LLM Compression

Compressing large language models (LLMs) for deployment on commodity GPUs remains challenging: conventional scalar quantization is limited to fixed bit-widths (e.g., 8/4/3-bit), offers only a few discrete compression points, and typically requires calibration data. We present FASQ (Flexible Accelerated Subspace Quantization), a calibration-free framework that applies product quantization to LLM weight matrices. By tuning two parameters, sub-vector size and codebook cardinality, FASQ exposes a continuous design space spanning 27-49% of the original FP16 model size, filling compression gaps that fixed-bit schemes cannot reach. On Meta-Llama-3-8B, FASQ surpasses 4-bit GPTQ and AWQ in accuracy (67.1-67.7 avg.) at 37-42% model size, with consistent results on Qwen3-8B and Qwen3.5-9B-Base. To make product quantization practical at inference time, we design custom CUDA kernels: a LUT-free direct-compute GEMV for decode and an output-stationary double-buffered LUT GEMM for prefill, both with split-K parallelism. On an RTX~3090, FASQ achieves 45.2 tok/s decode at effective 4-bit (2.56x memory reduction) and 51.8 tok/s at effective 3-bit (2.80x), both surpassing FP16 tensor-core performance (43.9 tok/s) and delivering 1.6 to 1.8x the throughput of AWQ, 2.5 to 2.5x of GPTQ, and 4.3 to 5x of RTN. FASQ is the only compressed method that accelerates decode beyond FP16, offering calibration-free compression, continuous size-quality trade-offs, and real-time inference on a single consumer GPU.

We present a systematic empirical study of transformer compression through over 40 experiments on GPT-2 (124M parameters) and Mistral 7B (7.24B parameters). Our analysis covers spectral compression, block-level function replacement, rotation-based quantization, activation geometry, and adaptive early exit. We identify five structural properties relevant to compression. (1) Variance is not importance: high-variance activation directions are approximately 96 percent uncorrelated with predictive directions (measured via CCA), and projecting onto these subspaces preserves over 90 percent of variance while degrading perplexity. (2) Block linearity is conditional: transformer blocks are approximately linear (R^2 ~ 0.95 on GPT-2, 0.93 on Mistral block 31) only under the correct upstream distribution; modifying earlier blocks induces distribution shift that degrades downstream approximations. (3) The reconstruction wall: approaches that factor weights into quantized components amplify errors through cross-terms, making direct quantization strictly superior. (4) Linearity increases with depth: Mistral 7B exhibits a progression from R^2 = 0.17 (block 0) to R^2 = 0.93 (block 31), indicating a division between nonlinear feature construction and linear refinement. (5) Approximately 30 percent of tokens are computationally easy, confirmed via exit heads and KL divergence sensitivity. We demonstrate that single-block linear replacement achieves 34x compression with a 1.71 perplexity increase on the final block of Mistral 7B, while multi-block replacement fails due to residual error accumulation and distribution shift. These findings suggest fundamental limits to static post-training compression and motivate adaptive, per-token computation as a more effective direction.

Large language models (LLMs) are expensive to serve because model parameters, attention computation, and KV caches impose substantial memory and latency costs. We present GRASPrune, a structured pruning framework applied after pretraining that jointly prunes FFN channels and KV head groups under a single global budget. Instead of learning importance scores without constraints and applying the budget only after training, GRASPrune learns lightweight gate scores with a projected straight-through estimator that enforces a hard mask satisfying the budget at every step while keeping the backbone weights frozen. After the mask is fixed, we calibrate scaling factors on the retained units to mitigate scale mismatch caused by pruning, and fold these factors into the pruned weights to obtain a smaller dense checkpoint with no extra parameters at inference. On LLaMA-2-7B, GRASPrune removes 50% of parameters and achieves 12.18 perplexity on WikiText-2 while maintaining competitive average zero-shot accuracy on five benchmarks, using four epochs on 512 unlabeled calibration sequences on a single NVIDIA A100 80GB GPU without any full model fine-tuning.

