Searcharxiv⌕ Search

arXiv subjects

Guoqi Li

Publications and source records attributed to Guoqi Li.

At least 19 recordsLinked to original sources
arXiv2026

Modern Transformers Are Implicit Hybrids: From Functional Differentiation to Principled Hybrid Architecture Design

Hybrid architectures combining Full Attention (FA) and Linear Attention (LA) are increasingly prominent, yet their allocation remains heuristic. We seek an evidence-grounded basis in head-level functional organization learned by RoPE-based Transformers. Behavioral probes do not yield a complete taxonomy, so we propose two intervention metrics: RoPE Frequency Importance Score (RFIS), measuring how each frequency affects a head's attention distribution, and RoPE Positional Dependence (RPD), isolating dependence on rotary positional modulation. On Qwen3-series models and Llama3.1, RFIS suggests and RPD verifies a complete taxonomy of retrieval and positional heads separated by a salient mid-low-frequency band. Controlled Transformers show that this boundary follows the training-length positional scale; we term it the Global Positional Band (GPBand). The analysis suggests a potential cause of zero-shot length-extrapolation failure and yields two principles: positional modeling should operate only locally, with global access through position-independent retrieval; and both functions should be assigned at head granularity with layer-specific allocation. We instantiate them in Head-wise Hybrid Architecture (HwH), using NoPE FA for global retrieval and LA for local positional modeling. With an FA-to-LA ratio below 1:3, HwH retains strong language modeling and commonsense reasoning while improving retrieval and substantially strengthening zero-shot long-context extrapolation over Transformer, LA, and a layer-wise hybrid baseline. Ablations validate both principles and component roles, highlighting principled hybrid architecture design as a promising route toward future foundation models.

cs.LG↗
arXiv2026

LoGo: Token-Level Dynamic Local-Global Attention

As context lengths scale, attention increasingly becomes a primary computational bottleneck in large language models. Standard Transformers remain powerful but computationally inefficient, as they allocate the same attention budget to every token regardless of its contextual demand. Existing local-global hybrids provide a more efficient alternative by mixing restricted- and full-context attention, but they typically allocate span statically across layers or heads. To address these limitations, we propose LoGo, a token-level dynamic local-global attention mechanism that uses attention span as a direct proxy for attention budget allocation. Each LoGo layer contains coupled local and global branches: all tokens receive efficient local attention over a restricted context window, while a learned gate activates global attention with full-context access only for tokens requiring long-range information. A threshold-based budget controller maintains a target global ratio without auxiliary losses, and a progressive masking schedule stabilizes training before sparse routing takes effect. We further implement query-sparse Triton kernels that convert reduced global-attention computation into practical speedups. Extensive experiments validate LoGo's effectiveness, showing that it preserves the scaling behavior of full-attention Transformers across model sizes. In controlled comparisons, LoGo improves over the full-attention Transformer and matched-budget static local-global hybrids, with clear gains on long-range retrieval. Analysis further shows that LoGo learns interpretable span allocation patterns. These results suggest that learned token-level span allocation is an effective and scalable way to improve the long-context performance-compute trade-off.

cs.CL↗
arXiv2026

GeoMFD: Continual Drone-View Geo-Localization with Geometry-Aware Adapter and Margin-Field Distillation

Existing drone-view geo-localization (DVGL) methods are mainly developed under a static training paradigm, where models are optimized for fixed environments with all training data available in advance. However, this paradigm is difficult to extend to real-world deployment, where drones may encounter diverse environments and require multiple environment-specific models, resulting in additional storage and model-selection costs. Directly adapting a single model to new environments also risks distorting previously learned cross-view embedding geometry and causing forgetting. To address these challenges, we formalize the continual drone-view geo-localization (C-DVGL) setting and propose GeoMFD, a geometry-aware continual adaptation method for DVGL. GeoMFD combines a cold-start bootstrapping strategy (CBS), a geometry-aware adapter (Geo-Adapter), and margin-field distillation (MFD) to balance adaptation and cross-view geometry preservation. CBS initializes a stable embedding space, Geo-Adapter enables environment adaptation through controlled residual corrections, and MFD preserves similarity margins between positive pairs and hard negatives to alleviate cross-view geometry forgetting. Extensive experiments demonstrate that GeoMFD effectively mitigates forgetting and achieves competitive performance with environment-specific DVGL methods using a single continuously updated model.

