TANDEM: Bi-Level Data Mixture Optimization with Twin Networks

The capabilities of large language models (LLMs) significantly depend on training data drawn from various domains. Optimizing domain-specific mixture ratios can be modeled as a bi-level optimization problem, which we simplify into a single-level penalized form and solve with twin networks: a proxy model trained on primary data and a dynamically updated reference model trained with additional data. Our proposed method, Twin Networks for bi-level DatA mixturE optiMization (TANDEM), measures the data efficacy through the difference between the twin models and up-weights domains that benefit more from the additional data. TANDEM provides theoretical guarantees and wider applicability, compared to prior approaches. Furthermore, our bi-level perspective suggests new settings to study domain reweighting such as data-restricted scenarios and supervised fine-tuning, where optimized mixture ratios significantly improve the performance. Extensive experiments validate TANDEM's effectiveness in all scenarios.

Calibration Data Trade-offs Across Capability Dimensions: Why Multi-Source Mixing Matters for High-Sparsity LLM Pruning

Post-training pruning compresses large language models to high sparsity using a small unlabelled calibration set, and recent work has concluded that the choice of calibration source has only modest impact on averaged post-pruning accuracy. We ask whether this conclusion survives once calibration impact is evaluated separately across distinct capability dimensions rather than aggregated. Decomposing post-pruning capability into General, Commonsense, Code, and Math, and analysing $n{=}15$ calibration sources via Spearman correlations between OIT information metrics and per-dimension retention, we uncover an opposite-sign trade-off: calibration perplexity correlates positively with General retention ($ρ{=}{+}0.71$) but negatively with Math and Code retention ($ρ{=}{-}0.53,\,{-}0.59$; $p{<}0.05$), so no single source can preserve all capabilities. We respond with multi-source calibration mixing, and propose IGSP, an information-guided self-calibration protocol that automates multi-source construction without capability-aligned corpora by minimising 4-gram aggregation and balancing perplexity across dimensions. On LLaMA-3.1-8B at SparseGPT 60% sparsity, a uniform multi-source mix reaches 58.8% total retention, outperforming the best single source (MetaMath, 50.0%) by $+8.8$ and the C4 default (40.0%) by $+18.8$; IGSP improves over Self-Cal by $+2.4$ and SGS by $+4.8$.

NestPipe: Large-Scale Recommendation Training on 1,500+ Accelerators via Nested Pipelining

Modern recommendation models have increased to trillions of parameters. As cluster scales expand to O(1k), distributed training bottlenecks shift from computation and memory to data movement, especially lookup and communication latency associated with embeddings. Existing solutions either optimize only one bottleneck or improve throughput by sacrificing training consistency. This paper presents NestPipe, a large-scale decentralized embedding training framework that tackles both bottlenecks while preserving synchronous training semantics. NestPipe exploits two hierarchical sparse parallelism opportunities through nested pipelining. At the inter-batch level, Dual-Buffer Pipelining (DBP) constructs a staleness-free five-stage pipeline through dual-buffer synchronization, mitigating lookup bottlenecks without embedding staleness. At the intra-batch level, we identify the embedding freezing phenomenon, which inspires Frozen-Window Pipelining (FWP) to overlap All2All communication with dense computation via coordinated stream scheduling and key-centric sample clustering. Experiments on production GPU and NPU clusters with 1,536 workers demonstrate that NestPipe achieves up to 3.06x speedup and 94.07% scaling efficiency.

Rollout-Training Co-Design for Efficient LLM-Based Multi-Agent Reinforcement Learning

Despite algorithm-level innovations for multi-agent reinforcement learning (MARL), the underlying networked infrastructure for large-scale MARL training remains underexplored. Existing training frameworks primarily optimize for single-agent scenarios and fail to address the unique system-level challenges of MARL, including rollout-training synchronization barriers, rollout load imbalance, and training resource underutilization. To bridge this gap, we propose FlexMARL, the first end-to-end training framework that holistically optimizes rollout, training, and their orchestration for large-scale LLM-based MARL. Specifically, FlexMARL introduces the joint orchestrator to manage data flow under the rollout-training disaggregated architecture. Building upon the experience store, a novel micro-batch driven asynchronous pipeline eliminates the synchronization barriers while providing strong consistency guarantees. Rollout engine adopts a parallel sampling scheme combined with hierarchical load balancing, which adapts to skewed inter/intra-agent request patterns. Training engine achieves on-demand hardware binding through agent-centric resource allocation. The training states of different agents are swapped via unified and location-agnostic communication. Empirical results on a large-scale production cluster demonstrate that FlexMARL achieves up to 7.3x speedup and improves hardware utilization by up to 5.6x compared to existing frameworks.

Kimi K2: Open Agentic Intelligence

We introduce Kimi K2, a Mixture-of-Experts (MoE) large language model with 32 billion activated parameters and 1 trillion total parameters. We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability while enjoying the advanced token efficiency of Muon. Based on MuonClip, K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, with strengths in agentic capabilities. Notably, K2 obtains 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual -- surpassing most open and closed-sourced baselines in non-thinking settings. It also exhibits strong capabilities in coding, mathematics, and reasoning tasks, with a score of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, and 27.1 on OJBench, all without extended thinking. These results position Kimi K2 as one of the most capable open-source large language models to date, particularly in software engineering and agentic tasks. We release our base and post-trained model checkpoints to facilitate future research and applications of agentic intelligence.