Recently, large language models (LLMs) have advanced recommendation systems (RSs), and recent works have begun to explore how to integrate LLMs into industrial RSs. While most approaches deploy LLMs offline to generate and pre-cache augmented representations for RSs, high-dimensional representations from LLMs introduce substantial storage and computational costs. Thus, it is crucial to compress LLM representations effectively. However, we identify a counterintuitive phenomenon during representation compression: Mid-layer Representation Advantage (MRA), where representations from middle layers of LLMs outperform those from final layers in recommendation tasks. This degraded final layer renders existing compression methods, which typically compress on the final layer, suboptimal. We interpret this based on modularity theory that LLMs develop spontaneous internal functional modularity and force the final layer to specialize in the proxy training task. Thus, we propose \underline{M}odul\underline{a}r \underline{R}epresentation \underline{C}ompression (MARC) to explicitly control the modularity of LLMs. First, Modular Adjustment explicitly introduces compression and task adaptation modules, enabling the LLM to operate strictly as a representation-learning module. Next, to ground each module to its specific task, Modular Task Decoupling uses information constraints and different network structures to decouple tasks. Extensive experiments validate that MARC addresses MRA and produces efficient representations. Notably, MARC achieved a 2.82% eCPM lift in an online A/B test within a large-scale commercial search advertising scenario.

Matrix-level low-rank compression is a promising way to reduce the cost of large language models, but running compression and evaluating the resulting models on language tasks can be prohibitively expensive. Can compression-induced degradation be predicted before committing to this compute? We systematically analyze the Qwen3 and Gemma3 model families across four representative low-rank compression methods: vanilla SVD, two ASVD variants, and SVD-LLM. We find that stable rank and information density, measured in bits per parameter, dominate performance degradation. The interaction term $γ\cdot \barρ_s$, defined as compression ratio times stable rank, is a robust predictor of accuracy degradation, achieving leave-one-out cross-validation Pearson correlations of $0.890$ for attention layers and $0.839$ for MLP layers. We provide theoretical intuition for why this predictor succeeds by connecting it to standard SVD truncation bounds and error composition mechanisms in transformer layers. These findings enable a predict-then-compress workflow: compute $γ\cdot \barρ_s$ from weights, estimate degradation, and invest compute only in desirable configurations.

Chain-of-thought reasoning has driven striking advances in language model capability, yet every reasoning step grows the KV cache, creating a bottleneck to scaling this paradigm further. Current approaches manage these constraints on the model's behalf using hand-designed criteria. A more scalable approach would let end-to-end learning subsume this design choice entirely, following a broader pattern in deep learning. After all, if a model can learn to reason, why can't it learn to forget? We introduce Neural Garbage Collection (NGC), in which a language model learns to forget while learning to reason, trained end-to-end from outcome-based task reward alone. As the model reasons, it periodically pauses, decides which KV cache entries to evict, and continues to reason conditioned on the remaining cache. By treating tokens in a chain-of-thought and cache-eviction decisions as discrete actions sampled from the language model, we can use reinforcement learning to jointly optimize how the model reasons and how it manages its own memory: what the model evicts shapes what it remembers, what it remembers shapes its reasoning, and the correctness of that reasoning determines its reward. Crucially, the model learns this behavior entirely from a single learning signal - the outcome-based task reward - without supervised fine-tuning or proxy objectives. On Countdown, AMC, and AIME tasks, NGC maintains strong accuracy relative to the full-cache upper bound at 2-3x peak KV cache size compression and substantially outperforms eviction baselines. Our results are a first step towards a broader vision where end-to-end optimization drives both capability and efficiency in language models.

