{"total":31,"items":[{"citing_arxiv_id":"2607.02461","ref_index":9,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"OrbitQuant: Data-Agnostic Quantization for Image and Video Diffusion Transformers","primary_cat":"cs.CV","submitted_at":"2026-07-02T17:27:34+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"OrbitQuant is a data-agnostic PTQ technique for DiTs that uses RPBH rotation in a normalized basis to enable a single codebook across all inputs, achieving SOTA low-bit performance on FLUX.1, CogVideoX and similar models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.23406","ref_index":9,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"HyperQuant: A Rate-Distortion-Optimal Quantization Pipeline for Large Language and Diffusion Models","primary_cat":"cs.LG","submitted_at":"2026-06-22T14:30:47+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"HyperQuant unifies Hadamard transform, optimal lattice quantization, and entropy coding to outperform prior schemes on LLM weight and KV cache quantization down to 1.7 bits per scalar while preserving quality on a 19B DiT model.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.13054","ref_index":7,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"TWLA: Achieving Ternary Weights and Low-Bit Activations for LLMs via Post-Training Quantization","primary_cat":"cs.LG","submitted_at":"2026-06-11T08:37:20+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"TWLA is a PTQ method using E2M-ATQ, KOTMS, and ILA-AMP to enable W1.58A4 quantization for LLMs with maintained accuracy.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.06060","ref_index":14,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE","primary_cat":"cs.CV","submitted_at":"2026-06-04T11:59:56+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"ReCache learns recomputation schedules via policy gradients to maximize quality under a target compute budget for any caching mechanism in diffusion models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.04349","ref_index":27,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"MorphoQuant: Modality-Aware Quantization for Omni-modal Large Language Models","primary_cat":"cs.CV","submitted_at":"2026-06-03T02:05:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"MorphoQuant proposes DABC and MDQFO for 4-bit quantization of omni-modal LLMs, claiming superior performance over SOTA W4A4 methods and even W4A16 baselines on benchmarks like ScienceQA.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.03328","ref_index":8,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Calibration Data Trade-offs Across Capability Dimensions: Why Multi-Source Mixing Matters for High-Sparsity LLM Pruning","primary_cat":"cs.LG","submitted_at":"2026-06-02T08:38:14+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Analysis of 15 calibration sources shows opposite-sign Spearman correlations between perplexity and retention across General vs. Math/Code dimensions in LLM pruning, and multi-source mixing via IGSP raises total retention from 40-50% to 58.8%.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.04032","ref_index":75,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Do Transformers Need Three Projections? Systematic Study of QKV Variants","primary_cat":"cs.LG","submitted_at":"2026-06-01T20:59:05+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Q-K=V projection sharing in transformers matches standard QKV performance with 50% KV cache reduction and combines with GQA/MQA for up to 96.9% reduction across vision and language tasks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.09864","ref_index":16,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation","primary_cat":"cs.LG","submitted_at":"2026-06-01T02:02:20+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":8.0,"formal_verification":"none","one_line_summary":"KV cache quantization silently erodes LLM safety alignment via vulnerable low-dimensional subspaces, diagnosed by Per-Channel Reduction into three failure modes and mitigated training-free with up to 97% recovery.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.01412","ref_index":17,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"GPTQ-intrinsic LoRA: A Near-optimal Algorithm for Low-precision Quantization with Low-rank Adaptation","primary_cat":"cs.LG","submitted_at":"2026-05-31T19:17:39+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":8.0,"formal_verification":"none","one_line_summary":"GPTQ-intrinsic LoRA augments GPTQ with intrinsic low-rank compensation via Hessian modification to achieve layer-wise reconstruction bounds that match information-theoretic lower bounds under structural assumptions.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.23078","ref_index":5,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs","primary_cat":"cs.LG","submitted_at":"2026-05-21T22:23:38+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"GEMQ applies global LP-based expert importance estimation and router fine-tuning within progressive quantization to cut memory and speed inference in MoE LLMs with little accuracy loss.