Artificial and biological systems may evolve similar computational solutions despite fundamental differences in architecture and learning mechanisms – a form of convergent evolution. We demonstrate this phenomenon through large-scale analysis of alignment between human brain activity and internal representations of over 600 AI models spanning language and vision domains, from 1.33M to 72B parameters. Analyzing 60 million alignment measurements reveals that higher-performing models spontaneously develop stronger brain alignment without explicit neural constraints, with language models showing markedly stronger correlation (r=0.89, p<7.5e-13) than vision models (r=0.53, p<2.0e-44). Crucially, longitudinal analysis demonstrates that brain alignment consistently precedes performance improvements during training, suggesting that developing brain-like representations may be a necessary stepping stone toward higher capabilities. We find systematic patterns: language models exhibit strongest alignment with limbic and integrative regions, while vision models show progressive alignment with visual cortices; deeper processing layers converge across modalities; and as representational scale increases, alignment systematically shifts from primary sensory to higher-order associative regions. These findings provide compelling evidence that optimization for task performance naturally drives AI systems toward brain-like computational strategies, offering both fundamental insights into principles of intelligent information processing and practical guidance for developing more capable AI systems.

Ensuring Large Language Model (LLM) safety remains challenging due to the absence of universal standards and reliable content validators, making it difficult to obtain effective training signals. We discover that aligned models already possess robust internal safety beliefs: they consistently produce high-confidence refusals to harmful requests while exhibiting high entropy when generating potentially dangerous content. This entropy gap reveals an untapped signal—models intrinsically "know" when to refuse. We introduce Safety Instincts Reinforcement Learning (*SIRL*), which transforms this internal confidence into a self-generated reward signal, eliminating dependence on external validators or human annotations. *SIRL* teaches models to trust their safety instincts by reinforcing low-entropy refusal behaviors. Evaluated on Llama and Qwen models, *SIRL* maintains 89\%+ Defense Success Rates (DSRs) against 20+ jailbreak methods, from static prompts to automated attacks. Using only 15,000 unlabeled prompts, *SIRL* surpasses resource-intensive supervised methods while preserving performance on mathematics, coding, and conversation benchmarks. Our work demonstrates that effective alignment can emerge from within, paving the way for more autonomous and robust AI safety mechanisms that scale without extensive human oversight.

On-policy self-distillation has emerged as a promising paradigm for post-training language models, in which the model conditions on environment feedback to serve as its own teacher, providing dense token-level rewards without external teacher models or step-level annotations. Despite its empirical success, what this reward actually measures and what kind of credit it assigns remain unclear. Under a posterior-compatibility interpretation of feedback conditioning, standard in the implicit-reward literature, we show that the self-distillation token reward is a Bayesian filtering increment whose trajectory sum is exactly the pointwise mutual information between the response and the feedback given the input. This pMI can be raised by input-specific reasoning or by input-generic shortcuts, so we further decompose the teacher log-probability along the input axis. Based on this analysis, we propose CREDIT (Contrastive REward from DIsTillation), which isolates the input-specific component with a batch-contrastive baseline. At the sequence level, CREDIT is a teacher-side surrogate for a contrastive pMI objective that also penalizes responses remaining likely under unrelated inputs. Across coding, scientific reasoning, and tool-use benchmarks on two model families, CREDIT delivers the strongest aggregate performance at negligible additional compute.

Spiking Transformers model token interactions primarily through spiking self-attention (SSA). However, binary query and key representations map continuous similarities to sparse and discrete relation responses, which may suppress weak relations and limit the propagation of local spatial context. To address this limitation, we introduce Spiking Local Interaction (SLI) and Adaptive Complementary Fusion (ACF). SLI establishes an attention-independent pathway for direct information exchange among neighboring spiking tokens using lightweight depthwise–pointwise transformations. ACF integrates SSA and SLI through layer-specific, channel-wise coefficients that adaptively balance their contributions at different network depths. The proposed design preserves the original attention formulation and can be incorporated into different Spiking Transformer architectures with modest parameter overhead. Experiments on ImageNet-1K, CIFAR-10, CIFAR-100, CIFAR10-DVS, and ADE20K show consistent improvements across image classification, event-based recognition, and semantic segmentation. In particular, QKFormer with SLI and ACF achieves 84.37% Top-1 accuracy on ImageNet-1K and 37.5% mIoU on ADE20K, where the segmentation model is trained without ImageNet pretraining. Ablation studies and qualitative analyses further indicate that SSA and SLI capture complementary interaction patterns and that learnable fusion consistently outperforms fixed weighting.

