. Furthermore, real-world deployment necessitates predetermined threshold configuration, making detector stability (i.e. the maintenance of consistent performance across diverse domains and adversarial scenarios), a critical factor. These aspects have been largely ignored in previous research and benchmarks. Our benchmark, SHIELD, addresses these limitations by integrating both reliability and stability factors into a unified evaluation metric designed for practical assessment. Furthermore, we develop a post-hoc, model-agnostic humanification framework that modifies AI text to more closely resemble human authorship, incorporating a controllable hardness parameter. This hardness-aware approach effectively challenges current SOTA zero-shot detection methods in maintaining both reliability and stability. (Data and code: https://github.com/navid-aub/SHIELD-Benchmark)

Adversarial prompt attacks can significantly alter the reliability of Retrieval-Augmented Generation (RAG) systems by re-ranking them to produce incorrect outputs. In this paper, we present a novel method that applies Differential Evolution (DE) to optimize adversarial prompt suffixes for RAG-based question answering. Our approach is gradient-free, treating the RAG pipeline as a black box and evolving a population of candidate suffixes to maximize the retrieval rank of a targeted incorrect document to be closer to real world scenarios. We conducted experiments on the BEIR QA datasets to evaluate attack success at certain retrieval rank thresholds under multiple retrieving applications. Our results demonstrate that DE-based prompt optimization attains competitive (and in some cases higher) success rates compared to GGPP to dense retrievers and PRADA to sparse retrievers, while using only a small number of tokens (<=5 tokens) in the adversarial suffix. Furthermore, we introduce a readability-aware suffix construction strategy, validated by a statistically significant reduction in MLM negative log-likelihood with Welch's t-test. Through evaluations with a BERT-based adversarial suffix detector, we show that DE-generated suffixes evade detection, yielding near-chance detection accuracy.

As large language models (LLMs) become increasingly integrated into real-world applications, scalable and rigorous safety evaluation is essential. This paper introduces Aymara AI, a programmatic platform for generating and administering customized, policy-grounded safety evaluations. Aymara AI transforms natural-language safety policies into adversarial prompts and scores model responses using an AI-based rater validated against human judgments. We demonstrate its capabilities through the Aymara LLM Risk and Responsibility Matrix, which evaluates 20 commercially available LLMs across 10 real-world safety domains. Results reveal wide performance disparities, with mean safety scores ranging from 86.2% to 52.4%. While models performed well in well-established safety domains such as Misinformation (mean = 95.7%), they consistently failed in more complex or underspecified domains, notably Privacy & Impersonation (mean = 24.3%). Analyses of Variance confirmed that safety scores differed significantly across both models and domains (p < .05). These findings underscore the inconsistent and context-dependent nature of LLM safety and highlight the need for scalable, customizable tools like Aymara AI to support responsible AI development and oversight.

While Artificial Intelligence (AI) is not a new field, recent developments, especially with the release of generative tools like ChatGPT, have brought it to the forefront of the minds of industry workers and academic folk alike. There is currently much talk about AI and its ability to reshape many everyday processes as we know them through automation. It also allows users to expand their ideas by suggesting things they may not have thought of on their own and provides easier access to information. However, not all of the changes this technology will bring or has brought so far are positive; this is why it is extremely important for all modern people to recognize and understand the risks before using these tools and allowing them to cause harm. This work takes a position on better understanding many equity concerns and the spread of misinformation that result from new AI, in this case, specifically ChatGPT and deepfakes, and encouraging collaboration with law enforcement, developers, and users to reduce harm. Considering many academic sources, it warns against these issues, analyzing their cause and impact in fields including healthcare, education, science, academia, retail, and finance. Lastly, we propose a set of future-facing guidelines and policy considerations to solve these issues while still enabling innovation in these fields, this responsibility falling upon users, developers, and government entities.