model deployment teams

Abstract: Large Language Models (LLMs) based on Transformer architectures have achieved remarkable performance across artificial intelligence tasks, but their computational cost remains a major bottleneck. Structured sparsification provides a promising approach to reducing this cost. In particular, $\ell_{1,\infty}$ projections can induce column-wise sparsity in model matrices, but existing proximal methods (Prox) typically require $\mathcal{O}(nm\log(nm))$ operations for matrices in $\mathbb{R}^{n\times m}$, limiting their computational and energy efficiency. In this paper, we introduce Two-Stage Clipping (TSC), an efficient structured sparsification operator inspired by bilevel optimization. Rather than solving a bilevel optimization problem in the strict sense, TSC decomposes the computation into two successive stages. First, group-level quantities are projected to allocate the global sparsity budget and identify active groups. Second, the resulting values are used as group-wise clipping thresholds for the matrix coefficients. For square attention matrices ($n \times n$), the complexity of our TSC method is $O(n^2)$, which is linear in the number of matrix coefficients. We establish a theoretical bound on the discrepancy between TSC and the exact proximal projection and show on synthetic data that the additional Euclidean approximation cost remains small in practice. Extensive benchmarks demonstrate substantial runtime improvements over proximal baselines and validate the theoretical norm identities underlying the proposed construction.

Abstract: The increasing release and reuse of Large Language Models (LLMs) as open-weight checkpoints make model fingerprinting important for ownership protection and offline auditing. Existing fingerprinting methods typically construct evidence from query responses, generated outputs, or input-induced activations. Although effective in representation-based settings, they can be affected by prompt selection, decoding configuration, inference environment, and forward-pass implementation, which may limit their applicability. Here we introduce a model-intrinsic weight-level framework that derives compact fingerprints from spectral structures encoded in LLM parameters. Specifically, the framework constructs GATE spectral scalar features from the singular-value distributions of GATE projection weights in gated feed-forward networks (FFNs), and aggregates the resulting layer-wise features through depth-wise segmented pooling to form a model-level fingerprint. Such representation preserves coarse layer-depth organization while reducing dependence on strict layer-to-layer alignment, enabling comparison across LLMs with different depths. The similarity between fingerprints is computed by linear centered kernel alignment. The entire pipeline is training-free and does not require input samples, generated outputs, forward activations, or an auxiliary verifier. Experiments on representative LLM families and post-training variants have demonstrated a clear separation between source-related models and unrelated ones, and stable similarity under instruction tuning, quantization, pruning, and distillation-related modifications, which verify that spectral features are robust evidence for LLM fingerprinting.

Abstract: Although post-training significantly enhances model performance, the use of massive undisclosed data raises concerns about benchmark contamination and copyright infringement. This drives growing interest in training data detection for post-trained models. Existing evaluations for training data detection often leverage benchmarks constructed post hoc, which can introduce distribution shifts. Alternatively, controlled evaluations rely on small-scale training setups that diverge from realistic post-training settings. As a result, their reported performance may not faithfully reflect the effectiveness of detection methods in real-world scenarios. To this end, we introduce \textsc{\method}, a reliable benchmark designed to evaluate detection methods in realistic, large-scale post-training settings. Built on a multi-stage pipeline, \method\space encompasses stage-specific models throughout the post-training lifecycle, spanning SFT, DPO, and RL. To perform a reliable evaluation, we construct model-specific evaluation sets across four diverse tasks, featuring distribution-matched members and non-members with verifiable membership labels. Our results reveal that even the best-performing method achieves an average AUC of only 0.529 on our benchmark, which approaches random-guessing level (0.5). Furthermore, our analysis suggests that evaluation protocols can significantly affect measured detection performance. These findings highlight the significant potential for developing effective, practical, and robust detection methods under realistic scenarios.

Abstract: Low-Rank Adaptation (LoRA) has become a standard approach for parameter-efficient fine-tuning, yet a fundamental practical question remains unresolved: how should the adapter rank be chosen? An overly small rank may lead to a poorly conditioned optimization landscape, whereas an unnecessarily large rank sacrifices the efficiency that motivates LoRA in the first place. Existing theoretical analyses provide only limited guidance on this trade-off, and their guarantees are typically established under restrictive theoretical settings. We address this gap by developing a substantially sharper landscape theory for LoRA, building on modern results from nonconvex low-rank matrix sensing. Our central insight is that the appropriate adapter rank should depend on the quality of the data-induced optimization geometry, rather than on the model alone. To formalize this connection, we introduce LoRA-RIP, a data-dependent restricted-isometry metric that characterizes the conditioning of the cross-entropy (CE) objective along LoRA-relevant low-rank directions. We prove that sufficient rank over-parameterization, with the required rank explicitly determined by the LoRA-RIP constant, eliminates spurious local minima, thereby extending existing RIP-based guarantees beyond the classical 1/3 regime. This characterization further enables principled data selection under a fixed rank budget. Experiments across language and vision tasks support these theoretical predictions, showing that rank and data quality are two coupled resources that should be jointly considered for more efficient and reliable LoRA fine-tuning.

