作者Jing Ding, Yash Nishant, Chandrish Ambati, Jyothsna Kamati, Trung Diep
LLM inference at scale faces a memory wall. The KV cache demands tens of terabytes at hundreds of gigabytes per second, yet no current memory tier delivers both at once. Characterization across multi-generation GPU systems with various LLaMA models shows host memory retrieval achieves up to 100x speedup over re-computation but supports only tens of concurrent long-context users. Electrical CXL pooling theoretically bridges this gap, but switch latency, cable reach limits, and power-scaling issues prevent practical TB-scale deployments. We present the Marvell Photonic Fabric Memory Appliance, a photonic-CXL hybrid architecture replacing electrical switches with a passive fiber shuffle to deliver 32 TB shared memory across 16 hosts via a switch-free full- crossbar topology. Emulation results demonstrate over 50 percent latency reduction versus electrical CXL pools. Simulation results show that the PF Memory Appliance eliminates cache eviction cliffs by improving time-to-first-token by 6.6x for multi-turn conversations workloads.
展开完整摘要收起摘要↓
LLM inference at scale faces a memory wall. The KV cache demands tens of terabytes at hundreds of gigabytes per second, yet no current memory tier delivers both at once. Characterization across multi-generation GPU systems with various LLaMA models shows host memory retrieval achieves up to 100x speedup over re-computation but supports only tens of concurrent long-context users. Electrical CXL pooling theoretically bridges this gap, but switch latency, cable reach limits, and power-scaling issues prevent practical TB-scale deployments. We present the Marvell Photonic Fabric Memory Appliance, a photonic-CXL hybrid architecture replacing electrical switches with a passive fiber shuffle to deliver 32 TB shared memory across 16 hosts via a switch-free full- crossbar topology. Emulation results demonstrate over 50 percent latency reduction versus electrical CXL pools. Simulation results show that the PF Memory Appliance eliminates cache eviction cliffs by improving time-to-first-token by 6.6x for multi-turn conversations workloads.
作者Sookyung Choi, Seungyong Lee, Kangkyu Park, Yunseo Chun, Junseok Lee, Hyeongseok Gwak, Myunghyun Rhee, Euiseok Kim, Donguk Moon, Kwangsik Shin, Guseul Heo, Youngpyo Joo, Hoshik Kim, Jongse Park
Modern LLMs and their agentic applications are broadening the range of serving workloads, spanning context lengths from a few hundred tokens to hundreds of thousands. As these requests frequently interleave within the same serving window, LLM serving systems must handle highly heterogeneous mixed-length workloads. Such mixed-length workloads expose fundamental inefficiencies in GPU-centric serving architectures, whose throughput depends on large, memory-constrained batches. In this paper, we present NELSSA, an LLM serving system that integrates GPUs with real-world Processing-near-Memory (PNM) accelerator devices to efficiently support mixed-length workloads. NELSSA employs length-based request placement to route short-context requests to GPUs and long-context requests to the PNM tier, incorporating runtime migration to accommodate dynamic context growth without recomputation. We prototype NELSSA as an end-to-end system, implementing device-level sparse attention on PNM, GPU decode kernels, and a host-side runtime that orchestrates scheduling and cross-tier memory movement over a CXL-enabled infrastructure with RPC and RDMA support. Across mixed-length LLM workloads, NELSSA improves decode throughput by up to 5.5x in tokens/sec and reduces P99 latency by up to 15x compared to GPU-only baselines. Our end-to-end prototype and experimental results suggest that integrated GPU-PNM serving, enabled by CXL-based disaggregation, is a promising system paradigm for scalable and flexible LLM infrastructures that support evolving workloads.
展开完整摘要收起摘要↓
Modern LLMs and their agentic applications are broadening the range of serving workloads, spanning context lengths from a few hundred tokens to hundreds of thousands. As these requests frequently interleave within the same serving window, LLM serving systems must handle highly heterogeneous mixed-length workloads. Such mixed-length workloads expose fundamental inefficiencies in GPU-centric serving architectures, whose throughput depends on large, memory-constrained batches. In this paper, we present NELSSA, an LLM serving system that integrates GPUs with real-world Processing-near-Memory (PNM) accelerator devices to efficiently support mixed-length workloads. NELSSA employs length-based request placement to route short-context requests to GPUs and long-context requests to the PNM tier, incorporating runtime migration to accommodate dynamic context growth without recomputation. We prototype NELSSA as an end-to-end system, implementing device-level sparse attention on PNM, GPU decode kernels, and a host-side runtime that orchestrates scheduling and cross-tier memory movement over a CXL-enabled infrastructure with RPC and RDMA support. Across mixed-length LLM workloads, NELSSA improves decode throughput by up to 5.5x in tokens/sec and reduces P99 latency by up to 15x compared to GPU-only baselines. Our end-to-end prototype and experimental results suggest that integrated GPU-PNM serving, enabled by CXL-based disaggregation, is a promising system paradigm for scalable and flexible LLM infrastructures that support evolving workloads.
Long-context inference is becoming a fundamental capability for modern LLM serving, especially driven by emerging agentic applications. Yet it faces a severe memory wall that the KV cache scales proportionally with increasing context length and request concurrency. Existing sparse KV cache methods offload most KV entries to host memory and retrieve only the critical KV entries needed by each decoding step. However, they commonly introduce substantial auxiliary states in GPU memory for KV retrieval management. Our measurements show that these often-overlooked auxiliary states introduce significant memory overhead and become a new bottleneck under high-concurrency workloads. In this paper, we present DualDecoder, a lightweight serving system for long-context LLM inference that enables efficient sparse KV cache retrieval from host memory. Our key insight is that the critical KV entries required for decoding the next token can be accurately predicted from the preceding speculated token. This predictability enables KV retrieval to be proactively prefetched and overlapped with decoding computation, effectively eliminating the GPU memory overhead of auxiliary states. To achieve this prefetching efficiently, DualDecoder leverages a novel dual-token decoding pipeline that accurately identifies critical KV entries with negligible computational overhead, and designs a layer-aware transfer schedule to overlap KV prefetching with model computation and a layer-scoped memory manager to reduce the GPU runtime buffer. Experimental results show that DualDecoder improves decoding throughput by up to 2.62$\times$ over state-of-the-art systems while preserving decoding latency and model quality.
展开完整摘要收起摘要↓
Long-context inference is becoming a fundamental capability for modern LLM serving, especially driven by emerging agentic applications. Yet it faces a severe memory wall that the KV cache scales proportionally with increasing context length and request concurrency. Existing sparse KV cache methods offload most KV entries to host memory and retrieve only the critical KV entries needed by each decoding step. However, they commonly introduce substantial auxiliary states in GPU memory for KV retrieval management. Our measurements show that these often-overlooked auxiliary states introduce significant memory overhead and become a new bottleneck under high-concurrency workloads. In this paper, we present DualDecoder, a lightweight serving system for long-context LLM inference that enables efficient sparse KV cache retrieval from host memory. Our key insight is that the critical KV entries required for decoding the next token can be accurately predicted from the preceding speculated token. This predictability enables KV retrieval to be proactively prefetched and overlapped with decoding computation, effectively eliminating the GPU memory overhead of auxiliary states. To achieve this prefetching efficiently, DualDecoder leverages a novel dual-token decoding pipeline that accurately identifies critical KV entries with negligible computational overhead, and designs a layer-aware transfer schedule to overlap KV prefetching with model computation and a layer-scoped memory manager to reduce the GPU runtime buffer. Experimental results show that DualDecoder improves decoding throughput by up to 2.62$\times$ over state-of-the-art systems while preserving decoding latency and model quality.
Mixture-of-experts (MoE) inference first aligns routed tokens into padded expert blocks, then executes packed quantized matrix multiplication over those blocks. This preprocessing is often treated as bookkeeping. In one pre-specified Qwen3-Coder AWQ layer-6 fixture on a pinned vLLM/Marlin build and RTX 3090 runtime, we show that the tested route-block interventions select exact packed arithmetic trajectories. Two fixed preconstruction histories produced distinct native alignments and exact trajectories. Injecting the opposite alignment transferred W13, activation, routed-W2, and final outputs. Permuting two routes within one block preserved each native trajectory, while exchanging two prior-data-selected routes across the boundary between expert-106 blocks 40 and 41 transferred the complete opposite trajectory. Source- and binary-derived schedule geometry maps those blocks to direct/full-K and split/global-reduction classes. Forcing a single-slice 200-block grid made W13 bitwise equal. Stable canonical construction made both histories converge to a third exact trajectory. The confirmatory cohort contains 70 valid cold processes and seven required perturbation rejections. This is a causal mechanism result for one fixture, not a prevalence, allocator, portability, or serving-impact claim.