cs.CV↗
arXiv2026

Neuromorphic Object Detection: An In-Depth Study and Future Directions

Conventional frame-based cameras face significant challenges in detecting objects under high-speed motion blur or in low-light environments. Neuromorphic cameras provide asynchronous visual streams with high temporal resolution and a wide dynamic range, offering a promising solution for object detection under challenging conditions. Despite the development of numerous models and the emergence of various applications in neuromorphic object detection, there is still a lack of deep understanding and standardized benchmarks to assess progress and address key challenges. In this paper, we provide a comprehensive survey and benchmark of existing neuromorphic object detection algorithms. Specifically, we first present a problem description, review the available datasets, and revisit the evaluation metrics. We then explore existing neuromorphic object detection approaches from various perspectives, including event representation, temporal modeling, multimodal fusion, asynchronous processing, low-latency processing, and energy-efficient computing. Furthermore, we evaluate a wide range of representative neuromorphic object detection models and offer detailed analyses of the comparative results. Finally, we discuss unresolved issues in neuromorphic object detection and propose potential future research directions. We hope this survey and benchmark will be a valuable resource for researchers and provide guidance for future advancements in neuromorphic object detection.

cs.CV↗
arXiv2026

Beyond Transformers: Linear Attention Policy for Open-Vocabulary Object Goal Navigation

Open-Vocabulary Object Goal Navigation (OVON) requires agents to operate under partial observability, making effective internal state updates critical for navigation performance. This update is implemented by the policy network, where recent approaches adopt Transformer-based backbones with self-attention over a context window to integrate temporal information. However, our controlled experiments show that performance does not scale with context length under Transformer-based policies, questioning the suitability of self-attention for state integration in navigation. To this end, we propose Linear Attention-based Navigation (LANav), which adopts linear attention (LA) as the policy backbone to maintain a structured state update rather than self-attention over the context window. Across multiple LA variants evaluated under identical settings, LANav consistently outperforms Transformer-based baselines. Performance improves as state update mechanisms become more structured and regulated, highlighting the importance of state update design. To improve state update effectiveness, we introduce Weighted State-Expansion Linear Attention (WSLA), which expands each attention head's state into multiple sub-states and uses learnable weighted readout to aggregate expanded sub-states. Equipped with WSLA, LANav achieves 36.4% average success rate (SR) on HM3D-OVON, outperforming Transformer-based counterparts by 6.3 percentage points in macro-averaged SR, while maintaining computational efficiency. Distance-stratified results show larger gains in long-distance episodes, while HSSD transfer and fine-tuning demonstrate robustness across scene distributions. Real-world deployment on a Unitree Go2 further achieves an 82% success rate over 50 trials, supporting the practical feasibility and sim-to-real transfer of LANav.

cs.RO↗
arXiv2026

Spiking Neural Networks for fMRI-Based Visual Semantic Decoding

Functional magnetic resonance imaging (fMRI)-based visual decoding aims to recover visual information from measured brain activity, commonly by mapping fMRI responses into latent visual features for downstream decoding tasks. Most existing methods learn mappings from fMRI responses to visual features extracted by artificial neural networks (ANNs), yet it remains unclear whether ANN-derived features provide suitable targets for brain decoding. In this study, we investigate spiking neural network (SNN)-derived visual features as alternative targets for fMRI-based visual decoding. We compare an ANN baseline with four SNN variants from the same architectural family, which differ in their spiking dynamics. To isolate the effect of the target features, all models use the same L2-regularized linear fMRI-to-feature decoder, while only the feature vectors used as regression targets are varied. Compared with the ANN baseline, SNN-derived features exhibit stronger alignment with fMRI responses and improve visual semantic decoding performance. For instance, on the GoD dataset, SNN-derived features reduce feature-prediction error from 0.7707 to 0.0282 and improve top-1 semantic decoding accuracy from 0.1800 to 0.4400. Ablation results further indicate that both spiking neural dynamics and temporal simulation steps contribute to the observed advantage. These findings support SNN-derived features as effective brain-decodable visual representations and highlight target feature design as an important component of fMRI-based visual decoding.

