#quantization
12 papers match
CACHE-UK: A Stability-Aware Memory Editor for Sequentially Updated Quantized LLMs in Finance
Anubhav Lakra, Yue Feng
The paper introduces CACHE-UK, a stability-aware memory editing framework for 4-bit quantized large language models used in UK finance, which reduces knowledge degradation during s…
HiFloat4 Format for End-To-End Reinforcement Learning Post-Training of Large Language Models
Hei Yi Mak, Shadan Golestan, Hoang Le +10
The paper introduces HiFloat4, a 4-bit floating-point format and a Rollout Residual Quantization technique that enable end-to-end reinforcement learning post‑training of large lang…
Flat Score, Amplified Failures: How the Error Budget Masks Damage in Quantized LLM Agents
Jiwon Jang, Kisu Yang, Heuiseok Lim +1
The paper evaluates 4-bit post‑training quantization of multi‑turn, tool‑calling LLM agents and finds that while standard scores remain unchanged, quantization substantially increa…
How Small Can You Go? A Controlled Study of LoRA Rank, Target Modules, and Quantization Trade-offs for Text-to-SQL on a 60M-Parameter Model
Mahendra Singh Rathor, Anagheem Azzam
The paper investigates how LoRA rank, adapted modules, and low‑bit quantization trade off accuracy and resource usage when fine‑tuning a 60 M‑parameter T5‑small model for the WikiS…
When W4A4 Breaks Camouflaged Object Detection: Token-Group Dual-Constraint Activation Quantization
Tianqi Li, Wenyu Fang, Xin He +3
The paper proposes a post‑training 4‑bit activation quantization method for transformer‑based camouflaged object detection that mitigates token‑level range domination to preserve s…
The Joint Effect of Quantization and Sampling Temperature on LLM Safety Alignment: A Factorial Analysis
Hari Prasad, Ritam Pal
The paper investigates how model quantization and higher sampling temperatures jointly affect the safety alignment of instruction-tuned large language models, finding that quantiza…
ExTernD: Expanded-Rank Ternary Decomposition Ternary LLM PTQ with Accuracy Approaching Any Quantization Level
Chethan Reddy G. P
The paper presents ExTernD, a post‑training factorization that expands the rank of ternary matrix decompositions to correct quantization errors, enabling large language models to a…
Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levels
Roman Prosvirnin, Victor Minchenkov, Alexey Soldatov +1
The paper introduces JADR, a protocol that examines a language model's internal Jacobian representations (J-space) to assess danger recognition before any response is generated, en…
Symmetries and Conservation Laws in Lie-Poisson Electrodynamics
M. A. Kurkov
The paper studies Lie-Poisson electrodynamics, a non‑Abelian nonlinear deformation of Maxwell theory, showing how a field redefinition maps its dynamics to standard electrodynamics…
Calibrated e-CUSUM Decoding for Quantized Reasoning Models: Why Token Log-Probability Is the Wrong Observable for Decoding Monitors
El Hassane Ettifouri, Ayoub Belfatmi, Mahaman Sanoussi Yahaya Alassan +1
The paper shows that using centered token log‑probability as a monitor for quantized reasoning language models is ineffective, and proposes a training‑free calibrated e‑CUSUM decod…
Decomposing Runtime, Kernel, and Quantization Speedups via a Matched FP16 Intermediate: A Hardware-Conditioned Case Study on Four NVIDIA RTX A5000 GPUs
Weijia Han, Lisha Qu
The paper analyzes how much of the reported speedup from quantized inference on NVIDIA RTX A5000 GPUs comes from reduced runtime versus kernel and quantization changes, using a mat…
A Replicate-and-Quantize Strategy for Plug-and-Play Load Balancing of Sparse Mixture-of-Experts LLMs
Zijie Liu, Jie Peng, Jinhao Duan +7
The paper proposes a training‑free method that replicates heavily used experts and quantizes less important ones to rebalance workload in sparse mixture‑of‑experts large language m…
One search, two signals: results blend meaning (embedding similarity, so papers that never use your words still surface) with keyword matches on titles, abstracts and summaries. Free, no sign-in needed.