Jailbreaking large language models (LLMs) is being intensively studied to evaluate the safety of LLMs [1, 2]. Existing jailbreaking methods, as adversarial attacks targeting small-scale learning models, can be categorized into white-box and black-box approaches. In white-box methods, gradient-based optimization techniques are employed to identify adversarial suffixes [3]. For black-box methods, various optimization strategies, such as genetic algorithms, are used to refine jailbreak prompt templates through rephrasing, word replacement, and other modifications [4, 5]. Recent research [6] has demonstrated that LLMs can function as powerful optimizers when provided with sufficient contextual information. Indeed, both the persuasive adversarial prompt (PAP) [7] and Tree of Attacks with Pruning (TAP) [8] methods leverage multiple LLMs to optimize malicious instructions. For the PAP method [7], two LLMs are utilized for the whole jailbreak optimization and for the TAP method [8], three LLMs are utilized.

Reward models have been increasingly critical for improving the reasoning capability of LLMs. Existing research has shown that a well-trained reward model can substantially improve model performances at inference time via search. However, the potential of reward models during RL training time still remains largely under-explored. It is currently unclear whether these reward models can provide additional training signals to enhance the reasoning capabilities of LLMs in RL training that uses sparse success rewards, which verify the correctness of solutions. In this work, we evaluate popular reward models for RL training, including the Outcome-supervised Reward Model (ORM) and the Process-supervised Reward Model (PRM), and train a collection of LLMs for math problems using RL by combining these learned rewards with success rewards. Surprisingly, even though these learned reward models have strong inference-time performances, they may NOT help or even hurt RL training, producing worse performances than LLMs trained with the success reward only. Our analysis reveals that an LLM can receive high rewards from some of these reward models by repeating correct but unnecessary reasoning steps, leading to a severe reward hacking issue. Therefore, we introduce two novel reward refinement techniques, including Clipping and Delta. The key idea is to ensure the accumulative reward of any reasoning trajectory is upper-bounded to keep a learned reward model effective without being exploited. We evaluate our techniques with multiple reward models over a set of 1.5B and 7B LLMs on MATH and GSM8K benchmarks and demonstrate that with a carefully designed reward function, RL training without any additional supervised tuning can improve all the evaluated LLMs, including the state-of-the-art 7B LLM Qwen2.5-Math-7B-Instruct on MATH and GSM8K benchmarks.

Training large language models to acquire specific skills remains a challenging endeavor. Conventional training approaches often struggle with data distribution imbalances and inadequacies in objective functions that do not align well with task-specific performance. To address these challenges, we introduce CycleQD, a novel approach that leverages the Quality Diversity framework through a cyclic adaptation of the algorithm, along with a model merging based crossover and an SVD-based mutation. In CycleQD, each task's performance metric is alternated as the quality measure while the others serve as the behavioral characteristics. This cyclic focus on individual tasks allows for concentrated effort on one task at a time, eliminating the need for data ratio tuning and simplifying the design of the objective function. Empirical results from AgentBench indicate that applying CycleQD to LLAMA3-8B-INSTRUCT based models not only enables them to surpass traditional fine-tuning methods in coding, operating systems, and database tasks, but also achieves performance on par with GPT-3.5-TURBO, which potentially contains much more parameters, across these domains. Crucially, this enhanced performance is achieved while retaining robust language capabilities, as evidenced by its performance on widely adopted language benchmark tasks. We highlight the key design choices in CycleQD, detailing how these contribute to its effectiveness. Furthermore, our method is general and can be applied to image segmentation models, highlighting its applicability across different domains.

Large language models (LLMs) have brought significant changes to many aspects of our lives. However, assessing and ensuring their chronological knowledge remains challenging. Existing approaches fall short in addressing the temporal adaptability of knowledge, often relying on a fixed time-point view. To overcome this, we introduce ChroKnowBench, a benchmark dataset designed to evaluate chronologically accumulated knowledge across three key aspects: multiple domains, time dependency, temporal state. Our benchmark distinguishes between knowledge that evolves (e.g., personal history, scientific discoveries, amended laws) and knowledge that remain constant (e.g., mathematical truths, commonsense facts). Building on this benchmark, we present ChroKnowledge (Chronological Categorization of Knowledge), a novel sampling-based framework for evaluating LLMs' non-parametric chronological knowledge. Our evaluation led to the following observations: (1) The ability of eliciting temporal knowledge varies depending on the data format that model was trained on. (2) LLMs partially recall knowledge or show a cut-off at temporal boundaries rather than recalling all aspects of knowledge correctly. Thus, we apply ourChroKnowPrompt, an in-depth prompting to elicit chronological knowledge by traversing step-by-step through the surrounding time spans. We observe that it successfully recalls objects across both open-source and proprietary LLMs, demonstrating versatility, though it faces challenges with dynamic datasets and unstructured formats.

