paperlist.ai

AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Semantic Class Distribution Learning for Debiasing Semi-Supervised Medical Image Segmentation

Yingxue Su, Yiheng Zhong, Keying Zhu, Zimu Zhang, Zhuoru Zhang, Yifang Wang, Yuxin Zhang, Xinyuan Zheng, Jingxin Liu, Xiaofeng Liu

Medical image segmentation is critical for computer-aided diagnosis. However, dense pixel-level annotation is time-consuming and costly, and medical datasets often exhibit severe class imbalance. Such an imbalance causes minority structures to be overwhelmed by dominant classes in feature representations, hindering the learning of discriminative features and making reliable segmentation particularly challenging. To address this, we propose the Semantic Class Distribution Learning (SCDL) framework, a plug-and-play module that mitigates supervision and representation biases by learning structured class-conditional feature distributions. SCDL integrates Class Distribution Bidirectional Alignment (CDBA) to align embeddings with learnable class proxies and leverages Semantic Anchor Constraints (SAC) to guide proxies using labeled data. Experiments on the Synapse and AMOS datasets demonstrate that SCDL largely improves segmentation performance across both overall and class-level metrics, with particularly notable gains for several low-frequency organs. Our anonymous code is released at https://anonymous.4open.science/r/SCDL.

arXiv

VTONQA: A Multi-Dimensional Quality Assessment Dataset for Virtual Try-on

Xinyi Wei, Sijing Wu, Zitong Xu, Yunhao Li, Huiyu Duan, Jia Wang, Ning Liu

With the rapid development of e-commerce and digital fashion, image-based virtual try-on (VTON) has attracted increasing attention. However, existing VTON models often suffer from artifacts such as garment distortion and body inconsistency, highlighting the need for reliable quality evaluation of VTON-generated images. To this end, we construct \textbf{VTONQA}, the first multi-dimensional quality assessment dataset specifically designed for VTON, which contains 8,132 images generated by 11 representative VTON models, along with 24,396 mean opinion scores (MOSs) across three evaluation dimensions (\textit{i.e.}, clothing fit, body compatibility, and overall quality). Based on VTONQA, we benchmark both VTON models and a diverse set of image quality assessment (IQA) metrics, revealing the limitations of existing methods and highlighting the value of the proposed dataset. We believe that the VTONQA dataset and corresponding benchmarks will provide a solid foundation for perceptually aligned evaluation, benefiting both the development of quality assessment methods and the advancement of VTON models. The dataset we proposed in this paper is publicly available at: https://huggingface.co/datasets/weixiny0408/vtonqa.

arXiv

LM-CartSeg: Automated Segmentation of Lateral and Medial Cartilage and Subchondral Bone for Radiomics Analysis

Tongxu Zhang, Zongpan Li, Aaron Kam Lun Leung, Siu Ngor Fu

Background and Objective: Radiomics of knee MRI requires robust, anatomically meaningful regions of interest (ROIs) that jointly capture cartilage and subchondral bone. Most existing work relies on manual ROIs and rarely reports quality control (QC). We present LM-CartSeg, a fully automatic pipeline for cartilage/bone segmentation, geometric lateral/medial (L/M) compartmentalization and radiomics analysis. Methods:Two 3D nnU-Net models were trained on SKM-TEA (138 knees) and OAIZIB-CM (404 knees). At test time, zero-shot predictions were fused and refined by simple geometric rules: connected-component cleaning,construction of 10mm subchondral bone bands in physical space, and a data-driven tibial L/M split based on PCA and $k$-means. Segmentation was evaluated on an OAIZIB-CM test set (103 knees) and on SKI-10 (100 knees). QC used volume and thickness signatures. From 10 ROIs we extracted 4,650 non-shape radiomic features to study inter-compartment similarity, dependence on ROI size, and OA vs. non-OA classification on OAIZIB-CM and a clinical Po-OA cohort (185 knees). Results: Post-processing improved macro ASSD on OAIZIB-CM from 2.63 to 0.36mm and HD95 from 25.2 to 3.35mm, with DSC approx 0.91; zero-shot DSC on SKI-10 was approx 0.80. The geometric L/M rule produced stable compartments across datasets, whereas a direct L/M nnU-Net showed domain-dependent side swaps. Only 6-12% of features per ROI were strongly correlated with volume or thickness. Radiomics-based models achieved AUC up to 0.91 (OAIZIB-CM) and 0.83 (Po-OA), clearly exceeding models restricted to size-linked features. Conclusions: LM-CartSeg yields automatic, QC'd ROIs and radiomic features that carry discriminative information beyond simple morphometry, providing a practical foundation for multi-centre knee OA radiomics studies. Code is available at https://github.com/jukieCheung/LM-CartSeg.