Field-Programmable Gate Arrays (FPGAs) have been shown to be viable for Large Language Model (LLM) deployment, but they remain less competitive than embedded GPUs and NPUs for final edge products. This is largely because existing FPGA-based LLM accelerator prototypes rely on large, expensive FPGA devices to provide sufficient hardware resources for satisfactory performance, whereas edge products are highly cost-sensitive. In this work, we move beyond pure architectural prototyping to evaluate the feasibility of using low-cost FPGAs as the final implementation medium for LLM deployment. We propose Hummingbird+, which encompasses: (1) a compact embedded FPGA-based LLM accelerator designed to deliver comparable inference performance compared to embedded GPUs and NPUs, and (2) a custom Printed Circuit Board (PCB) built around a Zynq UltraScale XCZU2CG/3EG SoC, equipped with 24GB of memory and an expected Bill of Materials (BOMs) under $150 in mass production. Through extensive FPGA-centric optimizations, we significantly reduce the accelerator's resource consumption, enabling deployment on entry-level FPGAs with exceptional cost efficiency. On this platform, we successfully deploy the GPTQ 4-bit Qwen3-30B-A3B LLM, achieving a decoding speed of over 18 token/s and a prefill speed of over 50 token/s without further model compression. To our knowledge, this is the first demonstration of an FPGA-based edge product serving as a practical and cost-effective final implementation medium for LLM deployment.

In recent years, Spiking Neural Networks (SNNs) have achieved remarkable progress, with Spiking Transformers emerging as a promising architecture for energy-efficient sequence modeling. However, existing Spiking Transformers still lack a principled mechanism for effective temporal fusion, limiting their ability to fully exploit spatiotemporal dependencies. Inspired by feedforward–feedback modulation in the human visual pathway, we propose , the first Spiking Transformer framework that achieves bidirectional temporal fusion by decoupling temporal modeling across its core components. Specifically, TEFormer employs a lightweight and hyperparameter-free , enabling fully parallel computation, while incorporating a to aggregate temporal information in reverse order and reinforce temporal consistency. Extensive experiments across a wide range of benchmarks demonstrate that TEFormer consistently and significantly outperforms strong SNN and Spiking Transformer baselines under diverse datasets. Moreover, through the first systematic evaluation of Spiking Transformers under different neural encoding schemes, we show that the performance gains of TEFormer remain stable across encoding choices, indicating that the improved temporal modeling directly translates into reliable accuracy improvements across varied spiking representations. These results collectively establish TEFormer as an effective and general framework for temporal modeling in Spiking Transformers. Code:

Spiking Neural Networks (SNNs) utilize spike-based activations to mimic the brain's energy-efficient information processing. However, the binary and discontinuous nature of spike activations causes vanishing gradients, making adversarial robustness evaluation via gradient descent unreliable. While improved surrogate gradient methods have been proposed, their effectiveness under strong adversarial attacks remains unclear. We propose a more reliable framework for evaluating SNN adversarial robustness. We theoretically analyze the degree of gradient vanishing in surrogate gradients and introduce the Adaptive Sharpness Surrogate Gradient (ASSG), which adaptively evolves the shape of the surrogate function according to the input distribution during attack iterations, thereby enhancing gradient accuracy while mitigating gradient vanishing. In addition, we design an adversarial attack with adaptive step size under the L_∞ constraint-Stable Adaptive Projected Gradient Descent (SA-PGD), achieving faster and more stable convergence under imprecise gradients. Extensive experiments show that our approach substantially increases attack success rates across diverse adversarial training schemes, SNN architectures and neuron models, providing a more generalized and reliable evaluation of SNN adversarial robustness. The experimental results further reveal that the robustness of current SNNs has been significantly overestimated and highlighting the need for more dependable adversarial training methods.

