Pathology foundation models improve transferable representation learning for histopathology, but recent gains often rely on encoders with hundreds of millions of parameters and high inference cost. We propose TAP-Path, a task-adaptive compression framework that directly restructures a pretrained Virchow2 encoder rather than distilling it into a separate student. TAP-Path combines validation-driven transformer-block selection, physical removal of redundant blocks, input-adaptive patch-token pruning, multi-depth feature recovery, and a lightweight gated task head. The final model retains 24 of 32 transformer blocks and 70% of patch tokens after pruning, reducing encoder parameters by 24.96% (631.24M to 473.70M) and analytical encoder compute by 35.20% (340.13G to 220.40G FLOPs). Across three task-head optimization seeds, TAP-Path achieved $87.98 \pm 0.067%$ test accuracy, $81.26 \pm 0.49%$ balanced accuracy, and $82.38 \pm 0.48%$ macro-F1 on a 32-class histopathology benchmark, compared with 86.89% for full Virchow2 and 87.67% for UNI2-h. TAP-Path achieved a Brier score of $0.1800 \pm 0.0005$ and failure-detection AUROC of $0.9047 \pm 0.0060$. A validation-only rare-aware objective improved rare-class balanced accuracy in a secondary operating analysis. Frozen external evaluation on 433 CPTAC samples yielded $91.22 \pm 0.83%$ accuracy and $91.10 \pm 0.81%$ balanced accuracy. These results show that task-adaptive structural and token sparsification can improve the accuracy-efficiency trade-off of large pathology foundation models while preserving reliability under internal and external evaluation.
Video diffusion models (VDMs) have achieved impressive progress in text-to-video generation, but their high memory and computational costs hinder practical deployment. Quantization-aware training (QAT) is an effective solution for compressing and accelerating advanced generative models without runtime overhead at inference. However, existing QAT methods suffer from a distinctive challenge in VDMs: while they often preserve prompt semantics, global layout, and coarse motion, the quantized model severely degrades visual details, texture fidelity, and sharpness. In this paper, we trace this degradation to the timestep-agnostic design of conventional quantization pipelines, which overlooks the stage-wise functionality of video denoising. In VDMs, early denoising steps mainly establish global structure and motion, whereas middle and late steps refine local appearance and high-frequency details. Based on this insight, we propose DSAQuant, a Denoising-Stage-Aligned Quantization-aware training framework for VDMs. During training, Denoising-Stage Oriented Supervision preserves teacher distillation in early steps for stable structure planning, while shifting later steps toward target-driven optimization to enhance detail reconstruction. During inference, Denoising-Stage Gated Guidance disables CFG in the final denoising steps to prevent it from amplifying quantization-induced errors into high-frequency artifacts. Extensive experiments on the Wan and CogVideoX families under W4A4 and W3A3 settings show that DSAQuant consistently outperforms the SOTA QAT baseline, improving the VBench average score by up to 6.60 under aggressive W3A3 quantization while preserving strong text-video alignment. These results demonstrate that effective VDM quantization requires not only reducing quantization error, but also aligning quantization training and inference with the stage-wise nature of video diffusion.
Francesco Corti, Dong Wang, Young D. Kwon +2cs.LG cs.AI
Deep neural networks deployed in the wild must be both efficient and adaptable, requiring model compression and test-time adaptation (TTA). While both are well studied in isolation, their interaction remains poorly understood. We systematically analyze how structured compression affects a model's ability to adapt under distribution shift. Using ResNet-18 and ViT-Base on CIFAR-10-C and ImageNet-C, we evaluate multiple compression methods combined with standard TTA techniques. We introduce a diagnostic framework that examines representational expressivity and adaptation subspace compatibility. Our results reveal a consistent gap: although compressed models retain high accuracy under supervised adaptation, their TTA performance degrades significantly with increasing compression. We show that this stems from reduced representational diversity and structural constraints that limit recoverability. These effects strongly depend on the compression method, highlighting the need to design compression strategies that preserve adaptability.