LLMs can now act as autonomous agents that interact with digital environments and complete specific objectives (e.g., arranging an online meeting). However, accuracy is still far from satisfactory, partly due to a lack of large-scale, direct demonstrations for digital tasks. Obtaining supervised data from humans is costly, and automatic data collection through exploration or reinforcement learning relies on complex environmental and content setup, resulting in datasets that lack comprehensive coverage of various scenarios. On the other hand, there is abundant knowledge that may indirectly assist task completion, such as online tutorials that were created for human consumption. In this work, we present Synatra, an approach that effectively transforms this indirect knowledge into direct supervision at scale. We define different types of indirect knowledge, and carefully study the available sources to obtain it, methods to encode the structure of direct demonstrations, and finally methods to transform indirect knowledge into direct demonstrations. We use 100k such synthetically-created demonstrations to finetune a 7B CodeLlama, and demonstrate that the resulting agent surpasses all comparably sized models on three web-based task benchmarks Mind2Web, MiniWoB++ and WebArena, as well as surpassing GPT-3.5 on WebArena and Mind2Web. In addition, while synthetic demonstrations prove to be only 3% the cost of human demonstrations (at $0.031 each), we show that the synthetic demonstrations can be more effective than an identical number of human demonstrations collected from limited domains.

Retrieval-augmented generation (RAG) is increasingly recognized as an effective approach for mitigating the hallucination of large language models (LLMs) through the integration of external knowledge. While numerous efforts, most studies focus on a single type of externeal knowledge source. However, in real-world applications, most situations involve diverse knowledge from various sources, yet this area has been less explored. The main dilemma is the lack of a suitable dataset containing multiple knowledge sources and pre-exploration of the associated issues. To address these challenges, we standardize a benchmark dataset that combines structured and unstructured knowledge across diverse and complementary domains. Based on this dataset, we further develop a plug-and-play RAG framework, PruningRAG, whose main characteristic is to employ multi-granularity pruning strategies for optimizing the integration of relevant information and minimizing misleading context. Building upon the standardized dataset and PruningRAG, we also report a series of experimental results, as well as insightful findings. Our dataset and code are publicly available\footnote{https://github.com/USTCAGI/PruningRAG}, with the aim of advancing future research in the RAG community.

Across science and policy, decision-makers often need to draw conclusions about the best candidate among competing alternatives. For instance, researchers may seek to infer the effectiveness of the most successful treatment or determine which demographic group benefits most from a specific treatment. Similarly, in machine learning, practitioners are often interested in the population performance of the model that performs best empirically. However, cherry-picking the best candidate leads to the winner's curse: the observed performance for the winner is biased upwards, rendering conclusions based on standard measures of uncertainty invalid. We introduce the zoom correction, a novel approach for valid inference on the winner. Our method is flexible: it can be employed in both parametric and nonparametric settings, can handle arbitrary dependencies between candidates, and automatically adapts to the level of selection bias. The method easily extends to important related problems, such as inference on the top k winners, inference on the value and identity of the population winner, and inference on "near-winners."

AI-generated writing is now all over the internet. The introduction of automated prose can sometimes change a website's character, like when once beloved publications get purchased and overhauled into AI content mills. Other times, however, it's harder to argue that AI really changed anything. For example, look at LinkedIn. The Microsoft-owned social media site for business professionals has embraced AI, even offering LinkedIn Premium subscribers access to its own in-house AI writing tools that can "rewrite" posts, profiles, and direct messages.

An artificial intelligence model can be made to spout gibberish if a single one of the many billions of numbers that compose it is altered. Large language models (LLMs) like the one behind OpenAI's ChatGPT contain billions of parameters or weights, which are the numerical values used to represent each "neuron" of their neural network. These are what get tuned and tweaked during training so the AI can learn abilities such as generating text. Input is passed through these weights, which determine the most statistically likely output.…

In 2025, entrepreneurs will unleash a flood of AI-powered apps. Finally, generative AI will deliver on the hype with a new crop of affordable consumer and business apps. This is not the consensus view today. OpenAI, Google, and xAI are locked in an arms race to train the most powerful large language model (LLM) in pursuit of artificial general intelligence, known as AGI, and their gladiatorial battle dominates the mindshare and revenue share of the fledgling Gen AI ecosystem. For example, Elon Musk raised 6 billion to launch the newcomer xAI and bought 100,000 Nvidia H100 GPUs, the costly chips used to process AI, costing north of 3 billion to train its model, Grok.