Yingchun Wang

Kexin Huang, Xiangyang Liu, Qianyu Guo et al.

The widespread adoption of large language models (LLMs) across various regions underscores the urgent need to evaluate their alignment with human values. Current benchmarks, however, fall short of effectively uncovering safety vulnerabilities in LLMs. Despite numerous models achieving high scores and 'topping the chart' in these evaluations, there is still a significant gap in LLMs' deeper alignment with human values and achieving genuine harmlessness. To this end, this paper proposes a value alignment benchmark named Flames, which encompasses both common harmlessness principles and a unique morality dimension that integrates specific Chinese values such as harmony. Accordingly, we carefully design adversarial prompts that incorporate complex scenarios and jailbreaking methods, mostly with implicit malice. By prompting 17 mainstream LLMs, we obtain model responses and rigorously annotate them for detailed evaluation. Our findings indicate that all the evaluated LLMs demonstrate relatively poor performance on Flames, particularly in the safety and fairness dimensions. We also develop a lightweight specified scorer capable of scoring LLMs across multiple dimensions to efficiently evaluate new models on the benchmark. The complexity of Flames has far exceeded existing benchmarks, setting a new challenge for contemporary LLMs and highlighting the need for further alignment of LLMs. Our benchmark is publicly available at https://github.com/AIFlames/Flames.

Yingchun Wang, Jingcai Guo, Song Guo et al.

Mixed-precision quantization mostly predetermines the model bit-width settings before actual training due to the non-differential bit-width sampling process, obtaining sub-optimal performance. Worse still, the conventional static quality-consistent training setting, i.e., all data is assumed to be of the same quality across training and inference, overlooks data quality changes in real-world applications which may lead to poor robustness of the quantized models. In this paper, we propose a novel Data Quality-aware Mixed-precision Quantization framework, dubbed DQMQ, to dynamically adapt quantization bit-widths to different data qualities. The adaption is based on a bit-width decision policy that can be learned jointly with the quantization training. Concretely, DQMQ is modeled as a hybrid reinforcement learning (RL) task that combines model-based policy optimization with supervised quantization training. By relaxing the discrete bit-width sampling to a continuous probability distribution that is encoded with few learnable parameters, DQMQ is differentiable and can be directly optimized end-to-end with a hybrid optimization target considering both task performance and quantization benefits. Trained on mixed-quality image datasets, DQMQ can implicitly select the most proper bit-width for each layer when facing uneven input qualities. Extensive experiments on various benchmark datasets and networks demonstrate the superiority of DQMQ against existing fixed/mixed-precision quantization methods.

Yingchun Wang, Jingcai Guo, Song Guo et al.

Recent studies show that even highly biased dense networks contain an unbiased substructure that can achieve better out-of-distribution (OOD) generalization than the original model. Existing works usually search the invariant subnetwork using modular risk minimization (MRM) with out-domain data. Such a paradigm may bring about two potential weaknesses: 1) Unfairness, due to the insufficient observation of out-domain data during training; and 2) Sub-optimal OOD generalization, due to the feature-untargeted model pruning on the whole data distribution. In this paper, we propose a novel Spurious Feature-targeted model Pruning framework, dubbed SFP, to automatically explore invariant substructures without referring to the above weaknesses. Specifically, SFP identifies in-distribution (ID) features during training using our theoretically verified task loss, upon which, SFP can perform ID targeted-model pruning that removes branches with strong dependencies on ID features. Notably, by attenuating the projections of spurious features into model space, SFP can push the model learning toward invariant features and pull that out of environmental features, devising optimal OOD generalization. Moreover, we also conduct detailed theoretical analysis to provide the rationality guarantee and a proof framework for OOD structures via model sparsity, and for the first time, reveal how a highly biased data distribution affects the model's OOD generalization. Extensive experiments on various OOD datasets show that SFP can significantly outperform both structure-based and non-structure OOD generalization SOTAs, with accuracy improvement up to 4.72% and 23.35%, respectively.

Yingchun Wang, Song Guo, Jingcai Guo et al.