展开完整摘要收起摘要↓
Mixture-of-experts (MoE) inference first aligns routed tokens into padded expert blocks, then executes packed quantized matrix multiplication over those blocks. This preprocessing is often treated as bookkeeping. In one pre-specified Qwen3-Coder AWQ layer-6 fixture on a pinned vLLM/Marlin build and RTX 3090 runtime, we show that the tested route-block interventions select exact packed arithmetic trajectories. Two fixed preconstruction histories produced distinct native alignments and exact trajectories. Injecting the opposite alignment transferred W13, activation, routed-W2, and final outputs. Permuting two routes within one block preserved each native trajectory, while exchanging two prior-data-selected routes across the boundary between expert-106 blocks 40 and 41 transferred the complete opposite trajectory. Source- and binary-derived schedule geometry maps those blocks to direct/full-K and split/global-reduction classes. Forcing a single-slice 200-block grid made W13 bitwise equal. Stable canonical construction made both histories converge to a third exact trajectory. The confirmatory cohort contains 70 valid cold processes and seven required perturbation rejections. This is a causal mechanism result for one fixture, not a prevalence, allocator, portability, or serving-impact claim.
作者Siqi Zeng, Sewoong Lee, Han Zhao, Julia Hockenmaier
Instruction hierarchies are a core safety assumption of language model deployment: higher priority inputs, such as system prompts, should override conflicting lower priority inputs from users or tools. Yet frontier LLMs often violate this hierarchy. We introduce V-Steer, a training-free inference time method that restores privileged influence by editing cached value vectors at prompt positions. Using direct logit attribution on the first next token prediction, V-Steer identifies heads where lower priority spans dominate privileged ones, then boosts privileged spans and suppresses conflicting lower priority spans through in-place multiplicative edits to cached V tensors. Since the method acts only on cached values, it remains compatible with fused attention backends and adds only a one time prefill overhead. Across models from 7B to 70B, this attribution guided intervention raises primary constraint accuracy from under 18% up to 92% on controlled role conflict benchmarks, and on broader instruction hierarchy evaluations substantially outperforms prompt only baselines while matching or exceeding SoTA training based methods on 3 of 4 scales of LLMs, with negligible decoding-speed overhead. The code is available at https://github.com/cindy2000sh/v-steer.
展开完整摘要收起摘要↓
Instruction hierarchies are a core safety assumption of language model deployment: higher priority inputs, such as system prompts, should override conflicting lower priority inputs from users or tools. Yet frontier LLMs often violate this hierarchy. We introduce V-Steer, a training-free inference time method that restores privileged influence by editing cached value vectors at prompt positions. Using direct logit attribution on the first next token prediction, V-Steer identifies heads where lower priority spans dominate privileged ones, then boosts privileged spans and suppresses conflicting lower priority spans through in-place multiplicative edits to cached V tensors. Since the method acts only on cached values, it remains compatible with fused attention backends and adds only a one time prefill overhead. Across models from 7B to 70B, this attribution guided intervention raises primary constraint accuracy from under 18% up to 92% on controlled role conflict benchmarks, and on broader instruction hierarchy evaluations substantially outperforms prompt only baselines while matching or exceeding SoTA training based methods on 3 of 4 scales of LLMs, with negligible decoding-speed overhead. The code is available at https://github.com/cindy2000sh/v-steer.
We present fabric_ext, an eBPF middleware compiler and runtime for extensible OS policies over GPU--CXL fabrics. fabric_ext lets one policy program execute across GPU hooks, driver/runtime hooks, DPU/NIC hooks, and CXL switch or near-memory hooks. The key abstraction is a semantic movement graph: edges describe bytes, stride, reuse distance, read/write ratio, source and destination, ordering requirement, alias set, ownership, and transformations such as Move, Quantize, Compress, Checksum, Filter, Reduce, Scatter/Gather, Replicate, and Persist. The compiler lowers this graph into per-device eBPF programs, verifier obligations, consistency-classed BPF maps, and artifacts for bpftime and dputime. At the fabric edge, fabric_ext treats a near-Type-2 small core as a hardware-JIT and state manager: it specializes verified movement descriptors into local copy, placement, ordering, and transformation commands, while the surrounding Von Neumann island of memory, DMA, and compute engines performs the dataflow. Because this dataflow is data-driven, fabric_ext also places observation beside the island, where queues, DMA completions, memory placement, and ownership transitions are visible as they happen. The canonical stress case is LLM prefill: attention streams KV blocks and reductions while FFN streams activations, weights, and compressible intermediates, forcing one request to cross GPU tensor execution, DPU/NIC event execution, and CXL or switch-local dataflow islands.
展开完整摘要收起摘要↓
We present fabric_ext, an eBPF middleware compiler and runtime for extensible OS policies over GPU--CXL fabrics. fabric_ext lets one policy program execute across GPU hooks, driver/runtime hooks, DPU/NIC hooks, and CXL switch or near-memory hooks. The key abstraction is a semantic movement graph: edges describe bytes, stride, reuse distance, read/write ratio, source and destination, ordering requirement, alias set, ownership, and transformations such as Move, Quantize, Compress, Checksum, Filter, Reduce, Scatter/Gather, Replicate, and Persist. The compiler lowers this graph into per-device eBPF programs, verifier obligations, consistency-classed BPF maps, and artifacts for bpftime and dputime. At the fabric edge, fabric_ext treats a near-Type-2 small core as a hardware-JIT and state manager: it specializes verified movement descriptors into local copy, placement, ordering, and transformation commands, while the surrounding Von Neumann island of memory, DMA, and compute engines performs the dataflow. Because this dataflow is data-driven, fabric_ext also places observation beside the island, where queues, DMA completions, memory placement, and ownership transitions are visible as they happen. The canonical stress case is LLM prefill: attention streams KV blocks and reductions while FFN streams activations, weights, and compressible intermediates, forcing one request to cross GPU tensor execution, DPU/NIC event execution, and CXL or switch-local dataflow islands.
Remote sensing change detection aims to identify land-cover changes from bi-temporal images. Most existing methods follow a one-shot dense prediction paradigm, directly regressing a change mask from fused features. However, such approaches overlook the intrinsic frequency characteristics of change patterns. We propose Freq-RemoteVAR, a frequency autoregressive framework that reformulates change detection as a structured generation problem in the frequency domain. Instead of predicting the change mask in a single step, we introduce a next-frequency prediction paradigm, where change information is progressively generated from coarse to fine. We design a frequency-aware mask tokenization strategy that decomposes change supervision into multi-frequency token targets via Fourier transformation and quantization. We develop a Frequency VAR Transformer, which performs causal autoregressive modeling over frequency tokens. The model starts from learned mask queries and progressively predicts frequency-level tokens conditioned on previously generated tokens and bi-temporal image features, effectively capturing long-range dependencies across frequency scales. We introduce Scale-Aligned RoPE Cross Attention (SRCA) module, which aligns frequency-domain mask queries with spatial-domain bi-temporal features under a unified coordinate system, enhancing spatial-frequency consistency during generation. We propose a Change-quality Control module that adaptively modulates the generation process through dynamic normalization, attention biasing, and spatial offset adjustment, thereby suppressing pseudo-change responses and improving robustness. Extensive experiments on CDD, GZ-CD, and LEVIR-CD demonstrate that Freq-RemoteVAR consistently outperforms existing methods, particularly in challenging scenarios with complex appearance variations and noisy disturbances.
展开完整摘要收起摘要↓
Remote sensing change detection aims to identify land-cover changes from bi-temporal images. Most existing methods follow a one-shot dense prediction paradigm, directly regressing a change mask from fused features. However, such approaches overlook the intrinsic frequency characteristics of change patterns. We propose Freq-RemoteVAR, a frequency autoregressive framework that reformulates change detection as a structured generation problem in the frequency domain. Instead of predicting the change mask in a single step, we introduce a next-frequency prediction paradigm, where change information is progressively generated from coarse to fine. We design a frequency-aware mask tokenization strategy that decomposes change supervision into multi-frequency token targets via Fourier transformation and quantization. We develop a Frequency VAR Transformer, which performs causal autoregressive modeling over frequency tokens. The model starts from learned mask queries and progressively predicts frequency-level tokens conditioned on previously generated tokens and bi-temporal image features, effectively capturing long-range dependencies across frequency scales. We introduce Scale-Aligned RoPE Cross Attention (SRCA) module, which aligns frequency-domain mask queries with spatial-domain bi-temporal features under a unified coordinate system, enhancing spatial-frequency consistency during generation. We propose a Change-quality Control module that adaptively modulates the generation process through dynamic normalization, attention biasing, and spatial offset adjustment, thereby suppressing pseudo-change responses and improving robustness. Extensive experiments on CDD, GZ-CD, and LEVIR-CD demonstrate that Freq-RemoteVAR consistently outperforms existing methods, particularly in challenging scenarios with complex appearance variations and noisy disturbances.
