Searcharxiv⌕ Search

arXiv subjects

Wenhe Zhang

Publications and source records attributed to Wenhe Zhang.

6 recordsLinked to original sources

Lost in a Few Heads: Diagnosing and Repairing Block Selection in Sparse Attention

Sparse attention models read a few blocks of keys for every query at a fixed budget and can lose much of their long-context retrieval accuracy at that budget. We ask where this loss is made. Using exact attention mass selection at the shipped budget as an oracle, we show that the selector under-ranks the needle rather than running out of budget. We find attention heads that miss the needle by running a calibration probe that teacher-forces the answer and reads two quantities in every head. The first is the needle attention that the selector forfeits, and the second is the share of the answer's log-probability that fixing the head's selection restores, integrated along the path from the shipped selector to full attention. Four checkpoints of the InfLLM-V2 family and NOSA-8B lose multi-key retrieval in the same three blind heads, which attend to the needle while a mean-pooled key summary dilutes its single decisive token. Full attention in those heads or a 4-bit scorer that keeps the shipped budget recovers most of the loss without training. On MiniCPM4.1-8B at 64K, either fix lifts multi-key retrieval accuracy from 54.1 to 97.5%, level with full attention in every head. In production serving engines, full attention in those heads keeps 84 to 89% of the time to first token (TTFT) that sparse attention saves at 32K to 128K context length. Beyond the InfLLM-V2 selector lineage, one unified rule flags attention heads above a threshold swept on each checkpoint until they close most of the gap to full attention on calibration prompts. A similar small set of blind attention heads carry the loss for the SeerAttention-R gates, NSA, and DashAttention, whereas many such heads exist in MoBA. On Qwen3-8B with the SeerAttention-R gates, full attention in 18 of 288 key-value heads lifts multi-key retrieval accuracy from 29.8 to 97.7%.

cs.LG↗

WirelessMathBench-XL: An Auditable Benchmark for Wireless Mathematical Reasoning

Benchmarks built from arXiv papers can overlap the arXiv text used in LLM pretraining, yet technical-domain releases rarely ship the per-item evidence needed to check this. We introduce a reverse-probe n-gram audit that indexes the benchmark side and streams the corpus past it, so memory scales with the benchmark rather than the corpus and flags on the distributed problems can be inspected and recomputed. We instantiate it in WirelessMathBench-XL, a 4,027-problem benchmark for wireless mathematical reasoning built from 836 arXiv papers across 20 subfields, with multiple-choice, progressive fill-in, and full-equation-completion formats. Streaming 12.9 B RedPajama-arXiv 13-grams yields a strict zero-hit view $\mathcal{S}_0$ covering 95.7% of the problems, and shows that the Dolma and ProofPile-2 arXiv slices repackage the same RedPajama dump. Filtering to $\mathcal{S}_0$ moves every evaluated model by less than 1 pp. With 30/800 flagged test items, this comparison bounds detected-overlap inflation at 0.31-0.51 pp for frontier models rather than testing for contamination, and it does not cover paraphrase, target-answer, alternate-version, post-training, or closed-corpus exposure. Frontier models reach 86.5-91.3%, and GRPO on the released training split lifts a 7B model 25.1 pp above a size-matched general-math control. We release source-paper identifiers, per-problem audit and threshold metadata, license-aware data files, audit and bootstrap scripts, Croissant metadata, and a Datasheet.

cs.LG↗

An Analysis of Streaming Deep Reinforcement Learning for Adaptive Continual Learning in Robotics

Over the course of a lifetime, robots may encounter novel scenarios unaccounted for in its original training that result in performance degradation. One common approach to mitigating this issue is to further grow the offline training dataset in hopes of producing a policy robust to these changes. In contrast, biological learning occurs moment-to-moment via a stream of experience, unlike the predominantly batch-based and offline nature of deep learning. Although recent works show the feasibility of stream-based deep reinforcement learning, where updates use only the latest experience, none have shown it to be a viable continual learning framework for adapting robotic policies to unseen changes. In this paper, we present the first analysis of streaming deep reinforcement learning for adaptive continual learning in robotics. In particular, we show that, following an initial pretraining phase, streaming deep RL can enable a robot to successfully adapt to unforeseen changes to itself, its environment, or goals. Our primary experiments within quadruped locomotion demonstrate that a deep neural network robotic policy with certain optimizers and plasticity loss mitigation techniques can successfully leverage domain task knowledge from its pretraining to quickly adapt online to diverse changes via stream learning, outperforming batch-based on-policy methods and improving task success rates by up to 90% over the pretrained policy. Furthermore, we perform additional evaluations on robotic manipulation tasks to determine if our previous observations extend to different robotic morphologies and scenarios. Our results show that the successes observed in quadruped locomotion can be partially realized in manipulation with stability and performance limitations. We conclude with a discussion on the limitations of our work and its implications for the future of continual robot learning.