Network pruning is a promising way to generate light but accurate models and enable their deployment on resource-limited edge devices. However, the current state-of-the-art assumes that the effective sub-network and the other superfluous parameters in the given network share the same distribution, where pruning inevitably involves a distribution truncation operation. They usually eliminate values near zero. While simple, it may not be the most appropriate method, as effective models may naturally have many small values associated with them. Removing near-zero values already embedded in model space may significantly reduce model accuracy. Another line of work has proposed to assign discrete prior over all possible sub-structures that still rely on human-crafted prior hypotheses. Worse still, existing methods use regularized point estimates, namely Hard Pruning, that can not provide error estimations and fail reliability justification for the pruned networks. In this paper, we propose a novel distribution-lossless pruning method, named DLLP, to theoretically find the pruned lottery within Bayesian treatment. Specifically, DLLP remodels the vanilla networks as discrete priors for the latent pruned model and the other redundancy. More importantly, DLLP uses Stein Variational Inference to approach the latent prior and effectively bypasses calculating KL divergence with unknown distribution. Extensive experiments based on small Cifar-10 and large-scaled ImageNet demonstrate that our method can obtain sparser networks with great generalization performance while providing quantified reliability for the pruned model.

Yingchun Wang, Jingcai Guo, Jie Zhang et al.

Federated learning (FL) is an emerging technique that trains massive and geographically distributed edge data while maintaining privacy. However, FL has inherent challenges in terms of fairness and computational efficiency due to the rising heterogeneity of edges, and thus usually results in sub-optimal performance in recent state-of-the-art (SOTA) solutions. In this paper, we propose a Customized Federated Learning (CFL) system to eliminate FL heterogeneity from multiple dimensions. Specifically, CFL tailors personalized models from the specially designed global model for each client jointly guided by an online trained model-search helper and a novel aggregation algorithm. Extensive experiments demonstrate that CFL has full-stack advantages for both FL training and edge reasoning and significantly improves the SOTA performance w.r.t. model accuracy (up to 7.2% in the non-heterogeneous environment and up to 21.8% in the heterogeneous environment), efficiency, and FL fairness.

Tianle Gu, Kexin Huang, Ruilin Luo et al.

Large Language Models (LLMs) can memorize sensitive information, raising concerns about potential misuse. LLM Unlearning, a post-hoc approach to remove this information from trained LLMs, offers a promising solution to mitigate these risks. However, previous practices face three key challenges: 1. Utility: successful unlearning often causes catastrophic collapse on unrelated tasks. 2. Efficiency: many methods either involve adding similarly sized models, which slows down unlearning or inference, or require retain data that are difficult to obtain. 3. Robustness: even effective methods may still leak data via extraction techniques. To address these challenges, we propose MEOW, a simple yet effective gradient descent-based unlearning method. Specifically, we use an offline LLM to generate a set of inverted facts. Then, we design a new metric, MEMO, to quantify memorization in LLMs. Finally, based on the signals provided by MEMO, we select the most appropriate set of inverted facts and finetune the model based on them. We evaluate MEOW on the commonly used unlearn benchmark, ToFU, with Llama2-7B-Chat and Phi-1.5B, and test it on both NLU and NLG tasks. Results demonstrate significant improvement of MEOW in forget quality without substantial loss in model utility. Meanwhile, MEOW does not exhibit significant degradation in NLU or NLG capabilities, and there is even a slight improvement in NLU performance.

Lingyu Li, Yixu Wang, Haiquan Zhao et al.

With large language models (LLMs) increasingly deployed as cognitive engines for AI agents, the reliability and effectiveness critically hinge on their intrinsic epistemic agency, which remains understudied. Epistemic agency, the ability to flexibly construct, adapt, and monitor beliefs about dynamic environments, represents a base-model-level capacity independent of specific tools, modules, or applications. We characterize the holistic process underlying epistemic agency, which unfolds in seven interrelated dimensions: prediction, decision-making, perception, memory, counterfactual thinking, belief updating, and meta-reflection. Correspondingly, we propose Reflection-Bench, a cognitive-psychology-inspired benchmark consisting of seven tasks with long-term relevance and minimization of data leakage. Through a comprehensive evaluation of 16 models using three prompting strategies, we identify a clear three-tier performance hierarchy and significant limitations of current LLMs, particularly in meta-reflection capabilities. While state-of-the-art LLMs demonstrate rudimentary signs of epistemic agency, our findings suggest several promising research directions, including enhancing core cognitive functions, improving cross-functional coordination, and developing adaptive processing mechanisms. Our code and data are available at https://github.com/AI45Lab/ReflectionBench.

