Large language models (LLMs) can solve arithmetic word problems with high accuracy, but little is known about how well they generalize to problems that are more complex than the ones on which they have been trained. Empirical investigations of such questions are impeded by two major flaws of current evaluations: (i) much of the evaluation data is contaminated, in the sense that it has already been seen during training, and (ii) benchmark datasets do not capture how problem proofs may be arbitrarily complex in various ways. As a step towards addressing these issues, we present a framework for evaluating LLMs on problems with arbitrarily complex arithmetic proofs, called MathGAP. MathGAP generates problems that follow fixed proof specifications -- along with chain-of-thought reasoning annotations -- enabling systematic studies on generalization with respect to arithmetic proof complexity. We apply MathGAP to analyze how in-context learning interacts with generalization to problems that have more complex proofs. We find that among the models tested, most show a significant decrease in performance as proofs get deeper and wider. This effect is more pronounced in complex, nonlinear proof structures, which are challenging even for GPT-4o. Surprisingly, providing in-context examples from the same distribution as the test set is not always beneficial for performance. In particular, zero-shot prompting as well as demonstrating a diverse range of examples that are less complex than the test data sometimes yield similar or higher accuracies.

Recent large visual-language action models pretrained on diverse robot datasets have demonstrated the potential for generalizing to new environments with a few in-domain data. However, those approaches usually predict discretized or continuous actions by a small action head, which limits the ability in handling diverse action spaces. In contrast, we model the continuous action with a large multi-modal diffusion transformer, dubbed as Diffusion Transformer Policy, in which we directly denoise action chunks by a large transformer model rather than a small action head. By leveraging the scaling capability of transformers, the proposed approach can effectively model continuous end-effector actions across large diverse robot datasets, and achieve better generalization performance. Extensive experiments demonstrate Diffusion Transformer Policy pretrained on diverse robot data can generalize to different embodiments, including simulation environments like Maniskill2 and Calvin, as well as the real-world Franka arm. Specifically, without bells and whistles, the proposed approach achieves state-ofthe-art performance with only a single third-view camera stream in the Calvin novel task setting (ABC D), improving the average number of tasks completed in a row of 5 to 3.6, and the pretraining stage significantly facilitates the success sequence length on the Calvin by over 1.2. The code will be publicly available. Traditional robot learning paradigm usually relies on large-scale data collected for a specific robot and task, but collecting robot data for generalist tasks is time-consuming and expensive due to the limitations of robot hardware in the real world. Nowadays, the foundational models OpenAI (2022; 2023; 2021); Rombach et al. (2021) in Natural Language Process and Computer Vision, pretrained on broad, diverse, task-agnostic datasets, have demonstrated powerful ability in solving downstream tasks either zero-shot or with a few task-specific samples. It is principally possible that a general robot policy exposed to large scale diverse robot datasets improves generalization and performance on downstream tasks Brohan et al. (2022; 2023).

Large Language Models (LLMs) have demonstrated impressive capabilities in a wide range of natural language processing tasks when leveraging in-context learning. To mitigate the additional computational and financial costs associated with in-context learning, several prompt compression methods have been proposed to compress the in-context learning prompts. Despite their success, these methods face challenges with transferability due to model-specific compression, or rely on external training data, such as GPT-4. In this paper, we investigate the ability of LLMs to develop a unified compression method that discretizes uninformative tokens, utilizing a self-supervised pre-training technique. By introducing a small number of parameters during the continual pre-training, the proposed Selection-p produces a probability for each input token, indicating whether to preserve or discard it. Experiments show Selection-p achieves state-of-the-art performance across numerous classification tasks, achieving compression rates of up to 10 times while experiencing only a marginal 0.8% decrease in performance. Moreover, it exhibits superior transferability to different models compared to prior work. Additionally, we further analyze how Selection-p helps maintain performance on in-context learning with long contexts.

