OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond

The rapid advancement toward long-context reasoning and multi-modal intelligence has made the memory footprint of the Key-Value (KV) cache a dominant memory bottleneck for efficient deployment. While the established per-channel quantization effectively accommodates intrinsic channel-wise outliers in Key tensors, its efficacy diminishes under extreme compression. In this work, we revisit the inherent limitations of the per-channel quantization paradigm from both empirical and theoretical perspectives. Our analysis identifies Token Norm Imbalance (TNI) as the primary bottleneck to quantization fidelity. We demonstrate that TNI systematically amplifies errors when shared quantization parameters are required to span token groups exhibiting substantial norm disparities. Instead of relying on intricate quantization pipelines (e.g., TurboQuant), we propose OScaR (Omni-Scaled Canalized Rotation), an accurate and lightweight KV cache compression framework for X-LLMs (i.e., text-only, multi-modal, and omni-modal LLMs). Advancing the per-channel paradigm, OScaR employs Canalized Rotation followed by Omni-Token Scaling to mitigate TNI-induced sequence-dimensional variance both effectively and efficiently, further supported by our optimized system design and CUDA kernels. Extensive evaluations across X-LLMs show that OScaR consistently outperforms existing methods and achieves near-lossless performance under INT2 quantization, establishing it as a robust, low-complexity, and universal framework that defines a new Pareto front. Compared with the BF16 FlashDecoding-v2 baseline, our OScaR implementation achieves a notable up to 3.0x speedup in decoding, reduces memory footprint by 5.3x, and increases throughput by 4.1x. The code for OScaR is publicly available at https://github.com/ZunhaiSu/OScaR-KV-Quant.

AIS: Adaptive Importance Sampling for Quantized RL

Reinforcement learning (RL) for large language models (LLMs) is dominated by the cost of rollout generation, which has motivated the use of low-precision rollouts (e.g., FP8) paired with a BF16 trainer to improve throughput and reduce memory pressure. This introduces a rollout-training mismatch that biases the policy gradient and can cause training to collapse outright on reasoning benchmarks. We show that the mismatch is non-stationary and acts as a double-edged sword: early in training it provides a stochastic exploration bonus, exposing the gradient to trajectories the trainer would otherwise under-sample, but the same perturbation transitions into a destabilizing source of bias as the policy concentrates. To solve this, we propose Adaptive Importance Sampling (AIS), a correction framework that adjusts the strength of its intervention on a per-batch basis. AIS combines three real-time diagnostics, namely weight reliability, divergence severity, and variance amplification, into a single mixing coefficient that interpolates between the uncorrected and fully importance-weighted gradients, suppressing the destabilizing component of the mismatch while preserving its exploratory benefit. We integrate AIS into GRPO and evaluate it on the diffusion-based LLaDA-8B-Instruct and the autoregressive Qwen3-8B and Qwen3.5-9B across mathematical reasoning and planning benchmarks. AIS matches the BF16 baseline on most tasks while retaining the 1.5 to 2.76x rollout speedup of FP8.

ROMER: Expert Replacement and Router Calibration for Robust MoE LLMs on Analog Compute-in-Memory Systems

Large language models (LLMs) with mixture-of-experts (MoE) architectures achieve remarkable scalability by sparsely activating a subset of experts per token, yet their frequent expert switching creates memory bandwidth bottlenecks that compute-in-memory (CIM) architectures are well-suited to mitigate. However, analog CIM systems suffer from inherent hardware imperfections that perturb stored weights, and its negative impact on MoE-based LLMs in noisy CIM environments remains unexplored. In this work, we present the first systematic investigation of MoE-based LLMs under noise model calibrated with real chip measurements, revealing that hardware noise critically disrupts expert load balance and renders clean-trained routing decisions consistently suboptimal. Based on these findings, we propose ROMER, a post-training calibration framework that (1) replaces underactivated experts with high-frequency ones to restore load balance, and (2) recalibrates router logits via percentile-based normalization to stabilize routing under noise. Extensive experiments across multiple benchmarks demonstrate that ROMER achieves up to 58.6\%, 58.8\%, and 59.8\% reduction in perplexity under real-chip noise conditions for DeepSeek-MoE, Qwen-MoE, and OLMoE, respectively, establishing its effectiveness and generalizability across diverse MoE architectures.

Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation

As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work provides a pivotal contribution by clarifying key concepts and guiding researchers through the evolution and trends of the field. We envision this survey as a definitive resource, empowering researchers and practitioners to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.

CodeComp: Structural KV Cache Compression for Agentic Coding

Agentic code tasks such as fault localization and patch generation require processing long codebases under tight memory constraints, where the Key-Value (KV) cache becomes the primary inference bottleneck. Existing compression methods rely exclusively on attention signals to estimate token importance, systematically discarding structurally critical tokens such as call sites, branch conditions, and assignments that are essential for code understanding. We present CodeComp, a training-free KV cache compression framework that incorporates static program analysis into LLM inference via Code Property Graph priors extracted by Joern. Across bug localization and code generation benchmarks, CodeComp consistently outperforms attention-only compression baselines under equal memory budgets, recovering the majority of full-context accuracy under aggressive KV cache compression, while matching the patch generation quality of uncompressed full-context inference and integrating seamlessly into SGLang-based agentic coding pipelines without model modification.

