Zeyuan Liu

Jiarui Yuan, Tailin Jin, Weize Chen et al.

True self-evolution requires agents to act as lifelong learners that internalize novel experiences to solve future problems. However, rigorously measuring this foundational capability is hindered by two obstacles: the entanglement of prior knowledge, where ``new'' knowledge may appear in pre-training data, and the entanglement of reasoning complexity, where failures may stem from problem difficulty rather than an inability to recall learned knowledge. We introduce SE-Bench, a diagnostic environment that obfuscates the NumPy library and its API doc into a pseudo-novel package with randomized identifiers. Agents are trained to internalize this package and evaluated on simple coding tasks without access to documentation, yielding a clean setting where tasks are trivial with the new API doc but impossible for base models without it. Our investigation reveals three insights: (1) the Open-Book Paradox, where training with reference documentation inhibits retention, requiring "Closed-Book Training" to force knowledge compression into weights; (2) the RL Gap, where standard RL fails to internalize new knowledge completely due to PPO clipping and negative gradients; and (3) the viability of Self-Play for internalization, proving models can learn from self-generated, noisy tasks when coupled with SFT, but not RL. Overall, SE-Bench establishes a rigorous diagnostic platform for self-evolution with knowledge internalization. Our code and dataset can be found at https://github.com/thunlp/SE-Bench.

Xiaobin Song, Zeyuan Liu, Benben Jiang

Machine learning methods based on AdaBoost have been widely applied to various classification problems across many mission-critical applications including healthcare, law and finance. However, there is a growing concern about the unfairness and discrimination of data-driven classification models, which is inevitable for classical algorithms including AdaBoost. In order to achieve fair classification, a novel fair AdaBoost (FAB) approach is proposed that is an interpretable fairness-improving variant of AdaBoost. We mainly investigate binary classification problems and focus on the fairness of three different indicators (i.e., accuracy, false positive rate and false negative rate). By utilizing a fairness-aware reweighting technique for base classifiers, the proposed FAB approach can achieve fair classification while maintaining the advantage of AdaBoost with negligible sacrifice of predictive performance. In addition, a hyperparameter is introduced in FAB to show preferences for the fairness-accuracy trade-off. An upper bound for the target loss function that quantifies error rate and unfairness is theoretically derived for FAB, which provides a strict theoretical support for the fairness-improving methods designed for AdaBoost. The effectiveness of the proposed method is demonstrated on three real-world datasets (i.e., Adult, COMPAS and HSLS) with respect to the three fairness indicators. The results are accordant with theoretic analyses, and show that (i) FAB significantly improves classification fairness at a small cost of accuracy compared with AdaBoost; and (ii) FAB outperforms state-of-the-art fair classification methods including equalized odds method, exponentiated gradient method, and disparate mistreatment method in terms of the fairness-accuracy trade-off.

Songqi Zhou, Zeyuan Liu, Benben Jiang

Ensuring fairness in machine learning models is a critical challenge. Existing debiasing methods often compromise performance, rely on static correction strategies, and struggle with data sparsity, particularly within minority groups. Furthermore, their utilization of sensitive attributes is often suboptimal, either depending excessively on complete attribute labeling or disregarding these attributes entirely. To overcome these limitations, we propose FairNet, a novel framework for dynamic, instance-level fairness correction. FairNet integrates a bias detector with conditional low-rank adaptation (LoRA), which enables selective activation of the fairness correction mechanism exclusively for instances identified as biased, and thereby preserve performance on unbiased instances. A key contribution is a new contrastive loss function for training the LoRA module, specifically designed to minimize intra-class representation disparities across different sensitive groups and effectively address underfitting in minority groups. The FairNet framework can flexibly handle scenarios with complete, partial, or entirely absent sensitive attribute labels. Theoretical analysis confirms that, under moderate TPR/FPR for the bias detector, FairNet can enhance the performance of the worst group without diminishing overall model performance, and potentially yield slight performance improvements. Comprehensive empirical evaluations across diverse vision and language benchmarks validate the effectiveness of FairNet.