Will AI Tell Lies to Save Sick Children? Litmus-Testing AI Values Prioritization with AIRiskDilemmas

Detecting AI risks becomes more challenging as stronger models emerge and find novel methods such as Alignment Faking to circumvent these detection attempts. Inspired by how risky behaviors in humans (i.e., illegal activities that may hurt others) are sometimes guided by strongly-held values, we believe that identifying values within AI models can be an early warning system for AI's risky behaviors. We create LitmusValues, an evaluation pipeline to reveal AI models' priorities on a range of AI value classes. Then, we collect AIRiskDilemmas, a diverse collection of dilemmas that pit values against one another in scenarios relevant to AI safety risks such as Power Seeking. By measuring an AI model's value prioritization using its aggregate choices, we obtain a self-consistent set of predicted value priorities that uncover potential risks. We show that values in LitmusValues (including seemingly innocuous ones like Care) can predict for both seen risky behaviors in AIRiskDilemmas and unseen risky behaviors in HarmBench.

HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages

Preference datasets are essential for training general-domain, instruction-following language models with Reinforcement Learning from Human Feedback (RLHF). Each subsequent data release raises expectations for future data collection, meaning there is a constant need to advance the quality and diversity of openly available preference data. To address this need, we introduce HelpSteer3-Preference, a permissively licensed (CC-BY-4.0), high-quality, human-annotated preference dataset comprising of over 40,000 samples. These samples span diverse real-world applications of large language models (LLMs), including tasks relating to STEM, coding and multilingual scenarios. Using HelpSteer3-Preference, we train Reward Models (RMs) that achieve top performance on RM-Bench (82.4%) and JudgeBench (73.7%). This represents a substantial improvement (~10% absolute) over the previously best-reported results from existing RMs. We demonstrate HelpSteer3-Preference can also be applied to train Generative RMs and how policy models can be aligned with RLHF using our RMs. Dataset (CC-BY-4.0): https://huggingface.co/datasets/nvidia/HelpSteer3#preference

Llama-Nemotron: Efficient Reasoning Models

We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.

SEE: Continual Fine-tuning with Sequential Ensemble of Experts

Continual fine-tuning of large language models (LLMs) suffers from catastrophic forgetting. Rehearsal-based methods mitigate this problem by retaining a small set of old data. Nevertheless, they still suffer inevitable performance loss. Although training separate experts for each task can help prevent forgetting, effectively assembling them remains a challenge. Some approaches use routers to assign tasks to experts, but in continual learning, they often require retraining for optimal performance. To address these challenges, we introduce the Sequential Ensemble of Experts (SEE) framework. SEE removes the need for an additional router, allowing each expert to independently decide whether a query should be handled. The framework employs distributed routing, and during continual fine-tuning, SEE only requires the training of new experts for incoming tasks rather than retraining the entire system. Experiments reveal that the SEE outperforms prior approaches, including multi-task learning, in continual fine-tuning. It also demonstrates remarkable generalization ability, as the expert can effectively identify out-of-distribution queries, which can then be directed to a more generalized model for resolution. This work highlights the promising potential of integrating routing and response mechanisms within each expert, paving the way for the future of distributed model ensembling.

Adversarial Training of Reward Models

Reward modeling has emerged as a promising approach for the scalable alignment of language models. However, contemporary reward models (RMs) often lack robustness, awarding high rewards to low-quality, out-of-distribution (OOD) samples. This can lead to reward hacking, where policies exploit unintended shortcuts to maximize rewards, undermining alignment. To address this challenge, we introduce Adv-RM, a novel adversarial training framework that automatically identifies adversarial examples -- responses that receive high rewards from the target RM but are OOD and of low quality. By leveraging reinforcement learning, Adv-RM trains a policy to generate adversarial examples that reliably expose vulnerabilities in large state-of-the-art reward models such as Nemotron 340B RM. Incorporating these adversarial examples into the reward training process improves the robustness of RMs, mitigating reward hacking and enhancing downstream performance in RLHF. We demonstrate that Adv-RM significantly outperforms conventional RM training, increasing stability and enabling more effective RLHF training in both synthetic and real-data settings.

Lost in Literalism: How Supervised Training Shapes Translationese in LLMs

Large language models (LLMs) have achieved remarkable success in machine translation, demonstrating impressive performance across diverse languages. However, translationese, characterized by overly literal and unnatural translations, remains a persistent challenge in LLM-based translation systems. Despite their pre-training on vast corpora of natural utterances, LLMs exhibit translationese errors and generate unexpected unnatural translations, stemming from biases introduced during supervised fine-tuning (SFT). In this work, we systematically evaluate the prevalence of translationese in LLM-generated translations and investigate its roots during supervised training. We introduce methods to mitigate these biases, including polishing golden references and filtering unnatural training instances. Empirical evaluations demonstrate that these approaches significantly reduce translationese while improving translation naturalness, validated by human evaluations and automatic metrics. Our findings highlight the need for training-aware adjustments to optimize LLM translation outputs, paving the way for more fluent and target-language-consistent translations. We release the data and code at https://github.com/yafuly/LLM_Translationese.