arXiv

PLASMA: A Layout-Aware Benchmark Reveals Memory Layout Matters for Graph-based ANNS on GPU

Yutaro Oguri, Mai Nishimura, Yusuke Matsui

We propose a $\textbf{P}$latform for $\textbf{L}$ayout-$\textbf{A}$ware $\textbf{S}$earch and $\textbf{M}$emory $\textbf{A}$rrangement ($\textbf{PLASMA}$), a unified evaluation framework for graph-based Approximate Nearest Neighbor Search (ANNS) on GPU that isolates the effects of graph index topology and memory layout. Graph-based ANNS is essential in modern AI applications such as RAG, and GPU utilization is attracting attention for datasets of millions or more vectors. Our framework extracts the topology of arbitrary graph-based indices and enables execution under a unified, GPU-optimized search algorithm, specifying the correspondence between vertex IDs and positions on memory to allow arbitrary vertex orderings. Through comprehensive experiments, we demonstrate that vertex reordering yields up to $80\%$ (typically $10-30\%$) QPS improvement while preserving search accuracy.

arXiv

The Linear Geometry of Interpretable Tokens: Jailbreaking Attacks and Defenses for Unlearned Diffusion Models

Siyi Chen, Yimeng Zhang, Sijia Liu, Qing Qu

Diffusion models excel at generating high-quality images but can memorize and reproduce harmful concepts when prompted. Although fine-tuning methods have been proposed to unlearn a target concept, they struggle to fully erase it while maintaining generation quality on other concepts, leaving models vulnerable to jailbreak attacks. Existing jailbreak methods demonstrate this vulnerability but offer limited insight into how unlearned models retain harmful concepts, limiting progress on effective defenses. In this work, we show that the erased concept persists as a coherent, interpretable linear subspace of the token embedding space, and that both an attack and a defense follow directly from this structure. We introduce SubAttack, a novel jailbreaking attack that reads out this subspace by learning an orthogonal set of attack token embeddings, each being a linear combination of human-interpretable textual elements, revealing that unlearned models still retain the target concept through related textual components. Furthermore, our attack is also more powerful and transferable across text prompts, initial noises, and unlearned models than prior attacks. Conversely, projecting out the same subspace yields SubDefense, a lightweight plug-and-play defense mechanism that suppresses the residual concept in unlearned models. SubDefense provides stronger robustness than existing defenses while better preserving safe generation quality. Extensive experiments across multiple unlearning methods, concepts, and attack types demonstrate that our approach advances both understanding and mitigation of vulnerabilities in diffusion unlearning.

arXiv

Generalizable AI-Generated Image Detection Based on Fractal Self-Similarity in the Spectrum

Shengpeng Xiao, Yuanfang Guo, Heqi Peng, Hui Miao, Zeming Liu, Liang Yang, Jiantao Zhou, Yunhong Wang

With the rapid development of image synthesis techniques, AI-generated images have become increasingly realistic, which heightens the potential risk associated with their misuse and creates a growing need for reliable detection. However, the growing diversity of generative models makes it increasingly difficult for detectors to generalize to images produced by unseen generators. Most existing methods rely on artifacts associated with specific generators, which limits their generalization to images produced by unseen models. To address this problem, we investigate structural characteristics arising from the image generation process itself. Image generation fundamentally involves constructing spatially rich content from more compact representations, while preserving the semantic identity of structures across different spatial locations. We formalize these properties through dimension-increasing shift-equivariant transformations and show that such transformations induce a self-similar structure in the Fourier spectrum. Across successive generation stages, this structure can propagate recursively and form a hierarchical fractal self-similar pattern. Consequently, different spectral sub-regions exhibit consistent structural correspondences inherited from the generation process, providing a generator-agnostic cue for detection. Based on this observation, we propose Fractal-CNN, which captures spectral self-similarity rather than generator-specific spectral values. Extensive experiments across diverse GAN- and diffusion-based generators demonstrate that Fractal-CNN achieves strong cross-generator generalization, with an average detection accuracy of 93.93% across 16 test generators.