The Primacy of Magnitude in Low-Rank Adaptation

Low-Rank Adaptation (LoRA) offers a parameter-efficient paradigm for tuning large models. While recent spectral initialization methods improve convergence and performance over the naive "Noise & Zeros" scheme, their extra computational and storage overhead undermines efficiency. In this paper, we establish update magnitude as the fundamental driver of LoRA performance and propose LoRAM, a magnitude-driven "Basis & Basis" initialization scheme that matches spectral methods without their inefficiencies. Our key contributions are threefold: (i) Magnitude of weight updates determines convergence. We prove low-rank structures intrinsically bound update magnitudes, unifying hyperparameter tuning in learning rate, scaling factor, and initialization as mechanisms to optimize magnitude regulation. (ii) Spectral initialization succeeds via magnitude amplification. We demystify that the presumed knowledge-driven benefit of the spectral component essentially arises from the boost in the weight update magnitude. (iii) A novel and compact initialization strategy, LoRAM, scales deterministic orthogonal bases using pretrained weight magnitudes to simulate spectral gains. Extensive experiments show that LoRAM serves as a strong baseline, retaining the full efficiency of LoRA while matching or outperforming spectral initialization across benchmarks.

RLHGNN: Reinforcement Learning-driven Heterogeneous Graph Neural Network for Next Activity Prediction in Business Processes

Next activity prediction represents a fundamental challenge for optimizing business processes in service-oriented architectures such as microservices environments, distributed enterprise systems, and cloud-native platforms, which enables proactive resource allocation and dynamic service composition. Despite the prevalence of sequence-based methods, these approaches fail to capture non-sequential relationships that arise from parallel executions and conditional dependencies. Even though graph-based approaches address structural preservation, they suffer from homogeneous representations and static structures that apply uniform modeling strategies regardless of individual process complexity characteristics. To address these limitations, we introduce RLHGNN, a novel framework that transforms event logs into heterogeneous process graphs with three distinct edge types grounded in established process mining theory. Our approach creates four flexible graph structures by selectively combining these edges to accommodate different process complexities, and employs reinforcement learning formulated as a Markov Decision Process to automatically determine the optimal graph structure for each specific process instance. RLHGNN then applies heterogeneous graph convolution with relation-specific aggregation strategies to effectively predict the next activity. This adaptive methodology enables precise modeling of both sequential and non-sequential relationships in service interactions. Comprehensive evaluation on six real-world datasets demonstrates that RLHGNN consistently outperforms state-of-the-art approaches. Furthermore, it maintains an inference latency of approximately 1 ms per prediction, representing a highly practical solution suitable for real-time business process monitoring applications. The source code is available at https://github.com/Joker3993/RLHGNN.

GPR Full-Waveform Inversion through Adaptive Filtering of Model Parameters and Gradients Using CNN

GPR full-waveform inversion optimizes the subsurface property model iteratively to match the entire waveform information. However, the model gradients derived from wavefield continuation often contain errors, such as ghost values and excessively large values at transmitter and receiver points. Furthermore, models updated based on these gradients frequently exhibit unclear characterization of anomalous bodies or false anomalies, making it challenging to obtain accurate inversion results. To address these issues, we introduced a novel full-waveform inversion (FWI) framework that incorporates an embedded convolutional neural network (CNN) to adaptively filter model parameters and gradients. Specifically, we embedded the CNN module before the forward modeling process and ensured the entire FWI process remains differentiable. This design leverages the auto-grad tool of the deep learning library, allowing model values to pass through the CNN module during forward computation and model gradients to pass through the CNN module during backpropagation. Experiments have shown that filtering the model parameters during forward computation and the model gradients during backpropagation can ultimately yield high-quality inversion results.

Logic-of-Thought: Injecting Logic into Contexts for Full Reasoning in Large Language Models

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks but their performance in complex logical reasoning tasks remains unsatisfactory. Although some prompting methods, such as Chain-of-Thought, can improve the reasoning ability of LLMs to some extent, they suffer from an unfaithful issue where derived conclusions may not align with the generated reasoning chain. To address this issue, some studies employ the approach of propositional logic to further enhance logical reasoning abilities of LLMs. However, the potential omissions in the extraction of logical expressions in these methods can cause information loss in the logical reasoning process, thereby generating incorrect results. To this end, we propose Logic-of-Thought (LoT) prompting which employs propositional logic to generate expanded logical information from input context, and utilizes the generated logical information as an additional augmentation to the input prompts, thereby enhancing the capability of logical reasoning. The LoT is orthogonal to existing prompting methods and can be seamlessly integrated with them. Extensive experiments demonstrate that LoT boosts the performance of various prompting methods with a striking margin across five logical reasoning tasks. In particular, the LoT enhances Chain-of-Thought's performance on the ReClor dataset by +4.35%; moreover, it improves Chain-of-Thought with Self-Consistency's performance on LogiQA by +5%; additionally, it boosts performance of Tree-of-Thoughts on ProofWriter dataset by +8%.

Breaking the Hourglass Phenomenon of Residual Quantization: Enhancing the Upper Bound of Generative Retrieval

Generative retrieval (GR) has emerged as a transformative paradigm in search and recommender systems, leveraging numeric-based identifier representations to enhance efficiency and generalization. Notably, methods like TIGER employing Residual Quantization-based Semantic Identifiers (RQ-SID), have shown significant promise in e-commerce scenarios by effectively managing item IDs. However, a critical issue termed the "\textbf{Hourglass}" phenomenon, occurs in RQ-SID, where intermediate codebook tokens become overly concentrated, hindering the full utilization of generative retrieval methods. This paper analyses and addresses this problem by identifying data sparsity and long-tailed distribution as the primary causes. Through comprehensive experiments and detailed ablation studies, we analyze the impact of these factors on codebook utilization and data distribution. Our findings reveal that the "Hourglass" phenomenon substantially impacts the performance of RQ-SID in generative retrieval. We propose effective solutions to mitigate this issue, thereby significantly enhancing the effectiveness of generative retrieval in real-world E-commerce applications.