Lingyu Li, Yang Yao, Yixu Wang et al.

As Large Language Models (LLMs) continue to advance, they exhibit certain cognitive patterns similar to those of humans that are not directly specified in training data. This study investigates this phenomenon by focusing on temporal cognition in LLMs. Leveraging the similarity judgment task, we find that larger models spontaneously establish a subjective temporal reference point and adhere to the Weber-Fechner law, whereby the perceived distance logarithmically compresses as years recede from this reference point. To uncover the mechanisms behind this behavior, we conducted multiple analyses across neuronal, representational, and informational levels. We first identify a set of temporal-preferential neurons and find that this group exhibits minimal activation at the subjective reference point and implements a logarithmic coding scheme convergently found in biological systems. Probing representations of years reveals a hierarchical construction process, where years evolve from basic numerical values in shallow layers to abstract temporal orientation in deep layers. Finally, using pre-trained embedding models, we found that the training corpus itself possesses an inherent, non-linear temporal structure, which provides the raw material for the model's internal construction. In discussion, we propose an experientialist perspective for understanding these findings, where the LLMs' cognition is viewed as a subjective construction of the external world by its internal representational system. This nuanced perspective implies the potential emergence of alien cognitive frameworks that humans cannot intuitively predict, pointing toward a direction for AI alignment that focuses on guiding internal constructions. Our code is available at https://TheOtherMind.github.io.

Meilin Li, Ji He, Yi Yu et al.

The rapid proliferation of Artificial Intelligence Generated Content has precipitated a crisis of trust and urgent regulatory demands. However, existing identification tools suffer from fragmentation and a lack of support for visible compliance marking. To address these gaps, we introduce the \textbf{UniMark}, an open-source, unified framework for multimodal content governance. Our system features a modular unified engine that abstracts complexities across text, image, audio, and video modalities. Crucially, we propose a novel dual-operation strategy, natively supporting both \emph{Hidden Watermarking} for copyright protection and \emph{Visible Marking} for regulatory compliance. Furthermore, we establish a standardized evaluation framework with three specialized benchmarks (Image/Video/Audio-Bench) to ensure rigorous performance assessment. This toolkit bridges the gap between advanced algorithms and engineering implementation, fostering a more transparent and secure digital ecosystem.

Yang Yao, Xuan Tong, Ruofan Wang et al.

Large Reasoning Models (LRMs) have significantly advanced beyond traditional Large Language Models (LLMs) with their exceptional logical reasoning capabilities, yet these improvements introduce heightened safety risks. When subjected to jailbreak attacks, their ability to generate more targeted and organized content can lead to greater harm. Although some studies claim that reasoning enables safer LRMs against existing LLM attacks, they overlook the inherent flaws within the reasoning process itself. To address this gap, we propose the first jailbreak attack targeting LRMs, exploiting their unique vulnerabilities stemming from the advanced reasoning capabilities. Specifically, we introduce a Chaos Machine, a novel component to transform attack prompts with diverse one-to-one mappings. The chaos mappings iteratively generated by the machine are embedded into the reasoning chain, which strengthens the variability and complexity and also promotes a more robust attack. Based on this, we construct the Mousetrap framework, which makes attacks projected into nonlinear-like low sample spaces with mismatched generalization enhanced. Also, due to the more competing objectives, LRMs gradually maintain the inertia of unpredictable iterative reasoning and fall into our trap. Success rates of the Mousetrap attacking o1-mini, Claude-Sonnet and Gemini-Thinking are as high as 96%, 86% and 98% respectively on our toxic dataset Trotter. On benchmarks such as AdvBench, StrongREJECT, and HarmBench, attacking Claude-Sonnet, well-known for its safety, Mousetrap can astonishingly achieve success rates of 87.5%, 86.58% and 93.13% respectively. Attention: This paper contains inappropriate, offensive and harmful content.