KV cache has traditionally been stored in GPU memory to accelerate the decoding phase of large language model (LLM) inference. However, it is increasingly necessary to move KV caches outside GPU devices, to enable cache reuse across different queries and inference engines. Our real-world usage statistics confirm this trend: over time, the total KV cache stored by users has grown rapidly, far exceeding the capacity of GPU memory. Despite this need, there lacks an efficient solution for offloading and transferring KV caches. We present LMCACHE, the first and so far the most efficient open-source KV caching solution, which extracts and stores KV caches generated by modern LLM engines (vLLM and SGLang) out of the GPU memory and shares them across engines and queries. LMCACHE supports both cache offloading (prefix reuse across queries) and prefill-decode (PD) disaggregation (cross-engine/GPU cache transfer). LMCACHE's high performance and wide adoption stem from the following contributions: (1) highly optimized KV cache data movement powered by batched data movement operations, compute and I/O pipelining; (2) a modular KV cache connector component, decoupling LMCACHE from the rapid evolution of inference engines; (3) a first-class control API for flexible cache orchestration across GPU, CPU, storage, and network layers. Our evaluation shows that combining LMCACHE with vLLM achieves up to 15x improvement in throughput across workloads such as multi-round question answering and document analysis. Large-scale adoption of LMCACHE in enterprise settings provides us valuable insights, for example, fetching KV cache from remote storage has unsurprisingly benefits to prefill delay, and that context truncation, which is a widely applied technique in industry, can greatly reduce prefix cache hit ratio by half. The source code of LMCACHE is at: https://github.com/LMCache/LMCache.

Abstract--Amid the rapid rise of agentic dialogue models, realistic user-simulator studies are essential for tuning effective conversation strategies. This work investigates a sales-oriented agent that adapts its dialogue based on user profiles spanning age, gender, and occupation. While age and gender influence overall performance, occupation produces the most pronounced differences in conversational intent. Leveraging this insight, we introduce a lightweight, occupation-conditioned strategy that guides the agent to prioritize intents aligned with user preferences, resulting in shorter and more successful dialogues. Our findings highlight the importance of rich simulator profiles and demonstrate how simple persona-informed strategies can enhance the effectiveness of sales-oriented dialogue systems. With the ongoing evolution of Agentic AI, researchers have begun to explore its application across diverse domains. Among these, dialogue systems designed for business recommendation tasks have attracted significant attention.

Large-scale Mixture of Experts (MoE) Large Language Models (LLMs) have recently become the frontier open weight models, achieving remarkable model capability similar to proprietary ones. But their random expert selection mechanism introduces significant data movement overhead that becomes the dominant bottleneck in multi-unit LLM serving systems. To understand the patterns underlying this data movement, we conduct comprehensive data-movement-centric profiling across four state-of-the-art large-scale MoE models released in 2025 (200B-1000B) using over 24,000 requests spanning diverse workloads. We perform systematic analysis from both temporal and spatial perspectives and distill six key insights to guide the design of diverse future serving systems. With our insights, we then demonstrate how to improve wafer-scale GPUs as a case study, and show that minor architectural modifications leveraging the insights achieve substantial performance gains, delivering 5.3x and 3.1x average speedups on DeepSeek V3 and Qwen3, respectively. Our work presents the first comprehensive data-centric analysis of large-scale MoE models and a concrete design study using the learned lessons, with profiling traces and simulation framework already open-sourced with $>$1k downloads. Our traces and results are publicly available at https://huggingface.co/datasets/core12345/MoE_expert_selection_trace

Recent advances in generative video models have enabled the creation of high-quality videos based on natural language prompts. However, these models frequently lack fine-grained temporal control, meaning they do not allow users to specify when particular visual elements should appear within a generated sequence. In this work, we introduce TempoControl, a method that allows for temporal alignment of visual concepts during inference, without requiring retraining or additional supervision. TempoControl utilizes cross-attention maps, a key component of text-to-video diffusion models, to guide the timing of concepts through a novel optimization approach. Our method steers attention using three complementary principles: aligning its temporal pattern with a control signal (correlation), adjusting its strength where visibility is required (magnitude), and preserving semantic consistency (entropy). TempoControl provides precise temporal control while maintaining high video quality and diversity. We demonstrate its effectiveness across various applications, including temporal reordering of single and multiple objects, action timing, and audio-aligned video generation. Please see our project page for more details: https://shira-schiber.github.io/TempoControl/.