Long Chain-of-Thought (CoT) reasoning is pivotal for the success of recent reasoning models but suffers from high computational overhead and latency. While prior works attempt to compress CoT via external compressor, they often fail to align with the model's internal reasoning dynamics, resulting in the loss of critical logical steps. This paper presents \textbf{C}ompressing \textbf{R}edundancy in Chain-of-Thought via \textbf{I}ntrinsic \textbf{S}aliency \textbf{P}runing (\textbf{CRISP}), a framework that compresses CoT by exploiting the model's intrinsic saliency. Our analysis reveals a distinct phenomenon: the reasoning termination token \texttt{[object Object]} acts as an information anchor, where its attention pattern effectively demarcates essential reasoning from redundancy. Based on this finding, we design a policy that utilizes these intrinsic attention signals to guide atomic compression operations. In contrast to coarse-grained pruning strategies, CRISP strategically distills the reasoning chain to maximize information density while preserving logical coherence. Empirical results across various backbone models and mathematical datasets demonstrate that CRISP achieves a 50-60% reduction in token count without compromising accuracy, effectively mitigating the efficiency bottleneck of long-context reasoning. We open-source our implementation to facilitate further research in efficient reasoning.

Supervised Fine-Tuning (SFT) accelerates taskspecific large language models (LLMs) development, but the resulting proliferation of finetuned models incurs substantial memory overhead. Delta compression addresses this by retaining a single pre-trained LLM with multiple compressed delta weights. However, existing methods fail on models fine-tuned with largescale datasets. We find that larger SFT data scale amplifies delta parameter magnitude, singular values, and entropy, exacerbating compression errors. To tackle this, we propose DQRELO (Delta Compression via Quantization and Residual Low-Rank), a novel training- and data-free delta compression method. It combines coarse-grained one-bit quantization to capture the dominant structure of the delta, followed by compensated residual low-rank approximation to recover fine-grained details from the smaller residual error. Experiments on various LLMs spanning dense and MoE architectures across multiple domains under this challenging setting demonstrate that DQRELO outperforms existing methods. Moreover, we establish key design principles for delta compression through extensive empirical analysis, demonstrating how task difficulty, architecture, and layer positioning create predictable patterns that can guide optimal compression strategies in production systems.

This paper presents a systematic benchmark of state-of-the-art multilingual large language models (LLMs) adapted via token pruning - a compression technique that eliminates tokens and embedding parameters corresponding to languages irrelevant to the target application. Focusing on Korean-centric natural language processing (NLP) tasks, we evaluate architectures including Qwen3, Gemma-3, Llama-3, and Aya across three vocabulary configurations: Original, English-Korean (EnKo), and English-Korean-Chinese (EnKoZh). Performance is assessed using established benchmarks for general aptitude, cultural literacy, instruction following, and machine translation. Our findings indicate that token pruning significantly improves generation stability by eliminating language confusion, and in the case of machine translation, frequently enhances performance on Korean-specific tasks. While instruction-following capabilities display architecture-dependent variance linked to latent cross-lingual representations, the significant reduction in vocabulary size validates token pruning as a highly effective optimization strategy for memory-constrained, domain-specific deployments, despite modest gains in inference latency.

Large Language Models (LLMs) have achieved remarkable success, underpinning diverse AI applications. However, they often suffer from performance degradation due to factors such as catastrophic forgetting during Supervised Fine-Tuning (SFT), quantization, and pruning. In this work, we introduce a performance recovery framework based on Self-Distillation Fine-Tuning (SDFT) that effectively restores model capabilities. Complementing this practical contribution, we provide a rigorous theoretical explanation for the underlying recovery mechanism. We posit that an LLM's generative capability fundamentally relies on the high-dimensional manifold constructed by its hidden layers. To investigate this, we employ Centered Kernel Alignment (CKA) to quantify the alignment between student and teacher activation trajectories, leveraging its invariance to orthogonal transformations and scaling. Our experiments demonstrate a strong correlation between performance recovery and manifold alignment, substantiating the claim that self-distillation effectively aligns the student's high-dimensional manifold with the optimal structure represented by the teacher. This study bridges the gap between practical recovery frameworks and geometric representation theory, offering new insights into the internal mechanisms of self-distillation.