cs.NE↗
arXiv2026

Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention

Recent NVFP4 pretraining work has primarily optimized Transformer linear projections, leaving persistent optimizer states, optimizer computation, and low-precision attention forward--backward paths less explored. We present \textbf{Full-Stack FP4}, a modular NVFP4 framework with separate recipes for projections, AdamW states, Root/Muon computation, and attention. \textbf{LoRA-SVD} protects a compact projection subspace in BF16 while retaining full-shape NVFP4 computation, reducing the linear-only loss gap from \textbf{1.40\%} to \textbf{0.61\%}. An ordered square-root, tile-mean, and Hadamard pipeline enables stable NVFP4 AdamW momentum storage; shape-dependent coefficients and clipping stabilize direct NVFP4 Root iterations; and mixed-precision attention retains softmax-sensitive operations in BF16. On 3B pretraining with 64B tokens, BF16 and Full-Stack FP4 reach losses of \textbf{2.267} and \textbf{2.286}, a \textbf{0.838\%} gap. Their average zero-shot perplexities are 26.675 and 26.665, respectively, with Full-Stack FP4 averaging 0.10 percentage points lower in accuracy. Native four-block measurements on one RTX 5090 show 2.50--2.83$\times$ Root speedups over optimized BF16 and 37.9--42.5\% lower AdamW peak memory.

cs.LG↗
arXiv2026

Burst Spiking Neural Networks

A central goal of current Spiking Neural Network (SNN) research is to improve their accuracy toward becoming low-power alternatives to Artificial Neural Networks (ANNs). This work further argues that realizing this ambition requires improving not only accuracy but also robustness, defined as the ability to maintain correct predictions under input perturbations. We identify two key issues in existing SNN methods that undermine robustness. First, binary spiking activations can produce large activation-state changes under small perturbations. Second, the lack of effective weight constraints makes network outputs more sensitive to input variations. To this end, we propose Burst Spiking Neural Networks (BuSNNs), built upon Burst-enhanced Spiking Neurons (BSNs) and a Dynamic Weight Constraint (DWC) mechanism. BSNs incorporate burst firing to provide a graded spiking pattern. This spiking mechanism mitigates perturbation-induced transitions in activation states and thereby enhances robustness. DWC penalizes connection weights based on activation states, effectively reducing weight magnitudes and improving robustness while preserving accuracy. We provide theoretical analyses to support these robustness effects. Experimental results further show that, on smaller-scale benchmarks such as CIFAR-10, BuSNNs outperform both SNN and ANN counterparts in accuracy and robustness. On large-scale ImageNet, BuSNN with the MS ResNet-34 backbone further improves top-1 accuracy and corruption robustness over the corresponding SNN baseline by 3.18% and 2.66%, respectively. Despite using spike-based activations, BuSNNs surpass 4-bit activation-quantized ANN baselines and approach 8-bit ANN baselines on ImageNet. They also preserve SNNs' low-power advantage. This work studies the accuracy-robustness problem in SNNs, advancing their practical viability in robust and energy-efficient applications.

cs.NE↗
arXiv2026

NeuroCogMap Reveals Cognitive Organization of Large Language Models

Understanding how complex cognitive functions are organized within artificial systems is central to interpreting large language models (LLMs) and relating them to biological cognition. Yet although LLMs exhibit broad cognitive-like behaviours, it remains unclear whether their internal representations form reproducible functional systems that explain behaviour, failure and links to human cognition. Here we present NeuroCogMap, a cognitive neuroscience-inspired framework that organizes internal features of LLMs into functional parcels and links them to interpretable functions, cognitive capabilities and a cognitive hierarchy. These parcels form a stable and semantically coherent organization that is partly conserved across models and functionally linked to model outputs. Within this organization, major LLM failures, including hallucination, bias, refusal failure and sycophancy, correspond to distinct disruptions in representational and behavioural-control systems, yielding internal signatures for mechanism-guided detection and targeted intervention. Beyond model behaviour, NeuroCogMap improves prediction of human cortical responses during naturalistic language comprehension, with the strongest correspondence in higher-order association cortex. At the cognitive level, its internal signatures expose latent strategies that guide refinements of classical models of human decision-making. Together, these findings establish NeuroCogMap as a system-level framework for mapping functional organization in artificial systems and for relating this organization to human cortical function and cognitive behaviour.

q-bio.NC↗
arXiv2026

MindAdapter: Few-Shot Parameter-Efficient Residual Calibration of Cross-Subject Brain-to-Visual Decoding Models