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.24011","ref_index":5,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models","primary_cat":"cs.CV","submitted_at":"2026-05-19T19:57:26+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"ActQuant achieves sub-4-bit (down to 2.5 bpw) quantization of VLA models via action-contribution bit allocation and curvature-based scale tuning, retaining over 90% performance on LIBERO and physical robot tasks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.19929","ref_index":9,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Breaking Modality Heterogeneity in Low-Bit Quantization for Large Vision-Language Models","primary_cat":"cs.CV","submitted_at":"2026-05-19T14:49:57+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SplitQ improves low-bit PTQ for VLMs by isolating modality-specific outlier channels via MOCD and applying dual-branch adaptive calibration via ACC, outperforming prior methods on six datasets across W4A8 to W3A2 settings.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.17745","ref_index":27,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"StatQAT: Statistical Quantizer Optimization for Deep Networks","primary_cat":"stat.ML","submitted_at":"2026-05-18T01:56:12+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"A statistical error analysis framework yields iterative and analytic quantizers that improve accuracy and stability when incorporated into quantization-aware training for integer and floating-point formats.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.14844","ref_index":4,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"XFP: Quality-Targeted Adaptive Codebook Quantization with Sparse Outlier Separation for LLM Inference","primary_cat":"cs.LG","submitted_at":"2026-05-14T13:52:31+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"XFP introduces quality-targeted adaptive codebook quantization with sparse outlier separation that auto-selects parameters from cosine similarity floors, achieving high throughput and accuracy on Qwen3.5 models at low effective bits without calibration data.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.11222","ref_index":6,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"ADMM-Q: An Improved Hessian-based Weight Quantizer for Post-Training Quantization of Large Language Models","primary_cat":"cs.LG","submitted_at":"2026-05-11T20:33:41+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"ADMM-Q is a new post-training quantization method using ADMM operator splitting that reduces WikiText-2 perplexity compared to GPTQ on Qwen3-8B across W3A16, W4A8, and W2A4KV4 settings.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.06052","ref_index":11,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"XtraMAC: An Efficient MAC Architecture for Mixed-Precision LLM Inference on FPGA","primary_cat":"cs.AR","submitted_at":"2026-05-07T11:37:52+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"XtraMAC unifies mixed-precision MAC on FPGA via shared integer mantissa products, delivering 1.4-2.0x higher compute density and up to 1.9x better energy efficiency.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.06014","ref_index":17,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Quantizing With Randomized Hadamard Transforms: Efficient Heuristic Now Proven","primary_cat":"cs.LG","submitted_at":"2026-05-07T11:11:54+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Two randomized Hadamard transforms suffice to make coordinate marginals O(d^{-1/2})-close to Gaussian for most quantization methods, with three needed for vector quantization to match uniform random rotations asymptotically.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.02262","ref_index":11,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"WindowQuant: Mixed-Precision KV Cache Quantization based on Window-Level Similarity for VLMs Inference Optimization","primary_cat":"cs.CV","submitted_at":"2026-05-04T06:17:13+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"WindowQuant performs window-adaptive mixed-precision KV cache quantization guided by similarity to the text prompt, with reordering to enable efficient inference in VLMs.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"from the original FP16 type to a lower bit-width type (such as INT8, INT4, or even INT2). However, most existing KV cache quantization efforts [24, 28, 35, 49] uniformly quantize the KV cache to a single bit-width. In reality, the importance of tokens within the KV cache is hierarchical, meaning some tokens are more critical than others. Therefore, uniform quantization can easily cause accuracy degradation. Some papers [11, 16, 17, 46] have proposed implementing mixed-precision quantization based on the importance of the tokens to address this issue, but their quantization search strategies are token-level and very time-consuming. Even worse, mixed-precision quantization can cause hardware inefficiency during the inference computation, such as increased latency and memory usage."},{"citing_arxiv_id":"2604.24273","ref_index":1,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"BitRL: Reinforcement Learning with 1-bit Quantized Language Models for Resource-Constrained Edge Deployment","primary_cat":"cs.LG","submitted_at":"2026-04-27T10:03:37+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"BitRL enables on-device RL agents via 1-bit quantized language models, delivering 10-16x memory reduction and 3-5x energy efficiency gains with 85-98% retained performance.