Large Language Models (LLMs) incur significant computational and memory costs when processing long prompts, as full self-attention scales quadratically with input length. Token compression aims to address this challenge by reducing the number of tokens representing inputs. However, existing prompt-compression approaches primarily operate in token space and overlook inefficiencies in the latent embedding space. In this paper, we propose K-Token Merging, a latent-space compression framework that merges each contiguous block of K token embeddings into a single embedding via a lightweight encoder. The compressed sequence is processed by a LoRA-adapted LLM, while generation remains in the original vocabulary. Experiments on structural reasoning (Textualized Tree), sentiment classification (Amazon Reviews), and code editing (CommitPackFT) show that K-Token Merging lies on the Pareto frontier of performance vs. compression, achieving up to 75% input length reduction with minimal performance degradation. Code is available at https://github.com/shsjxzh/K-Token-Merging.

Communication in Large Language Model (LLM)-based multi-agent systems is moving beyond discrete tokens to preserve richer context. Recent work such as LatentMAS enables agents to exchange latent messages through full key-value (KV) caches. However, full KV relay incurs high memory and communication cost. We adapt KV-cache eviction methods to this setting and introduce \textbf{Orthogonal BackFill (OBF)} to mitigate information loss from hard eviction. OBF injects a low-rank orthogonal residual from discarded KV states into the retained KV states. We evaluate OBF against full KV relay on nine benchmarks spanning mathematical reasoning, expert and commonsense QA, and coding. With only 9.9%-20.2% of the prompt KV states retained, H-OBF delivers between $97$ and $120$ of full KV relay's per-benchmark accuracy across the nine benchmarks. This suggests that more information does not necessarily lead to better communication; preserving the most useful information matters more. Our codebase is included in the supplementary material. Our codebase is publicly available on https://github.com/markli404/When-Less-Latent-Leads-to-Better-Relay.

There have been numerous attempts to distill quadratic attention-based large language models (LLMs) into sub-quadratic linearized architectures. However, despite extensive research, such distilled models often fail to match the performance of their teacher LLMs on various downstream tasks. We set out the goal of lossless distillation, which we define in terms of tolerance-corrected Win-and-Tie rates between student and teacher on sets of tasks. To this end, we introduce an effective distillation pipeline for xLSTM-based students. We propose an additional merging stage, where individually linearized experts are combined into a single model. We show the effectiveness of this pipeline by distilling base and instruction-tuned models from the Llama, Qwen, and Olmo families. In many settings, our xLSTM-based students recover most of the teacher's performance, and even exceed it on some downstream tasks. Our contributions are an important step towards more energy-efficient and cost-effective replacements for transformer-based LLMs.

Scalar quantization of large language models (LLMs) is fundamentally limited by information-theoretic bounds. While vector quantization (VQ) overcomes these limits by encoding blocks of parameters jointly, practical implementations must avoid the need for expensive lookup mechanisms or other explicit codebook storage. Lattice approaches address this through highly structured and dense packing. This paper explores the Leech lattice, which, with its optimal sphere packing and kissing configurations at 24 dimensions, is the highest dimensional lattice known with such optimal properties. To make the Leech lattice usable for LLM quantization, we extend an existing search algorithm based on the extended Golay code construction, to i) support indexing, enabling conversion to and from bitstrings without materializing the codebook, ii) allow angular search over union of Leech lattice shells, iii) propose fully-parallelisable dequantization kernel. Lastly, we provide a geometric reinterpretation of combining shape--gain quantization with GPTQ-style Hessian corrections: the standard scale-correction step of shape--gain acts as a retraction onto a product of spheres, yielding a Spherical GPTQ primarily acting on directions. We find that low-angular-distortion LLVQ reduces sensitivity to Hadamard/rotation preprocessing, and enables a strong Hadamard-free PTQ in practice. LLVQ delivers state-of-the-art LLM quantization performance, outperforming recent methods such as Quip#, QTIP, and PVQ. The results highlight the effectiveness of high-dimensional lattices for scalable, theoretically grounded model compression.