MathGen: Revealing the Illusion of Mathematical Competence through Text-to-Image Generation

Modern generative models have demonstrated the ability to solve challenging mathematical problems. In many real-world settings, however, mathematical solutions must be expressed visually through diagrams, plots, geometric constructions, and structured symbolic layouts, where correctness depends on precise visual composition. This naturally raises the question of whether generative models can still do so when the answer must be rendered visually rather than written in text? To study this problem, we introduce MathGen, a rigorous benchmark of 900 problems spanning seven core domains, each paired with an executable verifier under a Script-as-a-Judge protocol for deterministic and objective evaluation. Experiments on representative open-source and proprietary text-to-image models show that mathematical fidelity remains a major bottleneck: even the best closed-source model reaches only 42.0% overall accuracy, while open-source models achieve just ~ 1-11%, often near 0% on structured tasks. Overall, current T2I models remain far from competent at even elementary mathematical visual generation.

Model Evolution Under Zeroth-Order Optimization: A Neural Tangent Kernel Perspective

Zeroth-order (ZO) optimization enables memory-efficient training of neural networks by estimating gradients via forward passes only, eliminating the need for backpropagation. However, the stochastic nature of gradient estimation significantly obscures the training dynamics, in contrast to the well-characterized behavior of first-order methods under Neural Tangent Kernel (NTK) theory. To address this, we introduce the Neural Zeroth-order Kernel (NZK) to describe model evolution in function space under ZO updates. For linear models, we prove that the expected NZK remains constant throughout training and depends explicitly on the first and second moments of the random perturbation directions. This invariance yields a closed-form expression for model evolution under squared loss. We further extend the analysis to linearized neural networks. Interpreting ZO updates as kernel gradient descent via NZK provides a novel perspective for potentially accelerating convergence. Extensive experiments across synthetic and real-world datasets (including MNIST, CIFAR-10, and Tiny ImageNet) validate our theoretical results and demonstrate acceleration when using a single shared random vector.

Beyond Outliers: A Data-Free Layer-wise Mixed-Precision Quantization Approach Driven by Numerical and Structural Dual-Sensitivity

Layer-wise mixed-precision quantization (LMPQ) enables effective compression under extreme low-bit settings by allocating higher precision to sensitive layers. However, existing methods typically treat all intra-layer weight modules uniformly and rely on a single numerical property when estimating sensitivity, overlooking their distinct operational roles and structural characteristics. To address this, we propose NSDS, a novel calibration-free LMPQ framework driven by Numerical and Structural Dual-Sensitivity. Specifically, it first mechanistically decomposes each layer into distinct operational roles and quantifies their sensitivity from both numerical and structural perspectives. These dual-aspect scores are then aggregated into a unified layer-wise metric through a robust aggregation scheme based on MAD-Sigmoid and Soft-OR to guide bit allocation. Extensive experiments demonstrate that NSDS consistently achieves superior performance compared to various baselines across diverse models and downstream tasks, without relying on any calibration data.

Can We Trust LLMs on Memristors? Diving into Reasoning Ability under Non-Ideality

Memristor-based analog compute-in-memory (CIM) architectures provide a promising substrate for the efficient deployment of Large Language Models (LLMs), owing to superior energy efficiency and computational density. However, these architectures suffer from precision issues caused by intrinsic non-idealities of memristors. In this paper, we first conduct a comprehensive investigation into the impact of such typical non-idealities on LLM reasoning. Empirical results indicate that reasoning capability decreases significantly but varies for distinct benchmarks. Subsequently, we systematically appraise three training-free strategies, including thinking mode, in-context learning, and module redundancy. We thus summarize valuable guidelines, i.e., shallow layer redundancy is particularly effective for improving robustness, thinking mode performs better under low noise levels but degrades at higher noise, and in-context learning reduces output length with a slight performance trade-off. Our findings offer new insights into LLM reasoning under non-ideality and practical strategies to improve robustness.

Efficient Generative Modeling with Unitary Matrix Product States Using Riemannian Optimization

Tensor networks, which are originally developed for characterizing complex quantum many-body systems, have recently emerged as a powerful framework for capturing high-dimensional probability distributions with strong physical interpretability. This paper systematically studies matrix product states (MPS) for generative modeling and shows that unitary MPS, which is a tensor-network architecture that is both simple and expressive, offers clear benefits for unsupervised learning by reducing ambiguity in parameter updates and improving efficiency. To overcome the inefficiency of standard gradient-based MPS training, we develop a Riemannian optimization approach that casts probabilistic modeling as an optimization problem with manifold constraints, and further derive an efficient space-decoupling algorithm. Experiments on Bars-and-Stripes and EMNIST datasets demonstrate fast adaptation to data structure, stable updates, and strong performance while maintaining the efficiency and expressive power of MPS.