作者Frank Nie, Ethan B Liu, Yuan Zhu, Wei Fan, Jindong Han
Question answering (QA) over irregular clinical time series (ICTS) plays a pivotal role in a wide range of healthcare applications. Although recent multimodal time-series large language models (LLMs) have shown considerable promise in general-purpose time-series QA, they remain poorly equipped to model the sparsity, asynchrony, and irregular sampling patterns of clinical observations. To fill this gap, we propose ClinPRISM, a cost-effective multimodal LLM reasoning framework for question answering over ICTS data. First, we devise an irregularity-aware multi-scale encoder to capture sparse clinical evidence at diverse temporal scales. Then, we propose a temporal evidence distiller to integrate representations across these scales and compress them into a small number of LLM-compatible tokens. Moreover, we introduce a progressive alignment strategy that sequentially aligns the irregular trajectories with the LLM's textual embedding space. To facilitate training, we construct 30,000 clinical time series paired with multi-scale descriptions, together with 41,000 instruction-tuning instances spanning 11 tasks. Using a 4-billion-parameter LLM backbone, ClinPRISM achieves state-of-the-art performance on the held-out evaluation benchmark while using only 16 time-series tokens and achieving an average inference latency of 0.15 seconds per question.
展开完整摘要收起摘要↓
Question answering (QA) over irregular clinical time series (ICTS) plays a pivotal role in a wide range of healthcare applications. Although recent multimodal time-series large language models (LLMs) have shown considerable promise in general-purpose time-series QA, they remain poorly equipped to model the sparsity, asynchrony, and irregular sampling patterns of clinical observations. To fill this gap, we propose ClinPRISM, a cost-effective multimodal LLM reasoning framework for question answering over ICTS data. First, we devise an irregularity-aware multi-scale encoder to capture sparse clinical evidence at diverse temporal scales. Then, we propose a temporal evidence distiller to integrate representations across these scales and compress them into a small number of LLM-compatible tokens. Moreover, we introduce a progressive alignment strategy that sequentially aligns the irregular trajectories with the LLM's textual embedding space. To facilitate training, we construct 30,000 clinical time series paired with multi-scale descriptions, together with 41,000 instruction-tuning instances spanning 11 tasks. Using a 4-billion-parameter LLM backbone, ClinPRISM achieves state-of-the-art performance on the held-out evaluation benchmark while using only 16 time-series tokens and achieving an average inference latency of 0.15 seconds per question.
Memory is essential for LLM agents to accumulate task experience and reuse task-specific execution strategies. However, real-world deployment over boundary-agnostic and evolving task streams exposes a fundamental stability-plasticity dilemma. External retrieval-based memory can rapidly absorb new evidence, but it often fails to internalize recurring execution patterns and incurs inference-time retrieval overhead. Parametric memory enables stable and efficient execution once learned, but typically relies on explicit task boundaries and fixed parameter budgets. Inspired by the human brain, which balances plasticity and stability through complementary episodic storage and gradual consolidation, we propose UniMem, a self-routing framework for autonomous memory management. UniMem uses learnable routing tokens as memory controllers, enabling adaptive coordination between complementary memory pathways: novel or sparse tasks are retained in an episodic buffer for retrieval-augmented execution, while recurring and reliable patterns are consolidated into expandable parametric memory. By decoupling task identification from task execution with routing tokens and parametric memory blocks, UniMem expands memory on demand without task labels during deployment or uncontrolled parameter growth. Experiments on long-horizon streaming task sequences show that UniMem consistently outperforms baselines while maintaining execution fidelity, achieving an average gain of 4.0 EM points across three backbone models.
展开完整摘要收起摘要↓
Memory is essential for LLM agents to accumulate task experience and reuse task-specific execution strategies. However, real-world deployment over boundary-agnostic and evolving task streams exposes a fundamental stability-plasticity dilemma. External retrieval-based memory can rapidly absorb new evidence, but it often fails to internalize recurring execution patterns and incurs inference-time retrieval overhead. Parametric memory enables stable and efficient execution once learned, but typically relies on explicit task boundaries and fixed parameter budgets. Inspired by the human brain, which balances plasticity and stability through complementary episodic storage and gradual consolidation, we propose UniMem, a self-routing framework for autonomous memory management. UniMem uses learnable routing tokens as memory controllers, enabling adaptive coordination between complementary memory pathways: novel or sparse tasks are retained in an episodic buffer for retrieval-augmented execution, while recurring and reliable patterns are consolidated into expandable parametric memory. By decoupling task identification from task execution with routing tokens and parametric memory blocks, UniMem expands memory on demand without task labels during deployment or uncontrolled parameter growth. Experiments on long-horizon streaming task sequences show that UniMem consistently outperforms baselines while maintaining execution fidelity, achieving an average gain of 4.0 EM points across three backbone models.
Generative recommendation commonly represents items using fixed-length semantic identifiers (SIDs) constructed through clustering and quantization. However, these artificial codes may overcompress item semantics, remain misaligned with pretrained LLM vocabularies, and require costly autoregressive decoding. In light of this, we propose VaLiDRec, a generative recommendation framework based on variable-length, LLM-aligned semantic identifiers. VaLiDRec constructs SIDs directly from informative native LLM vocabulary tokens via token importance estimation, semantic-quality-aware pruning, and collision-aware refinement, allowing identifier lengths to adapt to item semantic complexity. To model user preferences, VaLiDRec incorporates graph-aware soft prompts and reformulates recommendation as token-set prediction with token-level item scoring, eliminating autoregressive SID generation and beam search. Experiments on four real-world datasets show that VaLiDRec consistently outperforms strong sequential and generative recommendation baselines across all evaluation metrics. It further achieves superior zero-shot item cold-start performance and 87.49$\times$ faster inference than LC-Rec. These results demonstrate that LLM-native variable-length semantic identifiers provide a more expressive and efficient paradigm for generative recommendation.
展开完整摘要收起摘要↓
Generative recommendation commonly represents items using fixed-length semantic identifiers (SIDs) constructed through clustering and quantization. However, these artificial codes may overcompress item semantics, remain misaligned with pretrained LLM vocabularies, and require costly autoregressive decoding. In light of this, we propose VaLiDRec, a generative recommendation framework based on variable-length, LLM-aligned semantic identifiers. VaLiDRec constructs SIDs directly from informative native LLM vocabulary tokens via token importance estimation, semantic-quality-aware pruning, and collision-aware refinement, allowing identifier lengths to adapt to item semantic complexity. To model user preferences, VaLiDRec incorporates graph-aware soft prompts and reformulates recommendation as token-set prediction with token-level item scoring, eliminating autoregressive SID generation and beam search. Experiments on four real-world datasets show that VaLiDRec consistently outperforms strong sequential and generative recommendation baselines across all evaluation metrics. It further achieves superior zero-shot item cold-start performance and 87.49$\times$ faster inference than LC-Rec. These results demonstrate that LLM-native variable-length semantic identifiers provide a more expressive and efficient paradigm for generative recommendation.
作者Hong Chen, Kang Chen, Yuxuan Fan, Bo Wang, Yubo Gao, Yuanlin Chu, Xuming Hu
Stateful multimodal assistants encode an image once but may answer questions about it many turns later. Attention-guided visual-KV eviction assumes that evidence irrelevant now will remain dispensable, although future questions are unknown. We ask when a visual fact is actually safe to forget and introduce the Causal Visual Memory Audit (CVMA), a paired single-prefill framework that tests what later answers lose when a visual region, the whole image, or prior assistant text becomes unavailable. On VisDial and ConvBench, current attention can rank future-useful regions worse than random even though a diagnostic marginal-utility control shows substantial selection headroom. Aggregate scores hide this failure when later turns do not need vision; controlled and stock-generated histories reveal a second escape route, in which assistant-text KV replaces image KV for facts already stated but not reliably for unstated facts. In the tested stacks, safe forgetting is supported by low future visual dependence or fact-specific verbalization---not by low current attention.
展开完整摘要收起摘要↓
Stateful multimodal assistants encode an image once but may answer questions about it many turns later. Attention-guided visual-KV eviction assumes that evidence irrelevant now will remain dispensable, although future questions are unknown. We ask when a visual fact is actually safe to forget and introduce the Causal Visual Memory Audit (CVMA), a paired single-prefill framework that tests what later answers lose when a visual region, the whole image, or prior assistant text becomes unavailable. On VisDial and ConvBench, current attention can rank future-useful regions worse than random even though a diagnostic marginal-utility control shows substantial selection headroom. Aggregate scores hide this failure when later turns do not need vision; controlled and stock-generated histories reveal a second escape route, in which assistant-text KV replaces image KV for facts already stated but not reliably for unstated facts. In the tested stacks, safe forgetting is supported by low future visual dependence or fact-specific verbalization---not by low current attention.
