Publications
publications by categories in reversed chronological order. * means equal contribution.
2026
- Tencent HYWhen Do Larger Batches Help Scale LLM Reinforcement Learning?Ziniu Li, Jinbo Wang, Guanhua Huang, Feiyuan Zhang, Pengbo Li, and Alex ChenTencent HunYuan Research Blog, 2026
We study when larger batches accelerate LLM reinforcement learning, separating sample efficiency from system throughput and providing practical guidance for batch-size selection and learning-rate retuning.
Larger batches reduce the variance of stochastic gradients per update and are therefore often expected to accelerate training. Yet whether this statistical benefit translates into lower wall-clock time-to-target remains unclear, because each update consumes more samples and may take longer to execute. We study this tradeoff in reinforcement learning for large language models. We separate its algorithmic and systems effects by comparing learning and execution along their natural axes. At the algorithmic level, we compare configurations at equal cumulative sample counts while retuning batch-dependent hyperparameters. Over a bounded range of batch sizes, this procedure yields an approximately batch-size-invariant family whose members follow similar sample-indexed learning trajectories. At the systems level, we exploit the computational asymmetry between rollout generation and training: autoregressive generation is often memory-bandwidth-bound at low concurrency, whereas training work scales approximately with the number of processed tokens. Combining these two views yields a direct decision rule: a larger-batch configuration reduces time-to-target only when its throughput gain exceeds its samples-to-target penalty. Experiments with GRPO and PPO support both sides of this decomposition. At the algorithmic level, square-root learning-rate scaling with Adam produces approximately batch-size-invariant learning curves over a bounded range of batch sizes. At the systems level, larger batches improve generation throughput by up to 2.29x on fixed hardware. In GRPO, combining higher throughput with learning-rate retuning reduces time-to-target by up to 29%, whereas increasing the batch without retuning is slower despite its higher throughput.
- arXivInformation-Optimal Quantum Measurements: Posterior Algebra and the Limits of Orthogonal ReadoutarXiv preprint arXiv:2609.27992, In submission, 2026
We prove Shor’s orthogonal-measurement conjecture, revealing the structure of information-optimal quantum measurements and advancing quantum information theory.
When do orthogonal measurements extract all accessible information? We introduce a posterior algebra for full-rank ensembles. Its commutativity determines whether an information bound is attained; its joint eigenspaces then classify optimal finite-outcome measurements. For two states, we constructively prove Shor’s conjecture in finite dimension, including singular states. Orthogonal measurements can fail at fixed guessing error: a discontinuous basis switch forces some qubit pairs to need three outcomes. A Tsallis information score can require three outcomes even without an error constraint.
- arXivComplementary Quantum Correlations Are Universal for QubitsJinbo Wang, Qihang Wang, and Kun ChenarXiv preprint arXiv:2608.04916, In submission, 2026
We prove complementary correlation bounds for all two-qubit states, enabling tomography-free certification of quantum resources.
Extracting total correlations from a quantum system usually requires reconstructing its state, whereas many experiments access only a few measurement settings. A possible shortcut is to add the mutual informations obtained from complementary measurements; in dimensions above two, however, this procedure can count the same classical correlation twice. We establish that qubits are protected from such overcounting. For every two-qubit state, the correlations observed in two complementary local bases are bounded by the premeasurement quantum mutual information. The proof traces this protection to binary-entropy curvature on the Bloch ball and combines a qubit information-exclusion tradeoff with data processing under local dephasing. Consequently, two correlation tables give a tomography-free lower bound on total correlation. A score above one bit also certifies a quantitative one-way entanglement-distillation rate; when applied to the Choi state of a qubit channel, the same data lower bound its quantum capacity. The theorem therefore identifies both an operational use of complementarity and the trusted two-dimensional setting in which its correlation accounting is valid.
- arXivWhen Complementary Measurements Count the Same Classical Bit Twice: Counterexamples to CQC, ECQC, and Complementarity-Based CertificationJinbo Wang, Qihang Wang, and Kun ChenarXiv preprint arXiv:2608.03828, In submission, 2026
We expose classical double counting in complementary measurements and disprove CQC, ECQC, and related quantum-correlation certification.
