The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can reason about our safety policies in context when responding to potentially unsafe prompts, through deliberative alignment. This leads to state-of-the-art performance on certain benchmarks for risks such as generating illicit advice, choosing stereotyped responses, and succumbing to known jailbreaks. Training models to incorporate a chain of thought before answering has the potential to unlock substantial benefits, while also increasing potential risks that stem from heightened intelligence. Our results underscore the need for building robust alignment methods, extensively stress-testing their efficacy, and maintaining meticulous risk management protocols. This report outlines the safety work carried out for the OpenAI o1 and OpenAI o1-mini models, including safety evaluations, external red teaming, and Preparedness Framework evaluations.

OpenAI's recent introduction of Reinforcement Fine-Tuning (RFT) showcases the potential of reasoning foundation model and offers a new paradigm for fine-tuning beyond simple pattern imitation. This technical report presents OpenRFT, our attempt to fine-tune generalist reasoning models for domain-specific tasks under the same settings as RFT. The evaluation is conducted on Sci-KnowEval, where OpenRFT achieves notable performance gains with only 100 domain-specific samples for each task. More experimental results will be updated continuously in later versions. OpenAI's o1 model has shown strong reasoning abilities in mathematics and programming, but its generalization to other tasks remains uncertain. The recent introduction of Reinforcement Fine-Tuning (RFT) (OpenAI, 2024) has provided a promising avenue for reasoning generalization. With only dozens of high-quality (question, answer) pairs, RFT enables the creation of customized reasoning models excelling at domain-specific tasks. The significance of RFT is at least two-fold: (1) It demonstrates the promise of using generalist reasoning models, like o1, as reasoning foundation models. By enabling the efficient creation of domain-specific reasoning models, RFT practically expands the applicability of reasoning models across diverse tasks. Unlike Supervised Fine-Tuning (SFT), which merely mimics patterns in training data, RFT leverages reasoning capabilities to facilitate thinking and trial-and-error learning.

Offline Reinforcement Learning (RL), which operates solely on static datasets without further interactions with the environment, provides an appealing alternative to learning a safe and promising control policy. The prevailing methods typically learn a conservative policy to mitigate the problem of Q-value overestimation, but it is prone to overdo it, leading to an overly conservative policy. Moreover, they optimize all samples equally with fixed constraints, lacking the nuanced ability to control conservative levels in a fine-grained manner. Consequently, this limitation results in a performance decline. To address the above two challenges in a united way, we propose a framework, Adaptive Conservative Level in Q-Learning (ACL-QL), which limits the Q-values in a mild range and enables adaptive control on the conservative level over each state-action pair, i.e., lifting the Q-values more for good transitions and less for bad transitions. We theoretically analyze the conditions under which the conservative level of the learned Q-function can be limited in a mild range and how to optimize each transition adaptively. Motivated by the theoretical analysis, we propose a novel algorithm, ACL-QL, which uses two learnable adaptive weight functions to control the conservative level over each transition. Subsequently, we design a monotonicity loss and surrogate losses to train the adaptive weight functions, Q-function, and policy network alternatively. We evaluate ACL-QL on the commonly used D4RL benchmark and conduct extensive ablation studies to illustrate the effectiveness and state-of-the-art performance compared to existing offline DRL baselines.

Reinforcement learning (RL) plays a major role in solving complex sequential decision-making tasks. Hierarchical and goal-conditioned RL are promising methods for dealing with two major problems in RL, namely sample inefficiency and difficulties in reward shaping. These methods tackle the mentioned problems by decomposing a task into simpler subtasks and temporally abstracting a task in the action space. One of the key components for task decomposition of these methods is subgoal discovery. We can use the subgoal states to define hierarchies of actions and also use them in decomposing complex tasks. Under the assumption that subgoal states are more unpredictable, we propose a free energy paradigm to discover them. This is achieved by using free energy to select between two spaces, the main space and an aggregation space. The $model \; changes$ from neighboring states to a given state shows the unpredictability of a given state, and therefore it is used in this paper for subgoal discovery. Our empirical results on navigation tasks like grid-world environments show that our proposed method can be applied for subgoal discovery without prior knowledge of the task. Our proposed method is also robust to the stochasticity of environments.

With the rapid advancement of LLM models, they have become widely useful in various fields. While these AI systems can be used for code generation, significantly simplifying and accelerating the tasks of developers, their use for students to do assignments has raised ethical questions in the field of education. In this context, determining the author of a particular code becomes important. In this study, we introduce AIGCodeSet, a dataset for AI-generated code detection tasks, specifically for the Python programming language. We obtain the problem descriptions and human-written codes from the CodeNet dataset. Using the problem descriptions, we generate AI-written codes with CodeLlama 34B, Codestral 22B, and Gemini 1.5 Flash models in three approaches: i) generating code from the problem description alone, ii) generating code using the description along with human-written source code containing runtime errors, and iii) generating code using the problem description and human-written code that resulted in wrong answers. Lastly, we conducted a post-processing step to eliminate LLM output irrelevant to code snippets. Overall, AIGCodeSet consists of 2,828 AI-generated and 4,755 human-written code snippets. We share our code with the research community to support studies on this important topic and provide performance results for baseline AI-generated code detection methods.