On-premise assistants can give factory workers conversational access to machine documentation, but models capable of the task rarely fit shop-floor hardware. We show that after structural compression and retrieval-grounded adaptation, model size is no longer a reliable predictor of adapted answer quality: general capability falls almost linearly with parameter count, while judged retrieval-augmented answer quality does not. We therefore treat deployment as a post-adaptation selection problem, committing one sub-network per device on judged answer quality and measured on-device throughput under a configurable general-capability floor and memory budget; rules that optimize size, speed, or quality alone each give up capability or throughput. A weight-shared supernetwork trained with sandwich-style in-place distillation keeps this selection inexpensive. In a manufacturing-manual case study, extraction costs 13.7 percent of the unpruned model's judged quality and retrieval-grounded distillation returns it to within 4.6 percent, recovering two thirds of the loss, and the same assistant runs across three heterogeneous edge tiers at 1.3 to 5 watts standby.
Deploying 3D point cloud models on edge hardware such as the NVIDIA Jetson Orin Nano is severely constrained by compute and memory budgets. Existing compression methods require access to the model's original source code, rendering them inapplicable to the Open Neural Network Exchange (ONNX) binaries commonly distributed by vendors and model repositories. We present \textbf{H3DNAS}, a hardware-aware model compression framework that operates directly on ONNX computational graphs without requiring original source code, architecture class definition, or gradient access during search. H3DNAS makes three contributions: (1) a \textbf{Channel Dependency Graph (CDG)} that classifies ONNX operators into four constraint classes and formally establishes that the free parameter fraction $ρ_f$ is topological invariant, a provable compression ceiling computable in $\mathcal{O}(|V|+|E|)$; (2) a \textbf{Two-Stage Hierarchical Search} that prunes candidate architectures by $L_1$-importance channel selection, ranks them by output fidelity as a zero-shot label-free proxy, and applies GhostConv structural mutation to Pareto-optimal candidates; and (3) the \textbf{first source-code-free compression pipeline for 3D point cloud models}, operating entirely via ONNX graph surgery with no original architecture definition required. On ModelNet40, H3DNAS reduces the number of parameters in PointNet, PointNet++, and PointMLP by $65.5\%$, $43.2\%$, and $49.1\%$, respectively, while achieving $1.99\times$, $1.29\times$, and $1.67\times$ inference speedups with negligible loss in accuracy. The source code is publicly available\footnote{https://github.com/ClarityLab-Org/h3dnas}.
Leech-lattice vector quantization holds the strongest reported 2-bit quality under its own evaluation protocol. Its kernel decodes one shell; we found no implementation of the multi-shell decoder the rate requires. This paper supplies one and measures its serving cost for decode-phase GEMV at batch 1. First, a serving path for the full 301-class codebook: an offline expansion into GPU layouts and a fused dequantize-plus-matvec kernel reading them without warp divergence, verified against f64. Second, the in-VRAM rate is a design axis distinct from the on-disk rate. Four bit-exact layouts timed in one process show binary bit planes beating one-hot masks on size and speed at constant bandwidth (4.80 bits per weight, 2.15x FP16). Below 4.3 bits a second, irregular stream enters; at 3.6 the decode stops being shifts and masks. Third, deployed four-bit (AWQ) and two-bit (QTIP) GEMV kernels run in the same process. The trellis kernel reads 2.40x fewer bytes than our served layout and runs 2.27x faster at near-equal fractions of their byte bounds: the time gap tracks the traffic gap, the price of unfolding a codebook too large for a lookup table. Fourth, the validity envelope: the trellis kernel outruns our no-weights control, so our launch geometry sets that floor, and on a second memory hierarchy every lattice arm falls below FP16. With the output head held identical across arms, the kernel-and-format path gains 1.11x, 1.29x and 1.41x end to end at 4B, 8B and 14B; with an int8 output head the served 4B reaches 87.0 tok/s in 2.60 GB. The quality cost, 1.38x perplexity and 14.7 MMLU points at 4B, shrinks across the three sizes measured.
Irina Proskurina, Guillaume Metzler, Antoine Gourru +1cs.CL
Model compression techniques such as pruning and quantization facilitate the efficient deployment and acceleration of Large Language Models (LLMs). However, recent studies show that weight sparsification methods, such as SparseGPT, can amplify existing biases in models, with outputs varying significantly depending on persona cues in the prompt. In this paper, we introduce Debias-SparseGPT, a post-training pruning method incorporating representational debiasing using a second-order term defined over demographically contrasting inputs. We perform empirical validation of our method over a wide range of generative LLMs. Across models and sparsity regimes (25%, 50%, and structured 2:4 sparsity), Debias-SparseGPT consistently reduces pruning-induced bias compared to SparseGPT while preserving model perplexity and zero-shot accuracy. Under the most restrictive 2:4 structured sparsity pattern, which most aggressively degrades model quality, augmenting the calibration set with long-context, content-rich examples further improves both downstream performance and fairness. Overall, Debias-SparseGPT advances the bias-performance trade-off while preserving the computational efficiency of sparse models.