arXiv

Iterative Flow Matching: Path Correction and Gradual Refinement for Enhanced Generative Modeling

Eldad Haber, Shadab Ahamed, Md. Shahriar Rahim Siddiqui, Niloufar Zakariaei, Moshe Eliasof

Generative models for image generation are now commonly used for a wide variety of applications, ranging from guided image generation for entertainment to solving inverse problems. Nonetheless, training a generator is a non-trivial feat that requires fine-tuning and can lead to so-called hallucinations, that is, the generation of images that are unrealistic. In this work, we explore image generation using flow matching. We explain and demonstrate why flow matching can generate hallucinations, and propose an iterative process to improve the generation process. Our iterative process can be integrated into virtually any generative modeling technique, thereby enhancing the performance and robustness of image synthesis systems.

arXiv

Interactive Stroke-based Neural SDF Sculpting

Fizza Rubab, Yiying Tong

Recent advances in implicit neural representations have made them a popular choice for modeling 3D geometry. However, directly editing these representations presents challenges due to the complex relationship between model weights and surface geometry, as well as the slow optimization required to update neural fields. Among various editing tools, sculpting stands out as a valuable operation for the graphics and modeling community. While traditional mesh-based tools like ZBrush enable intuitive edits, a comparable high-performance toolkit for sculpting neural SDFs is currently lacking. We introduce a framework that enables interactive surface sculpting directly on neural implicit representations with optimized performance. Unlike previous methods, which are limited to spot edits, our approach allows users to perform stroke-based modifications on the fly, ensuring intuitive shape manipulation without switching representations. By employing tubular neighborhoods to sample strokes and customizable brush profiles, we achieve smooth deformations along user-defined curves, providing intuitive control over the sculpting process. Our method demonstrates that versatile edits can be achieved while preserving the smooth nature of implicit representations, all without compromising interactive performance.

arXiv

Squeezing the Cache, Preserving the Truth: Monotonic Equipotential Allocation with Geodesia-KV

Vincenzo Dentamaro, Pancrazio Auteri, Giuseppe Pirlo

Current assessment of KV-cache compression performance confuses resident bits with read bandwidth and is affected by the artifacts of chunked teacher-forcing. We present Geodesia-KV, a family of training-free KV cache policies based on monotonic block-wise precision allocation, exact rate-distortion residuals, and query-sparse reading, enabling proper hardware-ready compression. With proper separation of resident and read bits and causal evaluation, we show that Geodesia-KV significantly outperforms other approaches. Specifically, on WikiText-2 with 16k context, the 5-bit operating point of Geodesia-KV results in lower perplexity at lower bitrate than KIVI-4 on Qwen. In addition, our compressed-Quest version delivers improved perplexity and reduces resident (9.83 vs 16.25 bits/value) and read rates (1.95 vs 2.32 bits/value) over baseline sparse methods on PG-19. As Geodesia-KV is implemented as native GeodesiaKVCacheManager plug-in of vLLM, Geodesia-KV fully removes the need for dense cache residency via monotonic bit demotion. With the full consumer hardware evaluation, Geodesia-KV leads to 1M-token context generation on a single 16 GiB GPU with up to 71.7% peak VRAM savings on all leading architectures (Qwen, Llama, DeepSeek).

arXiv

Activation-Keyed Momentum: An Anisotropic Momentum Update via the Delta Rule

Euijin Hong, Guannan Qu

Most modern optimizers form their momentum as an exponential moving average (EMA) of past gradients, forgetting every direction at one fixed rate. However, the inputs a deep network sees during training can be highly anisotropic, with a few directions queried frequently while most are seen rarely. Preconditioning methods address this anisotropy by wrapping extra processing around this buffer and leave the momentum update itself unchanged. We propose Activation-Keyed Momentum (AK-Momentum), which builds direction-awareness into the momentum update rule. The gradient of a linear layer splits into an input activation that acts as a key and an output-side error that acts as a value. Keying on that activation, AK-Momentum updates the momentum buffer by the canonical delta rule, so each direction is forgotten at a rate set by how often it appears. We prove that it is a valid momentum, that it applies the input-side curvature correction without matrix inversion, and that it clears stale directions faster than EMA under both a fixed and a drifting optimum. It is a drop-in replacement for the momentum buffer of any optimizer, its coefficient transfers across widths under $μ$P, and its extra compute stays between $22.2\%$ and $25.0\%$ of a gated-MLP block's linear cost with no persistent memory. In FineWeb-Edu pretraining, AdamW with AK-Momentum (AK-AdamW) reaches AdamW's validation loss in up to $46.39 \pm 4.32\%$ fewer steps at 67M and $22.12 \pm 0.80\%$ at 370M over three seeds, and the gain persists at 1B on a Chinchilla-optimal budget. A Muon baseline tuned under the same protocol sits above AK-AdamW at both language-model scales, and the gain holds for SGD, ResNet-18, and ViT-Tiny on CIFAR-10. Training-time diagnostics confirm the predicted mechanism, better gradient tracking and healthier input directions.

