Mixed-precision quantization must decide which parts of a model to keep at higher precision. A common premise, shared by sensitivity-based methods such as HAWQ and CoopQ, is that the loss from quantizing a set of layers can be reconstructed from per-layer or pairwise sensitivities measured in isolation. We test this premise at the 4-bit weight-and-activation precisions now being deployed, treating the change in loss $f(S)$ from quantizing a layer set $S$ as a set function on the Boolean cube and analyzing it through two classical changes of basis. This analysis yields two findings. First, across configurations drawn from the deployment distribution, 85--93% of the variance of $f$ is explained by per-layer effects alone. Second, a monotone transform of a sum of per-layer terms reproduces $f$'s ranking of configurations, misordering at most 2% of pairs. We propose the coverage model $f(S)=c\bigl(1-\prod_{i\in S}(1-a_i)\bigr)$, which reproduces the measured variance profile of $f$ to within a few percent from its $L$ fitted break-rates. This structure supports two predictors of a configuration's loss, each with $L+1$ parameters. The additive model is the optimal first-order predictor. By Parseval's identity its mean-squared error equals the variance of $f$ left unexplained by per-layer effects, which we measure on full lattices, estimate out of sample at full-network scale, and report with every result as a certificate of how well any additive model can do. The coverage model itself is the second predictor. As allocators at matched memory, they attain the lowest KL divergence among the compared allocators on models from 30B to 355B parameters. Below four bits, the resulting allocations continue to solve code and reasoning tasks at budgets where allocations from gradient sensitivities no longer produce terminating generations.
展开完整摘要收起摘要↓
Mixed-precision quantization must decide which parts of a model to keep at higher precision. A common premise, shared by sensitivity-based methods such as HAWQ and CoopQ, is that the loss from quantizing a set of layers can be reconstructed from per-layer or pairwise sensitivities measured in isolation. We test this premise at the 4-bit weight-and-activation precisions now being deployed, treating the change in loss $f(S)$ from quantizing a layer set $S$ as a set function on the Boolean cube and analyzing it through two classical changes of basis. This analysis yields two findings. First, across configurations drawn from the deployment distribution, 85--93% of the variance of $f$ is explained by per-layer effects alone. Second, a monotone transform of a sum of per-layer terms reproduces $f$'s ranking of configurations, misordering at most 2% of pairs. We propose the coverage model $f(S)=c\bigl(1-\prod_{i\in S}(1-a_i)\bigr)$, which reproduces the measured variance profile of $f$ to within a few percent from its $L$ fitted break-rates. This structure supports two predictors of a configuration's loss, each with $L+1$ parameters. The additive model is the optimal first-order predictor. By Parseval's identity its mean-squared error equals the variance of $f$ left unexplained by per-layer effects, which we measure on full lattices, estimate out of sample at full-network scale, and report with every result as a certificate of how well any additive model can do. The coverage model itself is the second predictor. As allocators at matched memory, they attain the lowest KL divergence among the compared allocators on models from 30B to 355B parameters. Below four bits, the resulting allocations continue to solve code and reasoning tasks at budgets where allocations from gradient sensitivities no longer produce terminating generations.
作者Md Mahfuzur Rahman, Pengzhan Zhou, A F M Abdun Noor, Md Imam Ahasan, Kah Ong Michael Goh, S. M. Hasan Mahmud, Md Mustafizur Rahman, Kaixin Gao
Predicting pedestrian crossing intention is a safety-critical task for autonomous driving, yet existing approaches often rely on single-modal inputs or dense multimodal fusion strategies that inadequately capture complementary visual and kinematic information while introducing redundant inter-modal interactions. We propose ADAPT (Adaptive Domain-Aware Pedestrian Crossing Transformer), a multimodal framework that jointly models local and global visual context together with temporal motion dynamics for accurate pedestrian crossing intention prediction. ADAPT processes four spatially aligned visual modalities, including RGB images, local depth maps, global semantic maps, and global depth maps, together with ego-vehicle speed, pedestrian bounding boxes, and skeleton pose information through five specialized modules: a weight-shared Swin Transformer V2 backbone for visual feature extraction, a Cross-Modality Guided Attention module for hierarchical visual fusion, a Mamba-based Motion Feature Encoding module for efficient temporal modeling, a Sparse Cross-Modal Attention module that selectively preserves the most informative inter-modal interactions, and a Vision Transformer-based Temporal Feature Fusion module for sequence-level prediction. Extensive experiments on the JAAD and PIE benchmark datasets demonstrate that ADAPT consistently outperforms existing state-of-the-art methods while maintaining low computational complexity. On JAAD, the proposed method achieves an AUC of 0.73 on JAADbeh and 0.85 on JAADall, while on PIE it achieves an accuracy of 0.92 and an AUC of 0.90. Furthermore, ADAPT performs inference in only 17.23 ms per sample, offering an effective balance between predictive accuracy and real-time deployment efficiency for intelligent transportation and autonomous driving applications.
展开完整摘要收起摘要↓
Predicting pedestrian crossing intention is a safety-critical task for autonomous driving, yet existing approaches often rely on single-modal inputs or dense multimodal fusion strategies that inadequately capture complementary visual and kinematic information while introducing redundant inter-modal interactions. We propose ADAPT (Adaptive Domain-Aware Pedestrian Crossing Transformer), a multimodal framework that jointly models local and global visual context together with temporal motion dynamics for accurate pedestrian crossing intention prediction. ADAPT processes four spatially aligned visual modalities, including RGB images, local depth maps, global semantic maps, and global depth maps, together with ego-vehicle speed, pedestrian bounding boxes, and skeleton pose information through five specialized modules: a weight-shared Swin Transformer V2 backbone for visual feature extraction, a Cross-Modality Guided Attention module for hierarchical visual fusion, a Mamba-based Motion Feature Encoding module for efficient temporal modeling, a Sparse Cross-Modal Attention module that selectively preserves the most informative inter-modal interactions, and a Vision Transformer-based Temporal Feature Fusion module for sequence-level prediction. Extensive experiments on the JAAD and PIE benchmark datasets demonstrate that ADAPT consistently outperforms existing state-of-the-art methods while maintaining low computational complexity. On JAAD, the proposed method achieves an AUC of 0.73 on JAADbeh and 0.85 on JAADall, while on PIE it achieves an accuracy of 0.92 and an AUC of 0.90. Furthermore, ADAPT performs inference in only 17.23 ms per sample, offering an effective balance between predictive accuracy and real-time deployment efficiency for intelligent transportation and autonomous driving applications.
We present WanToFight, a generative game engine that simulates real-time, two-player The King of Fighters '97 (KOF~'97) gameplay from keyboard input. Prior generative game engines target either single-player first-person settings or non-real-time cooperative scenarios; multi-player control, real-time inference, complex physical interaction, and adversarial gameplay have not been jointly addressed. WanToFight closes this gap with three components built on the Wan-1.3B video diffusion transformer: a streaming autoregressive generator with block-causal attention and a rolling KV cache; a visually grounded Player Association module that binds each player's keyboard signal to a character identity; and a gated, locally causal keyboard injection module trained with a single-player-to-full-gameplay curriculum. A four-step DMD-distilled student paired with a pruned VAE decoder sustains 30FPS at 512x384 on a single NVIDIA RTX 5090 over the duration of a complete match. To our knowledge, WanToFight is the first generative game engine to combine multi-player control, real-time inference, complex physical interaction, and adversarial gameplay in one system.
展开完整摘要收起摘要↓
We present WanToFight, a generative game engine that simulates real-time, two-player The King of Fighters '97 (KOF~'97) gameplay from keyboard input. Prior generative game engines target either single-player first-person settings or non-real-time cooperative scenarios; multi-player control, real-time inference, complex physical interaction, and adversarial gameplay have not been jointly addressed. WanToFight closes this gap with three components built on the Wan-1.3B video diffusion transformer: a streaming autoregressive generator with block-causal attention and a rolling KV cache; a visually grounded Player Association module that binds each player's keyboard signal to a character identity; and a gated, locally causal keyboard injection module trained with a single-player-to-full-gameplay curriculum. A four-step DMD-distilled student paired with a pruned VAE decoder sustains 30FPS at 512x384 on a single NVIDIA RTX 5090 over the duration of a complete match. To our knowledge, WanToFight is the first generative game engine to combine multi-player control, real-time inference, complex physical interaction, and adversarial gameplay in one system.
Vision Transformers (ViTs) achieve strong accuracy but incur high inference latency. Semi-structured N:M sparsity can reduce arithmetic cost, yet its theoretical savings often fail to translate into proportional end-to-end speedups on modern GPUs. This mismatch arises because deployment latency depends not only on arithmetic reduction but also on execution regularity and hardware scheduling under sparsity. Achieving practical acceleration, therefore, requires coordinated design across sparse execution and sparsity configuration. To this end, we propose a hardware-software co-design framework for N:M sparse ViT inference. On the hardware side, we design MD-SpMM, an N:M sparse CUDA kernel that reorganizes sparse GEMM into micro-dense, Tensor-Core-aligned dataflow and uses inference-aware adaptive parallelism to sustain utilization. On the software side, we perform layer-wise sparsity search under explicit end-to-end latency budgets using a three-stage heuristic search with constraint relaxation to avoid premature convergence and enable deployment-aware sparsity allocation. Experiments on multiple ViT/Swin models and GPU platforms show that the framework achieves over 2.2x latency speedup while maintaining comparable accuracy and delivering superior accuracy under the same latency constraint. The source code is publicly available at https://github.com/liuganhuo/realizable-nm-sparse-transformer.