Layer-wise capacity in large language models is highly non-uniform: some layers contribute disproportionately to loss reduction, whereas others are nearly redundant. Existing layer-scoring methods provide sensitivity estimates but do not give a principled rule for converting those estimates into allocation or pruning decisions under a global hardware budget. We introduce a curvature-aware, MDL-inspired framework built around the layer gain $ζ_k^2=g_k^\top\widetilde H_{kk}^{-1}g_k$. This quantity equals twice the maximal decrease predicted by the regularized layer-restricted quadratic model and incorporates inverse local curvature; it is therefore a local surrogate for reducible risk, not a universal dominance claim over gradient-norm scores. After normalizing the gains into scores $q_k$, we formulate two convex programs: one allocates expert slots under diminishing returns, and the other assigns layer-wise pruning ratios while protecting high-score layers. Both continuous programs have unique globally optimal solutions characterized by one dual variable and computable in $O(K\log(1/\varepsilon))$ time by bisection. We also prove a quadratic transfer-regret bound: when source and target score vectors differ by at most $δ$, the target surrogate cost of the transferred decision is within $O(δ^2)$ of the target optimum. Experiments on Mistral-7B and Gemma-7B show clear allocation gains in some settings and competitive, though mixed, pruning performance. The framework therefore replaces an empirical score-to-decision heuristic with a budget-feasible optimization procedure whose guarantees apply to the stated continuous surrogates. Code is available on github repo - TKAI-LAB-Mali/Curvature-Weighted-Capacity-Allocation

Large language models (LLMs) demand substantial computational and memory resources, posing challenges for efficient deployment. Two complementary approaches have emerged to address these issues: token-adaptive layer execution, which reduces floating-point operations (FLOPs) by selectively bypassing layers, and quantization, which lowers memory footprint by reducing weight precision. However, naively integrating these techniques leads to additional accuracy degradation due to reduced redundancy in token-adaptive models. We propose QTALE (Quantization-Robust Token-Adaptive Layer Execution for LLMs), a novel framework that enables seamless integration of token-adaptive execution with quantization while preserving accuracy. Conventional token-adaptive methods reduce redundancy in two ways: (1) by limiting the diversity of training paths explored during fine-tuning, and (2) by lowering the number of parameters actively involved in inference. To overcome these limitations, QTALE introduces two key components: (1) a training strategy that ensures diverse execution paths are actively explored during fine-tuning, and (2) a post-training mechanism that allows flexible adjustment of the execution ratio at inference to reintroduce redundancy when needed. Experimental results show that QTALE enables seamless integration of token-adaptive layer execution with quantization, while keeping the accuracy gap to quantization-only models below 0.5% on CommonsenseQA benchmarks. By combining tokenadaptive execution for FLOPs reduction and quantization for memory savings, QTALE provides an effective solution for efficient LLM deployment.

In current Mixture of Experts (MoE) architectures, linear memory scaling is present, the memory grows as the number of experts increases. $N$ independent expert weight matrices require $\mathcal{O}(N \cdot d^2)$ memory which exceeds the memory budget of edge devices. Current compression methods like quantization, pruning, and low-rank factorization reduce constant factors, but the scaling bottleneck is still unresolved. We introduce ButterflyMoE, a method which treats experts not as independent matrices but as geometric reorientations of a shared quantized substrate. Diversity amongst the experts arises from viewing different angles of the shared capacity and not from redundant storage. Learned rotations are applied to a shared ternary prototype. With this, each expert yields $\mathcal{O}(d^2 + N \cdot d \log d)$ memory-reducing per-expert cost from $\mathcal{O}(d^2)$ to $\mathcal{O}(d \log d)$. The key insight is that training these rotations with quantization reduces activation outliers and stabilizes extreme low-bit training where other static methods collapse. Across language modeling benchmarks, ButterflyMoE achieves 80$\times$ memory reduction at 8 experts with a highly favorable memory-accuracy tradeoff.At this 80x compression ButterflyMoE outperforms an equal memory dense baseline, showing that orbital parameterization extracts fundamentally more utility per byte. When scaled up to 256 experts, ButterflyMoE asymptotically compresses the memory by 150 $\times$. ButterflyMoE reduces the constant factor of linear scaling with compression ratio growing with the expert count.