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.24008","ref_index":7,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Coverage-Based Calibration for Post-Training Quantization via Weighted Set Cover over Outlier Channels","primary_cat":"cs.LG","submitted_at":"2026-04-27T03:43:29+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"COVERCAL selects PTQ calibration samples via weighted set cover over outlier channels, with a stylized clipping model showing missed coverage upper-bounds surrogate loss, yielding gains over random and other baselines on LLaMA and Mistral models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.22906","ref_index":31,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Network Edge Inference for Large Language Models: Principles, Techniques, and Opportunities","primary_cat":"cs.DC","submitted_at":"2026-04-24T16:56:53+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"A survey synthesizing challenges, system architectures, model optimizations, deployment methods, and resource management techniques for large language model inference at the network edge.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"• Encoder-only LLMs: These models consist of a stack of Transformer encoder bloc ks [151] that map input text to contextual vector representations, without an explicit decoding phase for free-form generation. They are typ- ically pretrained using a masked language modeling paradigm, i.e., predicting masked words in a sentence using bidirectional self-attention (no causal mask) conditioni ng on both left and right context. Canonical examples include BERT [31], RoBERTa [99], and ALBERT [86]. • Encoder-decoder LLMs : These models pair a bidirectional encoder with an autoregr essive decoder, linked via cross-attention from the decoder to the encoder. The enc oder converts the input text into context-rich em- beddings, while the decoder generates the target sequence l eft-to-right using causal self-attention plus cross-"},{"citing_arxiv_id":"2604.20682","ref_index":3,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Variance Is Not Importance: Structural Analysis of Transformer Compressibility Across Model Scales","primary_cat":"cs.LG","submitted_at":"2026-04-22T15:31:46+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"High-variance activation directions are uncorrelated with predictions, transformer blocks grow more linear with depth, and single-block linear replacement yields 34x compression on Mistral's final block at a 1.71 perplexity cost.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.19884","ref_index":8,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization","primary_cat":"cs.CL","submitted_at":"2026-04-21T18:03:38+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"LLM 2-bit quantization fails via either cumulative signal degradation or early computation collapse in key components.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.19167","ref_index":66,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"LBLLM: Lightweight Binarization of Large Language Models via Three-Stage Distillation","primary_cat":"cs.LG","submitted_at":"2026-04-21T07:25:02+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"LBLLM achieves better accuracy than prior binarization methods for LLMs by decoupling weight and activation quantization through initialization, layer-wise distillation, and learnable activation scaling.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.18556","ref_index":7,"ref_count":2,"confidence":0.9,"is_internal_anchor":false,"paper_title":"GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling","primary_cat":"cs.CL","submitted_at":"2026-04-20T17:45:47+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"GSQ uses Gumbel-Softmax to optimize scalar quantization grids for LLMs, closing most of the accuracy gap to vector methods like QTIP at 2-3 bits per parameter while using symmetric scalar grids compatible with existing kernels.","context_count":1,"top_context_role":"method","top_context_polarity":"use_method","context_text":"(6) Instead of introducing 2b logits for coordinate i, we introduce only 5 logits corresponding to a discrete shift δi ∈ {−2,−1,0,1,2} . Let ℓδ ∈R d×5 denote the corresponding trainable logits. At each training step, δi is obtained by applying Gumbel-Softmax sampling to the i-th row of ℓδ. The resulting grid index is then ji = clip(j0 i +δ i,1,2 b),(7) where clip(x, a, b) = min{max{x, a}, b} clips the value into the valid range. The final quantized value is qi = (Gb)ji .(8) Equivalently, the constraint set becomes Cshift b-bit = n ¯w ¯w=s·q, s∈R, q i = (Gb)clip(j0 i +δi,1,2 b), δ i ∈ {−2,−1,0,1,2} o .(9) Under this parameterization, each coordinate requires only5 trainable logits rather than 2b."},{"citing_arxiv_id":"2604.17104","ref_index":25,"ref_count":2,"confidence":0.9,"is_internal_anchor":false,"paper_title":"TStore: Rethinking AI Model Hub with Tensor-Centric Compression","primary_cat":"cs.DC","submitted_at":"2026-04-18T18:40:29+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"TStore reduces AI model storage via tensor-level fingerprinting, clustering, and compression without annotations while claiming to preserve usability.