作者Ya Zhang, Tong Lei, Yao Chen, Yonggang Che, Chuanfu Xu, Haozhong Qiu, Yusong Tan
Helper-thread prefetching hides the latency of irregular memory accesses by executing address dependency chains ahead of the main thread. However, its effectiveness depends on the range of future iterations covered by the helper thread. A fixed coverage range cannot consistently accommodate different workloads and processors, whereas exhaustively evaluating candidate configurations incurs substantial configuration cost. This paper presents CW-Ghost, which uses a single offline profiling run to estimate the average demand cache line fill volume generated per target iteration in a target region. CW-Ghost combines this estimate with a cache capacity budget to derive a Capacity Window, which determines the iteration granularity of each prefetch chunk. In addition, bounded chunk-level synchronization limits the number of chunks by which the helper thread may run ahead of the main thread. Across 14 workload instances evaluated on Intel and AMD CPU platforms, CW-Ghost achieves geometric mean speedups of 1.54x and 1.33x, respectively, over the original programs. Compared with Ghost Threading, it improves geometric mean performance by 15.8% and 10.8%, respectively, while achieving more than 99% of the empirically optimal performance within the candidate set on both platforms. These results demonstrate that cache capacity constraints can effectively guide the selection of granularity for helper-thread prefetching.
展开完整摘要收起摘要↓
Helper-thread prefetching hides the latency of irregular memory accesses by executing address dependency chains ahead of the main thread. However, its effectiveness depends on the range of future iterations covered by the helper thread. A fixed coverage range cannot consistently accommodate different workloads and processors, whereas exhaustively evaluating candidate configurations incurs substantial configuration cost. This paper presents CW-Ghost, which uses a single offline profiling run to estimate the average demand cache line fill volume generated per target iteration in a target region. CW-Ghost combines this estimate with a cache capacity budget to derive a Capacity Window, which determines the iteration granularity of each prefetch chunk. In addition, bounded chunk-level synchronization limits the number of chunks by which the helper thread may run ahead of the main thread. Across 14 workload instances evaluated on Intel and AMD CPU platforms, CW-Ghost achieves geometric mean speedups of 1.54x and 1.33x, respectively, over the original programs. Compared with Ghost Threading, it improves geometric mean performance by 15.8% and 10.8%, respectively, while achieving more than 99% of the empirically optimal performance within the candidate set on both platforms. These results demonstrate that cache capacity constraints can effectively guide the selection of granularity for helper-thread prefetching.
Compilers are fundamental software tools that translate high-level programs into machine code. Modern compilers expose hundreds of optimizations, each turned on or off through an optimization flag, to improve the performance of the generated code. However, the number of possible flag combinations grows exponentially, making it difficult to find a flag configuration well suited to a given target program. Existing compiler auto-tuning techniques reduce tuning cost by pruning the search space, injecting search biases, or predicting configuration performance. Although some exploit program features, the knowledge they extract from historical data is frozen once search begins; runtime feedback then guides only the search itself, never the prior. As a result, when this prior mismatches the target program, these methods waste much of the limited online budget before the search reaches good configurations. We propose WarmTuner, an offline-to-online reinforcement learning framework that instead turns historical records into a program-conditioned policy that predicts each flag's setting over the full flag space and remains adaptable on the target program. Offline, WarmTuner learns this program-conditioned policy over the full flag space from historical good configurations. Online, it refines the same policy on the target program using real compile-run feedback, so that the policy is driven by measured speedups rather than limited to the historical data. We instantiate the online update with Group Relative Policy Optimization (GRPO), which compares candidates in the same round and avoids a separate value model. We evaluate WarmTuner on GCC 15.2.0 with cBench and PolyBench. The results show that WarmTuner achieves an average speedup of 1.732x over GCC -O3 and obtains the best result on 14/30 programs, significantly outperforming the compared techniques.
展开完整摘要收起摘要↓
Compilers are fundamental software tools that translate high-level programs into machine code. Modern compilers expose hundreds of optimizations, each turned on or off through an optimization flag, to improve the performance of the generated code. However, the number of possible flag combinations grows exponentially, making it difficult to find a flag configuration well suited to a given target program. Existing compiler auto-tuning techniques reduce tuning cost by pruning the search space, injecting search biases, or predicting configuration performance. Although some exploit program features, the knowledge they extract from historical data is frozen once search begins; runtime feedback then guides only the search itself, never the prior. As a result, when this prior mismatches the target program, these methods waste much of the limited online budget before the search reaches good configurations. We propose WarmTuner, an offline-to-online reinforcement learning framework that instead turns historical records into a program-conditioned policy that predicts each flag's setting over the full flag space and remains adaptable on the target program. Offline, WarmTuner learns this program-conditioned policy over the full flag space from historical good configurations. Online, it refines the same policy on the target program using real compile-run feedback, so that the policy is driven by measured speedups rather than limited to the historical data. We instantiate the online update with Group Relative Policy Optimization (GRPO), which compares candidates in the same round and avoids a separate value model. We evaluate WarmTuner on GCC 15.2.0 with cBench and PolyBench. The results show that WarmTuner achieves an average speedup of 1.732x over GCC -O3 and obtains the best result on 14/30 programs, significantly outperforming the compared techniques.
Coding agents repeatedly search, navigate, and retain context from evolving repositories, but disconnected indexes, language servers, and task-local histories force repeated discovery and obscure lifecycle costs. CodeNib builds reusable lexical, dense, and structural views per repository commit, maps outputs to repository-relative source ranges, maintains selected views across edits, and serves ranked search, symbol navigation, and bounded context through one runtime. Across 100 snapshots, we map quality-cost frontiers across the repository-context lifecycle. When outputs match an independent rebuild, graph and vector updates are $8.7\times$ and $25.4\times$ faster at the median. On the static-navigation subset matching normalized live-server locations (63% of 1,000 requests), the median per-request live/static latency ratio is $4.7\times$. Across five models, selected context policies preserve localization with 50--87% fewer trajectory tokens than paired grep/read. Together, these results support multi-view repository-context serving with explicit, operation-specific validity boundaries.
展开完整摘要收起摘要↓
Coding agents repeatedly search, navigate, and retain context from evolving repositories, but disconnected indexes, language servers, and task-local histories force repeated discovery and obscure lifecycle costs. CodeNib builds reusable lexical, dense, and structural views per repository commit, maps outputs to repository-relative source ranges, maintains selected views across edits, and serves ranked search, symbol navigation, and bounded context through one runtime. Across 100 snapshots, we map quality-cost frontiers across the repository-context lifecycle. When outputs match an independent rebuild, graph and vector updates are $8.7\times$ and $25.4\times$ faster at the median. On the static-navigation subset matching normalized live-server locations (63% of 1,000 requests), the median per-request live/static latency ratio is $4.7\times$. Across five models, selected context policies preserve localization with 50--87% fewer trajectory tokens than paired grep/read. Together, these results support multi-view repository-context serving with explicit, operation-specific validity boundaries.
Language models are almost always quantized before they are deployed, and a growing line of work asks whether quantization also lowers their privacy risk. That work measures privacy almost entirely with membership inference. We think this is the wrong thing to measure for the risk that most people actually worry about, namely a model reproducing its training data word for word, and we measure that directly. Using the Pythia models and the public set of sequences each of them is known to have memorized, we track verbatim extraction across five precision levels, from full precision down to four bits, and across three model sizes, while measuring general capability (perplexity) at every point. We find two things. Quantization is a selective forgetter: verbatim memorization falls off faster than capability at every precision and every model size we tried, and this holds under two unrelated quantization algorithms and two evaluation corpora. But the selectivity is not enough to make quantization a privacy defense, which cuts against the optimistic reading of earlier membership-inference results. At the largest model we study, four-bit quantization still reproduces most of the memorized sequences while giving up only a few percent of capability, and the fraction of memorized data that survives quantization grows with model size. We conclude that compression should not be treated as a way to remove memorized training data, and that extraction, not membership inference, is the number practitioners should be watching. All code, sampled evaluation data, and per-configuration results are released.