Cross-subject brain-to-visual decoding remains a core challenge in brain-computer interfaces due to severe inter-individual variability that induces systematic subject-specific functional misalignment. To address this issue, we propose MindAdapter, a parameter-efficient few-shot calibration framework for pretrained brain-to-visual decoding models. MindAdapter adopts a decoupled linear-residual cascade alignment paradigm by freezing a pretrained explicit brain functional alignment backbone (coarse) and introducing a lightweight nonlinear residual adapter (fine), thereby disentangling global cross-subject correspondence from subject-specific residual corrections for fine-grained spatial and semantic calibration. To further preserve global representational stability, we design a topology-anchored dual-stream manifold constraint, where a small set of shared stimuli serves as topological pins with voxel-level paired supervision, while a semantic stream enforces consistency through a frozen vision-language decoder on unpaired brain data. Together, MindAdapter efficiently injects subject-specific corrections while maintaining the global representational geometry learned during pretraining. Experiments on the Natural Scenes Dataset (NSD) demonstrate that MindAdapter substantially improves cross-subject visual reconstruction and retrieval accuracy using only a few shared stimuli, offering a practical and data-efficient solution for personalized brain-to-visual decoding.

cs.CV↗
arXiv2026

SpikingBrain2.0: Brain-Inspired Foundation Models for Efficient Long-Context and Cross-Platform Inference

Scaling context length is reshaping large-model development, yet full-attention Transformers suffer from prohibitive computation and inference bottlenecks at long sequences. A key challenge is to design foundation models that maintain performance and long-context efficiency with minimal training overhead. We introduce SpikingBrain2.0 (SpB2.0), a 5B model that advances both architecture and training efficiency of its predecessor. Our contributions are two-fold. (1) Architectural Innovation: We propose Dual-Space Sparse Attention (DSSA), an inter-layer hybrid of Sparse Softmax Attention (MoBA) and Sparse Linear Attention (SSE), achieving an improved performance-efficiency trade-off for long-context modeling. SpB2.0 further supports dual quantization paths: INT8-Spiking coding enables sparse event-driven computation, while FP8 coding accelerates inference on modern GPUs. (2) Enhanced Training Strategy: We develop an optimized Transformer-to-Hybrid (T2H) pipeline with dual conversion paths for LLMs and VLMs using curated open-source data. Empirically, SpB2.0-5B and SpB2.0-VL-5B recover most of the base Transformer (Qwen3-4B) capability with under 7k A100 GPU hours. SpB2.0 achieves a 10.13x TTFT speedup at 4M context and supports over 10M tokens on 8 A100 GPUs under vLLM, where full-attention models exceed memory limits. It also demonstrates strong cross-platform compatibility, enabling FP8 GPU inference (2.52x speedup at 250k) and efficient neuromorphic execution (64.31% sparsity, with 70.6% and 46.5% area and power reduction at 500MHz). Overall, SpikingBrain2.0 provides a practical pathway for lightweight, multimodal, spiking foundation models, highlighting the potential of combining brain-inspired mechanisms with efficient architectures for resource-constrained and edge scenarios.

cs.LG↗
arXiv2026

SpikeMLLM: Spike-based Multimodal Large Language Models via Modality-Specific Temporal Scales and Temporal Compression

Multimodal Large Language Models (MLLMs) have achieved remarkable progress but incur substantial computational overhead and energy consumption during inference, limiting deployment in resource-constrained environments. Spiking Neural Networks (SNNs), with their sparse event-driven computation, offer inherent energy efficiency advantages on neuromorphic hardware, yet extending them to MLLMs faces two key challenges: heterogeneous modalities make uniform spike encoding insufficient, and high-resolution image inputs amplify timestep unfolding overhead. We propose SpikeMLLM, the first spike-based framework for MLLMs, which unifies existing ANN quantization methods in the spiking representation space and incorporates Modality-Specific Temporal Scales (MSTS) guided by Modality Evolution Discrepancy (MED) and Temporally Compressed LIF (TC-LIF) for timestep compression from T=L-1 to T=log2(L)-1. Experiments on four representative MLLMs across diverse multimodal benchmarks show that SpikeMLLM maintains near-lossless performance under aggressive timestep compression (Tv/Tt=3/4), with average gaps of only 0.72% and 1.19% relative to the FP16 baseline on InternVL2-8B and Qwen2VL-72B. We further develop a dedicated RTL accelerator tailored to the spike-driven datapath, observing 9.06x higher throughput and 25.8x better power efficiency relative to an FP16 GPU baseline under a deployment-oriented co-design setting, suggesting the promise of algorithm-hardware co-design for efficient multimodal intelligence.