Abstract: Existing attacks on unlearned diffusion models assume that the erased concepts are known in advance and focus on recovering them. In practice, however, model providers may not disclose which concepts have been removed, and even with access to the original base model, an adversary may still lack a clear target to attack. In this paper, we aim to answer the following critical but overlooked questions: which concepts have been erased from the model, and how many have been erased in total? To this end, we present Tracer, a framework that rapidly and accurately identifies erased concepts and estimates their number. Tracer efficiently identifies erased concepts without generating and classifying images. By combining lightweight spectral analysis of weight footprints, it enables efficient search over large candidate vocabularies. To distinguish multiple erased concepts, we introduce a footprint coverage objective that guides sequential discovery. Tracer estimates the number of erased concepts by detecting a sharp decline in candidate confidence as the selected concepts account for the erasure footprint, without requiring labeled examples for calibration. The framework requires only lightweight linear algebra and limited forward probes, with no prior knowledge of the unlearning algorithm. Experiments across text-to-image and text-to-video backbones and diverse unlearning methods demonstrate that Tracer identifies erased concepts and estimates their number in seconds, achieving 150 to 137,000 times and 133 to 20,000 times speedups over MIA and brute-force search on image and video models, respectively, with substantially higher identification accuracy.

Abstract: Zero-knowledge proofs are emerging as a promising approach for enabling private, verifiable LLM governance and auditing, where regulators, users, and auditors need to verify claims about training-data usage or LLM inference-time behavior, while model providers must protect proprietary model parameters. However, despite the growing interest in ZK-LLMs, the understanding of ZK-friendly quantization remains limited. This gap matters because in the ZK setting, quantization directly shapes the arithmetic structure, constraint complexity, and proving cost of ZK inference. ZK protocols operate over finite fields and incur costs that depend heavily on the number and type of arithmetic operations, nonlinearities, and lookup constraints. Understanding ZK-friendly quantization is therefore essential for making ZK-LLMs practical. In this work, we present the first systematic study of ZK-friendly quantization for LLMs. We first formalize the definition of ZK-friendly quantization, capturing the properties required for ZK proof generation. We then evaluate nine language models, including Qwen2.5-14B and the mixture-of-experts model Qwen3-30B-A3B, across a broad design space of weight, activation, and nonlinear lookup table precision. Our results show that activation precision is substantially more sensitive than weight precision, while nonlinear lookup approximations can become the dominant source of utility degradation. Also, we identify RMSNorm inverse-square-root lookups as a recurring bottleneck in several large models and recover near-baseline utility by selectively increasing precision only at the bottleneck. Finally, we show that reducing bit-width or lookup-table size does not necessarily yield proportional end-to-end proving savings, showing that conventional low-bit quantization heuristics do not directly translate to ZK proving efficiency and motivating operator-aware precision selection.

Abstract: Semantic watermarks for diffusion models embed ownership information into the generative process while preserving perceptual quality, but Gaussian-Shading extraction conventionally requires multi-step diffusion inversion to recover the initial noise. Recent one-step methods show that this cost can be reduced substantially. We study this problem through extended flow matching and conditional regression. The key observation is that, near the high-SNR image endpoint, recovering a useful noise statistic given by the first-step output of the extended flow matching in the high-SNR regime is much simpler than reconstructing the full inverse trajectory, and Gaussian Shading only requires the recovered latent to remain in the correct watermark decision region. Based on this observation, we propose a lightweight, prompt-free extractor that decomposes endpoint recovery into an image-like anchor and a noise-oriented residual, which increases the capability of the model to utilize GPU parallel computation. The resulting method avoids iterative inversion and repeated evaluation of a diffusion-scale U-Net, providing an efficient one-step extraction pipeline with a concise theoretical interpretation. The computational cost of extracting noise is lower than that of both OSI and FARI. The github repo is there: https://github.com/TheLovesOfLadyPurple/RAIN-lightweight-NN-for-one-step-semantic-watermark-extraction