arXiv

Measuring What a Specification Determines: A Formal Semantic-Block Model and an Execution-Judged Benchmark

Oleg Grynets, Dmytro Kostetskyi, Vasyl Lyashkevych

This work introduces a formal semantic-block model for specifications and an execution-judged benchmark for evaluating specification quality independently of model capability. A specification is represented as a structure comprising semantic blocks, dependency relations, block-owned rules, decision points, and explicitly open questions, subject to four machine-checkable well-formedness conditions: acyclicity, single ownership, constraint domination, and totality or ambiguity-stop. Determinacy is defined model-theoretically as agreement among all conforming implementations and is estimated empirically through convergence across independent implementers. The model is instantiated on an Oracle-to-PostgreSQL migration specification containing 18 blocks and 19 dependency edges. Computational validation shows that the five-layer decomposition reduces mean per-task context by approximately 71% through dependency closures, covers 85.5% of the study-defined Oracle construct taxonomy with all identified gaps triaged, is not Pareto-dominated by the tested alternative partitions, and is recovered at the 99.9th percentile from citation-derived edges not used to define the original structure. The benchmark keeps the implementer panel fixed, includes a mandatory no-specification control arm, and uses PostgreSQL 16 and a live Oracle instance as deterministic execution judges. Six designed studies, including three pre-registered manipulations and three diagnostic analyses, further examine specification effects. Repeated runs on a 25-unit subsample reveal an empirical variability floor with a median arm-delta spread of 14.4 percentage points. The results support determinacy as a formal concept but not as a standalone empirical quality metric for the evaluated contemporary LLM implementers.

arXiv

Are LLMs becoming similarly creative? Evidence from three years of models

Nirav Patel, Josiah Crossman, Eva Aggarwal, Emily Wenger

Many benchmarks track Large Language Model (LLM) performance on tasks with verifiable answers, but less is known about how LLM performance is evolving on open-ended tasks, where creativity, originality and diversity may matter as much as quality. As LLMs increasingly support human ideation and creative work, understanding trends in LLM performance on open-ended tasks is critical. This paper presents a preliminary analysis of LLM creative outputs spanning three years of model releases, examining model responses to Infinity-Chat100, a real-world collection of open-ended user queries, and the Alternate Uses Task, an established psychometric creativity assessment. Using sentence-embedding similarity, we examine trends in LLM responses to these prompts. Our findings show a statistically significant decrease in model output diversity over time, suggesting that LLM outputs may be converging in creative substance across models. If this trend persists, LLM-driven homogenization may progressively diminish human agency in human-AI co-creative work, demanding careful consideration of LLMs' role in the human creative process.

arXiv

HARP: Hierarchical Adaptive Ranking with Preference-Adaptive Fusion for Query-Based CVE Prioritization

Haochen Liu, Zhengzhang Chen, Haoyu Wang, Yanchi Liu, Jundong Li, Haifeng Chen

Vulnerability prioritization is inherently preference dependent, since the same CVE can receive different remediation priority under different operational preference scenarios. Existing scoring systems and ranking methods typically assume a fixed criterion. In practice, organizations already operate under a preference scenario, but this preference is often implicit and difficult to express as a written prompt instruction, while triage queries usually do not encode it. Past validated triage cases under the current scenario are more readily available. We study query-based CVE prioritization in this setting and propose HARP, a graph-grounded multi-view framework that ranks candidates from a natural-language query together with a support bank of historical labeled examples from the current preference scenario, without requiring an explicit textual summary of that scenario. HARP retrieves evidence from a vulnerability knowledge graph, scores candidates with policy-conditioned global, enterprise, and user views, and fits view-fusion weights from sampled supports. Experiments across three preference scenarios and multiple backbone LLMs show that HARP outperforms multiple baselines, expressing our method's effectiveness.

