Exploring the Limit of Outcome Reward for Learning Mathematical Reasoning

Reasoning abilities, especially those for solving complex math problems, are crucial components of general intelligence. Recent advances by proprietary companies, such as o-series models of OpenAI, have made remarkable progress on reasoning tasks. However, the complete technical details remain unrevealed, and the techniques that are believed certainly to be adopted are only reinforcement learning (RL) and the long chain of thoughts. This paper proposes a new RL framework, termed OREAL, to pursue the performance limit that can be achieved through \textbf{O}utcome \textbf{RE}w\textbf{A}rd-based reinforcement \textbf{L}earning for mathematical reasoning tasks, where only binary outcome rewards are easily accessible. We theoretically prove that behavior cloning on positive trajectories from best-of-N (BoN) sampling is sufficient to learn the KL-regularized optimal policy in binary feedback environments. This formulation further implies that the rewards of negative samples should be reshaped to ensure the gradient consistency between positive and negative samples. To alleviate the long-existing difficulties brought by sparse rewards in RL, which are even exacerbated by the partial correctness of the long chain of thought for reasoning tasks, we further apply a token-level reward model to sample important tokens in reasoning trajectories for learning. With OREAL, for the first time, a 7B model can obtain 94.0 pass@1 accuracy on MATH-500 through RL, being on par with 32B models. OREAL-32B also surpasses previous 32B models trained by distillation with 95.0 pass@1 accuracy on MATH-500. Our investigation also indicates the importance of initial policy models and training queries for RL. Code, models, and data will be released to benefit future research\footnote{https://github.com/InternLM/OREAL}.

Condor: Enhance LLM Alignment with Knowledge-Driven Data Synthesis and Refinement

The quality of Supervised Fine-Tuning (SFT) data plays a critical role in enhancing the conversational capabilities of Large Language Models (LLMs). However, as LLMs become more advanced, the availability of high-quality human-annotated SFT data has become a significant bottleneck, necessitating a greater reliance on synthetic training data. In this work, we introduce Condor, a novel two-stage synthetic data generation framework that incorporates World Knowledge Tree and Self-Reflection Refinement to produce high-quality SFT data at scale. Our experimental results demonstrate that a base model fine-tuned on only 20K Condor-generated samples achieves superior performance compared to counterparts. The additional refinement stage in Condor further enables iterative self-improvement for LLMs at various scales (up to 72B), validating the effectiveness of our approach. Furthermore, our investigation into the scaling for synthetic data in post-training reveals substantial unexplored potential for performance improvements, opening promising avenues for future research.

Efficient Transmission of Radiomaps via Physics-Enhanced Semantic Communications

Enriching information of spectrum coverage, radiomap plays an important role in many wireless communication applications, such as resource allocation and network optimization. To enable real-time, distributed spectrum management, particularly in the scenarios with unstable and dynamic environments, the efficient transmission of spectrum coverage information for radiomaps from edge devices to the central server emerges as a critical problem. In this work, we propose an innovative physics-enhanced semantic communication framework tailored for efficient radiomap transmission based on generative learning models. Specifically, instead of bit-wise message passing, we only transmit the key "semantics" in radiomaps characterized by the radio propagation behavior and surrounding environments, where semantic compression schemes are utilized to reduce the communication overhead. Incorporating the novel concepts of Radio Depth Maps, the radiomaps are reconstructed from the delivered semantic information backboned on the conditional generative adversarial networks. Our framework is further extended to facilitate its implementation in the scenarios of multi-user edge computing, by integrating with federated learning for collaborative model training while preserving the data privacy. Experimental results show that our approach achieves high accuracy in radio coverage information recovery at ultra-high bandwidth efficiency, which has great potentials in many wireless-generated data transmission applications.

DualGFL: Federated Learning with a Dual-Level Coalition-Auction Game

Despite some promising results in federated learning using game-theoretical methods, most existing studies mainly employ a one-level game in either a cooperative or competitive environment, failing to capture the complex dynamics among participants in practice. To address this issue, we propose DualGFL, a novel Federated Learning framework with a Dual-level Game in cooperative-competitive environments. DualGFL includes a lower-level hedonic game where clients form coalitions and an upper-level multi-attribute auction game where coalitions bid for training participation. At the lower-level DualGFL, we introduce a new auction-aware utility function and propose a Pareto-optimal partitioning algorithm to find a Pareto-optimal partition based on clients' preference profiles. At the upper-level DualGFL, we formulate a multi-attribute auction game with resource constraints and derive equilibrium bids to maximize coalitions' winning probabilities and profits. A greedy algorithm is proposed to maximize the utility of the central server. Extensive experiments on real-world datasets demonstrate DualGFL's effectiveness in improving both server utility and client utility.

LaMI-GO: Latent Mixture Integration for Goal-Oriented Communications Achieving High Spectrum Efficiency

The recent rise of semantic-style communications includes the development of goal-oriented communications (GOCOMs) remarkably efficient multimedia information transmissions. The concept of GO-COMS leverages advanced artificial intelligence (AI) tools to address the rising demand for bandwidth efficiency in applications, such as edge computing and Internet-of-Things (IoT). Unlike traditional communication systems focusing on source data accuracy, GO-COMs provide intelligent message delivery catering to the special needs critical to accomplishing downstream tasks at the receiver. In this work, we present a novel GO-COM framework, namely LaMI-GO that utilizes emerging generative AI for better quality-of-service (QoS) with ultra-high communication efficiency. Specifically, we design our LaMI-GO system backbone based on a latent diffusion model followed by a vector-quantized generative adversarial network (VQGAN) for efficient latent embedding and information representation. The system trains a common feature codebook the receiver side. Our experimental results demonstrate substantial improvement in perceptual quality, accuracy of downstream tasks, and bandwidth consumption over the state-of-the-art GOCOM systems and establish the power of our proposed LaMI-GO communication framework.