展开完整摘要收起摘要↓
Vision Transformers (ViTs) achieve strong accuracy but incur high inference latency. Semi-structured N:M sparsity can reduce arithmetic cost, yet its theoretical savings often fail to translate into proportional end-to-end speedups on modern GPUs. This mismatch arises because deployment latency depends not only on arithmetic reduction but also on execution regularity and hardware scheduling under sparsity. Achieving practical acceleration, therefore, requires coordinated design across sparse execution and sparsity configuration. To this end, we propose a hardware-software co-design framework for N:M sparse ViT inference. On the hardware side, we design MD-SpMM, an N:M sparse CUDA kernel that reorganizes sparse GEMM into micro-dense, Tensor-Core-aligned dataflow and uses inference-aware adaptive parallelism to sustain utilization. On the software side, we perform layer-wise sparsity search under explicit end-to-end latency budgets using a three-stage heuristic search with constraint relaxation to avoid premature convergence and enable deployment-aware sparsity allocation. Experiments on multiple ViT/Swin models and GPU platforms show that the framework achieves over 2.2x latency speedup while maintaining comparable accuracy and delivering superior accuracy under the same latency constraint. The source code is publicly available at https://github.com/liuganhuo/realizable-nm-sparse-transformer.
Sparse Mixture-of-Experts (MoE) models have become an important approach for scaling Large Language Models (LLMs), but their inference efficiency depends strongly on expert activation patterns. Speculative decoding (SD) accelerates autoregressive generation by verifying multiple draft tokens in parallel, yet existing draft selection strategies primarily optimize acceptance likelihood. In large-scale MoE models, however, selecting draft tokens also determines the union of experts activated during verification. We observe that confidence-driven SD can introduce expert scattering: high-probability draft tokens may route to disjoint experts, increasing expert-weight memory traffic and reducing the speedup from speculation. Motivated by this observation, we revisit draft-tree selection under the non-uniform memory-cost structure of MoE inference. We propose \textsc{EcoSpec}, a cost-aware speculative decoding framework that incorporates predicted marginal expert activation cost into draft selection. With a lightweight expert predictor and a dynamic expert buffer, \textsc{EcoSpec} favors draft paths that preserve high acceptance likelihood while reusing experts already covered by the current verification set, without modifying the target-model verification rule. We evaluate \textsc{EcoSpec} on three large-scale MoE models, including DeepSeek-V3.1 (671B), Qwen3-235B-A22B, and GPT-OSS-120B, across reasoning, coding, question-answering, and dialogue benchmarks. \textsc{EcoSpec} consistently reduces active expert footprints and improves end-to-end decoding speed, achieving up to $1.62\times$ speedup. These results show that accounting for expert activation cost is important for efficient speculative decoding in large-scale MoE models.
展开完整摘要收起摘要↓
Sparse Mixture-of-Experts (MoE) models have become an important approach for scaling Large Language Models (LLMs), but their inference efficiency depends strongly on expert activation patterns. Speculative decoding (SD) accelerates autoregressive generation by verifying multiple draft tokens in parallel, yet existing draft selection strategies primarily optimize acceptance likelihood. In large-scale MoE models, however, selecting draft tokens also determines the union of experts activated during verification. We observe that confidence-driven SD can introduce expert scattering: high-probability draft tokens may route to disjoint experts, increasing expert-weight memory traffic and reducing the speedup from speculation. Motivated by this observation, we revisit draft-tree selection under the non-uniform memory-cost structure of MoE inference. We propose \textsc{EcoSpec}, a cost-aware speculative decoding framework that incorporates predicted marginal expert activation cost into draft selection. With a lightweight expert predictor and a dynamic expert buffer, \textsc{EcoSpec} favors draft paths that preserve high acceptance likelihood while reusing experts already covered by the current verification set, without modifying the target-model verification rule. We evaluate \textsc{EcoSpec} on three large-scale MoE models, including DeepSeek-V3.1 (671B), Qwen3-235B-A22B, and GPT-OSS-120B, across reasoning, coding, question-answering, and dialogue benchmarks. \textsc{EcoSpec} consistently reduces active expert footprints and improves end-to-end decoding speed, achieving up to $1.62\times$ speedup. These results show that accounting for expert activation cost is important for efficient speculative decoding in large-scale MoE models.
Vision-language models (VLMs) process large numbers of visual tokens, resulting in substantial inference latency and memory overhead. This has motivated extensive research on visual token compression. While training-free strategies rely on heuristic metrics and suffer significant performance degradation under high compression ratios, many training-based methods introduce external compression modules that force the VLM backbone to adapt, incurring substantial retraining cost and compromising VLMs' priors. Effective visual token compression hinges on strong information encoding, a capability already present in pretrained VLMs but underutilized by existing approaches. Motivated by this, we propose VisCo, a training-efficient self-compression framework that reuses the pretrained VLM itself as an intrinsic compressor. VisCo is a parameter-sharing autoencoder that compresses visual information using a small set of memory tokens and transfers hierarchical information from encoding to decoding. Experiments show that VisCo surpasses prior methods across all evaluated compression ratios, with larger gains under more aggressive compression, and remains stable even in the extreme single-token setting. Moreover, when combined with the original visual tokens, the learned memory tokens can even improve the base model, suggesting that VisCo captures complementary representations beyond compression.
展开完整摘要收起摘要↓
Vision-language models (VLMs) process large numbers of visual tokens, resulting in substantial inference latency and memory overhead. This has motivated extensive research on visual token compression. While training-free strategies rely on heuristic metrics and suffer significant performance degradation under high compression ratios, many training-based methods introduce external compression modules that force the VLM backbone to adapt, incurring substantial retraining cost and compromising VLMs' priors. Effective visual token compression hinges on strong information encoding, a capability already present in pretrained VLMs but underutilized by existing approaches. Motivated by this, we propose VisCo, a training-efficient self-compression framework that reuses the pretrained VLM itself as an intrinsic compressor. VisCo is a parameter-sharing autoencoder that compresses visual information using a small set of memory tokens and transfers hierarchical information from encoding to decoding. Experiments show that VisCo surpasses prior methods across all evaluated compression ratios, with larger gains under more aggressive compression, and remains stable even in the extreme single-token setting. Moreover, when combined with the original visual tokens, the learned memory tokens can even improve the base model, suggesting that VisCo captures complementary representations beyond compression.
The key-value (KV) cache has become the dominant memory cost of transformer inference: it grows with batch size, context length, and depth, and at long context it, rather than the model weights, sets the throughput ceiling. Existing reductions fall into two families. Low-rank methods factor two-dimensional slices of the cache, either per-head matrices or cross-layer feature blocks, and quantization methods lower the bit-width of every entry. Neither exploits the fact that the cache at a layer is naturally a third-order tensor whose three axes, the heads, the tokens, and the features, carry very different amounts of redundancy. We take this tensor view directly. Our method, JoLT (Joint Lagrangian Tucker), applies a partial Tucker decomposition that compresses only the token and feature axes while leaving the head and layer axes intact, then restores the energy that truncation discards with a rotated low-bit residual: a random orthogonal rotation followed by low-bit quantization. A single Lagrangian dual allocates the Tucker ranks and the residual bit-widths together, per layer group and separately for keys and values, under one byte budget. The result is a near-lossless 2-3x compression. Perplexity stays near-lossless on both a grouped-query-attention model (Mistral-7B-v0.3) and a multi-head-attention model (LLaMA-2-13B), and GSM8K accuracy and needle-in-a-haystack retrieval hold at the uncompressed baseline at 2x on both architectures and through 3x on the GQA model. At 2x, JoLT reconstructs the cache to relative Frobenius error 0.009 (K) and 0.006 (V) on both architectures. A randomized-SVD variant, FlashJoLT, delivers a 5-13x compression-time speedup at 1024-token context and matched quality.
展开完整摘要收起摘要↓
The key-value (KV) cache has become the dominant memory cost of transformer inference: it grows with batch size, context length, and depth, and at long context it, rather than the model weights, sets the throughput ceiling. Existing reductions fall into two families. Low-rank methods factor two-dimensional slices of the cache, either per-head matrices or cross-layer feature blocks, and quantization methods lower the bit-width of every entry. Neither exploits the fact that the cache at a layer is naturally a third-order tensor whose three axes, the heads, the tokens, and the features, carry very different amounts of redundancy. We take this tensor view directly. Our method, JoLT (Joint Lagrangian Tucker), applies a partial Tucker decomposition that compresses only the token and feature axes while leaving the head and layer axes intact, then restores the energy that truncation discards with a rotated low-bit residual: a random orthogonal rotation followed by low-bit quantization. A single Lagrangian dual allocates the Tucker ranks and the residual bit-widths together, per layer group and separately for keys and values, under one byte budget. The result is a near-lossless 2-3x compression. Perplexity stays near-lossless on both a grouped-query-attention model (Mistral-7B-v0.3) and a multi-head-attention model (LLaMA-2-13B), and GSM8K accuracy and needle-in-a-haystack retrieval hold at the uncompressed baseline at 2x on both architectures and through 3x on the GQA model. At 2x, JoLT reconstructs the cache to relative Frobenius error 0.009 (K) and 0.006 (V) on both architectures. A randomized-SVD variant, FlashJoLT, delivers a 5-13x compression-time speedup at 1024-token context and matched quality.