arXiv

Linguistic Holonomy and Statistical Watermarks: Inner Geometry of Meaning-Preserving Transformations

Daniele Corradetti

Statistical watermarks for language models live in the freedom of the signifier: they choose among tokens that are nearly equivalent in meaning, and they are therefore eroded by exactly those transformations which move the form of a text while leaving its content in place. The literature measures such transformations by their endpoint, through the semantic similarity between the original and the rewritten text. We show that the endpoint is the wrong statistic. Adapting the formalism of linguistic loops, we prove that the invariant of a chain of meaning-preserving transformations factorises canonically into an endpoint part and a holonomy in the stabiliser of the initial state, the second of which the semantic deficit cannot see; the loop rotation is parallel transport on the unit sphere of the embedding space, so that the analogy with the Wilson loop becomes a theorem rather than a figure of speech. On the side of the detector we prove an exact identity: the residual statistic is proportional to the number of positions whose seeding window survived intact, from which the decay law $ρ^{h+1}$ follows as the independent-edit corollary. The identity has a disconcerting consequence, which we confirm to three decimal places: at one and the same retention rate the surviving signal may be one half of the original, one quarter of it, or exactly nothing, according only to where the edits fall.

arXiv

A Speech Corpus for Mizo Automatic Speech Recognition: Whisper and SraVaani 1.0 Fine-Tuning with Morphology-Aware Evaluation

Priyankoo Sarmah, Sanasam Ranbir Singh, Lalhmingmawia

This study reports the development of an Automatic Speech Recognition (ASR) system in Mizo, a low-resource language. The development included collecting 17.62 hours of speech data, curating it, and fine-tuning the Mizo ASR system with three Whisper multilingual models and with the SraVaani 1.0 Indic multilingual model. Whisper-large-v3 achieved the lowest conventional WER (18.08%), while morphology-aware evaluation yielded a WER of 7.22%. Zero-shot evaluation of the SraVaani 1.0 Indic multilingual model yielded a WER of 58.27%, while Mizo-specific fine-tuning reduced the conventional WER to 29.45% and the morphology-aware WER to 17.93%. The results demonstrate that the Whisper model can achieve a substantially low WER, even when adapted to an unseen language. In contrast, SraVaani 1.0 supports the Mizo language in its multilingual model; however, fine-tuning with carefully curated Mizo speech data substantially improves its performance.

arXiv

Outcome Monitors: Recovery Affordances for Silent Tool Failures

Sugam Panthi, Rabab Abdelfattah

When a tool call times out, the agent sees the failure and can route around it. A cached error page or negative price can instead arrive in the expected format and be consumed as fact. We introduce Outcome Monitors, which detect violations of outcome contracts mined from task-disjoint traces or derived from public schemas. On a violation, the monitor preserves the result and issues a nonbinding receipt naming the violated property and public recovery tools. In frozen, prespecified evaluations with injected failures, Outcome Monitors raise ToolMaze completion from 10.9% to 28.1% across four models in two provider families and replicate in a third. In tau-bench retail, completion improves by 14.0 and 12.0 points on two tiers. In separate ToolMaze controls, removing the recovery-tool list eliminates the measured gain and restoring it recovers the effect; diagnostic detail and timing produce no detectable differences. Gains concentrate where the fault blocks completion. On a suite transcribed from a published incident taxonomy, detection outside the mined vocabulary falls to 46%, though delivery continues and completion is unchanged. Recovery tools are the active receipt content in these controls; extending detection beyond the contract vocabulary remains open.

arXiv

PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX

Genghan Zhang, Yixin Dong, Chengze Fan, Zhichen Zeng, Yueming Yuan, Shaowei Zhu, Kunle Olukotun

We introduce PTXBench, a benchmark for evaluating and adapting large language models (LLMs) to use architecture-specific PTX for GPU kernel optimization. PTXBench measures functional correctness, whether selected target instructions execute at runtime, and speedup over frontier libraries across GEMM and attention workloads on H100 and B200 GPUs. Our evaluation shows that architecture-specific PTX capability remains uneven: success rates fall substantially on complex attention backward workloads, and executing the target instructions does not necessarily translate into competitive performance. No evaluated model consistently matches frontier libraries across the suite. We further adapt Qwen3.6-27B using supervised fine-tuning. Repair-conditioned training improves several tasks, but generalization remains uneven; data coverage, balance, and the quality of the reasoning teacher matter in addition to dataset size. PTXBench provides an auditable testbed for measuring and improving LLMs' ability to exploit evolving GPU architectures.