Federated learning (FL) triggers intra-client and inter-client class imbalance, with the latter compared to the former leading to biased client updates and thus deteriorating the distributed models. Such a bias is exacerbated during the server aggregation phase and has yet to be effectively addressed by conventional re-balancing methods. To this end, different from the off-the-shelf label or loss-based approaches, we propose a gradient alignment (GA)-informed FL method, dubbed as FedGA, where the importance of error asymmetry (EA) in bias is observed and its linkage to the gradient of the loss to raw logits is explored. Concretely, GA, implemented by label calibration during the model backpropagation process, prevents catastrophic forgetting of rate and missing classes, hence boosting model convergence and accuracy. Experimental results on five benchmark datasets demonstrate that GA outperforms the pioneering counterpart FedAvg and its four variants in minimizing EA and updating bias, and accordingly yielding higher F1 score and accuracy margins when the Dirichlet distribution sampling factor $\alpha$ increases. The code and more details are available at \url{https://anonymous.4open.science/r/FedGA-B052/README.md}.

Language is the cornerstone of human communication, and the complexity of language lies in the diversity and recursion of its structure. Chomsky (1957) introduced the concept of recursion into natural language, arguing that the grammar in human natural language was a finite set of recursive rules by which an infinite number of linguistic expressions could be generated. In Corballis' (2014) words, the claim that recursion is the essence of natural language has been a continuing theme of Chomsky's work since his 1957 book Syntactic Structures. This theme is reiterated in Hauser et al. (2002), proposing that the faculty of language in the narrow sense only includes recursion, the only uniquely human component of the faculty of language. This proposal is summarized as the "recursion-only hypothesis" in Jackendoff and Pinker (2005: 212), which highlights the importance of recursion in linguistics. In spited of the lack of a consistent definition of (linguistic) recursion in the literature, most literature involves category recursion, which is defined as the "embedding of a category inside another of the same category". For instance, Martins and Fitch (2014) claim that recursion has been used to characterize the process of embedding a constituent of a certain kind of category inside another constituent of the same kind. This "embedding" process naturally generates hierarchical structures that display similar properties across different levels of embedding, and, thus, the feature of "self-similarity" is a signature of recursive structures. To illustrate that, they hold that the compound noun [[student] committee] (which has the structure [[[A]A] ]) is recursive since a noun phrase (NP) is embedded inside another NP, while a sentence with a noun plus a verb such as [[trees] grow] (which has the structure [[[A]B] ]) is non-recursive since a constituent of a given type of category is not embedded within a constituent of that same type.

With the rapid advancement of deep learning technologies, artificial intelligence has become increasingly prevalent in the research and application of dermatological disease diagnosis. However, this data-driven approach often faces issues related to decision bias. Existing fairness enhancement techniques typically come at a substantial cost to accuracy. This study aims to achieve a better trade-off between accuracy and fairness in dermatological diagnostic models. To this end, we propose a novel fair dermatological diagnosis network, named FairDD, which leverages domain incremental learning to balance the learning of different groups by being sensitive to changes in data distribution. Additionally, we incorporate the mixup data augmentation technique and supervised contrastive learning to enhance the network's robustness and generalization. Experimental validation on two dermatological datasets demonstrates that our proposed method excels in both fairness criteria and the trade-off between fairness and performance.

Traditional symbol-based AAC devices impose meta-linguistic and memory demands on individuals with complex communication needs and hinder conversation partners from stimulating symbolic language in meaningful moments. This work presents a prototype application that generates situation-specific communication boards formed by a combination of descriptive, narrative, and semantic related words and phrases inferred automatically from photographs. Through semi-structured interviews with AAC professionals, we investigate how this prototype was used to support communication and language learning in naturalistic school and therapy settings. We find that the immediacy of vocabulary reduces conversation partners' workload, opens up opportunities for AAC stimulation, and facilitates symbolic understanding and sentence construction. We contribute a nuanced understanding of how vocabularies generated automatically from photographs can support individuals with complex communication needs in using and learning symbolic AAC, offering insights into the design of automatic vocabulary generation methods and interfaces to better support various scenarios of use and goals.

Each year, a small group of PhD students are chosen to participate in the AAAI/SIGAI Doctoral Consortium. This initiative provides an opportunity for the students to discuss and explore their research interests and career objectives in an interdisciplinary workshop together with a panel of established researchers. During 2024, we met with some of the students to find out more about their research and the doctoral consortium experience. They also shared their advice for prospective PhD students. Changhoon Kim completed his PhD in Computer Engineering at Arizona State University.