Autonomous coding agents force engineering organizations to choose between API-based frontier models — strong reasoning at high token cost — and on-premise quantized open-weights models, which promise low-marginal-cost scaling and data sovereignty at some loss of reasoning fidelity. We study this trade-off through a single-developer, non-randomized longitudinal case study over two contiguous 28-day periods on a production monorepo: an API-based Claude Opus 4.7/4.8 configuration using Claude Code versus an on-premise GLM-5.1/5.2 configuration using Opencode, quantized to NVFP4, on NVIDIA Blackwell hardware. Analyzing LLM telemetry and Git history, we find that prompt caching (99.3% hit rate) cuts realized API cost by 88.6% to an effective \$0.57 per million tokens — below even the \$2.83 amortized unit cost of the shared on-premise slice (a utilization-dependent inversion; total realized spend and total cost of ownership (TCO) are the robust quantities). At comparable gross code churn, the local configuration was associated with a far higher defect-repair burden: a Fix Commit Ratio (FCR) of 74.9% versus 45.9%, with the odds of a commit being a repair 2.6 to 4.9 times higher within every difficulty tier (Mantel-Haenszel OR = 3.61). Under Taiwan-market parameters and a symmetric labor model, on-premise deployment nonetheless saves 40.1% of true TCO under shared GPU allocation, whereas dedicated reservation costs 43.8% more than the cached API. Under shared allocation, the genuine penalty is not monetary but a measurable developer-experience burden — timestamp indicators show more work trapped in debugging spirals and a slower commit cadence — and an offline replay shows hybrid routing gateways trade defect rate for infrastructure savings along a cost-quality frontier rather than dominate the pure-API baseline.
展开完整摘要收起摘要↓
Autonomous coding agents force engineering organizations to choose between API-based frontier models — strong reasoning at high token cost — and on-premise quantized open-weights models, which promise low-marginal-cost scaling and data sovereignty at some loss of reasoning fidelity. We study this trade-off through a single-developer, non-randomized longitudinal case study over two contiguous 28-day periods on a production monorepo: an API-based Claude Opus 4.7/4.8 configuration using Claude Code versus an on-premise GLM-5.1/5.2 configuration using Opencode, quantized to NVFP4, on NVIDIA Blackwell hardware. Analyzing LLM telemetry and Git history, we find that prompt caching (99.3% hit rate) cuts realized API cost by 88.6% to an effective \$0.57 per million tokens — below even the \$2.83 amortized unit cost of the shared on-premise slice (a utilization-dependent inversion; total realized spend and total cost of ownership (TCO) are the robust quantities). At comparable gross code churn, the local configuration was associated with a far higher defect-repair burden: a Fix Commit Ratio (FCR) of 74.9% versus 45.9%, with the odds of a commit being a repair 2.6 to 4.9 times higher within every difficulty tier (Mantel-Haenszel OR = 3.61). Under Taiwan-market parameters and a symmetric labor model, on-premise deployment nonetheless saves 40.1% of true TCO under shared GPU allocation, whereas dedicated reservation costs 43.8% more than the cached API. Under shared allocation, the genuine penalty is not monetary but a measurable developer-experience burden — timestamp indicators show more work trapped in debugging spirals and a slower commit cadence — and an offline replay shows hybrid routing gateways trade defect rate for infrastructure savings along a cost-quality frontier rather than dominate the pure-API baseline.
Large Language Models (LLMs) are rapidly moving from research settings into the wild, deployed on enterprise infrastructure, personal devices, and edge platforms. While cloud deployments offer scalable compute, concerns over data sovereignty, compliance, latency, and third-party dependence are driving organizations toward edge and on-premise LLMs. This shift introduces new security and privacy challenges: limited compute and memory force aggressive optimizations, including quantization, pruning, model partitioning, and parameter-efficient adaptation, each of which can introduce vulnerabilities and reshape the threat landscape. We describe this tension as the Security-Efficiency Paradox, mechanisms that improve efficiency may weaken robustness, expose new attack surfaces, or increase privacy risks. We examine how compression can degrade safety alignment, how partitioned inference enables reconstruction attacks, and how continuous local adaptation may cause privacy leakage and model drift. To analyze these risks, we introduce a deployment-centric taxonomy organized around three architectural constraints: the Memory Wall, the Quadratic Wall, and the Compute Wall. We derive a unified constraint model that quantifies when unsafe optimizations become unavoidable, linking each wall to specific attack surfaces. Building on this model, we propose the Secure Operational Efficiency Score (SOES), a holistic metric balancing task accuracy, jailbreak resistance, and privacy against energy, memory, and latency, enabling practitioners to configure edge LLMs under real-world hardware limits. We further present a practical decision procedure and targeted mitigations for each optimization-induced vulnerability. Together, these contributions provide a co-designed framework for jointly evaluating security, privacy, and efficiency, laying a foundation for securing edge-native intelligent systems.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are rapidly moving from research settings into the wild, deployed on enterprise infrastructure, personal devices, and edge platforms. While cloud deployments offer scalable compute, concerns over data sovereignty, compliance, latency, and third-party dependence are driving organizations toward edge and on-premise LLMs. This shift introduces new security and privacy challenges: limited compute and memory force aggressive optimizations, including quantization, pruning, model partitioning, and parameter-efficient adaptation, each of which can introduce vulnerabilities and reshape the threat landscape. We describe this tension as the Security-Efficiency Paradox, mechanisms that improve efficiency may weaken robustness, expose new attack surfaces, or increase privacy risks. We examine how compression can degrade safety alignment, how partitioned inference enables reconstruction attacks, and how continuous local adaptation may cause privacy leakage and model drift. To analyze these risks, we introduce a deployment-centric taxonomy organized around three architectural constraints: the Memory Wall, the Quadratic Wall, and the Compute Wall. We derive a unified constraint model that quantifies when unsafe optimizations become unavoidable, linking each wall to specific attack surfaces. Building on this model, we propose the Secure Operational Efficiency Score (SOES), a holistic metric balancing task accuracy, jailbreak resistance, and privacy against energy, memory, and latency, enabling practitioners to configure edge LLMs under real-world hardware limits. We further present a practical decision procedure and targeted mitigations for each optimization-induced vulnerability. Together, these contributions provide a co-designed framework for jointly evaluating security, privacy, and efficiency, laying a foundation for securing edge-native intelligent systems.
Context-reduction layers for API-based coding agents, including command-output compressors, retrieval rankers, and API-boundary proxies, are commonly evaluated by how much context or tool output they remove. We ask a different question: which interventions actually reduce end-to-end billed cost while preserving task success? Our primary evidence is a pre-specified, hash-frozen, paired campaign of 2,908 provider-billed Claude Code runs, of which 2,848 were analyzed, covering 103 tasks, seven repositories, and three models. The campaign compared a baseline with two generations of hook-based compression and an API-boundary proxy within a broader measured program of roughly 5,500 billed executions. Three findings emerge. First, prompt-cache traffic dominated cost composition, accounting for about 87% of reconstructed four-component cost (about 80% of the actual bill), with an 8.7% dollar-weighted residual not attributable from retained telemetry. Second, local payload reduction was not a reliable predictor of end-to-end billed cost. An arm that removed 38% of estimated raw tool-output tokens incurred 6.8% higher paired cost (95% CI: +2.8% to +11.3%), while per-task reduction showed only a weak association with cost change (Pearson r = 0.15). Third, aggressive compression can remove action-critical evidence: on SWE-bench-derived Go tasks, compression reduced successful patch application from 27/40 to 15/40 by corrupting verbatim edit anchors. We propose evaluating context-reduction systems by success-adjusted billed cost rather than token reduction alone.
展开完整摘要收起摘要↓
Context-reduction layers for API-based coding agents, including command-output compressors, retrieval rankers, and API-boundary proxies, are commonly evaluated by how much context or tool output they remove. We ask a different question: which interventions actually reduce end-to-end billed cost while preserving task success? Our primary evidence is a pre-specified, hash-frozen, paired campaign of 2,908 provider-billed Claude Code runs, of which 2,848 were analyzed, covering 103 tasks, seven repositories, and three models. The campaign compared a baseline with two generations of hook-based compression and an API-boundary proxy within a broader measured program of roughly 5,500 billed executions. Three findings emerge. First, prompt-cache traffic dominated cost composition, accounting for about 87% of reconstructed four-component cost (about 80% of the actual bill), with an 8.7% dollar-weighted residual not attributable from retained telemetry. Second, local payload reduction was not a reliable predictor of end-to-end billed cost. An arm that removed 38% of estimated raw tool-output tokens incurred 6.8% higher paired cost (95% CI: +2.8% to +11.3%), while per-task reduction showed only a weak association with cost change (Pearson r = 0.15). Third, aggressive compression can remove action-critical evidence: on SWE-bench-derived Go tasks, compression reduced successful patch application from 27/40 to 15/40 by corrupting verbatim edit anchors. We propose evaluating context-reduction systems by success-adjusted billed cost rather than token reduction alone.
作者Anton Firc, Martin Perešíni, Vojtěch Mrázek, Kamil Malinka, Vojtěch Staněk, Zbyněk Lička, Nouhaila Innan, Walid El Maouaki, Alberto Marchisio, Muhammad Shafique
Hybrid quantum-classical machine learning workflows repeatedly evaluate many small parametrized circuits during training and model exploration. In this regime, framework dispatch and orchestration overhead often dominate runtime. Prior simulators accelerate execution but leave open the question of when compile-once specialization is the right choice for static variational circuits. We answer this question with VQCSim, a compile-once, PyTorch-native statevector execution path with native autograd. In a systematic MQT Bench study, VQCSim compiles all tested static circuits and provides 87.7% end-to-end semantic validation. Across a five-GPU evaluation set, VQCSim delivers pooled median speedups of 4.49x for native inference and 26.78x for native training, while retaining a 3.31x advantage under matched finite-difference training. Ablation identifies native autograd as the dominant source of acceleration (27.6x), with compile-once caching and batch vectorization contributing additional gains. The speedup trades higher GPU memory (VQCSim is memory-limited at the high end) for lower runtime. We derive a hardware-aware regime map and release vqcsim-oracle, an open-source backend selector with 91.1%-97.7% top-1 agreement (including cross-GPU transfers), enabling automatic simulator selection in QML design loops.
