Large language models have demonstrated impressive reasoning capabilities, yet they often suffer from inefficiencies due to unnecessarily verbose or redundant outputs. While many works have explored reinforcement learning (RL) to enhance reasoning abilities, most primarily focus on improving accuracy, with limited attention to reasoning efficiency. Some existing approaches introduce direct length-based rewards to encourage brevity, but this often leads to noticeable drops in accuracy. In this paper, we propose Bingo, an RL framework that advances length-based reward design to boost efficient reasoning. Bingo incorporates two key mechanisms: a significance-aware length reward, which gradually guides the model to reduce only insignificant tokens, and a dynamic length reward, which initially encourages elaborate reasoning for hard questions but decays over time to improve overall efficiency. Experiments across multiple reasoning benchmarks show that Bingo improves both accuracy and efficiency. It outperforms the vanilla reward and several other length-based reward baselines in RL, achieving a favorable trade-off between accuracy and efficiency. These results underscore the potential of training LLMs explicitly for efficient reasoning.

LLM self-evaluation relies on the LLM's own ability to estimate response correctness, which can greatly improve its deployment reliability. In this research track, we propose the Chain-of-Embedding (CoE) in the latent space to enable LLMs to perform output-free self-evaluation. CoE consists of all progressive hidden states produced during the inference time, which can be treated as the latent thinking path of LLMs. We find that when LLMs respond correctly and incorrectly, their CoE features differ, these discrepancies assist us in estimating LLM response correctness. Experiments in four diverse domains and seven LLMs fully demonstrate the effectiveness of our method. Meanwhile, its label-free design intent without any training and millisecond-level computational cost ensure real-time feedback in large-scale scenarios. More importantly, we provide interesting insights into LLM response correctness from the perspective of hidden state changes inside LLMs. Large Language Models (LLMs) have significantly enhanced their ability to generalize across diverse scenarios (Brown et al., 2020; Achiam et al., 2023; GLM et al., 2024). However, their outputs can sometimes be unstable, leading to incorrect responses that may threaten social safety. Therefore, labelfree LLM self-evaluation -- estimating the correctness of LLM responses fully through LLMs' own capabilities -- has emerged as a crucial research area. It can provide real-time response monitoring and feedback in large-scale employments, enhancing the reliability of LLMs (Sun et al., 2024). Popular self-evaluation research in the era of LLMs focuses more on output-based forms (Zhang et al., 2023). Two typical paradigms that do not assess the internal states of LLMs involve directly asking LLMs to express confidence in their responses through well-designed prompts (Lin et al., 2022a; Tian et al., 2023), and generating multiple responses by perturbing prompts (Gao et al., 2024) or decoding sampling (Wang et al., 2023) to calculating the response consistency (Xiong et al., 2024). Besides the two types, other methods basically draw on uncertainty estimation concepts from the era of deep neural networks, leveraging output logits or probability distributions to gauge the confidence of model responses (Malinin & Gales, 2020; Si et al., 2022; Huang et al., 2023; Kuhn et al., 2023). Recently, some research has revealed that the latent space of LLMs contains a substantial amount of untapped hidden state information, they can largely reflect response correctness (Azaria & Mitchell, 2023; Liu et al., 2023; Duan et al., 2024), and are usually more interpretable than LLM output (Li et al., 2024a). However, these output-free research often require correctness labels 0/1 for training probing classifiers to extract features from hidden states (Burns et al., 2022; Sky et al., 2024; Su et al., 2024). This contradicts our goal of being "label-free" and limits the generalization capabilities on unseen data.

Reliable confidence estimation for the predictions is important in many safety-critical applications. However, modern deep neural networks are often overconfident for their incorrect predictions. Recently, many calibration methods have been proposed to alleviate the overconfidence problem. With calibrated confidence, a primary and practical purpose is to detect misclassification errors by filtering out low-confidence predictions (known as failure prediction). In this paper, we find a general, widely-existed but actually-neglected phenomenon that most confidence calibration methods are useless or harmful for failure prediction. We investigate this problem and reveal that popular confidence calibration methods often lead to worse confidence separation between correct and incorrect samples, making it more difficult to decide whether to trust a prediction or not. Finally, inspired by the natural connection between flat minima and confidence separation, we propose a simple hypothesis: flat minima is beneficial for failure prediction. We verify this hypothesis via extensive experiments and further boost the performance by combining two different flat minima techniques.

Given data with label noise (i.e., incorrect data), deep neural networks would gradually memorize the label noise and impair model performance. To relieve this issue, curriculum learning is proposed to improve model performance and generalization by ordering training samples in a meaningful (e.g., easy to hard) sequence. Previous work takes incorrect samples as generic hard ones without discriminating between hard samples (i.e., hard samples in correct data) and incorrect samples. Indeed, a model should learn from hard samples to promote generalization rather than overfit to incorrect ones. In this paper, we address this problem by appending a novel loss function DiscrimLoss, on top of the existing task loss. Its main effect is to automatically and stably estimate the importance of easy samples and difficult samples (including hard and incorrect samples) at the early stages of training to improve the model performance. Then, during the following stages, DiscrimLoss is dedicated to discriminating between hard and incorrect samples to improve the model generalization. Such a training strategy can be formulated dynamically in a self-supervised manner, effectively mimicking the main principle of curriculum learning. Experiments on image classification, image regression, text sequence regression, and event relation reasoning demonstrate the versatility and effectiveness of our method, particularly in the presence of diversified noise levels.