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"This section reviews existing storage reduction techniques. A high-level comparison of these methods is summarized in Fig. 3. 2.1 Traditional Storage Reduction General-purpose Compression.General-purpose com- pressors such as Zstandard [18] and Brotli [1] are employed in storage reduction to exploit local byte-level redundancy via dictionary-based [92, 101] and entropy coding [25, 28, 39, 62, 73, 87]. Further gains are possible when the data type is known, e.g., run-length encoding for low-entropy data [79], delta encoding for versioned files [59, 85], and specialized codecs for columnar and time-series workloads [13, 14, 50, 52, 57, 88]. 2 TensorHub: Rethinking AI Model Hub with Tensor-Centric Compression Conference'17, July 2017, Washington, DC, USA"},{"citing_arxiv_id":"2604.06916","ref_index":55,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"FP4 Explore, BF16 Train: Diffusion Reinforcement Learning via Efficient Rollout Scaling","primary_cat":"cs.LG","submitted_at":"2026-04-08T10:14:47+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Sol-RL decouples FP4-based candidate exploration from BF16 policy optimization in diffusion RL, delivering up to 4.64x faster convergence with maintained or superior alignment performance on models like FLUX.1 and SD3.5.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"and Ping Luo. Omniquant: Omnidirectionally calibrated quantization for large language models, 2024. URL https: //arxiv.org/abs/2308.13137. [54] Vage Egiazarian, Andrei Panferov, Denis Kuznedelev, Elias Frantar, Artem Babenko, and Dan Alistarh. Extreme compression of large language models via additive quantization, 2024. URLhttps://arxiv.org/abs/2401.06118. [55] Tim Dettmers, Ruslan Svirschevski, Vage Egiazarian, Denis Kuznedelev, Elias Frantar, Saleh Ashkboos, Alexander Borzunov, Torsten Hoefler, and Dan Alistarh. Spqr: A sparse-quantized representation for near-lossless llm weight compression, 2023. URLhttps://arxiv.org/abs/2306.03078. [56] Shih-yang Liu, Zechun Liu, Xijie Huang, Pingcheng Dong, and Kwang-Ting Cheng."},{"citing_arxiv_id":"2605.20193","ref_index":13,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification","primary_cat":"cs.CL","submitted_at":"2026-04-04T04:50:03+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"A multi-pass prompt verification method stabilizes and improves accuracy of 4-bit and lower quantized LLaMA-3.1 models for thematic extraction from interview data, though 8-bit versions remain closest to a human-plus-BF16 gold standard.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2507.03052","ref_index":8,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"From 2:4 to 8:16 sparsity patterns in LLMs for Outliers and Weights with Variance Correction","primary_cat":"cs.LG","submitted_at":"2025-07-03T12:17:45+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"8:16 sparsity with variance correction and outlier handling lets compressed LLMs match or exceed dense-model accuracy under fixed memory limits, outperforming the common 2:4 pattern in flexibility.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2505.02380","ref_index":26,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"EntroLLM: Entropy Encoded Weight Compression for Efficient Large Language Model Inference on Edge Devices","primary_cat":"cs.LG","submitted_at":"2025-05-05T05:42:14+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"EntroLLM applies tensor-level mixed quantization to reduce weight entropy then uses Huffman coding for up to 65% storage savings and faster inference on memory-limited edge devices without retraining.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2404.14294","ref_index":196,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"A Survey on Efficient Inference for Large Language Models","primary_cat":"cs.CL","submitted_at":"2024-04-22T15:53:08+00:00","verdict":"ACCEPT","verdict_confidence":"MODERATE","novelty_score":3.0,"formal_verification":"none","one_line_summary":"The paper surveys techniques to speed up and reduce the resource needs of LLM inference, organized by data-level, model-level, and system-level changes, with comparative experiments on representative methods.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"ZipLM [176], LoRAPrune [177], LoRAS- hear [178], SliceGPT [179], PLATON [180], CoFi [181], SIMPLE [182], ExpertSpar- sity [183], SEER-MoE [184], Pruner-Zero [185], DSØT [186] Quantization Quantization- aware Training LLM-QAT [187], Norm Tweaking [188], QLoRA [189], QA-LoRA [190], LoftQ [191] Post-Training Quantization GPTQ [192], LUT-GEMM [193], AWQ [194], OWQ [195], SpQR [196], SqueezeLLM [197], QuIP [198], FineQuant [199], QuantEase [200], LLM-MQ [201], ZeroQuant [202], Flex- Gen [203], LLM.int8() [204], Smoothquant [205], ZeroQuant-V2 [206], RPTQ [207], OliVe [208], OS+ [169], ZeroQuant-FP [209], Omniquant [210], QLLM [211], ATOM [212], LLM-FP4 [187], BiLLM [213], Li et.al. [214], AffineQuant [215], QuIP [198], QuIP# [216],"}],"limit":50,"offset":0}