arXiv

Geometric and Behavioral Stratification in Transformer Residual Streams

Nelson Guda

Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream. But what kind of direction does such a basis select? We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor. Measured with respect to this anchor, residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction. The stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned). A narrow, scale-invariant prediction interface concentrates readout-relevant structure, while the vast prediction-distal complement expands with model scale. Because the prediction direction sits nearly orthogonal to the principal variance axes, variance-based analyses recover this organization only partly, and the shortfall grows with prompt heterogeneity. Anchoring reveals a steep geometric gradient: prediction-proximal regions are highly structured and cluster related prompts, while the complement is flatter and anti-discriminates among prompt groups. The interface is a narrow slice but functionally decisive. Disrupting the variance directions closest to the prediction causes immediate divergence and frequent task-frame shifts; disrupting the next level down delays divergence and preserves framing. The complement is weakly readout-aligned per direction yet causally and temporally load-bearing, and behavior is driven by direction rather than magnitude. These results establish the prediction direction as a privileged anchor distinct from previously described coordinate axes, and give a geometric account of how high-dimensional computation coexists with linear readout.

arXiv

Beyond Recall: Behavioral Specification as an Interpretive Layer for AI Personalization

Aarik Gulaya

If an AI agent makes decisions on a person's behalf, those decisions must align with its user. We introduce representational accuracy to measure how faithfully a system captures a person's interpretation. An interpretive layer is operationalized as a Behavioral Specification. Our reference implementation aggressively compresses a person's data into interpretive patterns, served as context to a language model. We evaluate the Specification on a prototype benchmark of held-out behavioral predictions scored by a calibrated 5-judge LLM panel. We test it independently and in composition with a range of context conditions: full raw corpus, full extracted facts, and four commercial memory systems (Mem0, Letta, Supermemory, Zep). Across 14 public-domain autobiographical corpora, the Specification lifts representational accuracy in aggregate and nearly eliminates model hedging. It recovers most of what the raw corpus delivers, at ~25x less context cost. The Specification lifts subjects toward a common predictive level regardless of pretraining baseline; the lift in absolute points is therefore largest where the baseline is lowest, suggesting the population of relevance is anyone not adequately represented in pretraining. Lift is greatest on interpretation-required questions, where providing an interpretive layer enables model behavior that extracted facts or raw corpus do not. Conversely, on recall-required questions, this layer can interfere rather than help. We conclude that representational accuracy is distinct from recall and that human-AI alignment is dependent on how accurately the user is represented. Representational accuracy makes that alignment testable.

arXiv

RequestRouter: Request-Boundary Routing for Efficient Single-GPU LLM Inference

Aman Sunesh, Ali Alshehhi, Hivansh Dhakne

RequestRouter is a lightweight request-boundary controller for reducing the latency and energy cost of single-GPU large language model inference. Rather than serving all requests with one static configuration, the system uses cheap request-level features to select one fixed inference mode per request, including FP16, quantized inference, speculative decoding, prefix caching, continuous batching, and hybrid modes such as GPTQ plus prefix caching and INT8 plus continuous batching. We evaluate RequestRouter using an 8B instruction-tuned language model served through vLLM on NVIDIA A100 GPUs. Across the full-scale A100 evaluation--26,500 fixed-mode evaluations followed by 3,500 online-controller evaluations, for 30,000 measured inference executions in total--the controller achieves a 2.10x mean latency speedup over FP16 and a 0.48x energy ratio on deployment-style workloads. A smaller matched evaluation with repeated measurements provides a controlled statistical check of this result: RequestRouter retains a 1.93x latency speedup (95% CI: 1.88--1.98x) and a 0.523 energy ratio (95% CI: 0.506--0.540), showing that the gains persist under a more tightly controlled protocol. On a separate expanded automatic benchmark evaluation, the routed policy retains 99.6% of FP16 macro accuracy. A 100,000-call CPU microbenchmark measures only 0.00475 ms mean routing overhead (0.00532 ms p99). Thus, simple request-aware routing can recover substantial serving efficiency without retraining or modifying the underlying LLM.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.