Hardware-A ware Neural Architecture Search (HW-NAS) requires joint optimization of accuracy and latency under device constraints. Traditional supernet-based methods require multiple GPU days per dataset. Large Language Model (LLM)- driven approaches avoid training a large supernet and can provide quick feedback, but we observe an exploration bias: the LLM repeatedly proposes neural network designs within limited search space and fails to discover architectures across different latency ranges in the entire search space. To address this issue, we propose LLM-NAS: an LLM-driven Neural Architecture Search that can generate neural networks with high accuracy and low latency with reduced search cost. Our proposed LLM-NAS has three key components: 1) a complexity-driven partitioning engine that divides the search space by complexity to enforce diversity and mitigate exploration bias; 2) an LLM-powered architecture prompt co-evolution operator, in which the LLM first updates a knowledge base of design heuristics based on results from the previous round, then performs a guided evolution algorithm on architectures with prompts that incorporate this knowledge base. Prompts and designs improve together across rounds which avoids random guesswork and improve efficiency; 3) a zero-cost predictor to avoid training a large number of candidates from scratch. Experimental results show that on HW-NAS-Bench, LLM-NAS can achieve overall higher HV, lower IGD, and up to 54% lower latency than baselines at similar accuracy. Meanwhile, the search cost drops from days to minutes compared with traditional supernet baselines. As deep learning expands into resource-constrained environments such as the Internet of Things (IoT) devices, Hardware-A ware Neural Architecture Search (HW-NAS) becomes essential for discovering models that optimize the trade-off between accuracy and inference latency Benmeziane et al. (2021b;a).

Parallel LLM inference scaling involves sampling a set of $N>1$ responses for a single input prompt. However, these $N$ parallel responses tend to be generated independently from each other, partitioning compute resources and leaving potentially useful information in one generation untapped by others. This is in contrast to response length scaling where past computation is used in all future steps. For higher quality responses and response sets, we propose Bridge to generate interdependent responses in parallel by rethinking batched LLM hidden states as holistic tensors rather than independent slices. With only a small amount (2.8%-5.1%) of new parameters, Bridge improves the relative mean accuracy gains from reinforcement learning with verifiable rewards by up to 39% and boosts consistency of correct responses. Trained once, Bridge scales to any generation width, all with greater performance than independent generations, unlocking a more general mode of parallel scaling that effectively leverages information between sequences, compatible with any post-generation aggregation technique.

Hallucinations in Large Language Models (LLMs) pose a major barrier to their reliable use in critical decision-making. Although existing hallucination detection methods have improved accuracy, they still struggle with disentangling semantic and reasoning information and maintaining robustness. To address these challenges, we propose HARP (Hallucination detection via reasoning subspace projection), a novel hallucination detection framework. HARP establishes that the hidden state space of LLMs can be decomposed into a direct sum of a semantic subspace and a reasoning subspace, where the former encodes linguistic expression and the latter captures internal reasoning processes. Moreover, we demonstrate that the Unembedding layer can disentangle these subspaces, and by applying Singular Value Decomposition (SVD) to its parameters, the basis vectors spanning the semantic and reasoning subspaces are obtained. Finally, HARP projects hidden states onto the basis vectors of the reasoning subspace, and the resulting projections are then used as input features for hallucination detection in LLMs. By using these projections, HARP reduces the dimension of the feature to approximately 5% of the original, filters out most noise, and achieves enhanced robustness. Experiments across multiple datasets show that HARP achieves state-of-the-art hallucination detection performance; in particular, it achieves an AUROC of 92.8% on TriviaQA, outperforming the previous best method by 7.5%.