In this paper, we propose a scalable Kronecker-based approximation that captures cross-layer interactions without storing the entire Fisher matrix, enabling practical Hessian analysis for billion-parameter networks where full computation is infeasible. Our approach reveals consistent vulnerability patterns: value projection layers exhibit the highest sensitivity and strongest cross-layer correlations across multiple model families, while other components exhibit architecture-specific behaviors. Through extensive experiments on quantization, sparsification, inter-layer corruption, and post-corruption fine-tuning, we demonstrate that our approximation strongly correlates with both performance degradation and recovery. Our framework provides a practical, theoretically grounded tool for identifying fragile components in large models, opening new avenues for guided compression and optimization strategies, such as mixed-precision allocation, layer-wise sparsity, and adaptive low-rank decomposition across layers and even individual weight groups.
Removing complete transformer layers preserves a standard serving architecture, but existing depth-compression methods can lose substantial quality, and the loss varies unpredictably across models. We introduce XMerge, a post-training method with two components. Cross-axis selection identifies a block with low relative-magnitude and angular hidden-state change, and local boundary reconstruction re-fits the adjacent surviving block to match the original two-block output. XMerge uses no task labels or end-to-end fine-tuning, and it introduces neither architectural changes nor additional inference-time parameters. Across seven Llama and Qwen backbones (0.5B-8B), five published baselines, and three layer-reduction levels, its advantage over baselines is largest at the most aggressive removal: at k=4 it ranks first on six of seven backbones on CORE (a 22-task aggregate) and, separately, on six of seven on MMLU (five of seven on both at once), while avoiding the large perplexity increases of several competing operators. In a task-level bootstrap, the 95% confidence intervals for the three largest CORE margins exclude zero; the remaining margins are consistent with ties. Across the 14 (model, regime) cells it is also the only evaluated operator that never collapses, ranking top-2 in both zero-shot and in-context regimes; on a first calibration probe (one backbone) it is the best-calibrated operator. Ablations show that local reconstruction provides most of the gain, while cross-axis fusion helps when the two selection axes disagree. The additional construction cost is recovered through per-token decode savings after roughly tens of thousands of requests.
A compressed student has two shapes that need not agree: the weight it deploys at inference and the weight family its training can reach. We show that a state-of-the-art weight-inheritance distiller, Low-Rank Clone (LRC), deploys a full-width student MLP but ties training to a teacher-induced slice, leaving 62.5-81.4% of each deployed matrix's independent linear degrees of freedom unreachable-paid for at inference, never trainable. Our principle is one line: train what you deploy. From the identical LRC warm start, we make the training object the entire deployed matrix, with no change in deployed shape, deployed parameter count, or inference FLOPs, via two mergeable realizations (Dense-LRC and CORE-LRC) that both collapse to one deployed weight. This recovers stranded capacity: taking the stronger realization per teacher, +2.36/+2.71/+10.45 Avg9 over matched-budget plain-LRC baselines across three teachers (Llama3.2-3B, Llama3.1-8B, Qwen2.5-3B), with the largest gain on the widest teacher (Qwen), where it reaches the original recipe's approx. 20B-token accuracy at 10B tokens (2x token efficiency); there the strictly same-lineage arm still recovers +6.39, the fully controlled figure. Controls strongly support attributing the gain to the enlarged reachable set, rather than to added parameters or the recipe. From approx. 10B distillation tokens plus a short SFT, a half-parameter 1.5B student matches its approx. 9T-token teacher's 9-task macro-average, within evaluation noise and with a residual MMLU deficit, and a 2.7B student beats Meta's own official compression of Llama3.1-8B at ~900x fewer compression tokens (a token count under unmatched recipes, not a compute claim). All results are from single-seed runs on the LRC backbone.