Mutually unbiased measurements are commonly expected to expose independent facets of a quantum state: a correlation that is classical in one basis should disappear in a complementary basis. In higher dimensions, however, this intuition becomes particularly subtle because correlations recovered in different settings need not represent different information. To expose this loophole, we propose a two-branch classical null test: before the setting is chosen, a shared bit selects one of two orthogonal product preparations, producing a rank-two classical–classical state, and the candidate protocol then runs unchanged. Different settings can read the same bit through different outcome patterns. This two-branch classical architecture disproves the complementary-quantum correlation (CQC) conjecture in every dimension $d\geq3$. A distinct rank-two classical–classical state disproves its complete-basis extension (ECQC) at $d=7$, with an overrun that grows without bound along prime dimensions. Its qutrit CQC instance also gives classical false positives for a proposed quantum-correlation measure and a proposed one-sided semi-device-independent steering criterion, and refutes a conditional-probability conjecture. The failures identify the missing requirement: information read in different settings must be nonredundant. In experiments and applications, the same low-overhead architecture can serve as a calibration test before a multibasis score is assigned quantum meaning.
- NeurIPSMore Expressive Feedforward Layers: Part I. Token-Adaptive Mixing of ActivationsIn Advances in Neural Information Processing Systems (NeurIPS), Bytedance Seed, 2026
We propose Token-Adaptive Mixing, an architectural enhancement for feedforward layers that adaptively mixes activations based on token characteristics, improving expressiveness and performance in language models.
Feedforward network (FFN) layers account for a large fraction of parameters and nonlinear expressivity in Transformer-based large language models (LLMs). Despite the evolution from ReLU and GELU to gated variants such as SwiGLU, most FFN designs still use a single fixed activation function, applying the same nonlinear transformation to all tokens. In this work, we propose Mixture of Activations (MoA), a token-adaptive FFN design that mixes a dictionary of activation functions using lightweight input-dependent gates while sharing the same linear projections. As an input-independent counterpart, we also introduce learnable activations (LA), which form linear combinations of activation functions for both ReLU-type and SwiGLU-type FFNs. Theoretically, we establish strict finite-width expressive separations among fixed-activation FFNs, LA, and MoA: LA strictly contains fixed-activation FFNs, while MoA strictly contains LA, with the additional expressivity arising from input-dependent nonlinear hybridization. Empirically, we evaluate MoA through extensive pre-training experiments on dense and MoE language models ranging from 0.12B to 2B parameters under different token budgets, optimizers, and learning rate schedules. MoA consistently achieves lower terminal loss and exhibits more favorable scaling behavior than well-tuned baselines, with minimal parameter and computational overhead. These results suggest that token-adaptive activation mixing is a simple and effective mechanism for improving FFN expressivity in LLMs.
- Tencent HYStabilizing RLVR via Token-level Gradient Diagnosis and Layerwise ClippingGuanhua Huang*, Tingqiang Xu*, Jinbo Wang*, Guangming Sheng, Siheng Li, Evander Yang, Kejiao Li, Yunxiang Li, Zenan Xu, Qi Yi, Kyrierl Deng, Ziyuan Nan, Yuhao Jiang, Chenchen Zhang, Taiqiang Wu, Feiyuan Zhang, Junhao Wang, Bo Zhou, Alex Chen, Di Wang, and Shunyu YaoTencent HunYuan Research Blog, 2026
We present GradLoc: transferring RLVR training collapse diagnostic from black-box heuristics to white-box fine-grained token localization via a distributed bisection, paired with layerwise clipping as a practical safeguard.
Frequent gradient spikes and training collapse hinder reinforcement learning with verifiable rewards (RLVR). We introduce the Gradient Anomaly Localizer (GradLoc), a distributed diagnostic tool that traces global gradient spikes to individual tokens through binary search with adaptive thresholds. GradLoc activates only when an anomaly is detected and leaves the optimization objective and parameter updates unchanged. Token-level diagnosis reveals that semantically coherent tokens can have pathological numerical properties. Beyond training-inference mismatch, it identifies severe gradient heterogeneity across layers even when importance sampling ratios remain within a reasonable range. In this setting, a few anomalous layers dominate the global gradient norm, causing global clipping to suppress updates in otherwise healthy layers. We therefore introduce Layerwise Gradient Clipping (LayerClip), which constrains each layer separately using its historical gradient statistics and improves RLVR training stability. Together, GradLoc and LayerClip provide a practical framework for diagnosing and mitigating training collapse.