展开完整摘要收起摘要↓
Language models are almost always quantized before they are deployed, and a growing line of work asks whether quantization also lowers their privacy risk. That work measures privacy almost entirely with membership inference. We think this is the wrong thing to measure for the risk that most people actually worry about, namely a model reproducing its training data word for word, and we measure that directly. Using the Pythia models and the public set of sequences each of them is known to have memorized, we track verbatim extraction across five precision levels, from full precision down to four bits, and across three model sizes, while measuring general capability (perplexity) at every point. We find two things. Quantization is a selective forgetter: verbatim memorization falls off faster than capability at every precision and every model size we tried, and this holds under two unrelated quantization algorithms and two evaluation corpora. But the selectivity is not enough to make quantization a privacy defense, which cuts against the optimistic reading of earlier membership-inference results. At the largest model we study, four-bit quantization still reproduces most of the memorized sequences while giving up only a few percent of capability, and the fraction of memorized data that survives quantization grows with model size. We conclude that compression should not be treated as a way to remove memorized training data, and that extraction, not membership inference, is the number practitioners should be watching. All code, sampled evaluation data, and per-configuration results are released.
作者Jiacong Xu, Hanwen Jiang, Zhixin Shu, Kalyan Sunkavalli, Vishal M. Patel, Yiqun Mei
We present Wonder, a general-purpose video world model for real-time, camera-controllable world exploration. Given an image or a conditional video, Wonder constructs a playable world where users can navigate interactively by moving the camera, discovering unseen regions, and revisiting previously observed areas in real time and over a long-term horizon. Achieving this capability requires a system-level co-design of control method, memory mechanism, and training strategy. We introduce a novel camera conditioning with a dense coordinate field whose renderings provide spatially aligned motion and orientation cues, allowing the model to interpret camera motion directly as visual evidence. To support fast and precise memory retrieval over a growing generation context, we propose an efficient sparse attention-based memory mechanism, enabling the model to selectively attend to a small set of relevant context tokens at inference time, regardless of actual context length. We further develop several techniques to rectify the self-forcing-style distillation pipeline, improving the student model's ability to respect control signals, as well as maintaining diverse generation modes and long-term memory from the teacher. Together, these components enable Wonder to synthesize diverse, minute-scale videos at 16 FPS while preserving coherent geometry, appearance, and dynamics across long rollouts. Beyond image-to-video generation, Wonder naturally supports video-conditioned generation, allowing existing dynamic scenes to be re-shot in real time.
展开完整摘要收起摘要↓
We present Wonder, a general-purpose video world model for real-time, camera-controllable world exploration. Given an image or a conditional video, Wonder constructs a playable world where users can navigate interactively by moving the camera, discovering unseen regions, and revisiting previously observed areas in real time and over a long-term horizon. Achieving this capability requires a system-level co-design of control method, memory mechanism, and training strategy. We introduce a novel camera conditioning with a dense coordinate field whose renderings provide spatially aligned motion and orientation cues, allowing the model to interpret camera motion directly as visual evidence. To support fast and precise memory retrieval over a growing generation context, we propose an efficient sparse attention-based memory mechanism, enabling the model to selectively attend to a small set of relevant context tokens at inference time, regardless of actual context length. We further develop several techniques to rectify the self-forcing-style distillation pipeline, improving the student model's ability to respect control signals, as well as maintaining diverse generation modes and long-term memory from the teacher. Together, these components enable Wonder to synthesize diverse, minute-scale videos at 16 FPS while preserving coherent geometry, appearance, and dynamics across long rollouts. Beyond image-to-video generation, Wonder naturally supports video-conditioned generation, allowing existing dynamic scenes to be re-shot in real time.
作者Stephen Bauer, Sheila Seidel, Shanza Iftikhar, Scott Veidenheimer, Gorkem Ulkar
Voice activity detection (VAD) triggers downstream speech processing in always-on systems under strict memory, latency, and compute constraints. Recent compact models report strong accuracy but rely on components that are not widely supported: learnable filterbanks, recurrent layers, or non-causal post-processing. We propose kiloVAD, designed for embedded inference using standard Mel features, CNN-only layers, and tunable context/spectral parameters. We introduce per-layer structured pruning with self-distillation and angle-based quantization-aware training (QAT) that outperforms standard QAT by 1-4%. Evaluated per-frame under causal conditions, kiloVAD achieves 0.850 AUC on AVA-Speech with 2.1 k parameters and 200 ms context, establishing a new state of the art for causal, deployment-ready VAD.
展开完整摘要收起摘要↓
Voice activity detection (VAD) triggers downstream speech processing in always-on systems under strict memory, latency, and compute constraints. Recent compact models report strong accuracy but rely on components that are not widely supported: learnable filterbanks, recurrent layers, or non-causal post-processing. We propose kiloVAD, designed for embedded inference using standard Mel features, CNN-only layers, and tunable context/spectral parameters. We introduce per-layer structured pruning with self-distillation and angle-based quantization-aware training (QAT) that outperforms standard QAT by 1-4%. Evaluated per-frame under causal conditions, kiloVAD achieves 0.850 AUC on AVA-Speech with 2.1 k parameters and 200 ms context, establishing a new state of the art for causal, deployment-ready VAD.
Deploying deep neural networks on resource-constrained hardware relies on mixed-precision quantisation (MPQ). current deployment toolchains severely fragment this process. Quantisation typically occurs as a hardware-agnostic preprocessing step in front-end frameworks, disconnected from the downstream compilers that generate the physical machine code. This separation leads to suboptimal configurations where assigned bit-widths map poorly to the target machine's heterogeneous hardware execution blocks such as tensor cores and variable-width vector units, incurring severe runtime execution penalties. Furthermore, evaluating these configurations via exhaustive hardware-in-the-loop (HIL) testing is intractable due to the exponentially large search space. We present CONQuER, a unified compiler-integrated infrastructure for hardware-aware MPQ. CONQuER shifts quantisation into the compiler pipeline at the TOSA level, enabling intelligent configuration handling based on compiler support. To evaluate this combinatorial search space of different of model layers within practical compilation budgets, CONQuER couples an NSGA-II evolutionary algorithm with a dual-surrogate prescreening engine. This engine evaluates theoretical cache memory bounds and feature space isotropy to discard non-viable configurations. CONQuER then executes only the strongest candidate policies on hardware via IREE, feeding the execution metrics into an online calibrator. This calibrator aligns the surrogate models with the true hardware behaviour during an NSGA-II evolutionary search. Evaluation across mobile and laptop CPUs, and server GPUs demonstrates that optimal quantisation policies are hardware-dependent. By coupling quantisation with compiler lowering and physical execution, CONQuER discovers Pareto-optimal configurations up to 12.19x faster inference with top-1 accuracy within 1.44% of the unquantised baseline.
展开完整摘要收起摘要↓
Deploying deep neural networks on resource-constrained hardware relies on mixed-precision quantisation (MPQ). current deployment toolchains severely fragment this process. Quantisation typically occurs as a hardware-agnostic preprocessing step in front-end frameworks, disconnected from the downstream compilers that generate the physical machine code. This separation leads to suboptimal configurations where assigned bit-widths map poorly to the target machine's heterogeneous hardware execution blocks such as tensor cores and variable-width vector units, incurring severe runtime execution penalties. Furthermore, evaluating these configurations via exhaustive hardware-in-the-loop (HIL) testing is intractable due to the exponentially large search space. We present CONQuER, a unified compiler-integrated infrastructure for hardware-aware MPQ. CONQuER shifts quantisation into the compiler pipeline at the TOSA level, enabling intelligent configuration handling based on compiler support. To evaluate this combinatorial search space of different of model layers within practical compilation budgets, CONQuER couples an NSGA-II evolutionary algorithm with a dual-surrogate prescreening engine. This engine evaluates theoretical cache memory bounds and feature space isotropy to discard non-viable configurations. CONQuER then executes only the strongest candidate policies on hardware via IREE, feeding the execution metrics into an online calibrator. This calibrator aligns the surrogate models with the true hardware behaviour during an NSGA-II evolutionary search. Evaluation across mobile and laptop CPUs, and server GPUs demonstrates that optimal quantisation policies are hardware-dependent. By coupling quantisation with compiler lowering and physical execution, CONQuER discovers Pareto-optimal configurations up to 12.19x faster inference with top-1 accuracy within 1.44% of the unquantised baseline.
作者Jianfei Ma, Zhaoxin Feng, Emmanuele Chersoni, Si Chen
Prompt compression shortens LLM input to reduce inference cost, yet existing methods score token importance through LM forward passes. It remains questionable whether such nuanced, costly token selection is necessary. Compression requires identifying informative content, a problem that linguistic research has long addressed through cues that can be operationalized as deterministic rules. We therefore ask: can linguistic rules alone serve as effective prompt compressors, without LM-based scoring at compression time? To address this, we conduct offline evolutionary search over lexical, syntactic, semantic, and discourse seeds to find competitive rule combinations. The resulting linguistic compressor requires no LM forward pass at deployment and uses only CPU-side processing for compression. We evaluate it with a dual-path protocol to balance compression quality and reconstruction fidelity. Across short passages, multi-document reasoning, and dialogue-memory QA datasets, evolved compressors achieve performance similar to that of recent advanced prompt-compression strategies. Performance is strongest under light-to-moderate compression and degrades as compression becomes more aggressive, while the Direct and Reconstruction paths exhibit distinct patterns. Evolutionary analysis reveals that effective compression fuses signals across linguistic levels and, as the compression ratio increases, rules shift from token pruning to sentence extraction.