Spiking transformers are emerging as a promising architecture that combines the energy efficiency of Spiking Neural Networks (SNNs) with the powerful attention mechanisms of transformers. However, existing hardware accelerators lack support for spiking attention, exhibit limited throughput when exploiting fine-grained sparsity, and struggle with scalable parallelism in sparse computation. To address these challenges, we propose FireFly-T, a dual-engine overlay architecture that integrates a sparse engine for activation sparsity and a binary engine for spiking attention. In the sparse engine, we present a high-throughput sparse decoder that exploits fine-grained sparsity by concurrently extracting multiple non-zero spikes. To complement this, we introduce a scalable load balancing mechanism with weight dispatch and out-of-order execution, eliminating bank conflicts to support scalable multidimensional parallelism. In the binary engine, we leverage the byte-level write capability of SRAMs to efficiently manipulate the 3D dataflows required for spiking attention with minimal resource overhead. We also optimize the core AND-PopCount operation in spiking attention through a LUT6-based implementation, improving timing closure and reducing LUT utilization on Xilinx FPGAs. As an overlay architecture, FireFly-T further incorporates an orchestrator that dynamically manipulates input dataflows with flexible adaptation to diverse network topologies, while ensuring efficient resource utilization and maintaining high throughput. Experimental results demonstrate that our accelerator achieves 1.39 & times; and 2.40 & times; higher energy efficiency, as well as 4.21 & times; and 7.10 & times; greater DSP efficiency, compared to FireFly v2 and the transformer-enabled SpikeTA, respectively. These results highlight its potential as an efficient hardware platform for spiking transformers.

Spiking Neural Networks (SNNs) offer a biologically plausible learning mechanism through synaptic plasticity, enabling unsupervised adaptation without the computational overhead of backpropagation. To harness this capability for robotics, this paper presents FireFly-P, an FPGA-based hardware accelerator that implements a novel plasticity algorithm for real-time adaptive control. By leveraging on-chip plasticity, our architecture enhances the network's generalization, ensuring robust performance in dynamic and unstructured environments. The hardware design achieves an end-to-end latency of just 8 μs for both inference and plasticity updates, enabling rapid adaptation to unseen scenarios. Implemented on a tiny Cmod A7-35T FPGA, FireFly-P consumes only 0.713 W and ∼10K LUTs, making it ideal for power- and resource-constrained embedded robotic platforms. This work demonstrates that hardware-accelerated SNN plasticity is a viable path toward enabling adaptive, low-latency, and energy-efficient control systems.

This paper presents a novel approach leveraging Spiking Neural Networks (SNNs) to construct a Variational Quantized Autoencoder (VQ-VAE) with a temporal codebook inspired by hippocampal time cells. This design captures and utilizes temporal dependencies, significantly enhancing the generative capabilities of SNNs. Neuroscientific research has identified hippocampal "time cells" that fire sequentially during temporally structured experiences. Our temporal codebook emulates this behavior by triggering the activation of time cell populations based on similarity measures as input stimuli pass through it. We conducted extensive experiments on standard benchmark datasets, including MNIST, FashionMNIST, CIFAR10, CelebA, and downsampled LSUN Bedroom, to validate our model's performance. Furthermore, we evaluated the effectiveness of the temporal codebook on neuromorphic datasets NMNIST and DVS-CIFAR10, and demonstrated the model's capability with high-resolution datasets such as CelebA-HQ, LSUN Bedroom, and LSUN Church. The experimental results indicate that our method consistently outperforms existing SNN-based generative models across multiple datasets, achieving state-of-the-art performance. Notably, our approach excels in generating high-resolution and temporally consistent data, underscoring the crucial role of temporal information in SNN-based generative modeling.

Lateral connection is a fundamental feature of biological neural circuits, facilitating local information processing and adaptive learning. In this work, we integrate lateral connections with a substructure selection network to develop a novel diffusion model based on spiking neural networks (SNNs). Unlike conventional artificial neural networks, SNNs employ an intrinsic spiking inner loop to process sequential binary spikes. We leverage this spiking inner loop alongside a lateral connection mechanism to iteratively refine the substructure selection network, enhancing model adaptability and expressivity. Specifically, we design a lateral connection framework comprising a learnable lateral matrix and a lateral mapping function, both implemented using spiking neurons, to dynamically update lateral connections. Through mathematical modeling, we establish that the proposed lateral update mechanism, under a well-defined local objective, aligns with biologically plausible synaptic plasticity principles. Extensive experiments validate the effectiveness of our approach, analyzing the role of substructure selection and lateral connection during training. Furthermore, quantitative comparisons demonstrate that our model consistently surpasses state-of-the-art SNN-based generative models across multiple benchmark datasets.