- NeurIPS SpotlightWhen and How SignSGD Outperforms SGD: A Functional Scaling Law PerspectiveIn Advances in Neural Information Processing Systems (NeurIPS), Spotlight (Top 0.95%). Also presented at ICLR 2026 Sci4DL Workshop, 2026
We develop a scaling law analysis for SignSGD, to understand why it outperforms SGD in large-scale training.
Despite their widespread use in deep learning, the mechanisms underlying the effectiveness of adaptive gradient methods in large-scale training remain poorly understood. In this work, we provide a scaling-law analysis of SignSGD, a minimal yet expressive optimizer that captures the core coordinate-wise adaptivity shared by more sophisticated adaptive methods. We consider feature-space linear regression with power-law spectra, which allows us to precisely characterize the training dynamics of SignSGD. Specifically, we derive explicit scaling laws for SignSGD that accurately describe the loss dynamics. By further analyzing the data-limited regime, we characterize the phase diagram of SignSGD training and quantify the superiority of SignSGD in data scaling. We also show that SignSGD admits a substantially larger critical batch size than SGD, which gives SignSGD more benefits from large-batch training. Finally, we systematically validate our theoretical predictions through large-scale LLM pre-training experiments, demonstrating that the scaling laws uncovered here extend beyond the controlled model and are predictive of practical training behavior.
- ACL FindingsSWE-Mutation: Can LLMs Generate Reliable Test Suites in Software Engineering?Yuxuan Sun, Yuze Zhao, Yufeng Wang, Yao Du, Zhiyuan Ma, Jinbo Wang, Mengdi Zhang, Kai Zhang, and Zhenya HuangIn Findings of the Association for Computational Linguistics: ACL 2026, 2026
We introduce SWE-Mutation, a multilingual repository-level benchmark that uses agent-generated semantic mutants to evaluate LLM test generation and repair, revealing persistent weaknesses in bug detection.
Evaluating software engineering capabilities has become a core component of modern large language models (LLMs); however, the key bottleneck hindering further scaling lies not in the scarcity of high-quality solutions, but in the lack of high-quality test suites. Test suites are indispensable both for synthesizing program repair trajectories and for providing precise feedback signals in reinforcement learning. Unfortunately, due to the high cost and difficulty of annotation, high-quality test suites have long been hard to obtain, while those automatically generated by LLMs tend to be superficial and lack sufficient discriminative power. As a first step toward constructing high-quality test suites, we introduce SWE-Mutation, a benchmark for evaluating LLM-generated test suites. The benchmark characterizes test suites by introducing systematically mutated solutions that attempt to ``fool’’ the test suites and pass validation. We further propose an agentic, language-agnostic framework for automatically generating complex mutants. Our benchmark consists of 2,636 mutated variants derived from 800 original instances and includes a multilingual subset spanning nine programming languages. Experiments on seven LLMs reveal that even DeepSeek-V3.1 achieves only 10.20% verification and 36.15% detection rates, highlighting the inadequacy of current LLMs. Additionally, our agentic mutation strategy enhances realism, reducing average detection rates from 71.04% to 39.81% compared to conventional methods. These findings expose persistent deficiencies in the ability of current LLMs to generate reliable and discriminative test suites.
2025
- ICLRFast Catch-Up, Late Switching: Optimal Batch Size Scheduling via Functional Scaling LawsJinbo Wang*, Binghui Li*, Zhanpeng Zhou, Mingze Wang, Yuxuan Sun, Jiaqi Zhang, Xunliang Cai, and Lei WuIn International Conference on Learning Representations (ICLR), 2025 This work was completed in 2025 and formally published in ICLR 2026.
We study batch size schedule: uncover optimal batch size schedule, fast catch-up effect and later switch strategy from Functional Scaling Laws (FSL) theoretical framework, and bring our insights to LLM pre-training.
Batch size scheduling (BSS) plays a critical role in large-scale deep learning training, influencing both optimization dynamics and computational efficiency. Yet, its theoretical foundations remain poorly understood. In this work, we show that the functional scaling law (FSL) framework introduced in Li et al. (2025a) provides a principled lens for analyzing BSS. Specifically, we characterize the optimal BSS under a fixed data budget and show that its structure depends sharply on task difficulty. For easy tasks, optimal schedules keep increasing batch size throughout. In contrast, for hard tasks, the optimal schedule maintains small batch sizes for most of training and switches to large batches only in a late stage. To explain the emergence of late switching, we uncover a dynamical mechanism—the fast catch-up effect—which also manifests in large language model (LLM) pretraining. After switching from small to large batches, the loss rapidly aligns with the constant large-batch trajectory. Using FSL, we show that this effect stems from rapid forgetting of accumulated gradient noise, with the catch-up speed determined by task difficulty. Crucially, this effect implies that large batches can be safely deferred to late training without sacrificing performance, while substantially reducing data consumption. Finally, extensive LLM pretraining experiments—covering both Dense and MoE architectures with up to 1.1B parameters and 1T tokens—validate our theoretical predictions. Across all settings, late-switch schedules consistently outperform constant-batch and early-switch baselines.