Identifying offensive language is essential for maintaining safety and sustainability in the social media era. Though large language models (LLMs) have demonstrated encouraging potential in social media analytics, they lack thorough evaluation when in offensive language detection, particularly in multilingual environments. We for the first time evaluate multilingual offensive language detection of LLMs in three languages: English, Spanish, and German with three LLMs, GPT-3.5, Flan-T5, and Mistral, in both monolingual and multilingual settings. We further examine the impact of different prompt languages and augmented translation data for the task in non-English contexts. Furthermore, we discuss the impact of the inherent bias in LLMs and the datasets in the mispredictions related to sensitive topics.

We present an advanced approach to medical question-answering (QA) services, using fine-tuned Large Language Models (LLMs) to improve the accuracy and reliability of healthcare information. Our study focuses on optimizing models like LLaMA-2 and Mistral, which have shown great promise in delivering precise, reliable medical answers. By leveraging comprehensive datasets, we applied fine-tuning techniques such as rsDoRA+ and ReRAG. rsDoRA+ enhances model performance through a combination of decomposed model weights, varied learning rates for low-rank matrices, and rank stabilization, leading to improved efficiency. ReRAG, which integrates retrieval on demand and question rewriting, further refines the accuracy of the responses. This approach enables healthcare providers to access fast, dependable information, aiding in more efficient decision-making and fostering greater patient trust. Our work highlights the potential of fine-tuned LLMs to significantly improve the quality and accessibility of medical information services, ultimately contributing to better healthcare outcomes for all.

Efficient and accurate evaluation is crucial for the continuous improvement of large language models (LLMs). Among various assessment methods, subjective evaluation has garnered significant attention due to its superior alignment with real-world usage scenarios and human preferences. However, human-based evaluations are costly and lack reproducibility, making precise automated evaluators (judgers) vital in this process. In this report, we introduce \textbf{CompassJudger-1}, the first open-source \textbf{all-in-one} judge LLM. CompassJudger-1 is a general-purpose LLM that demonstrates remarkable versatility. It is capable of: 1. Performing unitary scoring and two-model comparisons as a reward model; 2. Conducting evaluations according to specified formats; 3. Generating critiques; 4. Executing diverse tasks like a general LLM. To assess the evaluation capabilities of different judge models under a unified setting, we have also established \textbf{JudgerBench}, a new benchmark that encompasses various subjective evaluation tasks and covers a wide range of topics. CompassJudger-1 offers a comprehensive solution for various evaluation tasks while maintaining the flexibility to adapt to diverse requirements. Both CompassJudger and JudgerBench are released and available to the research community athttps://github.com/open-compass/CompassJudger. We believe that by open-sourcing these tools, we can foster collaboration and accelerate progress in LLM evaluation methodologies.

Foundation models (FMs), large deep learning models pre-trained on vast, unlabeled datasets, exhibit powerful capabilities in understanding complex patterns and generating sophisticated outputs. However, they often struggle to adapt to specific tasks. Reinforcement learning (RL), which allows agents to learn through interaction and feedback, offers a compelling solution. Integrating RL with FMs enables these models to achieve desired outcomes and excel at particular tasks. Additionally, RL can be enhanced by leveraging the reasoning and generalization capabilities of FMs. This synergy is revolutionizing various fields, including robotics. FMs, rich in knowledge and generalization, provide robots with valuable information, while RL facilitates learning and adaptation through real-world interactions. This survey paper comprehensively explores this exciting intersection, examining how these paradigms can be integrated to advance robotic intelligence. We analyze the use of foundation models as action planners, the development of robotics-specific foundation models, and the mutual benefits of combining FMs with RL. Furthermore, we present a taxonomy of integration approaches, including large language models, vision-language models, diffusion models, and transformer-based RL models. We also explore how RL can utilize world representations learned from FMs to enhance robotic task execution. Our survey aims to synthesize current research and highlight key challenges in robotic reasoning and control, particularly in the context of integrating FMs and RL--two rapidly evolving technologies. By doing so, we seek to spark future research and emphasize critical areas that require further investigation to enhance robotics. We provide an updated collection of papers based on our taxonomy, accessible on our open-source project website at: https://github.com/clmoro/Robotics-RL-FMs-Integration.