cs.NE↗
arXiv2026

Spike-driven Large Language Model

Current Large Language Models (LLMs) are primarily based on large-scale dense matrix multiplications. Inspired by the brain's information processing mechanism, we explore the fundamental question: how to effectively integrate the brain's spiking-driven characteristics into LLM inference. Spiking Neural Networks (SNNs) possess spike-driven characteristics, and some works have attempted to combine SNNs with Transformers. However, achieving spike-driven LLMs with billions of parameters, relying solely on sparse additions, remains a challenge in the SNN field. To address the issues of limited representational capacity and sparsity in existing spike encoding schemes at the LLM level, we propose SDLLM, a spike-driven large language model that eliminates dense matrix multiplications through sparse addition operations. Specifically, we use the plug-and-play gamma-SQP two-step spike encoding method to ensure that the quantization process aligns with the model's semantic space, mitigating representation degradation caused by binary spikes. Furthermore, we introduce bidirectional encoding under symmetric quantization and membrane potential clipping mechanisms, leading to spike trains with no or low firing counts dominating, significantly reducing the model's spike firing rate, while halving the number of time steps. Experimental results show that SDLLM not only significantly reduces inference costs but also achieves state-of-the-art task performance under the spike-based paradigm. For example, compared to previous spike-based LLMs, SDLLM reduces energy consumption by 7x and improves accuracy by 4.2%. Our model provides inspiration for the architecture design of the next generation of event-driven neuromorphic chips.

cs.NE↗
arXiv2026

Efficient Sparse Selective-Update RNNs for Long-Range Sequence Modeling

Real-world sequential signals, such as audio or video, contain critical information that is often embedded within long periods of silence or noise. While recurrent neural networks (RNNs) are designed to process such data efficiently, they often suffer from ``memory decay'' due to a rigid update schedule: they typically update their internal state at every time step, even when the input is static. This constant activity forces the model to overwrite its own memory and makes it hard for the learning signal to reach back to distant past events. Here we show that we can overcome this limitation using Selective-Update RNNs (suRNNs), a non-linear architecture that learns to preserve its memory when the input is redundant. By using a neuron-level binary switch that only opens for informative events, suRNNs decouple the recurrent updates from the raw sequence length. This mechanism allows the model to maintain an exact, unchanged memory of the past during low-information intervals, creating a direct path for gradients to flow across time. Our experiments on the Long Range Arena, WikiText, and other synthetic benchmarks show that suRNNs match or exceed the accuracy of much more complex models such as Transformers, while remaining significantly more efficient for long-term storage. By allowing each neuron to learn its own update timescale, our approach resolves the mismatch between how long a sequence is and how much information it actually contains. By providing a principled approach to managing temporal information density, this work establishes a new direction for achieving Transformer-level performance within the highly efficient framework of recurrent modeling.

cs.LG↗
arXiv2026

Parallel Training in Spiking Neural Networks

The bio-inspired integrate-fire-reset mechanism of spiking neurons constitutes the foundation for efficient processing in Spiking Neural Networks (SNNs). Recent progress in large models demands that spiking neurons support highly parallel computation to scale efficiently on modern GPUs. This work proposes a novel functional perspective that provides general guidance for designing parallel spiking neurons. We argue that the reset mechanism, which induces complex temporal dependencies and hinders parallel training, should be removed. However, any such modification should satisfy two principles: 1) preserving the functions of reset as a core biological mechanism; and 2) enabling parallel training without sacrificing the serial inference ability of spiking neurons, which underpins their efficiency at test time. To this end, we identify the functions of the reset and analyze how to reconcile parallel training with serial inference, upon which we propose a dynamic decay spiking neuron. We conduct comprehensive testing of our method in terms of: 1) Training efficiency and extrapolation capability. On 16k-length sequences, we achieve a 25.6x training speedup over the pioneering parallel spiking neuron, and our models trained on 2k-length can stably perform inference on sequences as long as 30k. 2) Generality. We demonstrate the consistent effectiveness of the proposed method across five task categories (image classification, neuromorphic event processing, time-series forecasting, language modeling, and reinforcement learning), three network architectures (spiking CNN/Transformer/SSMs), and two spike activation modes (spike/integer activation). 3) Energy consumption. The spiking firing of our neuron is lower than that of vanilla and existing parallel spiking neurons.

cs.NE↗
arXiv2026

BrainFuse: a unified infrastructure integrating realistic biological modeling and core AI methodology