展开完整摘要收起摘要↓
Hybrid quantum-classical machine learning workflows repeatedly evaluate many small parametrized circuits during training and model exploration. In this regime, framework dispatch and orchestration overhead often dominate runtime. Prior simulators accelerate execution but leave open the question of when compile-once specialization is the right choice for static variational circuits. We answer this question with VQCSim, a compile-once, PyTorch-native statevector execution path with native autograd. In a systematic MQT Bench study, VQCSim compiles all tested static circuits and provides 87.7% end-to-end semantic validation. Across a five-GPU evaluation set, VQCSim delivers pooled median speedups of 4.49x for native inference and 26.78x for native training, while retaining a 3.31x advantage under matched finite-difference training. Ablation identifies native autograd as the dominant source of acceleration (27.6x), with compile-once caching and batch vectorization contributing additional gains. The speedup trades higher GPU memory (VQCSim is memory-limited at the high end) for lower runtime. We derive a hardware-aware regime map and release vqcsim-oracle, an open-source backend selector with 91.1%-97.7% top-1 agreement (including cross-GPU transfers), enabling automatic simulator selection in QML design loops.
Sensor-rich data-driven applications increasingly use Bayesian approaches to infer latent states of dynamic systems from noisy sensor measurements and physical models. Yet the computation of the likelihood remains an essential bottleneck for accurate posteriors and performant inference. This paper presents a Bayesian filtering technique that uses processor-native uncertainty tracking for both uncertainty propagation and inference. The technique implements deterministic hierarchical importance restructuring through a native operation, giving deterministic latency and bounded memory use for arbitrary models written as program code. Benchmarks across three nonlinear state-space systems compare the approach against particle filters and Monte-Carlo-based likelihood estimators. The technique enables deterministic approximate filtering with as high as 805$\times$ average speedup against direct Monte Carlo work at matched result quality for model evaluation, and Pareto-dominant accuracy-latency trade-offs for posterior inference while remaining competitive in RMSE with baseline particle filters.
展开完整摘要收起摘要↓
Sensor-rich data-driven applications increasingly use Bayesian approaches to infer latent states of dynamic systems from noisy sensor measurements and physical models. Yet the computation of the likelihood remains an essential bottleneck for accurate posteriors and performant inference. This paper presents a Bayesian filtering technique that uses processor-native uncertainty tracking for both uncertainty propagation and inference. The technique implements deterministic hierarchical importance restructuring through a native operation, giving deterministic latency and bounded memory use for arbitrary models written as program code. Benchmarks across three nonlinear state-space systems compare the approach against particle filters and Monte-Carlo-based likelihood estimators. The technique enables deterministic approximate filtering with as high as 805$\times$ average speedup against direct Monte Carlo work at matched result quality for model evaluation, and Pareto-dominant accuracy-latency trade-offs for posterior inference while remaining competitive in RMSE with baseline particle filters.
Feedforward network (FFN) blocks account for a large fraction of the parameters and computation in Transformer architectures, yet their internal structure remains difficult to interpret due to the additive superposition induced by the residual stream. We examine whether the activation of an FFN neuron can be explained by a sparse set of preceding neuron activations and attention outputs. We introduce a training-free attribution method that estimates the relative influence of upstream neurons and attention outputs on a target neuron's activation. Empirically, across models and layers, we find that small subsets of preceding activations and attention outputs suffice to preserve neuron activations with high fidelity when all remaining inputs are masked with their average values. Effective sparsity is even greater when accounting for the inherent activation sparsity of upstream layers. Moreover, applying the neuron-specific masks in all layers simultaneously, such that the induced deviations propagate through the network, leaves model perplexity largely unchanged at moderate sparsity levels. These results demonstrate that, despite dense parameterization, FFNs exhibit sparse and structured inter-layer dependencies at the neuron level. Our method provides a practical, scalable tool for circuit-level interpretability and identifies candidate sparse pathways with potential implications for efficient inference.
展开完整摘要收起摘要↓
Feedforward network (FFN) blocks account for a large fraction of the parameters and computation in Transformer architectures, yet their internal structure remains difficult to interpret due to the additive superposition induced by the residual stream. We examine whether the activation of an FFN neuron can be explained by a sparse set of preceding neuron activations and attention outputs. We introduce a training-free attribution method that estimates the relative influence of upstream neurons and attention outputs on a target neuron's activation. Empirically, across models and layers, we find that small subsets of preceding activations and attention outputs suffice to preserve neuron activations with high fidelity when all remaining inputs are masked with their average values. Effective sparsity is even greater when accounting for the inherent activation sparsity of upstream layers. Moreover, applying the neuron-specific masks in all layers simultaneously, such that the induced deviations propagate through the network, leaves model perplexity largely unchanged at moderate sparsity levels. These results demonstrate that, despite dense parameterization, FFNs exhibit sparse and structured inter-layer dependencies at the neuron level. Our method provides a practical, scalable tool for circuit-level interpretability and identifies candidate sparse pathways with potential implications for efficient inference.
Enterprise Retrieval-Augmented Generation (RAG) deployments face a critical governance gap: while LLM generation cost is metered per token, the retrieval layer - vector memory, similarity compute, and embedding API calls - remains an unattributed shared cost, enabling invisible cross-subsidization among tenants. We present Cost-Governed RAG, an architecture that integrates a codebook-oblivious vector index (TurboVec) with a multi-tenant LLM governance gateway, creating a unified observability stack where embedding, retrieval, and generation costs are jointly attributable per tenant. The architecture exploits TurboVec's deterministic, closed-form memory formula to enable near-exact per-tenant retrieval cost calculation - a property unavailable in graph-based indexes with non-linear memory overhead. Deployed on Snowpark Container Services within a cloud data platform's governance boundary, the system achieves 99.96% end-to-end cost attribution accuracy across 100 simulated tenants (10M vectors, log-normal size distribution) with telemetry overhead below 0.04% of query latency. The architecture reduces retrieval infrastructure cost by 3.1-9.0x compared to managed vector database services under the pricing assumptions detailed in Section IV. We formalize a three-layer cost model and demonstrate that codebook-oblivious quantization enables deterministic per-tenant cost attribution while also removing the shared-codebook leakage surface present in trained quantizers - the latter observation being exploratory and subject to the limitations described in Section VII.
展开完整摘要收起摘要↓
Enterprise Retrieval-Augmented Generation (RAG) deployments face a critical governance gap: while LLM generation cost is metered per token, the retrieval layer - vector memory, similarity compute, and embedding API calls - remains an unattributed shared cost, enabling invisible cross-subsidization among tenants. We present Cost-Governed RAG, an architecture that integrates a codebook-oblivious vector index (TurboVec) with a multi-tenant LLM governance gateway, creating a unified observability stack where embedding, retrieval, and generation costs are jointly attributable per tenant. The architecture exploits TurboVec's deterministic, closed-form memory formula to enable near-exact per-tenant retrieval cost calculation - a property unavailable in graph-based indexes with non-linear memory overhead. Deployed on Snowpark Container Services within a cloud data platform's governance boundary, the system achieves 99.96% end-to-end cost attribution accuracy across 100 simulated tenants (10M vectors, log-normal size distribution) with telemetry overhead below 0.04% of query latency. The architecture reduces retrieval infrastructure cost by 3.1-9.0x compared to managed vector database services under the pricing assumptions detailed in Section IV. We formalize a three-layer cost model and demonstrate that codebook-oblivious quantization enables deterministic per-tenant cost attribution while also removing the shared-codebook leakage surface present in trained quantizers - the latter observation being exploratory and subject to the limitations described in Section VII.
作者Induk Um, Youngung Han, Kyeonghun Kim, Yului Jeong, Jina Jeong, Hyunsu Go, Dohyun Kweon, Sungha Park, Junga Kim, Anna Jung, Suah Park, Hyuk-Jae Lee, Pa Hong, Woo Kyoung Jeong, Won Jae Lee, Ken Ying-Kai Liao, Nam-Joon Kim
Perineural invasion (PNI) is associated with poor prognosis in cholangiocarcinoma (CCA). However, its detection from 3D MRI remains challenging due to the subtle and spatially heterogeneous imaging signatures at the tumor periphery. Capturing such spatially sparse cues necessitates volumetric analysis of 3D MRI, but existing deep learning approaches incur prohibitive computational costs on volumetric medical images, limiting their clinical deployment. We propose Dual Sparsity Spikformer (SpikeDS), a spiking neural network architecture that jointly exploits activation sparsity from binary spike communication and spatial sparsity from window pruning based on firing rates. SpikeDS introduces Dual Sparsity Spiking Attention (DSSA), which combines two complementary mechanisms. The first is Window-based Expert Mixture Spiking Attention (W-EMSA), which selectively applies attention only to salient windows identified by their firing rates. The second is Cross-Window Spiking Self-Attention (CW-SSA), which enables global context exchange through an asymmetric scheme in which pruned windows still contribute as key-value sources. Evaluated on a clinical cohort of 139 CCA patients via 5-fold cross-validation, SpikeDS achieves an AUC of 0.753 while consuming only 14.4 mJ, surpassing the best baseline in both AUC and energy efficiency. These results suggest that dual sparsity provides an effective hardware-aware strategy for improving the efficiency of 3D spiking transformers without compromising diagnostic performance.
展开完整摘要收起摘要↓
Perineural invasion (PNI) is associated with poor prognosis in cholangiocarcinoma (CCA). However, its detection from 3D MRI remains challenging due to the subtle and spatially heterogeneous imaging signatures at the tumor periphery. Capturing such spatially sparse cues necessitates volumetric analysis of 3D MRI, but existing deep learning approaches incur prohibitive computational costs on volumetric medical images, limiting their clinical deployment. We propose Dual Sparsity Spikformer (SpikeDS), a spiking neural network architecture that jointly exploits activation sparsity from binary spike communication and spatial sparsity from window pruning based on firing rates. SpikeDS introduces Dual Sparsity Spiking Attention (DSSA), which combines two complementary mechanisms. The first is Window-based Expert Mixture Spiking Attention (W-EMSA), which selectively applies attention only to salient windows identified by their firing rates. The second is Cross-Window Spiking Self-Attention (CW-SSA), which enables global context exchange through an asymmetric scheme in which pruned windows still contribute as key-value sources. Evaluated on a clinical cohort of 139 CCA patients via 5-fold cross-validation, SpikeDS achieves an AUC of 0.753 while consuming only 14.4 mJ, surpassing the best baseline in both AUC and energy efficiency. These results suggest that dual sparsity provides an effective hardware-aware strategy for improving the efficiency of 3D spiking transformers without compromising diagnostic performance.