Spiking Neural Networks (SNNs) offer a promising direction for energy-efficient and brain-inspired computing, yet their vulnerability to adversarial perturbations remains poorly understood. In this work, we revisit the adversarial robustness of SNNs through the lens of temporal ensembling, treating the network as a collection of evolving sub-networks across discrete timesteps. This formulation uncovers two critical but underexplored challenges-the fragility of individual temporal sub-networks and the tendency for adversarial vulnerabilities to transfer across time. To overcome these limitations, we propose Robust Temporal self-Ensemble (RTE), a training framework that improves the robustness of each sub-network while reducing the temporal transferability of adversarial perturbations. RTE integrates both objectives into a unified loss and employs a stochastic sampling strategy for efficient optimization. Extensive experiments across multiple benchmarks demonstrate that RTE consistently outperforms existing training methods in robust-accuracy trade-off. Additional analyses reveal that RTE reshapes the internal robustness landscape of SNNs, leading to more resilient and temporally diversified decision boundaries. Our study highlights the importance of temporal structure in adversarial learning and offers a principled foundation for building robust spiking models.

Multimodal learning enhances the perceptual ability of intelligent systems by integrating information across sensory modalities. However, most artificial intelligence approaches still rely on static fusion schemes and do not account for the dynamic nature of multisensory integration observed in the brain. In biological systems, the principle of inverse effectiveness states that weaker unimodal cues contribute more strongly to multisensory integration, whereas strong cues reduce the relative benefit of fusion. Motivated by this mechanism, we investigate the relationship between multimodal outputs and modality-specific information and introduce an inverse effectiveness-driven multimodal fusion (IEMF) strategy. Integrating IEMF into neural networks yields more adaptive fusion behavior, improved accuracy, and substantial gains in computational efficiency. We validate the generality of IEMF across audiovisual perception, vision-language understanding, and trimodal sentiment analysis, demonstrating consistent improvements over state-of-the-art baselines in classification, continual learning, and question answering. The approach also transfers across architectures, including both artificial neural networks and spiking neural networks.

On-policy self-distillation, where a student is pulled toward a copy of itself conditioned on privileged context (e.g., a verified solution or feedback), offers a promising direction for advancing reasoning capability without a stronger external teacher. Yet in math reasoning the gains are inconsistent, even when the same approach succeeds elsewhere. A pointwise mutual information analysis traces the failure to the privileged context itself: it inflates the teacher's confidence on tokens already implied by the solution (structural connectives, verifiable claims) and deflates it on deliberation tokens ("Wait", "Let", "Maybe") that drive multi-step search. We propose Anti-Self-Distillation (AntiSD), which ascends a divergence between student and teacher rather than descending it: this reverses the per-token sign and yields a naturally bounded advantage in one step. An entropy-triggered gate disables the term once the teacher entropy collapses, completing a drop-in replacement for default self-distillation. Across five models from 4B to 30B parameters on math reasoning benchmarks, AntiSD reaches the GRPO baseline's accuracy in 2 to 10x fewer training steps and improves final accuracy by up to 11.5 points. AntiSD opens a path to scalable self-improvement, where a language model bootstraps its own reasoning through its training signal.

Large Language Models (LLMs) have achieved remarkable success across a wide range of applications. However, individual LLMs often produce inconsistent, biased, or hallucinated outputs due to limitations in their training corpora and model architectures. Recently, collaborative frameworks such as the Multi-LLM Network (MultiLLMN) have been introduced, enabling multiple LLMs to interact and jointly respond to user queries. Nevertheless, MultiLLMN architectures raise critical concerns regarding the reliability and security of the generated content, particularly in open environments where malicious or compromised LLMs may be present. Moreover, reliance on centralized coordination undermines system efficiency and introduces single points of failure. In this paper, we propose a novel Trusted MultiLLMN framework, driven by a Weighted Byzantine Fault Tolerance (WBFT) blockchain consensus mechanism, to ensure the reliability, security, and efficiency of multi-LLM collaboration. In WBFT, voting weights are adaptively assigned to each LLM based on its response quality and trustworthiness, incentivizing reliable behavior, and reducing the impact of malicious nodes. Extensive simulations demonstrate that WBFT significantly improves both consensus security and efficiency compared to classical and modern consensus mechanisms, particularly under wireless network conditions. Furthermore, our evaluations reveal that Trusted MultiLLMN supported by WBFT can deliver higher-quality and more credible responses than both single LLMs and conventional MultiLLMNs, thereby providing a promising path toward building robust, decentralized AI collaboration networks.