cs.RO↗

Elign: Equivariant Diffusion Model Alignment from Foundational Machine Learning Force Fields

Generative models for 3D molecular conformations must respect Euclidean symmetries and concentrate probability mass on thermodynamically favorable, mechanically stable structures. However, E(3)-equivariant diffusion models often reproduce biases from semi-empirical training data rather than capturing the equilibrium distribution of a high-fidelity Hamiltonian. While physics-based guidance can correct this, it faces two computational bottlenecks: expensive quantum-chemical evaluations (e.g., DFT) and the need to repeat such queries at every sampling step. We present Elign, a post-training framework that amortizes both costs. First, we replace expensive DFT evaluations with a faster, pretrained foundational machine-learning force field (MLFF) to provide physical signals. Second, we eliminate repeated run-time queries by shifting physical steering to the training phase. To achieve the second amortization, we formulate reverse diffusion as a reinforcement learning problem and introduce Force--Energy Disentangled Group Relative Policy Optimization (FED-GRPO) to fine-tune the denoising policy. FED-GRPO includes a potential-based energy reward and a force-based stability reward, which are optimized and group-normalized independently. Experiments show that Elign generates conformations with lower gold-standard DFT energies and forces, while improving stability. Crucially, inference remains as fast as unguided sampling, since no energy evaluations are required during generation.

cs.LG↗

Rethinking complexity for software code structures: A pioneering study on Linux kernel code repository

The recent progress of artificial intelligence(AI) has shown great potentials for alleviating human burden in various complex tasks. From the view of software engineering, AI techniques can be seen in many fundamental aspects of development, such as source code comprehension, in which state-of-the-art models are implemented to extract and express the meaning of code snippets automatically. However, such technologies are still struggling to tackle and comprehend the complex structures within industrial code, thus far from real-world applications. In the present work, we built an innovative and systematical framework, emphasizing the problem of complexity in code comprehension and further software engineering. Upon automatic data collection from the latest Linux kernel source code, we modeled code structures as complex networks through token extraction and relation parsing. Comprehensive analysis of complexity further revealed the density and scale of network-based code representations. Our work constructed the first large-scale dataset from industrial-strength software code for downstream software engineering tasks including code comprehension, and incorporated complex network theory into code-level investigations of software development for the first time. In the longer term, the proposed methodology could play significant roles in the entire software engineering process, powering software design, coding, debugging, testing, and sustaining by redefining and embracing complexity.

cs.SE↗

Machine Number Sense: A Dataset of Visual Arithmetic Problems for Abstract and Relational Reasoning

As a comprehensive indicator of mathematical thinking and intelligence, the number sense (Dehaene 2011) bridges the induction of symbolic concepts and the competence of problem-solving. To endow such a crucial cognitive ability to machine intelligence, we propose a dataset, Machine Number Sense (MNS), consisting of visual arithmetic problems automatically generated using a grammar model--And-Or Graph (AOG). These visual arithmetic problems are in the form of geometric figures: each problem has a set of geometric shapes as its context and embedded number symbols. Solving such problems is not trivial; the machine not only has to recognize the number, but also to interpret the number with its contexts, shapes, and relations (e.g., symmetry) together with proper operations. We benchmark the MNS dataset using four predominant neural network models as baselines in this visual reasoning task. Comprehensive experiments show that current neural-network-based models still struggle to understand number concepts and relational operations. We show that a simple brute-force search algorithm could work out some of the problems without context information. Crucially, taking geometric context into account by an additional perception module would provide a sharp performance gain with fewer search steps. Altogether, we call for attention in fusing the classic search-based algorithms with modern neural networks to discover the essential number concepts in future research.

cs.AI↗