Anirudh Malik, M Sparsh Mehra, Poojith Devancs.AI cs.LG
Ultra-low-bit language models can reduce storage and memory bandwidth, but a nominal "1.58-bit" label does not fully describe the stored representation, retained capability, or runtime behavior. We study an end-to-end post-training conversion of Qwen, an instruction-tuned 4B-parameter model, using KOTMS rotation, E2M-ATQ ternarization, and GPTQ-style error compensation from TWLA. The experiment is weight-only: activations remain at 16-bit precision, so ILA-AMP is omitted. We evaluate effective bit accounting, task capability retention, perplexity, calibration sensitivity, checkpoint composition, and deployment behavior. The final conversion uses 1.641 effective bits per weight for quantized linear weights, with 81.62% of model parameters targeted. Across ten scored capability comparisons, accuracy falls from 64.5% to 54.7%. Degradation is uneven: BoolQ retains 84.6% chance-corrected teacher performance, while ARC-Challenge retains 43.8%. Perplexity rises from 13.639 to 18.748 on WikiText-2, 24.700 to 31.992 on PTB, and 19.831 to 28.966 on C4. A subsequent packing run preserves the ternary planes and scales, reducing reported model size from 8.29 GiB to 3.96 GiB with essentially unchanged perplexity. A separate third-party packing attempt was lossy and is excluded from the primary artifact claim. The packed artifact has not been benchmarked end-to-end for task accuracy or generation throughput. A preliminary Triton GEMV microbenchmark is 4.6x slower than FP16 cuBLAS on one tested shape. We therefore do not claim that compression alone yields faster inference.
Osvaldo M Velarde, Lucas C Parra, Alireza Hashemi +1cs.LG
Artificial neural networks are often regarded as powerful yet opaque black boxes. Here, we demonstrate that learning in deep neural networks generates local symmetries known in graph theory as fibrations and coverings. We prove that covering symmetries are stable attractors of stochastic gradient descent. Consistent with this theory, we report the emergence of covering symmetries across major network architectures, including multilayer, convolutional, recurrent, and transformer networks. Exploiting these symmetries enables drastic model compression - reducing networks to 17% of their original size without sacrificing performance. Furthermore, controlled breaking of covering symmetry overcomes the loss of plasticity, achieving state-of-the-art performance in continual learning. The theoretical results provide a new foundation for AI systems based on symmetries that convert black boxes into interpretable colored graphs and enable more efficient inference and lifelong learning.
Ahmad Alfan Alfian Irfan, Nur Ahmad Khatim, Mansur Ariefcs.AI cs.CV
Many automobile and mobility companies deploy learned driving policies on embedded computers with limited memory and power. Pruning, knowledge distillation, and quantization are the standard methods to reduce the size and the inference cost of these policies. However, these methods are commonly assessed by aggregate numerical scores, and such scores may not reflect the ability of the policy to drive safely when interacting with other road users. In this study, we propose a stage-wise closed-loop evaluation approach to follow a driving policy through a compression pipeline. We formulate the driving task as a partially observable Markov decision process (POMDP) and train a belief-state policy with proximal policy optimization (PPO) in Gym-Duckietown. We then extract the actor, compress it one stage at a time, and evaluate it on five driving curricula. We show that structured pruning is the stage at which the driving capability is first lost. Meanwhile, distillation improves the pruned actor, but the improvement is limited by its rehearsal data. Integer quantization of the improved actor loses some of the curricula that require the vehicle to stop and then resume. Interestingly, the same procedure on the unpruned actor preserves all five curricula. Our study thus provides an empirical analysis aiming to answer the currently active discussions on how to accept a compressed driving policy, so as to achieve a safe and statistically reliable deployment of automated driving functions.
Edge-AI model selection is commonly driven by one isolated metric - accuracy, latency, memory, energy, or safety, even though a deployable language model must balance all five. Our work focuses on answering the question whether na- tively trained small language models (SLMs) or large language models (LLMs) compressed through post-training quantization offer the more sustainable edge- deployment trade-off. We introduce a reproducible Holistic Sustainability Score (HSS) organized around the triple bottom line: an economic pillar for capability and systems efficiency, an environmental pillar for operational GPU energy and a social pillar for harmful-prompt robustness. Five BF16 SLMs and five LLMs under different quantization approaches - BF16, INT8, NF4 4-bit, GPTQ 4-bit, and GGUF Q4 produce 30 measured configurations. Capability is assessed on five zero-shot benchmarks; efficiency uses latency, throughput, peak VRAM and energy; and safety is approximated by attack success rate on five harmful prompts. Qwen3-30B-A3B/GGUF Q4 ranks first in the combined pool (93.38), followed by Mistral-Small-24B/GGUF Q4 (92.40), while Phi-4-mini/BF16 is the highest- ranked SLM in that pool (89.49). Thus, the hypothesis that native SLMs must be the most sustainable edge choice is not supported universally; optimized quantized LLMs can win overall, while SLMs remain competitive through lower resource demand. Quantization is a systems-level choice rather than a monotonic precision- efficiency trade-off and HSS remains relative to its comparison pool and proxy definitions.