Exact induced motif enumeration is a fundamental operation in graph mining, but it remains challenging on GPUs because candidate expansion is irregular, repeated set intersections dominate execution, and induced counting must consider both the presence and absence of edges. We present HARP-ME, which stands for Hierarchical Anchor-Reuse Partitioned Motif Enumeration. It is a GPU framework for the exact enumeration of connected induced four-node motifs. HARP-ME introduces closure-aware compilation, which selects a set of anchors for explicit enumeration by considering traversal cost, algebraic derivation benefits, expected state reuse, and the additional halo overhead caused by graph partitioning. HARP-ME also introduces induced-signature reuse. This technique identifies reusable completion states using candidate-frontier information together with compact constraints representing both adjacency and non-adjacency relationships. For graphs that exceed GPU memory, a canonical anchor-owner rule ensures exact counting across partitions with overlapping halo regions. We do not claim that the individual graphlet closure identities are new. Instead, our contribution is their integration into a GPU execution framework that reduces both explicit candidate expansion and repeated induced-edge checks. Across six social, web, biological, and synthetic graphs, HARP-ME is the fastest among the evaluated methods. It achieves up to 2.11 times speedup over Pangolin, up to 1.83 times speedup over partitioned PBE, and up to 10.73 times speedup over the evaluated CPU baseline. Detailed measurements show cache-hit rates between 64% and 76%, along with substantially lower host-to-device data transfer overhead than PBE-style partitioning. These results demonstrate that optimizing anchor selection for algebraic derivation benefits can complement traversal-based and reuse-based GPU motif enumeration techniques.
展开完整摘要收起摘要↓
Exact induced motif enumeration is a fundamental operation in graph mining, but it remains challenging on GPUs because candidate expansion is irregular, repeated set intersections dominate execution, and induced counting must consider both the presence and absence of edges. We present HARP-ME, which stands for Hierarchical Anchor-Reuse Partitioned Motif Enumeration. It is a GPU framework for the exact enumeration of connected induced four-node motifs. HARP-ME introduces closure-aware compilation, which selects a set of anchors for explicit enumeration by considering traversal cost, algebraic derivation benefits, expected state reuse, and the additional halo overhead caused by graph partitioning. HARP-ME also introduces induced-signature reuse. This technique identifies reusable completion states using candidate-frontier information together with compact constraints representing both adjacency and non-adjacency relationships. For graphs that exceed GPU memory, a canonical anchor-owner rule ensures exact counting across partitions with overlapping halo regions. We do not claim that the individual graphlet closure identities are new. Instead, our contribution is their integration into a GPU execution framework that reduces both explicit candidate expansion and repeated induced-edge checks. Across six social, web, biological, and synthetic graphs, HARP-ME is the fastest among the evaluated methods. It achieves up to 2.11 times speedup over Pangolin, up to 1.83 times speedup over partitioned PBE, and up to 10.73 times speedup over the evaluated CPU baseline. Detailed measurements show cache-hit rates between 64% and 76%, along with substantially lower host-to-device data transfer overhead than PBE-style partitioning. These results demonstrate that optimizing anchor selection for algebraic derivation benefits can complement traversal-based and reuse-based GPU motif enumeration techniques.
Indexer-TopK, the operation to compute the scores and select the top-k candidates, is widely used by sparse attention kernels in large language models and vector retrieval in recommendation systems and vector databases. However, existing GPU-based Indexer-TopK kernels like DeepSeek Sparse Attention (DSA) remain inefficient due to excessive global memory traffic, costly synchronization, and prohibitive memory overhead. In this work, we exploit the curse of dimensionality in high-dimensional spaces, where distances between high-dimensional vectors tend to concentrate within a narrow range, to design LITETOPK, a novel and efficient fused Indexer-TopK kernel. LITETOPK first samples a small subset of data to estimate query-data score ranges, then uses these estimates to partition candidate results into bins online. This organization allows the LITETOPK kernel to maintain a tight approximate threshold, write back only promising candidates, reduce unnecessary I/O, substantially lower memory overhead, and still preserve exact Top-k correctness. Experimental results show that LITETOPK accelerates the prefill stage of GLM 5.2 by 1.2x in real-world deployment scenarios while incurring lower memory overhead.
展开完整摘要收起摘要↓
Indexer-TopK, the operation to compute the scores and select the top-k candidates, is widely used by sparse attention kernels in large language models and vector retrieval in recommendation systems and vector databases. However, existing GPU-based Indexer-TopK kernels like DeepSeek Sparse Attention (DSA) remain inefficient due to excessive global memory traffic, costly synchronization, and prohibitive memory overhead. In this work, we exploit the curse of dimensionality in high-dimensional spaces, where distances between high-dimensional vectors tend to concentrate within a narrow range, to design LITETOPK, a novel and efficient fused Indexer-TopK kernel. LITETOPK first samples a small subset of data to estimate query-data score ranges, then uses these estimates to partition candidate results into bins online. This organization allows the LITETOPK kernel to maintain a tight approximate threshold, write back only promising candidates, reduce unnecessary I/O, substantially lower memory overhead, and still preserve exact Top-k correctness. Experimental results show that LITETOPK accelerates the prefill stage of GLM 5.2 by 1.2x in real-world deployment scenarios while incurring lower memory overhead.
作者Yaqi Qiao, Ping He, Songrun Xie, Ayush Barik, Chensong Zhang, Zhengzhong Tu, Fan Lai
Diffusion models have become the central backbone for modern image, video, and audio generation, but their efficient service remains a challenge. Unlike autoregressive decoding, diffusion inference repeatedly updates high-dimensional spatial or temporal latents over many denoising steps. This all-region execution pattern makes generation latency high and limits serving throughput. Existing multi-GPU parallelization methods can reduce per-step computation, but often introduce substantial activation exchange overhead, causing communication to offset or even outweigh the benefits of parallel execution. This paper presents FlashDiff, a diffusion serving system that improves inference efficiency through adaptive regional execution and scheduling. FlashDiff is based on the observation that diffusion refinement is not uniform across latent regions or denoising steps: different regions often stabilize at different rates, while neighboring steps exhibit strong temporal correlation. FlashDiff leverages these properties to selectively execute only regions that require further refinement and to reallocate the resulting compute slack across concurrent serving requests. FlashDiff consists of three mechanisms. First, it decomposes the latent representation into coherent execution regions using early-stage attention signals, preserving semantic structure while exposing fine-grained parallelism. Second, it uses a lightweight runtime controller to estimate region activity and bypass low-impact updates when further refinement is unlikely to affect output quality. Third, it applies an affinity-aware online scheduler that co-locates dependent regions, balances residual load across GPUs, and reuses reclaimed compute capacity to improve serving efficiency. Across real-world image, video, and audio workloads, FlashDiff reduces end-to-end serving latency by 30-97% and improves throughput by 1.2-2.2x.
展开完整摘要收起摘要↓
Diffusion models have become the central backbone for modern image, video, and audio generation, but their efficient service remains a challenge. Unlike autoregressive decoding, diffusion inference repeatedly updates high-dimensional spatial or temporal latents over many denoising steps. This all-region execution pattern makes generation latency high and limits serving throughput. Existing multi-GPU parallelization methods can reduce per-step computation, but often introduce substantial activation exchange overhead, causing communication to offset or even outweigh the benefits of parallel execution. This paper presents FlashDiff, a diffusion serving system that improves inference efficiency through adaptive regional execution and scheduling. FlashDiff is based on the observation that diffusion refinement is not uniform across latent regions or denoising steps: different regions often stabilize at different rates, while neighboring steps exhibit strong temporal correlation. FlashDiff leverages these properties to selectively execute only regions that require further refinement and to reallocate the resulting compute slack across concurrent serving requests. FlashDiff consists of three mechanisms. First, it decomposes the latent representation into coherent execution regions using early-stage attention signals, preserving semantic structure while exposing fine-grained parallelism. Second, it uses a lightweight runtime controller to estimate region activity and bypass low-impact updates when further refinement is unlikely to affect output quality. Third, it applies an affinity-aware online scheduler that co-locates dependent regions, balances residual load across GPUs, and reuses reclaimed compute capacity to improve serving efficiency. Across real-world image, video, and audio workloads, FlashDiff reduces end-to-end serving latency by 30-97% and improves throughput by 1.2-2.2x.
作者Varun Ramesh Jois, Antonella DiLillo, James Storer
There's been a surge in adoption of video conferencing applications for both personal and business use cases. However, the bandwidth limitations faced by many users worldwide may restrict the optimal use of such applications. Although deep learning offers a solution for enhancing low bit rate videos, most models today are either hard to incorporate with modern compression standards or require specialized hardware to run such as significant GPUs making these models impractical. To address these issues, we introduce the Realtime Face Video Enhancement (RTFVE) model which can be easily incorporated with any video decoder and can run in realtime on ordinary CPUs. Experiments show that our model improves perceptual quality over the compressed video baseline at multiple low bitrate settings. The source code will be made available at https://github.com/varun-jois/RTFVE.