Sparse Autoencoders (SAEs) are widely employed for mechanistic interpretabil-ity and model steering. Within this context, steering is by design performed by means of decoding altered SAE intermediate representations. In contrast to existing literature, we forward an encoder-centric alternative to model steering which demonstrates a stronger cross-modal performance. We introduce S&P T op-K, a retraining-free and computationally lightweight Selection and Projection framework that identifies T op-K encoder features aligned with a sensitive attribute or behavior, optionally aggregates them into a single control axis, and computes an orthogonal projection to be subsequently applied directly in the model's native embedding space. In vision-language models, it improves fairness metrics on CelebA and FairFace by up to 3.2 times over conventional SAE usage, and in large language models, it substantially reduces aggressiveness and sycophancy in Llama-3 8B Instruct, achieving up to 3.6 times gains over masked reconstruction. These findings suggest that encoder-centric interventions provide a general, efficient, and more effective mechanism for shaping model behavior at inference time than the traditional decoder-centric use of SAEs.Figure 1: Sample generation demonstrating behavioral steering interventions on Llama 3 8B Instruct prompted to produce a sycophantic opinion. We apply two Sparse Autoencoder (SAE)-based methods to remove sycophancy: the conventional decoder-centric Masked Reconstruction approach and our proposed encoder-centric S&P Top-K protocol. Lower LLM-as-a-judge sycophancy scores indicate superior mitigation of the targeted behavioral pattern. The results illustrate that conventional Masked Reconstruction fails to suppress sycophantic behavior, while our S&P Top-K intervention successfully redirects the model's output, eliminating direct praise, repeatedly deferring endorsement, and leading the model to ultimately employ laudatory language in a sarcastic manner that subverts the original sycophantic intent. The main steps of our approach are highlighted in green. We first employ a selection mechanism to identify relevant SAE features.

Sparse Autoencoders (SAEs) extract features from LLM internal activations, meant to correspond to interpretable concepts. A core SAE training hyperparame-ter is L0: how many SAE features should fire per token on average. Existing work compares SAE algorithms using sparsity-reconstruction tradeoff plots, implying L0 is a free parameter with no single correct value aside from its effect on reconstruction. In this work we study the effect of L0 on SAEs, and show that if L0 is not set correctly, the SAE fails to disentangle the underlying features of the LLM. If L0 is too low, the SAE will mix correlated features to improve reconstruction. If L0 is too high, the SAE finds degenerate solutions that also mix features. Further, we present a proxy metric that can help guide the search for the correct L0 for an SAE on a given training distribution. We show that our method finds the correct L0 in toy models and coincides with peak sparse probing performance in LLM SAEs. We find that most commonly used SAEs have an L0 that is too low. Our work shows that L0 must be set correctly to train SAEs with correct features. It is theorized that Large Language Models (LLMs) represent concepts as linear directions in representation space, known as the Linear Representation Hypothesis (LRH) (Elhage et al., 2022; Park et al., 2024). These concepts are nearly orthogonal linear directions, allowing the LLM to represent many more concepts than there are neurons, a phenomenon known as superposition (Elhage et al., 2022). However, superposition poses a challenge for interpretability, as neurons in the LLM are polysemantic, firing on many different concepts. Sparse autoencoders (SAEs) are meant to reverse superposition, and extract interpretable, monose-mantic latent features (Cunningham et al., 2024; Bricken et al., 2023) using sparse dictionary learning (Olshausen & Field, 1997).

Large Language Models (LLMs) increasingly exhibit over-refusal - erroneously rejecting benign queries due to overly conservative safety measures - a critical functional flaw that undermines their reliability and usability. Current methods for testing this behavior are demonstrably inadequate, suffering from flawed benchmarks and limited test generation capabilities, as highlighted by our empirical user study. To the best of our knowledge, this paper introduces the first evolutionary testing framework, ORFuzz, for the systematic detection and analysis of LLM over-refusals. ORFuzz uniquely integrates three core components: (1) safety category-aware seed selection for comprehensive test coverage, (2) adaptive mutator optimization using reasoning LLMs to generate effective test cases, and (3) OR-Judge, a human-aligned judge model validated to accurately reflect user perception of toxicity and refusal. Our extensive evaluations demonstrate that ORFuzz generates diverse, validated over-refusal instances at a rate (6.98% average) more than double that of leading baselines, effectively uncovering vulnerabilities. Furthermore, ORFuzz's outputs form the basis of ORFuzzSet, a new benchmark of 1,855 highly transferable test cases that achieves a superior 63.56% average over-refusal rate across 10 diverse LLMs, significantly outperforming existing datasets. ORFuzz and ORFuzzSet provide a robust automated testing framework and a valuable community resource, paving the way for developing more reliable and trustworthy LLM-based software systems.