展开完整摘要收起摘要↓
Prompt compression shortens LLM input to reduce inference cost, yet existing methods score token importance through LM forward passes. It remains questionable whether such nuanced, costly token selection is necessary. Compression requires identifying informative content, a problem that linguistic research has long addressed through cues that can be operationalized as deterministic rules. We therefore ask: can linguistic rules alone serve as effective prompt compressors, without LM-based scoring at compression time? To address this, we conduct offline evolutionary search over lexical, syntactic, semantic, and discourse seeds to find competitive rule combinations. The resulting linguistic compressor requires no LM forward pass at deployment and uses only CPU-side processing for compression. We evaluate it with a dual-path protocol to balance compression quality and reconstruction fidelity. Across short passages, multi-document reasoning, and dialogue-memory QA datasets, evolved compressors achieve performance similar to that of recent advanced prompt-compression strategies. Performance is strongest under light-to-moderate compression and degrades as compression becomes more aggressive, while the Direct and Reconstruction paths exhibit distinct patterns. Evolutionary analysis reveals that effective compression fuses signals across linguistic levels and, as the compression ratio increases, rules shift from token pruning to sentence extraction.
作者Haoyang Huang, Wenjie Huang, Tianqi Xu, Hongyaoxing Gu, Kang Tan, Yikai Fu, Yuhao Shen, Tianyu Liu, Baolin Zhang, Jun Zhang, Xinyi Hu, Jun Dai, Shuang Ge, Lei Chen, Yue Li, Mingchen Wang, Meng Zhang
Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.
展开完整摘要收起摘要↓
Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.
作者Jiabao Ji, Yujian Liu, Li An, Rohit Jain, Gungor Polatkan, Siyu Zhu, Shiyu Chang
Large language model agents often spend substantial wall-clock time waiting for tool call results. Tool-call speculation can hide this latency by predicting and pre-executing an agent's next tool call if the prediction matches the agent's eventual tool call, but existing speculators are typically separate draft models or cached traces that are poorly aligned with the deployed agent's own behavior. We identify this speculator-agent gap and show that the target agent itself is a strong next-call speculator. This points to a simpler design: unifying the agent and speculator within the same model. In this paper, we introduce the self-speculating agent, a single model that both solves tasks in agent mode and predicts its next tool call from partial trajectories in speculator mode, fully reusing prefix KV cache. To enable this dual-mode agent without degrading performance, we propose a joint agent-speculator reinforcement learning method, which derives speculation targets from the agent's own rollouts and alternates agent and speculator updates. Across agentic search QA and conversational tool-use agentic tasks, our method improves average next tool-call Hit@1 from 44.1 to 61.2 for Qwen3-4B and from 48.9 to 66.3 for Qwen3.5-4B, while preserving agent task success.
展开完整摘要收起摘要↓
Large language model agents often spend substantial wall-clock time waiting for tool call results. Tool-call speculation can hide this latency by predicting and pre-executing an agent's next tool call if the prediction matches the agent's eventual tool call, but existing speculators are typically separate draft models or cached traces that are poorly aligned with the deployed agent's own behavior. We identify this speculator-agent gap and show that the target agent itself is a strong next-call speculator. This points to a simpler design: unifying the agent and speculator within the same model. In this paper, we introduce the self-speculating agent, a single model that both solves tasks in agent mode and predicts its next tool call from partial trajectories in speculator mode, fully reusing prefix KV cache. To enable this dual-mode agent without degrading performance, we propose a joint agent-speculator reinforcement learning method, which derives speculation targets from the agent's own rollouts and alternates agent and speculator updates. Across agentic search QA and conversational tool-use agentic tasks, our method improves average next tool-call Hit@1 from 44.1 to 61.2 for Qwen3-4B and from 48.9 to 66.3 for Qwen3.5-4B, while preserving agent task success.
Knowledge-intensive multimodal question answering (KI-MMQA) sits at the intersection of three expensive primitives: long visual token sequences, dense retrieval over large external corpora, and full cross-modal fusion. Existing systems pay all three costs uniformly per query, even though only a small fraction of visual content and retrieved knowledge is actually relevant to any given question. We introduce SKIP (Salient Knowledge-Injected Pathways), a unified inference architecture that routes computation along sparse pathways jointly conditioned on the question, the image, and a difficulty estimate. SKIP combines question-guided visual token pruning, region-conditional sparse retrieval, bipartite sparse cross-attention, and speculative knowledge verification with an adaptive budget controller that allocates compute proportional to predicted question difficulty. We derive an information-bottleneck bound showing that the optimal visual sparsity rate scales as $O(1/\sqrt{N})$ under realistic question-image mutual-information assumptions, with retained accuracy guarantees. Across five KI-MMQA benchmarks (OK-VQA, A-OKVQA, InfoSeek, Encyclopedic-VQA, and ViQuAE), SKIP matches or exceeds the accuracy of strong dense baselines while using $3.4$--$6.8\times$ fewer FLOPs and $2.7\times$ less end-to-end latency. Code available at: https://pmlrbd.github.io/skip/
展开完整摘要收起摘要↓
Knowledge-intensive multimodal question answering (KI-MMQA) sits at the intersection of three expensive primitives: long visual token sequences, dense retrieval over large external corpora, and full cross-modal fusion. Existing systems pay all three costs uniformly per query, even though only a small fraction of visual content and retrieved knowledge is actually relevant to any given question. We introduce SKIP (Salient Knowledge-Injected Pathways), a unified inference architecture that routes computation along sparse pathways jointly conditioned on the question, the image, and a difficulty estimate. SKIP combines question-guided visual token pruning, region-conditional sparse retrieval, bipartite sparse cross-attention, and speculative knowledge verification with an adaptive budget controller that allocates compute proportional to predicted question difficulty. We derive an information-bottleneck bound showing that the optimal visual sparsity rate scales as $O(1/\sqrt{N})$ under realistic question-image mutual-information assumptions, with retained accuracy guarantees. Across five KI-MMQA benchmarks (OK-VQA, A-OKVQA, InfoSeek, Encyclopedic-VQA, and ViQuAE), SKIP matches or exceeds the accuracy of strong dense baselines while using $3.4$--$6.8\times$ fewer FLOPs and $2.7\times$ less end-to-end latency. Code available at: https://pmlrbd.github.io/skip/
作者Yufei Xue, Lin Niu, Hong Liu, Siran Liu, Hanyong Shao, Wei Liu, Guanghua Yu, Jianchen Zhu, Jun Zhang
The quadratic cost of self-attention makes long-context inference prohibitively expensive, and proxy-based block-sparse attention has become a practical remedy. Existing methods typically rely on a proxy to predict a binary sparse mask and a kernel to consume this mask and perform sparse attention computation. Such an approach is effective under moderate budgets. However, as the budget tightens, the estimated proxy inevitably drops some salient blocks, while the kernel can only apply the sparse mask mechanically, leading to an evident drop in model accuracy. We propose CoSA, a two-stage training-free Sparse Attention under proxy-kernel CO-design, which couples a Kernel-Aware Proxy (KAP) with an Ordered-Skipping Kernel (OSK). In the first stage, the KAP selects blocks under a moderate budget and produces an ordered mask that prescribes the order in which KV pages are visited in the kernel inner loop. In the second stage, the OSK applies this mask and skips more blocks under a tightened budget given online-softmax statistics. Across mainstream LLM backbones and long-context benchmarks, CoSA attains higher accuracy at lower budgets. Impressively, CoSA achieves a 4.93$\times$ attention speedup and reduces end-to-end Time-to-First-Token by 2.53$\times$ under a context length of 128K with negligible performance degradation.
展开完整摘要收起摘要↓
The quadratic cost of self-attention makes long-context inference prohibitively expensive, and proxy-based block-sparse attention has become a practical remedy. Existing methods typically rely on a proxy to predict a binary sparse mask and a kernel to consume this mask and perform sparse attention computation. Such an approach is effective under moderate budgets. However, as the budget tightens, the estimated proxy inevitably drops some salient blocks, while the kernel can only apply the sparse mask mechanically, leading to an evident drop in model accuracy. We propose CoSA, a two-stage training-free Sparse Attention under proxy-kernel CO-design, which couples a Kernel-Aware Proxy (KAP) with an Ordered-Skipping Kernel (OSK). In the first stage, the KAP selects blocks under a moderate budget and produces an ordered mask that prescribes the order in which KV pages are visited in the kernel inner loop. In the second stage, the OSK applies this mask and skips more blocks under a tightened budget given online-softmax statistics. Across mainstream LLM backbones and long-context benchmarks, CoSA attains higher accuracy at lower budgets. Impressively, CoSA achieves a 4.93$\times$ attention speedup and reduces end-to-end Time-to-First-Token by 2.53$\times$ under a context length of 128K with negligible performance degradation.