Many self-attention sublayers in large language models (LLMs) can be removed with little to no loss. We attribute this to the Attention Suppression Hypothesis: during pre-training, some deep attention layers learn to mute their own contribution, leaving the residual stream and the MLP to carry the representation. We propose Gate-Norm, a one-shot, weight-only criterion that ranks attention sublayers by query-key coupling and removes the least coupled ones, requiring no calibration data, no forward passes, no fine-tuning, and no specialized kernels. On 40-layer, 13B-parameter LLaMA models, Gate-Norm prunes the model in under a second. Pruning 8-16 attention sublayers yields up to $1.30\times$ higher inference throughput while keeping average zero-shot accuracy within 1.5 percentage points of the unpruned baseline across BoolQ, RTE, HellaSwag, WinoGrande, ARC-Easy/Challenge, and OpenBookQA. Across these settings, Gate-Norm matches data-driven pruning methods in accuracy while being $\sim 1000\times$ faster to score layers, enabling practical, data-free compression of LLMs.

Processing long documents with large language models (LLMs) is expensive: a single query over a 100K-token document can cost from tens of cents to over a dollar in API fees, depending on the model, and memory grows linearly with context length. We introduce BudgetMem, a training-free architecture that keeps only high-salience content under an explicit memory budget. Unlike token-level neural compressors such as LLMLingua, BudgetMem makes chunk-level keep-or-discard decisions from interpretable features: entity density, TF-IDF importance, position, numerical density, discourse markers, and question presence. Across four benchmarks, BudgetMem matches the uncompressed baseline on template-generated structured documents (F1 = 0.859 vs. 0.855) while discarding 70% of chunks. As a pre-retrieval filter on the same documents it outperforms LLMLingua-2 by a wide margin (0.859 vs. 0.554), because token-level compression destroys the phrasal structure that retrieval depends on; the advantage holds across model scale and family (Llama-3.2-3B, Qwen2.5-7B), under dense retrieval (0.918 vs. 0.530), and in LLMLingua-2's intended direct-feed mode (0.797 vs. 0.720). On real NLP research papers (Qasper), BudgetMem is competitive but not free, trailing the full baseline by 7.4% and LLMLingua-2 by 5.5% relative F1 at a 72% storage reduction. Removing discourse markers or numerical density hurts most, although TF-IDF alone scores as well on template documents. The pipeline needs no trained models or GPU for compression and runs on a $10/month Colab instance.

Mixture-of-Experts (MoE) models have shown strong potential in scaling language models efficiently by activating only a small subset of experts per input. However, their widespread deployment remains limited due to the high memory overhead associated with storing all expert parameters, particularly as the number of experts increases. To address this challenge, prior works have explored expert dropping and merging strategies, yet they often suffer from performance drop at high compression ratios. In this paper, we introduce PuzzleMoE, a training-free MoE compression method that achieves both high accuracy and efficient inference through two key innovations: First, PuzzleMoE performs sparse expert merging by identifying element-wise weight redundancy and specialization. It uses a dual-mask to capture both shared and expert-specific parameters. Second, to avoid the overhead of storing binary masks and signs, PuzzleMoE introduces a bit-packed encoding scheme that reuses underutilized exponent bits, enabling efficient MoE inference on GPUs. Extensive experiments demonstrate that PuzzleMoE can compress MoE models by up to 50% while maintaining accuracy across various tasks. Specifically, it outperforms prior MoE compression methods by up to 16.7% on MMLU at 50% compression ratio, and achieves up to 1.28\times inference speedup.

Large language models (LLM) and vision-language models (VLM) have achieved state-of-the-art performance, but they impose significant memory and computing challenges in deployment. We present a novel low-rank compression framework to address this challenge. First, we upper bound the change of network loss via layer-wise activation-based compression errors, filling a theoretical gap in the literature. We then formulate low-rank model compression as a bi-objective optimization and prove that a single uniform tolerance yields surrogate Pareto-optimal heterogeneous ranks. Based on our theoretical insights, we propose Pareto-Guided Singular Value Decomposition (PGSVD), a zero-shot pipeline that improves activation-aware compression via Pareto-guided rank selection and alternating least-squares implementation. We apply PGSVD to both LLM and VLM, showing better accuracy at the same compression levels and inference speedup.