Mixture-of-experts (MoE) architectures scale large language models efficiently, but they demand massive GPU memory. To cope with such demand, models are commonly compressed to reduce their memory footprint. Residual sparsification is a representative compression technique that decomposes each projection matrix of an expert into a shared base matrix and per-expert residual matrix, and then compresses the residuals. Existing sparsification methods compress each residual matrix independently by minimizing its compression error, thereby minimizing the error of each projection matrix. However, our analysis shows that this objective is misaligned with preserving model accuracy after compression. In an expert, the final output is produced through computations coupled across multiple projections and hidden representations. Therefore, even small errors in individual matrices can propagate through hidden representations and projection interactions, leading to large expert output errors and accuracy degradation. To address this misalignment, we propose PARSER, a new residual sparsification method that shifts the compression objective from minimizing isolated matrix errors to preserving the expert output error. PARSER achieves this by introducing output importance, which measures the actual contribution to the expert output error. Our experiments show that, compared with existing methods, PARSER narrows the accuracy gap to the uncompressed model by 1.41$\times$ on Qwen and 1.44$\times$ on DeepSeek, while achieving the same peak memory reduction.
Maria Matveev, Pascal Esser, Ayush Bharadwaj +2cs.LG
A central question in modern machine learning is how much a trained model can be compressed without changing its behavior, to reduce the memory, compute and energy required to deploy it. To study this, we quantify functional degeneracy through the behavioral recovery rank, defined as the number of leading behavioral-Hessian eigendirections required to recover a trained model's performance. Using the behavioral recovery rank as a geometric benchmark for compression, we find that structural and magnitude pruning retain more degrees of freedom, even after the task is saturated. This gap suggests that functional redundancy is distributed across parameter directions and is not exposed by individual weights or neurons.
Mixed-precision quantization (MPQ) assigns a different bitwidth to each linear layer of a large language model (LLM) to minimize the quantization-induced quality loss under a fixed budget, but Mixture-of-Experts (MoE) models contain these layers in every expert of every MoE block, so the allocation space grows far larger than in a dense model. Existing methods either allocate within each block under a uniform per-block budget, or allocate across blocks through an additive proxy, and neither directly optimizes a model-level objective over the choices that couple the blocks. We propose Q-Strata, a bi-level allocator that ranks within-block assignments with a cheap proxy and allocates across blocks with a model-level objective evaluated on the assembled quantized model. Its inner stage caches a Pareto frontier of candidates per block over finely spaced budgets, leaving the outer stage to set one budget per block instead of a bitwidth for every linear layer. With the search reduced to one budget per block, the outer stage optimizes this model-level objective directly, capturing the inter-block coupling that additive proxies miss. On Mixtral-8x7B-Instruct, Qwen1.5-MoE-A2.7B, and DeepSeek-V2-Lite, Q-Strata consistently achieves lower WikiText2 perplexity than uniform-bitwidth GPTQ and the state-of-the-art MoE MPQ methods MxMoE and GEMQ in the low-bit regime. The code is available at https://github.com/snu-mllab/Q-Strata/tree/main.
Inference with transformer-based large language models (LLMs) is often limited by the memory-bound KV cache and quadratic attention cost. State-space models (SSMs) mitigate this through linear attention and fixed-size recurrent states, but their large dense linear projections remain computationally expensive even after quantization. We introduce a method that induces sparse neural activity in heavily quantized linear-attention models with minimal performance loss. Activations below a per-projection trainable threshold ($\pm Δ$) are nullified while preserving crucial outliers, achieving comparable performance to dense models with up to 4$\times$ fewer effective arithmetic operations. Targeting a multi-core, multi-chip neuromorphic platform, where event-driven execution converts unstructured sparsity into throughput at both the compute and communication levels, a capability GPU architectures fundamentally lack, we project up to 37$\times$ higher throughput and 16$\times$ lower power versus edge GPU inference of a comparable transformer-based model, and up to 5.4$\times$ improvements over the non-sparsified baseline. These results position sparse, quantized linear-attention models as a natural fit for deploying LLMs on event-driven multi-core platforms.
