The paper introduces LCPO, a lightweight preference‑optimization technique that uses only 800 training examples and 50 steps to teach large language models to produce concise, accurate answers, halving inference length while often improving accuracy and reducing training cost by up to two orders of magnitude.
The ICLR‑2026 paper introduces LCPO, a lightweight preference‑optimization technique that uses only 800 curated examples and 50 training steps to cut large‑model chain‑of‑thought generation length by about 50% while maintaining or even improving answer accuracy, dramatically reducing training and inference costs.
Recent open‑weight LLMs such as Gemma 4, Laguna XS.2, ZAYA1‑8B, and DeepSeek V4 introduce KV‑cache sharing, per‑layer embeddings, layer‑wise attention budgeting, and compressed attention mechanisms that dramatically reduce memory and compute overhead for very long contexts while preserving model quality.
Xiaomi secured three top spots at CVPR 2026 NTIRE—first in Efficient Super‑Resolution with SPANV2, first in Portrait Restoration using a dual‑stage cascade, and second in Reflection Removal via RDNet‑XL and diffusion‑model distillation—showcasing hardware‑software co‑design, ultra‑fast inference, and novel algorithmic innovations.
The paper introduces Heima, a framework that compresses chain‑of‑thought reasoning into a small set of abstract “thinking tokens” for multimodal large models, dramatically reducing generated tokens while preserving inference capability, and provides an adaptive interpreter to reconstruct human‑readable reasoning for analysis.
KSA inserts learnable summary tokens to compress KV cache by a factor of eight, enabling accurate long‑context retrieval with far lower memory and compute costs, and it consistently outperforms full‑attention and other hybrid methods on large‑scale benchmarks.
The paper introduces ReBalance, a training‑free two‑stage inference control framework that uses model confidence signals to dynamically balance reasoning depth, achieving up to a 10‑point accuracy gain and a 35.4% reduction in token length across multiple LLM sizes and benchmarks.
The article analyzes how reward‑shaping techniques can shorten the chain‑of‑thought outputs of Qwen 30‑parameter series models by 20‑40% while preserving or slightly improving performance on AIME‑25 and out‑of‑distribution benchmarks, and it details the experimental design, strategic considerations, and practical insights behind this efficient reasoning approach.
Researchers introduce Latent Reasoning Tuning (LRT), a lightweight inference network that encodes explicit reasoning chains into fixed‑length latent vectors, eliminating thousands of decoding steps; experiments reveal substantial redundancy in traditional chains and demonstrate that LRT achieves faster, more accurate inference and outperforms existing efficient reasoning methods.
This article examines the visual token redundancy in decoder-only multimodal large language models and introduces a training-free dynamic computation reduction framework—featuring Probe-Activated Dynamic FFN, Hollow Attention, and a Layer Ranking Algorithm—that significantly lowers inference cost while preserving performance.
The COMI framework introduces a marginal information gain metric and a coarse‑to‑fine adaptive compression strategy that preserves relevance and diversity, enabling 32× text compression while boosting downstream QA performance by up to 25 points and doubling inference speed.
FaithLens is an 8‑billion‑parameter model that surpasses GPT‑4.1 and other large models on twelve hallucination‑detection benchmarks while providing high‑quality natural‑language explanations, thanks to a novel data‑synthesis pipeline, three‑dimensional filtering, and rule‑based reinforcement learning.
DeepSeek’s newly released Engram module introduces a conditional memory mechanism that leverages O(1) N‑gram lookup to create a new sparsity axis for large language models, reducing early‑layer compute, improving inference efficiency, and delivering notable performance gains across reasoning and knowledge tasks, as demonstrated by extensive experiments on 27‑billion‑parameter models.
Xiaomi's MiMo-V2-Flash, a 309B MoE model with hybrid attention and Multi‑Token Prediction acceleration, delivers top‑2 global agent benchmark scores, up to 2× faster inference, and only 2.5% of the cost of comparable closed‑source models, while being fully open‑source.
LongCat-Flash-Omni, an open‑source 560 billion‑parameter multimodal model, combines efficient Shortcut‑Connected MoE architecture with advanced perception and speech modules to deliver low‑latency real‑time audio‑video interaction and state‑of‑the‑art performance across text, image, video, and audio tasks.
At the Ant Star Innovation Journey event, the Baoling team unveiled their roadmap for trillion‑parameter models, detailing the development of Ling‑1T, Ring‑1T and multimodal Ming series, the scaling‑law‑guided architecture, training innovations, evaluation methods, and open‑source releases that aim to advance efficient, high‑performance AI.
The paper introduces VSRM, a verifiable step‑reward mechanism that penalizes ineffective reasoning steps and rewards useful ones in large language model inference, dramatically shortening output length while preserving or even improving performance across multiple benchmarks and reinforcement‑learning algorithms.
Ling-mini-2.0, an open-source 16 B MoE language model that activates only 1.4 B parameters, achieves dense-level performance with 7× efficiency, generates over 300 tokens / s, and introduces the first FP8 mixed-precision training suite, offering multiple pre-training checkpoints for the AI community.
This article introduces DistilQwen2.5, a lightweight LLM series built on Qwen2.5 that uses a novel two‑layer distillation framework, instruction‑data optimization, and parameter‑fusion techniques to achieve higher performance while drastically reducing computational cost and deployment overhead.
