Jun 2026· International Conference on Mobile Data Management· pp. 363-366· 0 citations· 21 references
Abstract
Deploying large language models in resource-constrained environments such as mobile devices requires effective compression. Pruning and quantization are largely used, yet applying pruning to a new model typically demands manual inspection of its architecture to identify layer names, MLP structure, and projection patterns, a process that is tedious, error-prone, and must be repeated for every new model design. We present ShrinkLLM, an end-to-end compression system that eliminates the manual inspection of model structure during pruning by delegating architectural analysis to an AI agent. Given a causal LLM, the agent automatically identifies MLP layer organization, expansion layers, gate projections for Gated Linear Unit architectures, and down-projection layers, structural information that is then passed directly to a compression engine to perform neuron pruning, quantization, or a combination of both. We demonstrate the system on llama-3.1-8b-instruct, Qwen2.5-3b-instruct and Gemma-2-2b-it showcasing how the agent correctly resolves architecturespecific structural details and drives compression across all MLP layers automatically.
Large Language Models (LLMs) trained using Chain-of-Thought (CoT) supervision have achieved state-of-the-art performance on complex reasoning tasks. However, the generation of long reasoning chains introduces substantial computational overhead during inference, limiting their deployment in low-latency and resource-constrained environments. This paper proposes AdaptiReason, a novel framework that dynamically compresses intermediate reasoning steps based on task difficulty and model confidence without requiring retraining of the underlying base model. AdaptiReason employs a lightweight difficulty estimator to determine the appropriate reasoning depth for each input, followed by a learned token-pruning policy that eliminates redundant or low-information reasoning steps. Experimental evaluation on the MATH, GSM8K, and ARC-Challenge benchmarks demonstrates that AdaptiReason reduces the average number of generated tokens by 3.7× while preserving 98.2% of the baseline reasoning accuracy. Furthermore, the proposed framework is model-agnostic and can be seamlessly integrated with instruction-tuned LLMs without requiring access to model parameters, relying solely on output logits for adaptive reasoning compression. The results demonstrate that AdaptiReason significantly improves inference efficiency while maintaining high reasoning performance, making it suitable for real-time and resource-constrained LLM applications.
V. A, Mithaguru, Amrita Kundu et al.· 2026 4th International Confe...· 0 citations
Despite fixed sampling parameters and random seeds, Large Language Model (LLM) inference exhibits output inconsistency, which undermines downstream tasks such as model evaluation and reinforcement learning. A major source of this nondeterminism is batch-dependent GPU execution: dynamic input shapes change kernel tiling and floating-point reduction orders. Existing systems address this problem with batch-invariant kernels, but these kernels restrict optimized tiling and split reductions, increasing more than 2$\times$ latency and reducing serving throughput by up to 74 %. This paper observes that although most kernels are not batch-invariant, they are position-invariant. Leveraging this property, we present CoRun, a scheduling-based system that achieves deterministic inference without requiring batch invariance. CoRun employs isolated prefill and fixed-shape batched decode to handle the two stages of LLM inference, respectively, leveraging CUDA graphs for efficient execution and simplified implementation. Experiments on LLMs with diverse architectures, including Qwen and DeepSeek, show that CoRun ensures determinism while improving throughput by 15-324 % over batch-invariant approaches, reducing time-to-first-token by 51.8 % and time-per-output-token by 48.6 % on average.
Shiju Zhao, Jiacheng Yang, Qihang Chen et al.· 0 citations
This study introduces CodeGrep, a 14B retrieval agent trained end-to-end with GRPO to issue multi-turn parallel grep, glob, and read tool calls and return candidate files to a frozen downstream coding agent, and applies the efficiency signal at the advantage layer rather than the reward layer to reduce KL drift and translates cleanly into downstream efficiency.
Wu-Ya Chen, Yihao Yang, Yang Cao et al.· 0 citations
This work proposes a mechanism that bridges these two paradigms: an LLM produces a high-quality seed architecture, then decomposes it into a "slotted architecture", a scaffold with named, interchangeable module slots that automatically defines a bounded, task-specific search space for conventional NAS to explore, without manual engineering.
Seokhoon Jeong, Mijung Kim, Taehwan Kim· 0 citations
Training trillion-parameter models has made collective communication a dominant bottleneck in GPU clusters. Existing approaches face a fundamental tension: optimization-based synthesis techniques suffer from combinatorial complexity that can require minutes to hours, while analytic heuristics rely on rigid assumptions that often fail under production variability. To the best of our knowledge, we present OptiFlow, among the first LLM-driven frameworks for automated design of high-performance collective communication algorithms. Our key insight is a two-layer decomposition: the LLM generates compact data-movement intent expressed in a domain-specific language, while deterministic scheduling algorithms compile these programs into executable schedules. We further employ an iterative pipeline that uses real-hardware feedback to refine LLM-generated proposals, without updating model parameters or relying on manually designed search heuristics. Evaluated on a 32-GPU NVIDIA A100 cluster, OptiFlow discovers All-Gather schedules that outperform NCCL by up to 3.75 × , while it also surpasses TACCL and TE-CCL with gains up to 5.12 × and 3.28 × , and shows better optimization-time scalability.
Fei Long, Ziyue Yang, Kaihui Gao et al.· Asia-Pacific Workshop on Net...· 0 citations
Pruning long context for coding agents has been a vital technology for efficient context management. While existing context pruning methods such as SWE-Pruner realize this by attaching a separate code classifier, we find the agent itself encodes internal representations indicating the relevance of code context when reading tool output. Based on this finding, we propose SWE-Pruner Pro, which prunes tool outputs directly inside the agent. Concretely, a small head turns the agent's own internal representations into a keep-or-prune label for each line, with a length-aware embedding keyed to each tool output's line count. Across two open-weight backbones and four multi-turn benchmarks, SWE-Pruner Pro saves up to 39% of prompt and completion tokens while preserving task quality, with bounded inference overhead. Notably, on MiMo-V2-Flash SWE-Pruner Pro additionally raises the SWE-Bench Verified resolve rate by +3.8% and the long-context Oolong accuracy by +2.2 points.
Yuhang Wang, Yuling Shi, Shaoqiu Zhang et al.· 1 citation