Modular compression has enabled considerable parameter reduction in LLMs while preserving strong language understanding and downstream task accuracy. However, existing joint modular compression methods primarily rely on activation statistics, leaving loss-sensitivity information and its module-level characterization underexplored. We investigate addressing this gap with LaMoC, a loss-aware modular compression methodology that blends activation and Empirical Fisher statistics through gradient-error alignment. LaMoC improves joint compression by selecting compression statistics that better align local module reconstruction error with the downstream loss. Our contributions are three-fold: (1) We characterize the Empirical Fisher as a module-level loss-aware proxy that can be blended with the activation statistics required for compression. (2) We reformulate joint modular compression as a two-tiered optimization problem that minimizes module reconstruction error while tuning the activation and gradient information blending rate. (3) We implement an empirically driven methodology with statistical validation to solve the resulting compression problem. We evaluate LaMoC across four model families spanning eight models. On the 4-8B models, LaMoC achieves an average 2.5% reduction in perplexity and a 1% relative improvement in task accuracy over state-of-the-art modular compression methods.
Single-GPU deployment of 70B-parameter language models on an NVIDIA GPU is constrained by device memory, long-context throughput, and engineering integration cost. We cast single-GPU inference as a budget-aware design problem over these three axes and study how pruning, quantization, and KV-cache compression interact under realistic execution. Controlled ablations show that layer-wise pruning makes weight quantization more robust. KV-cache sparsification complements INT8 KV quantization by reducing memory without hurting decoding speed, while static vector quantizers often conflict with dynamic caching. Guided by these coupling results and explicit budget tracking, we assembled a practical pipeline and compressed a 70B model to about 33 GB, sustained about 57 tokens/s on 10k token prompts on a single A40, and kept absolute accuracy within 5% on common and reasoning benchmarks. We contribute design rules and a reproducible evaluation protocol that jointly report quality, memory, and end-to-end speed, and we provide a foundation for automated pipeline search under realistic single-GPU constraints.
Structured pruning uses surrogate objectives because direct task evaluation over every feasible mask is too expensive. Most evaluations report average surrogate error or rank correlation on broadly sampled masks. These summaries do not directly test the mask chosen by the surrogate. We introduce PruneShift, an evaluation framework that separates broad predictive fidelity, fidelity near selector outputs, and the quality of the selected pruning decision. We first prove that Spearman and Kendall agreement can approach one while normalized selection regret remains maximal. We then derive sufficient conditions based on uniform error, selector suboptimality, decision margin, density ratio, and comparison mass. The analysis also yields a finite pool certificate with an explicit excess cost bound. Four studies test different links in this argument. External TextbookQA confirmation is heterogeneous: 7 of 20 simultaneous intervals favor the surrogate-selected mask, 6 favor its fixed comparator, and 7 cross zero. On a fixed Natural Questions pool, strict improvement holds in one of four settings. A controlled QQP experiment supports the proposed coverage mechanism in all 16 prespecified endpoints, although the sufficient bounds are conservative. Finally, a restricted OSSCAR reconstruction study on OPT-125M shows better local than broad fidelity in 68 of 75 primary endpoints. Independent fixed-mask confirmation is inconclusive in 24 of 25 endpoints and favors the comparator in one. These results show why predictive fit, decision reliability, and pruning method quality require separate evidence.
The rapid development of LLMs incurs prohibitive memory footprints and intensive computational demands. Quantization-Aware Training (QAT) techniques have emerged as a promising solution to address these challenges by explicitly simulating quantization effects during model training, yielding low-bit models that achieve accuracy comparable to their full-precision counterparts. In this work, we provide a target-centric survey of QAT, aimed at clarifying both its theoretical foundations and its evolving implementation landscape. We systematically review existing QAT methods through a target-centric taxonomy and synthesize cross-target differences in error characteristics, numerical formats, and strategy transferability. We further summarize QAT evaluation paradigms and discuss challenges in optimization and deployment, outlining potential directions for future research.