Large language models (LLMs) deliver impressive performance but incur prohibitive memory and compute costs at deployment. Model pruning is an effective way to reduce these overheads, yet existing approaches face challenges: unstructured sparsity, where nonzeros can appear anywhere, preserves accuracy but yields irregular access patterns that prevent GPU acceleration, while semi-structured 2:4 sparsity is hardware-friendly but enforces a rigid 50% pattern that degrades model quality. To bridge this gap, we introduce PATCH, a hybrid sparsity framework that enables a continuous sparsity ratio between 0% and 50%. PATCH partitions weight matrices into tiles, assigning each tile to be either dense or 2:4 sparse via a learnable mask selection mechanism. This design provides fine-grained control over accuracy-acceleration tradeoffs and supports non-uniform sparsity across layers, leading to superior overall quality. Across models from 0.5B to 13B parameters, PATCH consistently narrows the gap to dense accuracy while delivering practical speedups. For instance, on LLaMA-2 7B with an A6000 GPU, PATCH achieves 1.18x-1.38x end-to-end speedup over dense baselines while improving accuracy by 0.37%-2.96% compared to the state-of-the-art 2:4 pruning method, MaskLLM.

Large Language Models (LLMs) face significant computational challenges when processing long contexts due to the quadratic complexity of self-attention. While soft context compression methods, which map input text to smaller latent representations, have shown promise, their real-world adoption is limited. Existing techniques typically compress the context as a single unit, which leads to quadratic compression complexity and an inability to reuse computations across queries with overlapping contexts. In this work, we introduce CompLLM, a soft compression technique designed for practical deployment. Instead of processing the context holistically, CompLLM divides it into segments and compresses each one independently. This simple design choice yields three critical properties: efficiency, as the compression step scales linearly with the context length; scalability, enabling models trained on short sequences (e.g., 1k tokens) to generalize to contexts of 100k tokens; and reusability, allowing compressed segments to be cached and reused across different queries. Our experiments show that with a 2x compression rate, at high context lengths CompLLM speeds up Time To First Token (TTFT) by up to 4x and reduces the KV cache size by 50%. Furthermore, CompLLM achieves performance comparable to that obtained with the uncompressed context, and even surpasses it on very long sequences, demonstrating its effectiveness and practical utility.

Large Language Models have achieved remarkable performance across a large number of tasks, but face critical deployment and usage barriers due to substantial computational requirements. Model compression methods, which aim to reduce model size while preserving its capacity, are an important means to mitigate these issues. Promising approaches along these lines, such as structured pruning, typically require costly manual hyperparameter exploration or rely on local heuristics that may run the risk of ignoring better solutions. In this work we introduce GeLaCo, an evolutionary approach to LLM compression via layer collapse. Our approach supports an efficient exploration of the compression solution space via population-based search and a novel layer collapse formulation based on parametrized weight merging, with a fitness function based on similarity over residual updates and language modeling KL divergence. GeLaCo also supports both single and multi-objective evolutionary compression search, establishing the first Pareto front estimation along compression and quality axes. We evaluate GeLaCo solutions via both perplexity-based and generative evaluations over foundational and instruction-tuned models, outperforming state-of-the-art alternatives.