展开完整摘要收起摘要↓
There's been a surge in adoption of video conferencing applications for both personal and business use cases. However, the bandwidth limitations faced by many users worldwide may restrict the optimal use of such applications. Although deep learning offers a solution for enhancing low bit rate videos, most models today are either hard to incorporate with modern compression standards or require specialized hardware to run such as significant GPUs making these models impractical. To address these issues, we introduce the Realtime Face Video Enhancement (RTFVE) model which can be easily incorporated with any video decoder and can run in realtime on ordinary CPUs. Experiments show that our model improves perceptual quality over the compressed video baseline at multiple low bitrate settings. The source code will be made available at https://github.com/varun-jois/RTFVE.
作者Xinchen Liu, Hang Zhou, Yingjie Zong, Yuchuan Tian, Liuyang Song, Shuo Zhang, Yulong Li, Wei He, Mengyu Zheng, Runke Liu, Siyang Cheng, Xiang Kuang, Hailin Hu, Kai Han, Yunhe Wang
Large language model agents are increasingly executed not by a single model call, but by an execution harness that manages observation, context, control, action, state, and verification. At the same time, frontier and open models are becoming structurally specialized: a model that is strong at code editing, long-context recovery, tool use, mathematical reasoning, or low-latency response may not dominate on the other axes. This makes model selection inside an agent a core systems problem rather than a per-query serving trick. Existing routing methods mostly optimize single-turn cost-quality trade-offs and therefore miss the execution state, intermediate failures, and feedback loops that make agents different from chat completion. We propose Harness-Native agentic routing, a step-level routing paradigm that selects either a single best-fit model for cost-effective execution or multiple complementary models for ensemble-style accuracy improvement, conditioned on the full harness state. The key insight is that every routing decision naturally produces a structured data record — consisting of the query, harness state, model choice or model set, execution trace, outcome, and cost — whose labels are supplied by the environment rather than by the router itself. These records form a harness-native data flywheel: execution traces train better routers and harness-native models, which improve cost-quality trade-offs and generate more traces under the same budget. We instantiate this idea in OpenSquilla with a four-layer routing stack, an open LightGBM cold-start ranker, and a staged router-model path that turns logged arena records into progressively stronger routing policies. The report studies singleton and multi-model routing on agentic benchmarks including DRACO and PinchBench, and argues that agentic routing is not merely cost control, but a data engine for agent-native training.
展开完整摘要收起摘要↓
Large language model agents are increasingly executed not by a single model call, but by an execution harness that manages observation, context, control, action, state, and verification. At the same time, frontier and open models are becoming structurally specialized: a model that is strong at code editing, long-context recovery, tool use, mathematical reasoning, or low-latency response may not dominate on the other axes. This makes model selection inside an agent a core systems problem rather than a per-query serving trick. Existing routing methods mostly optimize single-turn cost-quality trade-offs and therefore miss the execution state, intermediate failures, and feedback loops that make agents different from chat completion. We propose Harness-Native agentic routing, a step-level routing paradigm that selects either a single best-fit model for cost-effective execution or multiple complementary models for ensemble-style accuracy improvement, conditioned on the full harness state. The key insight is that every routing decision naturally produces a structured data record — consisting of the query, harness state, model choice or model set, execution trace, outcome, and cost — whose labels are supplied by the environment rather than by the router itself. These records form a harness-native data flywheel: execution traces train better routers and harness-native models, which improve cost-quality trade-offs and generate more traces under the same budget. We instantiate this idea in OpenSquilla with a four-layer routing stack, an open LightGBM cold-start ranker, and a staged router-model path that turns logged arena records into progressively stronger routing policies. The report studies singleton and multi-model routing on agentic benchmarks including DRACO and PinchBench, and argues that agentic routing is not merely cost control, but a data engine for agent-native training.
Recently, Vision Transformer (ViT)-based models have exhibited remarkable performance in image super-resolution. However, the quadratic computational complexity of ViTs with respect to spatial resolution severely constrains their efficiency, leading to high latency and massive memory consumption. To alleviate this, various window-based attention mechanisms have been proposed; yet, they inherently compromise the long-range dependency modeling that is the primary advantage of ViTs. To overcome these limitations, we propose the Clustered Unit-level Similarity Transformer (CUST), a novel architecture that efficiently integrates global and local information. Specifically, CUST enables each patch to aggregate and attend to similar patches within a broadened regional scope outside its local window, thereby capturing extensive contextual understanding. Furthermore, it employs overlapping attention windows to capture local dependencies, while explicitly extracting high-frequency details by computing the residual difference between the original features and their downsampled-upsampled counterparts. Comprehensive experiments demonstrate that our proposed model achieves a practical balance between computational efficiency and restoration performance. It achieves a lower memory footprint and faster inference speed compared to recent global context or lightweight models under realistic constraints. Code is available at [https://github.com/jwgdmkj/CUST].
展开完整摘要收起摘要↓
Recently, Vision Transformer (ViT)-based models have exhibited remarkable performance in image super-resolution. However, the quadratic computational complexity of ViTs with respect to spatial resolution severely constrains their efficiency, leading to high latency and massive memory consumption. To alleviate this, various window-based attention mechanisms have been proposed; yet, they inherently compromise the long-range dependency modeling that is the primary advantage of ViTs. To overcome these limitations, we propose the Clustered Unit-level Similarity Transformer (CUST), a novel architecture that efficiently integrates global and local information. Specifically, CUST enables each patch to aggregate and attend to similar patches within a broadened regional scope outside its local window, thereby capturing extensive contextual understanding. Furthermore, it employs overlapping attention windows to capture local dependencies, while explicitly extracting high-frequency details by computing the residual difference between the original features and their downsampled-upsampled counterparts. Comprehensive experiments demonstrate that our proposed model achieves a practical balance between computational efficiency and restoration performance. It achieves a lower memory footprint and faster inference speed compared to recent global context or lightweight models under realistic constraints. Code is available at [https://github.com/jwgdmkj/CUST].
Discovering the memory or nonlocal kernel governing an integro-differential equation (IDE) from sparse and noisy observations is an ill-posed inverse problem. Existing identification methods often rely on problem-specific analytical derivations, specialized observation requirements, or restrictive assumptions about the kernel, limiting their applicability across different classes of IDEs. In this work, we propose a differentiable-solver-based framework for discovering memory and nonlocal kernels directly from spatiotemporal observations. Within the solver, the unknown kernel is represented using a constrained Kolmogorov--Arnold Network (KAN) parameterization, with the physical constraints imposed through two different approaches: a Bernstein-polynomial-based Monotone--Convex KAN (MC-KAN), whose coefficient constraints enforce positivity, monotonic decrease, and convexity by construction, and a Chebyshev-based KAN (Cheb-KAN), in which the same properties are encouraged through soft penalty terms. After training, symbolic regression is applied to the learned kernels to obtain interpretable closed-form representations. We evaluate both methods on benchmarks spanning a one-dimensional Volterra equation, a one-dimensional viscoelastic wave partial integro-differential equation, and a two-dimensional nonlocal reaction-diffusion equation with an anisotropic coupled kernel. For the 1D problems, both methods recover the correct kernel functional form and achieve comparable solution-reconstruction accuracy. In contrast, for the sparse and noisy 2D nonlocal problem, the hard-constrained MC-KAN consistently achieves lower kernel reconstruction errors than the soft-constrained Cheb-KAN. Our results demonstrate that enforcing physically motivated shape constraints by construction provides greater robustness than soft penalties for multidimensional kernel discovery from sparse and noisy observations.
展开完整摘要收起摘要↓
Discovering the memory or nonlocal kernel governing an integro-differential equation (IDE) from sparse and noisy observations is an ill-posed inverse problem. Existing identification methods often rely on problem-specific analytical derivations, specialized observation requirements, or restrictive assumptions about the kernel, limiting their applicability across different classes of IDEs. In this work, we propose a differentiable-solver-based framework for discovering memory and nonlocal kernels directly from spatiotemporal observations. Within the solver, the unknown kernel is represented using a constrained Kolmogorov--Arnold Network (KAN) parameterization, with the physical constraints imposed through two different approaches: a Bernstein-polynomial-based Monotone--Convex KAN (MC-KAN), whose coefficient constraints enforce positivity, monotonic decrease, and convexity by construction, and a Chebyshev-based KAN (Cheb-KAN), in which the same properties are encouraged through soft penalty terms. After training, symbolic regression is applied to the learned kernels to obtain interpretable closed-form representations. We evaluate both methods on benchmarks spanning a one-dimensional Volterra equation, a one-dimensional viscoelastic wave partial integro-differential equation, and a two-dimensional nonlocal reaction-diffusion equation with an anisotropic coupled kernel. For the 1D problems, both methods recover the correct kernel functional form and achieve comparable solution-reconstruction accuracy. In contrast, for the sparse and noisy 2D nonlocal problem, the hard-constrained MC-KAN consistently achieves lower kernel reconstruction errors than the soft-constrained Cheb-KAN. Our results demonstrate that enforcing physically motivated shape constraints by construction provides greater robustness than soft penalties for multidimensional kernel discovery from sparse and noisy observations.