The LLM-as-a-Judge paradigm has emerged as a scalable alternative to human evaluation. However, single-model judges are limited by their inherent model biases, while multi-agent evaluation protocols that mitigate this through diverse deliberation are prohibitively expensive at inference time. To this end, we propose \textbf{\modelname}, which equips a compact \underline{Judge} model with multi-agent \underline{Panel} deliberation capability. Specifically, we first train on panel deliberation traces from an ensemble of strong evaluators, capturing structured patterns of discussion, disagreement, and resolution. To further improve judgment quality beyond SFT, we introduce \textit{AdaReward}, an adaptive multi-reward RL algorithm that dynamically rebalances reward component weights as different objectives saturate at different rates during RL training. For practical deployment, we further design a lightweight domain specialization module for rapid adaptation to new evaluation domains with few hundred labeled samples. As a result, (i) \textit{Novel}: the first framework to equip a single compact judge with multi-agent panel deliberation capability at single-model inference cost; (ii) \textit{Effective \& Reliable}: JudgePanel with a 14B backbone outperforms judge-specialized models up to 70B across four evaluation benchmarks, demonstrates strong position consistency, and rapidly specializes to new domains with few hundred samples.
Small language models (sLLMs) are nowadays hosted on devices with limited memory and computational budget. In an autoregressive setup, inference is memory-bandwidth bound: uniform quantization is often detrimental to such models, since their architecture has limited redundancies and only a few layers are not very sensitive to lower precision. We propose a composite metric that combines two orthogonal criteria: information retention (measured in terms of a normalized SQNR-based coefficient) and throughput gains (modeled using a roofline-based latency analysis). By profiling Gemma 3 1B, we find that Feed-Forward Network blocks and the embedding matrix are the most promising targets for acceleration. For each candidate, we estimate a normalized quality score based on simulated quantization and a normalized speed score based on roofline modeling with no actual execution needed. We combine the two scores in a composite priority coefficient, allowing us to tune the trade-off between speed and quality as needed. Our metric is general and can be used to prioritize individual blocks, their projection sublayers, or transformer layers as a whole. We evaluate our approach on several model architectures, showing that our estimates have at around 4% prediction error for the accelerated speedup. We find that our method generally allocates more resources to the most expressive layers compared to evolutionary search, specialized accelerators, or Shapley-value-based approaches that require expensive approximate inference. Our analytical approach makes sLLM quantization a predictable engineering task.
Rongfeng Wang, Shichao Weng, Zhiqiang Wang +4cs.CL
Mixture-of-Experts (MoE) models route each token to a subset of expert networks, increasing capacity while keeping per-token computation sparse. In many deployed MoEs, the number of active experts is fixed across layers and tasks, although layer roles and expert redundancy vary with depth and demand varies with difficulty. Existing approaches address only part of this setting: layer-wise allocations are usually determined offline and reused for all tasks, while token-level methods vary expert activation using local routing signals without task-level context. We propose MetaNet, a support-set controller that predicts, for each layer, an expert-retention threshold and a bounded routing bias. The backbone, experts, and router remain frozen. On DeepSeek-MoE-16B-Chat, MetaNet provides a tunable accuracy-expert-activation trade-off. Relative to fixed k=6, a conservative setting activates 3.61 experts on average (40% fewer) and achieves comparable MMLU accuracy (0.489 vs. 0.474), whereas an aggressive setting activates 2.28 experts on average (62% fewer) with accuracy approximately 3.7 percentage points lower. The MMLU-trained controller also transfers to C-Eval without retraining, activating 2.90 experts on average (52% fewer than fixed k=6) at 0.386 accuracy.
Extreme compression of deep neural networks, up to full binarization, dramatically reduces memory footprint and arithmetic complexity, facilitating deployment on constrained edge hardware with field-programmable gate arrays (FPGAs) and microcontrollers. Although combining binarization with pruning promises additional efficiency gains, existing pruning strategies are ill-suited to binarized representations and rarely translate into meaningful hardware savings. We introduce a PyTorch-based, research-oriented framework that incorporates freezing and pruning mechanisms for designing and optimizing binarized neural networks. The framework enables rapid and reproducible evaluation of state-of-the-art approaches and the fast prototyping of new ones. Leveraging this framework, we propose a novel pruning method that accounts for the relative importance of learned parameters across abstraction levels. Such a global weighting mechanism consistently achieves a superior trade-off between model accuracy and pruning rate, achieving a 70% pruning rate on VGG11 with constant accuracy, while state-of-the-art results reach only 41% in the binarized setting.