Dedicated Feedback and Edit Models Empower Inference-Time Scaling for Open-Ended General-Domain Tasks

Inference-Time Scaling has been critical to the success of recent models such as OpenAI o1 and DeepSeek R1. However, many techniques used to train models for inference-time scaling require tasks to have answers that can be verified, limiting their application to domains such as math, coding and logical reasoning. We take inspiration from how humans make first attempts, ask for detailed feedback from others and make improvements based on such feedback across a wide spectrum of open-ended endeavors. To this end, we collect data for and train dedicated Feedback and Edit Models that are capable of performing inference-time scaling for open-ended general-domain tasks. In our setup, one model generates an initial response, which are given feedback by a second model, that are then used by a third model to edit the response. We show that performance on Arena Hard, a benchmark strongly predictive of Chatbot Arena Elo can be boosted by scaling the number of initial response drafts, effective feedback and edited responses. When scaled optimally, our setup based on 70B models from the Llama 3 family can reach SoTA performance on Arena Hard at 92.7 as of 5 Mar 2025, surpassing OpenAI o1-preview-2024-09-12 with 90.4 and DeepSeek R1 with 92.3.

Unveiling Attractor Cycles in Large Language Models: A Dynamical Systems View of Successive Paraphrasing

Dynamical systems theory provides a framework for analyzing iterative processes and evolution over time. Within such systems, repetitive transformations can lead to stable configurations, known as attractors, including fixed points and limit cycles. Applying this perspective to large language models (LLMs), which iteratively map input text to output text, provides a principled approach to characterizing long-term behaviors. Successive paraphrasing serves as a compelling testbed for exploring such dynamics, as paraphrases re-express the same underlying meaning with linguistic variation. Although LLMs are expected to explore a diverse set of paraphrases in the text space, our study reveals that successive paraphrasing converges to stable periodic states, such as 2-period attractor cycles, limiting linguistic diversity. This phenomenon is attributed to the self-reinforcing nature of LLMs, as they iteratively favour and amplify certain textual forms over others. This pattern persists with increasing generation randomness or alternating prompts and LLMs. These findings underscore inherent constraints in LLM generative capability, while offering a novel dynamical systems perspective for studying their expressive potential.

From Drafts to Answers: Unlocking LLM Potential via Aggregation Fine-Tuning

Scaling data and model size has been proven effective for boosting the performance of large language models. In addition to training-time scaling, recent studies have revealed that increasing test-time computational resources can further improve performance. In this work, we introduce Aggregation Fine-Tuning (AFT), a supervised finetuning paradigm where the model learns to synthesize multiple draft responses, referred to as proposals, into a single, refined answer, termed aggregation. At inference time, a propose-and-aggregate strategy further boosts performance by iteratively generating proposals and aggregating them. Empirical evaluations on benchmark datasets show that AFT-trained models substantially outperform standard SFT. Notably, an AFT model, fine-tuned from Llama3.1-8B-Base with only 64k data, achieves a 41.3% LC win rate on AlpacaEval 2, surpassing significantly larger LLMs such as Llama3.1-405B-Instruct and GPT4. By combining sequential refinement and parallel sampling, the propose-and-aggregate framework scales inference-time computation in a flexible manner. Overall, These findings position AFT as a promising approach to unlocking additional capabilities of LLMs without resorting to increasing data volume or model size.

Diverging Preferences: When do Annotators Disagree and do Models Know?

We examine diverging preferences in human-labeled preference datasets. We develop a taxonomy of disagreement sources spanning 10 categories across four high-level classes -- task underspecification, response style, refusals, and annotation errors. We find that the majority of disagreements are in opposition with standard reward modeling approaches, which are designed with the assumption that annotator disagreement is noise. We then explore how these findings impact two areas of LLM development: reward modeling and evaluation. In our experiments, we demonstrate how standard reward modeling methods, like the Bradley-Terry model, fail to differentiate whether a given preference judgment is the result of unanimous agreement among annotators or the majority opinion among diverging user preferences. We also find that these tendencies are also echoed by popular LLM-as-Judge evaluation methods, which consistently identify a winning response in cases of diverging preferences. These findings highlight remaining challenges in LLM evaluations, which are greatly influenced by divisive features like response style, and in developing pluralistically aligned LLMs. To address these issues, we develop methods for identifying diverging preferences to mitigate their influence on evaluation and training.