Recommendation systems help users recommend relevant items from a large collection of choices. Present work on transformer-based sequential recommendation learns user preferences from interaction logs, but it mostly focuses on item identifiers and doesn't fully use the semantic meaning of items. This limitation becomes a major challenge in sparse and cold-start scenarios where historical interaction data is limited. To solve this problem, we introduce SISA-Rec (Semantically Integrated Sequential Recommendation), a transformer-based framework that embeds semantic context directly into sequential modeling. Our approach fuses item ID embeddings with BERT-based text embeddings via a gated fusion module, injects semantic similarity into the self-attention mechanism, and leverages an attention-based aggregation module to construct comprehensive user representations. Finally, a joint learning objective which combines Bayesian Personalized Ranking (BPR) and contrastive alignment loss, aligns the underlying behavioral and semantic spaces. Experiments were conducted on the two highly sparse Amazon Beauty and Amazon Toys & Games datasets, both having 99.93% sparsity. The results show that SISA-Rec outperforms state-of-the-art baseline models across all evaluation metrics. Compared with the BERT4Rec \cite{petrov2022systematic}, SISA-Rec improves HR@10 by 16.6% and NDCG@10 by 10.3% on Amazon Beauty, and HR@10 by 23.1% and NDCG@10 by 17.9% on Amazon Toys & Games. Cold-start analysis further shows that the proposed model achieves the largest improvements for users with limited interaction historical records. This showcases the value of semantic information when user behavior data is scarce. Overall, the results demonstrate that integrating semantic information into the attention mechanism leads to more accurate and reliable recommendations.
展开完整摘要收起摘要↓
Recommendation systems help users recommend relevant items from a large collection of choices. Present work on transformer-based sequential recommendation learns user preferences from interaction logs, but it mostly focuses on item identifiers and doesn't fully use the semantic meaning of items. This limitation becomes a major challenge in sparse and cold-start scenarios where historical interaction data is limited. To solve this problem, we introduce SISA-Rec (Semantically Integrated Sequential Recommendation), a transformer-based framework that embeds semantic context directly into sequential modeling. Our approach fuses item ID embeddings with BERT-based text embeddings via a gated fusion module, injects semantic similarity into the self-attention mechanism, and leverages an attention-based aggregation module to construct comprehensive user representations. Finally, a joint learning objective which combines Bayesian Personalized Ranking (BPR) and contrastive alignment loss, aligns the underlying behavioral and semantic spaces. Experiments were conducted on the two highly sparse Amazon Beauty and Amazon Toys & Games datasets, both having 99.93% sparsity. The results show that SISA-Rec outperforms state-of-the-art baseline models across all evaluation metrics. Compared with the BERT4Rec \cite{petrov2022systematic}, SISA-Rec improves HR@10 by 16.6% and NDCG@10 by 10.3% on Amazon Beauty, and HR@10 by 23.1% and NDCG@10 by 17.9% on Amazon Toys & Games. Cold-start analysis further shows that the proposed model achieves the largest improvements for users with limited interaction historical records. This showcases the value of semantic information when user behavior data is scarce. Overall, the results demonstrate that integrating semantic information into the attention mechanism leads to more accurate and reliable recommendations.
作者Liqian Feng, Lintao Wang, Xiaochen Liu, Anusha Withana, Ken-Tye Yong, Dehui Kong, Zhiyong Wang, Kun Hu
Most sign language translation (SLT) methods focus on isolated native sign-spoken pairs (e.g., American Sign Language - English). Extending language-specific SLT models to multilingual translation would improve accessibility by enabling communication across diverse sign and spoken language communities. However, existing multilingual SLT approaches still struggle to learn a unified model that minimizes cross-lingual conflicts while capturing shared cross-lingual semantics and preserving language-specific variations across different sign languages. Therefore, we propose Q-BridgeNet, a unified framework for multilingual SLT that jointly mitigates cross-lingual conflicts across both the sign language and spoken language sides. On the sign language side, Q-BridgeNet learns discrete Q-units via adaptive segmentation and residual vector quantization: a shared base codebook provides language-agnostic semantic primitives, while language-specific residual codebooks refine heterogeneous signing semantics. On the spoken language side, a multilingual LLM is fine-tuned to operate in the Q-unit space, leveraging cross-lingual priors to enable a unified SLT model. Experiments on PHOENIX14T, How2Sign, and CSL-Daily show that Q-BridgeNet effectively mitigates cross-lingual conflicts, achieving state-of-the-art performance on native sign-spoken pairs while also demonstrating strong generalization to non-native pairs. Our source code is publicly available at: https://github.com/FengLiQ/Q-BridgeNet
展开完整摘要收起摘要↓
Most sign language translation (SLT) methods focus on isolated native sign-spoken pairs (e.g., American Sign Language - English). Extending language-specific SLT models to multilingual translation would improve accessibility by enabling communication across diverse sign and spoken language communities. However, existing multilingual SLT approaches still struggle to learn a unified model that minimizes cross-lingual conflicts while capturing shared cross-lingual semantics and preserving language-specific variations across different sign languages. Therefore, we propose Q-BridgeNet, a unified framework for multilingual SLT that jointly mitigates cross-lingual conflicts across both the sign language and spoken language sides. On the sign language side, Q-BridgeNet learns discrete Q-units via adaptive segmentation and residual vector quantization: a shared base codebook provides language-agnostic semantic primitives, while language-specific residual codebooks refine heterogeneous signing semantics. On the spoken language side, a multilingual LLM is fine-tuned to operate in the Q-unit space, leveraging cross-lingual priors to enable a unified SLT model. Experiments on PHOENIX14T, How2Sign, and CSL-Daily show that Q-BridgeNet effectively mitigates cross-lingual conflicts, achieving state-of-the-art performance on native sign-spoken pairs while also demonstrating strong generalization to non-native pairs. Our source code is publicly available at: https://github.com/FengLiQ/Q-BridgeNet
作者Khaleelulla Khan Nazeer, Sirine Arfa, Matthias Jobst, Richard George, Christian Mayr
A substantial number of patients experience diminished mobility due to disabilities, diseases, or accidents. Although modern prostheses, powered by deep neural networks, hold the promise of significantly enhancing the quality of life for these individuals, their widespread adoption is hindered by significant latency, energy consumption, and spatial requirements. Wired connections to external high-performance processors restrict patient mobility, while wireless connections limit the volume of information that can be transmitted to these processors. Spiking neural networks offer the potential for compressed communication and low-power inference, yet they often lag behind state-of-the-art deep learning models in various applications. In this study, we propose a high-performance neural decoding method that effectively balances task performance and efficiency. An eventbased gated recurrent unit generates a sparse communication pattern with graded spikes, surpassing classical spiking neural networks in terms of task performance. Utilising an efficient training method and sparse inference, our model presents new opportunities for on-device neural decoding.
展开完整摘要收起摘要↓
A substantial number of patients experience diminished mobility due to disabilities, diseases, or accidents. Although modern prostheses, powered by deep neural networks, hold the promise of significantly enhancing the quality of life for these individuals, their widespread adoption is hindered by significant latency, energy consumption, and spatial requirements. Wired connections to external high-performance processors restrict patient mobility, while wireless connections limit the volume of information that can be transmitted to these processors. Spiking neural networks offer the potential for compressed communication and low-power inference, yet they often lag behind state-of-the-art deep learning models in various applications. In this study, we propose a high-performance neural decoding method that effectively balances task performance and efficiency. An eventbased gated recurrent unit generates a sparse communication pattern with graded spikes, surpassing classical spiking neural networks in terms of task performance. Utilising an efficient training method and sparse inference, our model presents new opportunities for on-device neural decoding.
作者Wencheng Ye, Yi Bin, Yujuan Ding, Hongye Fang, Zheng Wang, Xing Xu, Jingkuan Song, Yun Zhang, Sirui Da, Heng Tao Shen
Vision-language models increasingly succeed on multimodal reasoning benchmarks, yet their visual evidence often becomes unstable once it enters the language stack, weakening evidence-grounded reasoning. To understand this fragility, we examine the internal dynamics of VLMs through a mechanistic lens and uncover a stable three-stage redistribution of multimodal attention focus across depth: an early question-conditioned organization, a critical middle visual-dominant relay, and a late return to answer formation. We operationalize the middle phase as the Visual Relay Window (VRW), and show that its geometry varies with task demand, is causally tied to grounded generation, and distinguishes unsupported answers from stronger reasoning trajectories. Guided by this internal rhythm, we propose TRACE, a task-adaptive inference-time control framework with lightweight trained modules. It reshapes relay allocation during prefill and preserves assembled visual support after handoff during decoding. Across four open-weight VLM backbones and seven benchmarks, TRACE delivers large gains on grounding-sensitive settings, improving them by 4.33 points on average and by up to 6.6 points, while also improving reasoning-heavy tasks. These results show that explicitly controlling multimodal focus across depth offers a unified and effective mechanism for strengthening evidence-grounded multimodal reasoning.
展开完整摘要收起摘要↓
Vision-language models increasingly succeed on multimodal reasoning benchmarks, yet their visual evidence often becomes unstable once it enters the language stack, weakening evidence-grounded reasoning. To understand this fragility, we examine the internal dynamics of VLMs through a mechanistic lens and uncover a stable three-stage redistribution of multimodal attention focus across depth: an early question-conditioned organization, a critical middle visual-dominant relay, and a late return to answer formation. We operationalize the middle phase as the Visual Relay Window (VRW), and show that its geometry varies with task demand, is causally tied to grounded generation, and distinguishes unsupported answers from stronger reasoning trajectories. Guided by this internal rhythm, we propose TRACE, a task-adaptive inference-time control framework with lightweight trained modules. It reshapes relay allocation during prefill and preserves assembled visual support after handoff during decoding. Across four open-weight VLM backbones and seven benchmarks, TRACE delivers large gains on grounding-sensitive settings, improving them by 4.33 points on average and by up to 6.6 points, while also improving reasoning-heavy tasks. These results show that explicitly controlling multimodal focus across depth offers a unified and effective mechanism for strengthening evidence-grounded multimodal reasoning.