Large-kernel Convolutional Neural Networks (CNNs) deliver remarkable performance in vision tasks by significantly expanding receptive fields, yet their quadratic parameter growth critically impedes storage-efficient edge deployment. While existing efficient architectures adopt parameter-efficient depthwise separable convolution backbones that leverage techniques like low-rank approximation and weight sharing to compress depthwise convolutions, we identify a critical oversight: pointwise convolutions dominate parameter volume (>87% in models like RepLKNet-31B) and constitute the primary deployment bottleneck on resource-constrained edge devices. This results in prohibitive storage costs and severe memory-loading constraints on resource-limited devices (e.g., smartphones with 4-12 GB Random Access Memory (RAM)). To overcome this, we propose Channel Group-Shared (CGS) low-rank approximation, a novel Singular Value Decomposition (SVD)-based parameter-sharing strategy. CGS constructs a structured low-rank paradigm isomorphic to SVD decomposition, comprising shared (high-parameter-cost) down/up-projection matrices across channel groups within a layer and channel-group-specific (low-parameter-cost) scalable diagonal matrices. This group-sharing design achieves significant parameter reduction. Extensive experiments demonstrate that large-kernel CNNs (RepLKNet, ConvNeXt, SLaK) enhanced with CGS strike an empirically favorable balance between competitive performance and substantially reduced storage costs. Crucially, by alleviating storage constraints, reducing memory bandwidth pressure during loading, and minimizing model loading latency, CGS enables the feasible deployment of pre-trained large-kernel CNN models on edge devices, thereby bridging the gap between high-performance vision models and practical edge deployment.
Depth pruning removes entire Transformer blocks to reduce the inference cost of large language models, but disrupts the hidden-state distributions expected by downstream layers, leading to significant accuracy loss. We introduce SHIFT-LLM, a training-free post-pruning correction framework that inserts a Linear Residual Adapter (LRA) at each pruning site. Each LRA preserves the identity pathway of the original residual block and adds a lightweight affine residual correction. This correction is calibrated via closed-form least-squares regression on a small held-out set, without gradient computation, to approximate the missing residual update produced by the pruned block. Together with the preserved identity pathway, the resulting LRA output approximates the hidden state produced by the original block, thereby mitigating the distributional mismatch introduced by layer removal while avoiding the expensive attention and feed-forward computations of the removed blocks. The resulting LRAs support low-rank factorization and exact merging across consecutive pruned layers for additional compression, and combine naturally with parameter-efficient fine-tuning for further recovery beyond fine-tuning the pruned model alone. Experiments on five model families, six layer-selection criteria, and seven zero-shot benchmarks show that SHIFT-LLM consistently recovers accuracy lost to depth pruning across most configurations, achieving gains up to +15.7 points on Llama-3.1-8B-Instruct while requiring only a few hundred calibration samples and no gradient computation.
Alexandru-Dragos Manolache, Yunqiang Li, Jan van Gemertcs.CV cs.LG
Ternary transformers offer extreme memory and compute efficiency, but existing low-bit LoRA-based methods cannot directly fine-tune ternary weights. Current approaches either require dequantization, restoring low-bit base weights to higher precision to merge with adaptation weight, or update only quantization parameters, preventing a merged model that remains ternary. We propose ternary multiplicative adaptation, which represents discrete updates of ternary weights such as sign flips or zeroing through a low-rank Kronecker factorization into two small ternary matrices applied element-wise to ternary weights. This design is parameter-efficient and expressive, preserves the ternary domain, and supports direct merging without dequantization. Experiments on six models across language and vision, including ternarized LLaMA-3 1B and 3B and a ternary ViT-B/16, demonstrate that our method recovers much of the performance lost to quantization and outperforms strong low-bit and ternary baselines. Code is available at https://github.com/alexmanoo/ternary_adaptation.
Quantization correction methods usually optimize weights, quantization parameters, or reconstruction objectives, while the underlying parameter subspaces responsible for effective correction remain unclear. In this work, we study quantization correction from a parameter subspace perspective and reveal that correction capability is highly non-uniform across parameter groups. By decomposing trainable parameters into backbone weights, normalization-affine parameters, and quantization parameters, we show that the low-dimensional normalization-affine subspace provides a highly efficient correction direction under matched budgets. Based on this finding, we propose SandwichQuant, a two-stage normalization-affine correction framework that performs adaptation before and after quantization. The pre-stage improves quantization robustness, while the post-stage compensates residual errors after the quantized graph is fixed. Extensive experiments on vision models and large language models demonstrate consistent improvements under various low-bit quantization settings, validating the effectiveness of subspace-aligned correction.