Speculative decoding accelerates large language model inference without changing the target distribution, but no single drafting structure performs best across real-world workloads. Autoregressive multi-token prediction (MTP) is a lightweight, stable proposal mechanism, whereas block-parallel diffusion amortizes drafting latency over much longer candidate sequences; the better choice depends strongly on the output distribution. We present AngelSpec, a unified training framework for MTP and block-parallel speculative decoding that addresses this heterogeneity at three levels. At the training level, rather than fitting one universal drafter to a uniform data mixture, we co-specialize structure and data: the MTP drafter is trained on diverse conversational data for high-entropy open-ended chat, and the block-diffusion drafter on code and mathematics data for longer predictable continuations. At the architecture level, we propose DFly, a block-diffusion framework combining a hybrid target-conditioning backbone with a predecessor-conditioned autoregressive head, improving target-feature utilization and intra-block dependency modeling while keeping generation parallel. At the inference level, both acceptance length and verification cost vary with domain, request, online load, and hardware, so DFly treats verification as a shared batch-level resource: it reallocates compute toward high-confidence prefixes across requests and combines expected utility with a profiled cost model to adapt verification depth online. Across the Hy3 series, DFly raises the average accepted length on Hy3-A21B by roughly 30% and attains the highest average throughput at every tested concurrency from 4 to 64, a 1.98-2.40x speedup over autoregressive decoding and 10.5-11.8% higher throughput than DFlash. We release AngelSpec to support training and extending these methods.
展开完整摘要收起摘要↓
Speculative decoding accelerates large language model inference without changing the target distribution, but no single drafting structure performs best across real-world workloads. Autoregressive multi-token prediction (MTP) is a lightweight, stable proposal mechanism, whereas block-parallel diffusion amortizes drafting latency over much longer candidate sequences; the better choice depends strongly on the output distribution. We present AngelSpec, a unified training framework for MTP and block-parallel speculative decoding that addresses this heterogeneity at three levels. At the training level, rather than fitting one universal drafter to a uniform data mixture, we co-specialize structure and data: the MTP drafter is trained on diverse conversational data for high-entropy open-ended chat, and the block-diffusion drafter on code and mathematics data for longer predictable continuations. At the architecture level, we propose DFly, a block-diffusion framework combining a hybrid target-conditioning backbone with a predecessor-conditioned autoregressive head, improving target-feature utilization and intra-block dependency modeling while keeping generation parallel. At the inference level, both acceptance length and verification cost vary with domain, request, online load, and hardware, so DFly treats verification as a shared batch-level resource: it reallocates compute toward high-confidence prefixes across requests and combines expected utility with a profiled cost model to adapt verification depth online. Across the Hy3 series, DFly raises the average accepted length on Hy3-A21B by roughly 30% and attains the highest average throughput at every tested concurrency from 4 to 64, a 1.98-2.40x speedup over autoregressive decoding and 10.5-11.8% higher throughput than DFlash. We release AngelSpec to support training and extending these methods.
Prefill-decode disaggregation (PD) and roofline-based operator placement are common strategies for partitioning Large Language Model (LLM) inference across heterogeneous systems, but they are often insufficient in practice. End-to-end latency also depends on workload shape, runtime device contention, and persistent weight layout. We present DOPS (dynamic operator scheduling), a hardware-aware, closed-loop framework that jointly optimizes operator scheduling and blockwise weight layouts. DOPS constructs a stage-aware directed acyclic graph (DAG) and integrates two components: the Bifocal scheduler for dynamic operator-to-device placement and the Weight Layout Arbiter (WLA) for selecting hardware-efficient weight layouts under strict memory constraints. Across representative heterogeneous systems combining neural processing units (NPUs) and processing-in-memory (PIM) devices, Bifocal achieves geometric-mean speedups of 1.20$\times$ to 2.23$\times$ over the PD baseline. WLA provides an additional geometric-mean speedup of 1.28$\times$ to 1.33$\times$ over Bifocal/Linear. DOPS also supports systematic analysis of workload sensitivity and hardware scalability for LLM serving. The source code is available at https://github.com/YIAI-02/TriForm, and the visualization tool is demonstrated at https://youtu.be/Ya_oMCyYno0.
展开完整摘要收起摘要↓
Prefill-decode disaggregation (PD) and roofline-based operator placement are common strategies for partitioning Large Language Model (LLM) inference across heterogeneous systems, but they are often insufficient in practice. End-to-end latency also depends on workload shape, runtime device contention, and persistent weight layout. We present DOPS (dynamic operator scheduling), a hardware-aware, closed-loop framework that jointly optimizes operator scheduling and blockwise weight layouts. DOPS constructs a stage-aware directed acyclic graph (DAG) and integrates two components: the Bifocal scheduler for dynamic operator-to-device placement and the Weight Layout Arbiter (WLA) for selecting hardware-efficient weight layouts under strict memory constraints. Across representative heterogeneous systems combining neural processing units (NPUs) and processing-in-memory (PIM) devices, Bifocal achieves geometric-mean speedups of 1.20$\times$ to 2.23$\times$ over the PD baseline. WLA provides an additional geometric-mean speedup of 1.28$\times$ to 1.33$\times$ over Bifocal/Linear. DOPS also supports systematic analysis of workload sensitivity and hardware scalability for LLM serving. The source code is available at https://github.com/YIAI-02/TriForm, and the visualization tool is demonstrated at https://youtu.be/Ya_oMCyYno0.
Fine-grained weight pruning and activation sparsification have emerged as effective approaches for reducing the compute and memory cost of inference for Transformer models. In the moderate-sparsity regime, Gustavson's dataflow provides a natural execution model for exploiting both activation and weight sparsity on vector processors through metadata-driven indexed accumulation. However, existing RVV architectures lack native support for this pattern, forcing kernels to rely on software index decoding and L1-backed indexed memory operations that keep sparse tensor contractions far below their roofline performance bound. We present Ventaglio, a runtime-configurable sparse execution unit coupled with RVV ISA extensions that drives sparse tensor contractions toward their roofline through indexed gather-accumulate-scatter support. Integrated into an open-source vector processing cluster and implemented in 12nm FinFET, Ventaglio accelerates sparse tensor contraction kernels by $6.9\text{--}7.4\times$ over optimized RVV baselines, with only $3.1%$ area overhead for a cluster of tightly-L1 coupled vector processing elements. We build a performance-accurate instruction-level model of the Ventaglio extension, calibrate it against RTL implementation, and leverage it for scale-out performance analysis on a large $4\times4$ multi-cluster system. Using a DuoGPT-pruned LLaMA-3-8B model with practical $40\text{--}60%$ dual sparsity, Ventaglio achieves $2.40\text{--}5.25\times$ and $2.06\text{--}3.16\times$ speedup over dense baselines during prefill and autoregressive decoding, respectively.
展开完整摘要收起摘要↓
Fine-grained weight pruning and activation sparsification have emerged as effective approaches for reducing the compute and memory cost of inference for Transformer models. In the moderate-sparsity regime, Gustavson's dataflow provides a natural execution model for exploiting both activation and weight sparsity on vector processors through metadata-driven indexed accumulation. However, existing RVV architectures lack native support for this pattern, forcing kernels to rely on software index decoding and L1-backed indexed memory operations that keep sparse tensor contractions far below their roofline performance bound. We present Ventaglio, a runtime-configurable sparse execution unit coupled with RVV ISA extensions that drives sparse tensor contractions toward their roofline through indexed gather-accumulate-scatter support. Integrated into an open-source vector processing cluster and implemented in 12nm FinFET, Ventaglio accelerates sparse tensor contraction kernels by $6.9\text{--}7.4\times$ over optimized RVV baselines, with only $3.1%$ area overhead for a cluster of tightly-L1 coupled vector processing elements. We build a performance-accurate instruction-level model of the Ventaglio extension, calibrate it against RTL implementation, and leverage it for scale-out performance analysis on a large $4\times4$ multi-cluster system. Using a DuoGPT-pruned LLaMA-3-8B model with practical $40\text{--}60%$ dual sparsity, Ventaglio achieves $2.40\text{--}5.25\times$ and $2.06\text{--}3.16\times$ speedup over dense baselines during prefill and autoregressive decoding, respectively.