To handle unpredictable workloads, cloud providers typically over-provision memory to meet peak demand, resulting in substantial underutilization across datacenter clusters. At the same time, memory-constrained tenants may suffer elevated cache miss rates, even when idle capacity remains stranded elsewhere in the infrastructure. MemExchange is a cluster-wide, multi-tenant memory management system that dynamically right-sizes in-memory caching tenants according to workload demand. Leveraging marginal-utility-based allocation derived from online Miss Ratio Curve (MRC) estimation, MemExchange redistributes idle memory between tenants across physical nodes using RDMA. This approach transforms the dedicated caching memory scattered across servers into a logically aggregated pool, enabling cross-node memory exchange without centralized coordination or forced tenant co-location. To support efficient remote access, we design the MemExchange Tracker Communication (MTC) protocol, an application-layer mechanism that coordinates memory reallocation and enables one-sided RDMA operations without involving remote CPUs. We implement MemExchange in Memcached and evaluate it through microbenchmarks, medium and rack-scale deployments of up to 100 CloudLab servers. Our results show up to 2.3x lower remote-access overhead compared to TCP-based designs, a 13% increase in cluster-wide memory utilization at rack scale, and up to 63% reduction in miss rate for memory-constrained tenants under skewed workloads.
展开完整摘要收起摘要↓
To handle unpredictable workloads, cloud providers typically over-provision memory to meet peak demand, resulting in substantial underutilization across datacenter clusters. At the same time, memory-constrained tenants may suffer elevated cache miss rates, even when idle capacity remains stranded elsewhere in the infrastructure. MemExchange is a cluster-wide, multi-tenant memory management system that dynamically right-sizes in-memory caching tenants according to workload demand. Leveraging marginal-utility-based allocation derived from online Miss Ratio Curve (MRC) estimation, MemExchange redistributes idle memory between tenants across physical nodes using RDMA. This approach transforms the dedicated caching memory scattered across servers into a logically aggregated pool, enabling cross-node memory exchange without centralized coordination or forced tenant co-location. To support efficient remote access, we design the MemExchange Tracker Communication (MTC) protocol, an application-layer mechanism that coordinates memory reallocation and enables one-sided RDMA operations without involving remote CPUs. We implement MemExchange in Memcached and evaluate it through microbenchmarks, medium and rack-scale deployments of up to 100 CloudLab servers. Our results show up to 2.3x lower remote-access overhead compared to TCP-based designs, a 13% increase in cluster-wide memory utilization at rack scale, and up to 63% reduction in miss rate for memory-constrained tenants under skewed workloads.
Diffusion models are increasingly deployed as production visual-generation services, where serving high-resolution image and long video generation is often limited by GPU memory. Popular memory-saving techniques such as weight offloading, sharding, and VAE slicing are often not practical because they tend to introduce significant performance overhead. In this paper, we present Xema, a memory-efficient diffusion serving system that exploits predictable tensor lifetimes for trace-guided memory optimization. For each request template, Xema derives an offline memory trace to identify short memory-pressure intervals and applies memory mitigation only within these intervals and only by the amount needed to fit the target GPU budget. Xema further constructs a static memory layout for tensors with predictable lifetimes, reducing fragmentation-induced reserved memory and making offline memory reasoning reliable at runtime. Built on this memory optimization layer, Xema introduces an offline planner that jointly selects parallelism, concurrency, and memory control under GPU memory and SLO constraints. The selected plan is stored in a plan table and directly used by the online serving runtime. We implement Xema on production diffusion pipelines and evaluate it with Flux.2, CogVideoX-5B, and LTX-2. Compared with existing serving configurations, Xema improves SLO attainment by up to 3.7x and reduces planning cost from 6.3 hours to 197 seconds compared with grid search.
展开完整摘要收起摘要↓
Diffusion models are increasingly deployed as production visual-generation services, where serving high-resolution image and long video generation is often limited by GPU memory. Popular memory-saving techniques such as weight offloading, sharding, and VAE slicing are often not practical because they tend to introduce significant performance overhead. In this paper, we present Xema, a memory-efficient diffusion serving system that exploits predictable tensor lifetimes for trace-guided memory optimization. For each request template, Xema derives an offline memory trace to identify short memory-pressure intervals and applies memory mitigation only within these intervals and only by the amount needed to fit the target GPU budget. Xema further constructs a static memory layout for tensors with predictable lifetimes, reducing fragmentation-induced reserved memory and making offline memory reasoning reliable at runtime. Built on this memory optimization layer, Xema introduces an offline planner that jointly selects parallelism, concurrency, and memory control under GPU memory and SLO constraints. The selected plan is stored in a plan table and directly used by the online serving runtime. We implement Xema on production diffusion pipelines and evaluate it with Flux.2, CogVideoX-5B, and LTX-2. Compared with existing serving configurations, Xema improves SLO attainment by up to 3.7x and reduces planning cost from 6.3 hours to 197 seconds compared with grid search.
Speculative decoding accelerates autoregressive generation by letting a lightweight drafter propose multiple tokens that are verified by a larger target model. Although effective for text-only LLMs, speculative decoding yields limited gains in VLMs because drafters often diverge on vision-critical content, while existing multimodal acceleration methods do not directly address irrelevant visual evidence or optimize the verifier-accepted prefix length that governs speedup. We propose TIGER, a Text-conditioned vIsual GatEd Routing framework for multimodal speculative decoding. TIGER dynamically selects a sparse set of context-relevant visual tokens based on the drafter's current textual state, rather than expose the full visual token set or a fixed compressed interface. To better align training with inference-time efficiency, we optimize the drafter with acceptance-aligned group-based policy training using verifier-derived rewards based on accepted prefix length, built on top of distillation warm start with KL anchoring. This encourages the drafter not only to imitate the target model, but also to produce speculative continuations that survive verification for longer prefixes. Experiments show that TIGER yields consistent gains in accepted prefix length and speculative speedup under exact verifier-side speculative decoding, while achieving favorable quality-latency trade-offs with comparable downstream accuracy in visual-routing analyses.
展开完整摘要收起摘要↓
Speculative decoding accelerates autoregressive generation by letting a lightweight drafter propose multiple tokens that are verified by a larger target model. Although effective for text-only LLMs, speculative decoding yields limited gains in VLMs because drafters often diverge on vision-critical content, while existing multimodal acceleration methods do not directly address irrelevant visual evidence or optimize the verifier-accepted prefix length that governs speedup. We propose TIGER, a Text-conditioned vIsual GatEd Routing framework for multimodal speculative decoding. TIGER dynamically selects a sparse set of context-relevant visual tokens based on the drafter's current textual state, rather than expose the full visual token set or a fixed compressed interface. To better align training with inference-time efficiency, we optimize the drafter with acceptance-aligned group-based policy training using verifier-derived rewards based on accepted prefix length, built on top of distillation warm start with KL anchoring. This encourages the drafter not only to imitate the target model, but also to produce speculative continuations that survive verification for longer prefixes. Experiments show that TIGER yields consistent gains in accepted prefix length and speculative speedup under exact verifier-side speculative decoding, while achieving favorable quality-latency trade-offs with comparable downstream accuracy in visual-routing analyses.
Low-bit quantization makes small reasoning models inexpensive to deploy but can degrade their chains of thought. This motivates decoder-side monitors that intervene when generation becomes unreliable. We show that a natural candidate, the centered token log-probability increment $\log p(w_t)+H_t$, is the wrong observable for this purpose. Under the model's own sampling law it is a mean-zero martingale by construction, so it measures sampling self-consistency rather than trajectory health and is nearly silent during confident repetition, where both $\log p(w_t)$ and entropy are close to zero. We introduce a training-free decoding controller that combines (i) a degeneration-aware alarm score fusing token uncertainty with explicit verbatim repetition and (ii) a calibrated e-process-inspired sequential detector. The raw product process is Ville-valid under a conditional-mean null, while the deployed CUSUM-floored statistic is treated as an empirical change detector because the score is history-dependent and autocorrelated. On GSM8K with DeepSeek-R1-Distill-Qwen-1.5B in FP16 and INT4, calibration turns a monitor that fires on 93--95% of generations into a selective detector of failing traces ($φ\approx 0.3$, precision $\approx 0.6$ against a 0.38 base rate). In this pilot, the controller reduces measured verbatim-degeneration signals and yields a positive but statistically inconclusive INT4 accuracy change from 63% to 69% (paired McNemar $p=0.18$, $n=100$), at a 28% token-budget cost. We also find that non-termination, rather than looping, is the dominant failure mode on GSM8K. The main contribution is methodological: an explanation of why centered token log-probability is inadequate for decoder monitoring and a calibrated, cautiously evaluated replacement.
展开完整摘要收起摘要↓
Low-bit quantization makes small reasoning models inexpensive to deploy but can degrade their chains of thought. This motivates decoder-side monitors that intervene when generation becomes unreliable. We show that a natural candidate, the centered token log-probability increment $\log p(w_t)+H_t$, is the wrong observable for this purpose. Under the model's own sampling law it is a mean-zero martingale by construction, so it measures sampling self-consistency rather than trajectory health and is nearly silent during confident repetition, where both $\log p(w_t)$ and entropy are close to zero. We introduce a training-free decoding controller that combines (i) a degeneration-aware alarm score fusing token uncertainty with explicit verbatim repetition and (ii) a calibrated e-process-inspired sequential detector. The raw product process is Ville-valid under a conditional-mean null, while the deployed CUSUM-floored statistic is treated as an empirical change detector because the score is history-dependent and autocorrelated. On GSM8K with DeepSeek-R1-Distill-Qwen-1.5B in FP16 and INT4, calibration turns a monitor that fires on 93--95% of generations into a selective detector of failing traces ($φ\approx 0.3$, precision $\approx 0.6$ against a 0.38 base rate). In this pilot, the controller reduces measured verbatim-degeneration signals and yields a positive but statistically inconclusive INT4 accuracy change from 63% to 69% (paired McNemar $p=0.18$, $n=100$), at a 28% token-budget cost. We also find that non-termination, rather than looping, is the dominant failure mode on GSM8K. The main contribution is methodological: an explanation of why centered token log-probability is inadequate for decoder monitoring and a calibrated, cautiously evaluated replacement.