The Fisher information is a fundamental concept for characterizing the sensitivity of parameters in neural networks. However, leveraging the full observed Fisher information is too expensive for large models, so most methods rely on simple diagonal approximations. While efficient, this approach ignores parameter correlations, often resulting in reduced performance on downstream tasks. In this work, we mitigate these limitations and propose Generalized Fisher-Weighted SVD (GFWSVD), a post-training LLM compression technique that accounts for both diagonal and off-diagonal elements of the Fisher information matrix, providing a more accurate reflection of parameter importance. To make the method tractable, we introduce a scalable adaptation of the Kronecker-factored approximation algorithm for the observed Fisher information. We demonstrate the effectiveness of our method on LLM compression, showing improvements over existing compression baselines. For example, at a 20 compression rate on the MMLU benchmark, our method outperforms FWSVD, which is based on a diagonal approximation of the Fisher information, by 5 percent, SVD-LLM by 3 percent, and ASVD by 6 percent compression rate.

Training high-performing Small Language Models (SLMs) remains costly, even with knowledge distillation and pruning from larger teacher models. Existing work often faces three key challenges: (1) information loss from hard pruning, (2) inefficient alignment of representations, and (3) underutilization of informative activations, particularly from Feed-Forward Networks (FFNs). To address these challenges, we introduce Low-Rank Clone (LRC), an efficient pre-training method that constructs SLMs aspiring to behavioral equivalence with strong teacher models. LRC trains a set of low-rank projection matrices that jointly enable soft pruning by compressing teacher weights, and activation clone by aligning student activations, including FFN signals, with those of the teacher. This unified design maximizes knowledge transfer while removing the need for explicit alignment modules. Extensive experiments with open-source teachers (e.g., Llama-3.2-3B-Instruct, Qwen2.5-3B/7B-Instruct) show that LRC matches or surpasses state-of-the-art models trained on trillions of tokens--while using only 20B tokens, achieving over 1,000x training efficiency. Our codes and model checkpoints are available at https://github.com/CURRENTF/LowRankClone and https://huggingface.co/collections/JitaiHao/low-rank-clone-lrc-6828389e96a93f1d4219dfaf.

The emergence of Transformer-based Large Language Models (LLMs) has substantially augmented the capabilities of Natural Language Processing (NLP), thereby intensifying the demand for computational resources. Therefore, enhancing efficiency based on factors like computational requirements, energy consumption, carbon footprint and financial cost has become a vital area of research. This motivates us to conduct a survey on Transformer-based LLMs in NLP from the perspective of efficiency. In this survey of 312 articles, the efficiency-improvement endeavors have been systematically discussed targeting various aspects such as data curation, model design, model downsizing, and dynamic inferencing. This has been augmented with efficiency considerations in model adaptation strategies like pre-training, fine-tuning, prompt-engineering and Retrieval-Augmented Generation (RAG). Furthermore, a statistical analysis followed by an in-depth evaluation of the efficiency and efficacy of more than 30 renowned NLP models has been performed on 13 evaluation benchmarks. This paper offers valuable insights for researchers, professionals as well as scholars, and explores the trend of research toward sustainable practices in NLP.

In this paper, we formulate the compression of large language models (LLMs) by optimally deleting transformer blocks (``block removal'') as a constrained binary optimization (CBO) problem that can be mapped to a physical system (Ising glass), whose energies are a strong proxy for downstream model performance. This formulation enables an efficient ranking of a large number of candidate block-removal configurations yielding many high-quality, non-trivial solutions beyond those only removing consecutive regions. Our method performs strongly in the deep compression regime, such as for 50% compression of Llama-3.3-70B-Instruct, where we achieve an almost 23 percentage point increase on the MMLU benchmark compared to other state-of-the-art (SOTA) block-removal methods. For lighter compression, it performs on par with those methods across several benchmarks for Llama-3.1-8B-Instruct, Qwen3-14B (both before and after retraining), as well as Llama-3.3-70B-Instruct. The approach is computationally efficient and requires only forward and backward passes on a calibration dataset for a few active parameters. Additionally, we demonstrate that using good heuristic solvers for the CBO problem provides solutions that perform well on downstream tasks in negligible runtime when it is unfeasible to solve the problem exactly. The method can be readily applied to any architecture. We illustrate this generality on the recent NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 model, which exhibits a highly inhomogeneous and challenging block structure, and where we outperform SOTA for AIME25 and GPQA when removing either 2 attention layers or 3 mixture-of-experts layers.