The safety of large language models (LLMs) has increasingly emerged as a fundamental aspect of their development. Existing safety alignment for LLMs is predominantly achieved through post-training methods, which are computationally expensive and often fail to generalize well across different models. A small number of lightweight alignment approaches either rely heavily on prior-computed safety injections or depend excessively on the model's own capabilities, resulting in limited generalization and degraded efficiency and usability during generation. In this work, we propose a safety-aware decoding method that requires only low-cost training of an expert model and employs a single neuron as a gating mechanism. By effectively balancing the model's intrinsic capabilities with external guidance, our approach simultaneously preserves utility and enhances output safety. It demonstrates clear advantages in training overhead and generalization across model scales, offering a new perspective on lightweight alignment for the safe and practical deployment of large language models. Code: https://github.com/Beijing-AISI/NGSD.

The alignment of large language models (LLMs) with human values is critical for their safe and effective deployment across diverse user populations. However, existing benchmarks often neglect cultural and demographic diversity, leading to limited understanding of how value alignment generalizes globally. In this work, we introduce DiverValue-Bench, a benchmark that systematically evaluates LLMs’ alignment with multi-dimensional human value preferences across 74 countries/regions. DiverValue-Bench contains 23,763 high-quality instances annotated with fine-grained value labels, personalized questions, and rich demographic metadata, providing broad demographic and geographic coverage for population-aware value-alignment evaluation. Using DiverValue-Bench, we conduct an in-depth analysis of several representative LLMs, revealing substantial disparities in alignment performance across geographic and demographic lines. We further demonstrate that lightweight fine-tuning methods, such as Low-Rank Adaptation (LoRA) and Direct Preference Optimization (DPO), can significantly enhance value alignment in both in-domain and out-of-domain settings. Our findings underscore the necessity for population-aware alignment evaluation and provide actionable insights for building culturally adaptive and value-sensitive LLMs. DiverValue-Bench serves as a practical foundation for future research on global alignment, personalized value modeling, and equitable AI development.

Human beings often experience stress, which can significantly influence their performance. This study explores whether Large Language Models (LLMs) exhibit stress responses similar to those of humans and whether their performance fluctuates under different stress-inducing prompts. To investigate this, we developed a novel set of prompts, termed StressPrompt, designed to induce varying levels of stress. These prompts were derived from established psychological frameworks and carefully calibrated based on ratings from human participants. We then applied these prompts to several LLMs to assess their responses across a range of tasks, including instruction-following, complex reasoning, and emotional intelligence. The findings suggest that LLMs, like humans, perform optimally under moderate stress, consistent with the Yerkes-Dodson law. Notably, their performance declines under both low and high-stress conditions. Our analysis further revealed that these StressPrompts significantly alter the internal states of LLMs, leading to changes in their neural representations that mirror human responses to stress. This research provides critical insights into the operational robustness and flexibility of LLMs, demonstrating the importance of designing AI systems capable of maintaining high performance in real-world scenarios where stress is prevalent, such as in customer service, healthcare, and emergency response contexts. Moreover, this study contributes to the broader AI research community by offering a new perspective on how LLMs handle different scenarios and their similarities to human cognition.

The hierarchical architecture has become a mainstream design paradigm for Vision Transformers (ViTs), with Patch Merging serving as the pivotal component that transforms a columnar architecture into a hierarchical one. Drawing inspiration from the brain's ability to integrate global and local information for comprehensive visual understanding, we propose Stepwise Patch Merging (SPM), which enhances the subsequent attention mechanism's ability to 'see' better. SPM consists of Multi-Scale Aggregation (MSA) and Guided Local Enhancement (GLE) striking a proper balance between long-range dependency modeling and local feature enhancement. Extensive experiments conducted on benchmark datasets, including ImageNet-1K, COCO, and ADE20K, demonstrate that SPM significantly improves the performance of various models, particularly in dense prediction tasks such as object detection and semantic segmentation. Meanwhile, experiments show that combining SPM with different backbones can further improve performance. The code has been released at https://github.com/Yonghao-Yu/StepwisePatchMerging.

Deploying large language models (LLMs) on embedded devices remains a significant research challenge due to the high computational and memory demands of LLMs and the limited hardware resources available in such environments. While embedded FPGAs have demonstrated performance and energy efficiency in traditional deep neural networks, their potential for LLM inference remains largely unexplored. Recent efforts to deploy LLMs on FPGAs have primarily relied on large, expensive cloud-grade hardware and have only shown promising results on relatively small LLMs, limiting their real-world applicability. In this work, we present Hummingbird, a novel FPGA accelerator designed specifically for LLM inference on embedded FPGAs. Hummingbird is smaller, targeting embedded FPGAs such as the KV260 and ZCU104 with 67