Neuroscience and artificial intelligence represent distinct yet complementary pathways to general intelligence. However, amid the ongoing boom in AI research and applications, the translational synergy between these two fields has grown increasingly elusive-hampered by a widening infrastructural incompatibility: modern AI frameworks lack native support for biophysical realism, while neural simulation tools are poorly suited for gradient-based optimization and neuromorphic hardware deployment. To bridge this gap, we introduce BrainFuse, a unified infrastructure that provides comprehensive support for biophysical neural simulation and gradient-based learning. By addressing algorithmic, computational, and deployment challenges, BrainFuse exhibits three core capabilities: (1) algorithmic integration of detailed neuronal dynamics into a differentiable learning framework; (2) system-level optimization that accelerates customizable ion-channel dynamics by up to 3,000x on GPUs; and (3) scalable computation with highly compatible pipelines for neuromorphic hardware deployment. We demonstrate this full-stack design through both AI and neuroscience tasks, from foundational neuron simulation and functional cylinder modeling to real-world deployment and application scenarios. For neuroscience, BrainFuse supports multiscale biological modeling, enabling the deployment of approximately 38,000 Hodgkin-Huxley neurons with 100 million synapses on a single neuromorphic chip while consuming as low as 1.98 W. For AI, BrainFuse facilitates the synergistic application of realistic biological neuron models, demonstrating enhanced robustness to input noise and improved temporal processing endowed by complex HH dynamics. BrainFuse therefore serves as a foundational engine to facilitate cross-disciplinary research and accelerate the development of next-generation bio-inspired intelligent systems.

cs.NE↗
arXiv2025

Efficient Spike-driven Transformer for High-performance Drone-View Geo-Localization

Traditional drone-view geo-localization (DVGL) methods based on artificial neural networks (ANNs) have achieved remarkable performance. However, ANNs rely on dense computation, which results in high power consumption. In contrast, spiking neural networks (SNNs), which benefit from spike-driven computation, inherently provide low power consumption. Regrettably, the potential of SNNs for DVGL has yet to be thoroughly investigated. Meanwhile, the inherent sparsity of spike-driven computation for representation learning scenarios also results in loss of critical information and difficulties in learning long-range dependencies when aligning heterogeneous visual data sources. To address these, we propose SpikeViMFormer, the first SNN framework designed for DVGL. In this framework, a lightweight spike-driven transformer backbone is adopted to extract coarse-grained features. To mitigate the loss of critical information, the spike-driven selective attention (SSA) block is designed, which uses a spike-driven gating mechanism to achieve selective feature enhancement and highlight discriminative regions. Furthermore, a spike-driven hybrid state space (SHS) block is introduced to learn long-range dependencies using a hybrid state space. Moreover, only the backbone is utilized during the inference stage to reduce computational cost. To ensure backbone effectiveness, a novel hierarchical re-ranking alignment learning (HRAL) strategy is proposed. It refines features via neighborhood re-ranking and maintains cross-batch consistency to directly optimize the backbone. Experimental results demonstrate that SpikeViMFormer outperforms state-of-the-art SNNs. Compared with advanced ANNs, it also achieves competitive performance.Our code is available at https://github.com/ISChenawei/SpikeViMFormer

cs.CV↗
arXiv2025

PretrainZero: Reinforcement Active Pretraining

Mimicking human behavior to actively learning from general experience and achieve artificial general intelligence has always been a human dream. Recent reinforcement learning (RL) based large-thinking models demonstrate impressive expert-level abilities, i.e., software and math, but still rely heavily on verifiable rewards in specific domains, placing a significant bottleneck to extend the performance boundary of general reasoning capabilities. In this work, we propose PretrainZero, a reinforcement active learning framework built on the pretraining corpus to extend RL from domain-specific post-training to general pretraining. PretrainZero features the following characteristics: 1) Active pretraining: inspired by the active learning ability of humans, PretrainZero learns a unified reasoning policy to actively identify reasonable and informative contents from pretraining corpus, and reason to predict these contents by RL. 2) Self-supervised learning: without any verifiable labels, pretrained reward models, or supervised fine-tuning, we directly pretrain reasoners from 3 to 30B base models on the general Wikipedia corpus using RL, significantly breaking the verification data-wall for general reasoning. 3) Verification scaling: by tackling increasingly challenging masked spans, PretrainZero substantially enhances the general reasoning abilities of pretrained base models. In reinforcement pretraining, PretrainZero improves Qwen3-4B-Base for 8.43, 5.96 and 10.60 on MMLU-Pro, SuperGPQA and math average benchmarks. In post-training, the pretrained models can also serve as reasoning foundation models for downstream RLVR tasks.

cs.CL↗