- arXivHow Does Local Landscape Geometry Evolve in Language Model Pre-Training?arXiv preprint arXiv:2609.39767, In submission, 2025
We study how loss landscape geometry evolves during LLM pre-training, explaining learning-rate warmup and yielding batch-size scheduling that substantially improve data efficiency.
The scale and expense of pre-training language models make efficient hyperparameter tuning essential, yet a principled guidance is still missing. Recent work shows that the geometry of loss landscape shapes training dynamics of neural networks and further informs hyperparameter choices. In this work, we analyze language model pre-training dynamics from a local landscape geometry perspective. Our study reveals two distinct phases. In the early phase, sharpness of the local landscape is initially high, leading to instability and loss plateaus under large learning rates (LRs). Later, the landscape shifts from sharp to flatter regions. This dynamic explains the necessity of LR warmup and further suggests that larger peak LRs require proportionally longer warmup periods. In the late phase, the local landscape is governed by the gradient noise scale. Through diffusion-limit analysis, we prove a depth–flatness trade-off: high noise from smaller batches widens the loss basin, whereas reduced noise from larger batches deepens it. This theory motivates a dynamic batch-size (BS) scheduler that begins with a small BS and increases it late in training. Together, we provide a unified account of loss landscape evolution, which translates into actionable tuning strategies for large-scale pre-training.
- ICMLGradPower: Powering Gradients for Faster Language Model Pre-TrainingIn International Conference on Machine Learning (ICML), 2025 This work was completed in 2025 and formally published in ICML 2026.
We propose GradPower, a lightweight sign-power gradient transformation, to accelerate LLM pre-training.
We propose GradPower, a lightweight gradient-transformation technique for accelerating language model pre-training. Given a gradient vector $\boldsymbol{g}=(g_{i})_{i}$, GradPower first applies the elementwise
sign-powertransformation: $ \varphi_p(\boldsymbol{g}) = \left({\rm sign}(g_i)|g_i|^p\right)_{i} $ for a fixed $p\gt;0$, and then feeds the transformed gradient into a base optimizer. Notably, GradPower requires only a single-line code change and no modifications to the base optimizer’s internal logic, including the hyperparameters. When applied to AdamW (termed AdamWPower), GradPower consistently achieves lower terminal loss across diverse architectures (LLaMA, Qwen2MoE), parameter scales (66M to 2B), datasets (C4, OpenWebText), and learning-rate schedules (cosine, warmup-stable-decay). The most pronounced gains are observed when training modern mixture-of-experts models with warmup-stable-decay schedules. GradPower also integrates seamlessly with other state-of-the-art optimizers, such as Muon, yielding further improvements. Finally, we provide theoretical analyses that reveal the underlying mechanism of GradPower and highlight the influence of gradient noise. - ICMLThe Sharpness Disparity Principle in Transformers for Accelerating Language Model Pre-TrainingIn International Conference on Machine Learning (ICML), 2025
We uncover a persistent Hessian heterogeneity in Transformer and turn it into a practical blockwise learning rate strategy via Edge of Stability (EoS) theory. Our algorithm achieves lower terminal loss and faster pre-training across settings.
Transformers have become the cornerstone of modern AI. Unlike traditional architectures, transformers exhibit a distinctive characteristic: diverse types of building blocks, such as embedding layers, normalization layers, self-attention mechanisms, and point-wise feed-forward networks, work collaboratively. Understanding the disparities and interactions among these blocks is therefore important. In this paper, we uncover a clear sharpness disparity across these blocks, which intriguingly emerges early in training and persists throughout the training process. Building on this insight, we propose a novel Blockwise Learning Rate (LR) strategy to accelerate large language model (LLM) pre-training. Specifically, by integrating Blockwise LR into AdamW, we consistently achieve lower terminal loss and nearly $2\times$ speedup compared to vanilla AdamW. This improvement is demonstrated across GPT-2 and LLaMA models, with model sizes ranging from 0.12B to 1.1B and datasets including OpenWebText and MiniPile. Finally, we incorporate Blockwise LR into Adam-mini (Zhang et al., 2024), a recently proposed memory-efficient variant of Adam, achieving a combined $2\times$ speedup and $2\times$ memory savings. These results underscore the potential of leveraging the sharpness disparity principle to improve LLM training.