作者Shuyue Wei, Chang Liu, Zimu Zhou, Yongxin Tong, Lizhen Cui
Recently, memory management has become a key infrastructure for LLM-based agents, as it directly affects long-horizon reasoning, personalized responses, and knowledge reuse. However, existing LLM memory systems typically adopt a coarse-grained (utility-agnostic) manner that treats heterogeneous user-LLM interaction records uniformly, leading to redundant and low-impact records persisting in the memory repository. To address this challenge, we present MemLens, a value-aware memory management system that takes memory records as first-class data objects. MemLens provides an end-to-end interactive analytics dashboard that exposes the complete memory lifecycle, including Shapley-style memory evaluation, value-aware storage, and memory-assisted response. Through a study-copilot application, the system enables users to inspect memory values, visualize hierarchical memory structures, and compare various memory management strategies in terms of response quality, retrieval latency, and token consumption. Therefore, our MemLens can serve as an efficient, interpretable, and personalized long-term memory management system for LLM-based agents.
展开完整摘要收起摘要↓
Recently, memory management has become a key infrastructure for LLM-based agents, as it directly affects long-horizon reasoning, personalized responses, and knowledge reuse. However, existing LLM memory systems typically adopt a coarse-grained (utility-agnostic) manner that treats heterogeneous user-LLM interaction records uniformly, leading to redundant and low-impact records persisting in the memory repository. To address this challenge, we present MemLens, a value-aware memory management system that takes memory records as first-class data objects. MemLens provides an end-to-end interactive analytics dashboard that exposes the complete memory lifecycle, including Shapley-style memory evaluation, value-aware storage, and memory-assisted response. Through a study-copilot application, the system enables users to inspect memory values, visualize hierarchical memory structures, and compare various memory management strategies in terms of response quality, retrieval latency, and token consumption. Therefore, our MemLens can serve as an efficient, interpretable, and personalized long-term memory management system for LLM-based agents.
作者Yuchen Wang, Qihui Zhu, Yang Liu, Xiaoyan Sun, Siying Wu
Recent multimodal large language models (MLLMs), such as Qwen2.5-VL and InternVL3, generate large numbers of vision tokens for high-resolution inputs, leading to substantial computational cost. Existing vision token pruning methods either depend on cross-modal attention and cannot prune before the prefill stage, or rely on diversity estimation with high computational overhead. We observe that attention scores from both vision and text tokens peak at modality separator tokens, suggesting that these separators bridge the two modalities. Based on this observation, we propose SepPrune, an efficient, training-free, plug-and-play pruning method that uses the separator token as a unified query to rank and select informative vision tokens. SepPrune reuses the LLM's built-in projection parameters and requires no architectural changes. Experiments on Qwen2.5-VL-7B show that SepPrune achieves state-of-the-art performance, retaining 96.3% of the original accuracy while removing 80.2% of vision tokens.
展开完整摘要收起摘要↓
Recent multimodal large language models (MLLMs), such as Qwen2.5-VL and InternVL3, generate large numbers of vision tokens for high-resolution inputs, leading to substantial computational cost. Existing vision token pruning methods either depend on cross-modal attention and cannot prune before the prefill stage, or rely on diversity estimation with high computational overhead. We observe that attention scores from both vision and text tokens peak at modality separator tokens, suggesting that these separators bridge the two modalities. Based on this observation, we propose SepPrune, an efficient, training-free, plug-and-play pruning method that uses the separator token as a unified query to rank and select informative vision tokens. SepPrune reuses the LLM's built-in projection parameters and requires no architectural changes. Experiments on Qwen2.5-VL-7B show that SepPrune achieves state-of-the-art performance, retaining 96.3% of the original accuracy while removing 80.2% of vision tokens.
作者Harrison J. Goldwyn, Graham Johnson, Christopher Ibarra, Lace Padilla, Kenny Gruchalla
Empirical user studies are essential for evaluating visual encodings and can reveal perceptual and cognitive mechanisms, but they do not by themselves provide causal, predictive accounts of interpretation errors. Evaluations are therefore often post hoc: they measure performance after a design has been specified rather than predicting how attention, uncertainty, memory, and bias may produce accurate or erroneous judgments. To address this mechanistic gap, we translate a cognitive theory of visualization interpretation into executable simulation using Active Inference, a probabilistic framework for perception, learning, and action. We model chart reading as dynamic visual search in which agents update beliefs and choose actions that balance uncertainty reduction against cognitive effort. As a proof of concept, we implement Fast, heuristic (Type 1) and Slow, analytic (Type 2) agents for a bar-chart average-estimation task. The Fast agent is vulnerable to tick-salience bias, whereas the Slow agent is more vulnerable to working-memory decay. Both produce inspectable cognitive traces, including evolving belief uncertainty and fixation sequences. By expressing these hypothesized failure mechanisms as interpretable parameters, the architecture provides a framework for formalizing and testing mechanistic hypotheses about visualization interpretation. Empirical studies can then parameterize, refine, or falsify these simulations, supporting earlier and more predictive in silico evaluation of visualization efficacy.
展开完整摘要收起摘要↓
Empirical user studies are essential for evaluating visual encodings and can reveal perceptual and cognitive mechanisms, but they do not by themselves provide causal, predictive accounts of interpretation errors. Evaluations are therefore often post hoc: they measure performance after a design has been specified rather than predicting how attention, uncertainty, memory, and bias may produce accurate or erroneous judgments. To address this mechanistic gap, we translate a cognitive theory of visualization interpretation into executable simulation using Active Inference, a probabilistic framework for perception, learning, and action. We model chart reading as dynamic visual search in which agents update beliefs and choose actions that balance uncertainty reduction against cognitive effort. As a proof of concept, we implement Fast, heuristic (Type 1) and Slow, analytic (Type 2) agents for a bar-chart average-estimation task. The Fast agent is vulnerable to tick-salience bias, whereas the Slow agent is more vulnerable to working-memory decay. Both produce inspectable cognitive traces, including evolving belief uncertainty and fixation sequences. By expressing these hypothesized failure mechanisms as interpretable parameters, the architecture provides a framework for formalizing and testing mechanistic hypotheses about visualization interpretation. Empirical studies can then parameterize, refine, or falsify these simulations, supporting earlier and more predictive in silico evaluation of visualization efficacy.
This paper studies the behavior of language models in a multi-agent crowd simulation, focusing on how affect propagates among agents that perceive and appraise one another. Each agent perceives its neighbors through visual, auditory, and tactile channels, then appraises these perceptions in light of its prompted personality profile, memory, current affective state, and situational context. Appraisal is carried out by an LLM, which updates the agent's internal affective state and selects its outward expression. The architecture contains no hand-authored mechanism for directly transferring affective state between agents; instead, inter-agent influence arises through the perception-appraisal-expression loop. The agent representation draws on the Big Five personality model and Russell's circumplex model of affect. To limit latency, low-level steering and navigation are handled by a conventional crowd simulator operating independently of the LLM-based cognitive layer. We evaluate the architecture across five scenario environments spanning alarming, joyful, and neutral situations in different spatial layouts. The results show that the system produces emotional contagion dynamics with spatial, temporal, and personality-dependent structure in sparse, small crowds. Alarm spreads from seeded agents as a traveling front, the mean alarmed fraction settles at a nonzero plateau, and the distribution of prompted personality profiles determines whether an ambiguous alarm ignites panic and whether a provocation is interpreted as anger or fear. We further evaluate the appraisal step through controlled experiments across prompt variants, sampling temperatures, and four model backends, showing that the dynamics are backend-dependent.
展开完整摘要收起摘要↓
This paper studies the behavior of language models in a multi-agent crowd simulation, focusing on how affect propagates among agents that perceive and appraise one another. Each agent perceives its neighbors through visual, auditory, and tactile channels, then appraises these perceptions in light of its prompted personality profile, memory, current affective state, and situational context. Appraisal is carried out by an LLM, which updates the agent's internal affective state and selects its outward expression. The architecture contains no hand-authored mechanism for directly transferring affective state between agents; instead, inter-agent influence arises through the perception-appraisal-expression loop. The agent representation draws on the Big Five personality model and Russell's circumplex model of affect. To limit latency, low-level steering and navigation are handled by a conventional crowd simulator operating independently of the LLM-based cognitive layer. We evaluate the architecture across five scenario environments spanning alarming, joyful, and neutral situations in different spatial layouts. The results show that the system produces emotional contagion dynamics with spatial, temporal, and personality-dependent structure in sparse, small crowds. Alarm spreads from seeded agents as a traveling front, the mean alarmed fraction settles at a nonzero plateau, and the distribution of prompted personality profiles determines whether an ambiguous alarm ignites panic and whether a provocation is interpreted as anger or fear. We further evaluate the appraisal step through controlled experiments across prompt variants, sampling temperatures, and four model backends, showing that the dynamics are backend-dependent.