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context. Such conflicts can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. In this work, we investigate whether LLMs can identify knowledge conflicts and whether it is possible to know which source of knowledge the model will rely on by analysing the residual stream of the LLM. Through probing tasks, we find that LLMs can internally register the signal of knowledge conflict in the residual stream, which can be accurately detected by probing the intermediate model activations. This allows us to detect conflicts within the residual stream before generating the answers without modifying the input or model parameters. Moreover, we find that the residual stream shows significantly different patterns when the model relies on contextual knowledge versus parametric knowledge to resolve conflicts. This pattern can be employed to estimate the behaviour of LLMs when conflict happens and prevent unexpected answers before producing the answers. Our analysis offers insights into how LLMs internally manage knowledge conflicts and provides a foundation for developing methods to control the knowledge selection processes.

Recent progress in large language models (LLMs) offers promising new approaches for recommendation system (RecSys) tasks. While the current state-of-the-art methods rely on fine-tuning LLMs to achieve optimal results, this process is costly and introduces significant engineering complexities. Conversely, methods that bypass fine-tuning and use LLMs directly are less resource-intensive but often fail to fully capture both semantic and collaborative information, resulting in sub-optimal performance compared to their fine-tuned counterparts. In this paper, we propose a Simple Training-free Approach for Recommendation (STAR), a framework that utilizes LLMs and can be applied to various recommendation tasks without the need for fine-tuning. Our approach involves a retrieval stage that uses semantic embeddings from LLMs combined with collaborative user information to retrieve candidate items. We then apply an LLM for pairwise ranking to enhance next-item prediction. Experimental results on the Amazon Review dataset show competitive performance for next item prediction, even with our retrieval stage alone. Our full method achieves Hits@10 performance of +23.8% on Beauty, +37.5% on Toys and Games, and -1.8% on Sports and Outdoors relative to the best supervised models. This framework offers an effective alternative to traditional supervised models, highlighting the potential of LLMs in recommendation systems without extensive training or custom architectures.

Ensuring correctness is crucial for code generation. Formal verification offers a definitive assurance of correctness, but demands substantial human effort in proof construction and hence raises a pressing need for automation. The primary obstacle lies in the severe lack of data -- there is much less proof than code for LLMs to train upon. In this paper, we introduce SAFE, a novel framework that overcomes the lack of human-written proof to enable automated proof generation of Rust code. SAFE establishes a self-evolving cycle where data synthesis and fine-tuning collaborate to enhance the model capability, leveraging the definitive power of a symbolic verifier in telling correct proof from incorrect ones. SAFE also re-purposes the large number of synthesized incorrect proofs to train the selfdebugging capability of the fine-tuned models, empowering them to fix incorrect proofs based on the verifier's feedback. SAFE demonstrates superior efficiency and precision compared to GPT-4o. Through tens of thousands of synthesized proofs and the self-debugging mechanism, we improve the capability of opensource models, initially unacquainted with formal verification, to automatically write proof for Rust code. This advancement leads to a significant improvement in performance, achieving a 70.50% accuracy rate in a benchmark crafted by human experts, a significant leap over GPT-4o's performance of 24.46%. Large Language Models (LLMs) have recently exhibited impressive capabilities in code generation (Roziere et al., 2023; Guo et al., 2024; Lozhkov et al., 2024; Google, 2024). However, the correctness of generated code cannot be guaranteed. To tackle this issue, prior research (Chen et al., 2022; Zhang et al., 2024a; Huang et al., 2023) has explored assessing generated code by test cases, which sometimes are also generated by LLMs. Although helpful, testing cannot cover all possible program inputs; they can reveal the presence of bugs but cannot prove their absence (Dahl et al., 1972).