2024
- NeurIPSImproving Generalization and Convergence by Enhancing Implicit RegularizationMingze Wang, Jinbo Wang, Haotian He, Zilin Wang, Guanhua Huang, Feiyu Xiong, Zhiyu Li, Weinan E, and Lei WuIn Advances in Neural Information Processing Systems (NeurIPS), 2024
Our algorithm enhances implicit regularization to reach flatter minima faster, improving generalization while maintaining stable optimization across vision and LLM setups.
In this work, we propose an Implicit Regularization Enhancement (IRE) framework to accelerate the discovery of flat solutions in deep learning, thereby improving generalization and convergence. Specifically, IRE decouples the dynamics of flat and sharp directions, which boosts the sharpness reduction along flat directions while maintaining the training stability in sharp directions. We show that IRE can be practically incorporated with generic base optimizers without introducing significant computational overload. Experiments show that IRE consistently improves the generalization performance for image classification tasks across a variety of benchmark datasets (CIFAR-10/100, ImageNet) and models (ResNets and ViTs). Surprisingly, IRE also achieves a $2\times$ speed-up compared to AdamW in the pre-training of Llama models (of sizes ranging from 60M to 229M) on datasets including Wikitext-103, Minipile, and Openwebtext. Moreover, we provide theoretical guarantees, showing that IRE can substantially accelerate the convergence towards flat minima in Sharpness-aware Minimization (SAM).
- JMLMemory³: Language Modeling with Explicit MemoryHongkang Yang, Zehao Lin, Wenjin Wang, Hao Wu, Zhiyu Li, Bo Tang, Wenqiang Wei, Jinbo Wang, Zeyun Tang, Shichao Song, Chenyang Xi, Yu Yu, Kai Chen, Feiyu Xiong, Linpeng Tang, and Weinan EJournal of Machine Learning, 2024
We propose Memory³: Externalizes knowledge into explicit memory, reducing reliance on parameters and improving efficiency toward separating reasoning vs. memory in language modeling.
The training and inference of large language models (LLMs) are together a costly process that transports knowledge from raw data to meaningful computation. Inspired by the memory hierarchy of the human brain, we reduce this cost by equipping LLMs with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG). Conceptually, with most of its knowledge externalized to explicit memories, the LLM can enjoy a smaller parameter size, training cost, and inference cost, all proportional to the amount of remaining “abstract knowledge”. As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs as well as RAG models, and maintains higher decoding speed than RAG. The model is named Memory³, since explicit memory is the third form of memory in LLMs after implicit memory (model parameters) and working memory (context key-values). We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable and a two-stage pretraining scheme that facilitates memory formation.
2023
- ICONIPExploring the Integration of Large Language Models into Automatic Speech Recognition Systems: An Empirical StudyZeping Min and Jinbo WangIn International Conference on Neural Information Processing (ICONIP), 2023
We explore the integration of LLMs into automatic speech recognition (ASR) systems to improve accuracy.
This paper explores the integration of Large Language Models (LLMs) into Automatic Speech Recognition (ASR) systems to improve transcription accuracy. The increasing sophistication of LLMs, with their in-context learning capabilities and instruction-following behavior, has drawn significant attention in the field of Natural Language Processing (NLP). Our primary focus is to investigate the potential of using an LLM’s in-context learning capabilities to enhance the performance of ASR systems, which currently face challenges such as ambient noise, speaker accents, and complex linguistic contexts. We designed a study using the Aishell-1 and LibriSpeech datasets, with ChatGPT and GPT-4 serving as benchmarks for LLM capabilities. Unfortunately, our initial experiments did not yield promising results, indicating the complexity of leveraging LLM’s in-context learning for ASR applications. Despite further exploration with varied settings and models, the corrected sentences from the LLMs frequently resulted in higher Word Error Rates (WER), demonstrating the limitations of LLMs in speech applications. This paper provides a detailed overview of these experiments, their results, and implications, establishing that using LLMs’ in-context learning capabilities to correct potential errors in speech recognition transcriptions is still a challenging task at the current stage.