Are Your LLMs Capable of Stable Reasoning?

The rapid advancement of Large Language Models (LLMs) has demonstrated remarkable progress in complex reasoning tasks. However, a significant discrepancy persists between benchmark performances and real-world applications. We identify this gap as primarily stemming from current evaluation protocols and metrics, which inadequately capture the full spectrum of LLM capabilities, particularly in complex reasoning tasks where both accuracy and consistency are crucial. This work makes two key contributions. First, we introduce G-Pass@k, a novel evaluation metric that provides a continuous assessment of model performance across multiple sampling attempts, quantifying both the model's peak performance potential and its stability. Second, we present LiveMathBench, a dynamic benchmark comprising challenging, contemporary mathematical problems designed to minimize data leakage risks during evaluation. Through extensive experiments using G-Pass@k on state-of-the-art LLMs with LiveMathBench, we provide comprehensive insights into both their maximum capabilities and operational consistency. Our findings reveal substantial room for improvement in LLMs' "realistic" reasoning capabilities, highlighting the need for more robust evaluation methods. The benchmark and detailed results are available at: https://github.com/open-compass/GPassK.

Diff-GO$^\text{n}$: Enhancing Diffusion Models for Goal-Oriented Communications

The rapid expansion of edge devices and Internet-of-Things (IoT) continues to heighten the demand for data transport under limited spectrum resources. The goal-oriented communications (GO-COM), unlike traditional communication systems designed for bit-level accuracy, prioritizes more critical information for specific application goals at the receiver. To improve the efficiency of generative learning models for GO-COM, this work introduces a novel noise-restricted diffusion-based GO-COM (Diff-GO$^\text{n}$) framework for reducing bandwidth overhead while preserving the media quality at the receiver. Specifically, we propose an innovative Noise-Restricted Forward Diffusion (NR-FD) framework to accelerate model training and reduce the computation burden for diffusion-based GO-COMs by leveraging a pre-sampled pseudo-random noise bank (NB). Moreover, we design an early stopping criterion for improving computational efficiency and convergence speed, allowing high-quality generation in fewer training steps. Our experimental results demonstrate superior perceptual quality of data transmission at a reduced bandwidth usage and lower computation, making Diff-GO$^\text{n}$ well-suited for real-time communications and downstream applications.

CompassJudger-1: All-in-one Judge Model Helps Model Evaluation and Evolution

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.

ProSA: Assessing and Understanding the Prompt Sensitivity of LLMs

Large language models (LLMs) have demonstrated impressive capabilities across various tasks, but their performance is highly sensitive to the prompts utilized. This variability poses challenges for accurate assessment and user satisfaction. Current research frequently overlooks instance-level prompt variations and their implications on subjective evaluations. To address these shortcomings, we introduce ProSA, a framework designed to evaluate and comprehend prompt sensitivity in LLMs. ProSA incorporates a novel sensitivity metric, PromptSensiScore, and leverages decoding confidence to elucidate underlying mechanisms. Our extensive study, spanning multiple tasks, uncovers that prompt sensitivity fluctuates across datasets and models, with larger models exhibiting enhanced robustness. We observe that few-shot examples can alleviate this sensitivity issue, and subjective evaluations are also susceptible to prompt sensitivities, particularly in complex, reasoning-oriented tasks. Furthermore, our findings indicate that higher model confidence correlates with increased prompt robustness. We believe this work will serve as a helpful tool in studying prompt sensitivity of LLMs. The project is released at: https://github.com/open-compass/ProSA .

HelloBench: Evaluating Long Text Generation Capabilities of Large Language Models

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks (e.g., long-context understanding), and many benchmarks have been proposed. However, we observe that long text generation capabilities are not well investigated. Therefore, we introduce the Hierarchical Long Text Generation Benchmark (HelloBench), a comprehensive, in-the-wild, and open-ended benchmark to evaluate LLMs' performance in generating long text. Based on Bloom's Taxonomy, HelloBench categorizes long text generation tasks into five subtasks: open-ended QA, summarization, chat, text completion, and heuristic text generation. Besides, we propose Hierarchical Long Text Evaluation (HelloEval), a human-aligned evaluation method that significantly reduces the time and effort required for human evaluation while maintaining a high correlation with human evaluation. We have conducted extensive experiments across around 30 mainstream LLMs and observed that the current LLMs lack long text generation capabilities. Specifically, first, regardless of whether the instructions include explicit or implicit length constraints, we observe that most LLMs cannot generate text that is longer than 4000 words. Second, we observe that while some LLMs can generate longer text, many issues exist (e.g., severe repetition and quality degradation). Third, to demonstrate the effectiveness of HelloEval, we compare HelloEval with traditional metrics (e.g., ROUGE, BLEU, etc.) and LLM-as-a-Judge methods, which show that HelloEval has the highest correlation with human evaluation. We release our code in https://github.com/Quehry/HelloBench.