Goodtriever uses retrieval-augmented methods to reduce inference latency by 43% while maintaining state-of-the-art toxicity mitigation.

Luiza Pozzobon, Beyza Ermis, Patrick Lewis et al.

CacheGen compresses KV caches with custom encoding and adaptive streaming, reducing fetch delay by 3.2-3.7× while maintaining response quality.

Yuhan Liu, Hanchen Li, Yihua Cheng et al.

This paper investigates linear decompositions of Transformer embeddings, showing high correlation with model performance at corpus level but high variability across runs and tasks.

Timothee Mickus, Raúl Vázquez

Mistral 7B employs GQA and SWA, with 7B parameters, outperforming Llama 2 13B in reasoning and efficiency.

Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch et al.

This study demonstrates that large language models (LLMs) linearly encode truth/falsehood in their internal representations, validated through transfer and causal intervention experiments, especially at scale.

Samuel Marks, Max Tegmark

Uni3D employs end-to-end pretraining of a 2D ViT, scaled to 1 billion parameters, aligning 3D point clouds with image-text features, achieving state-of-the-art results.

Junsheng Zhou, Jinsheng Wang, Baorui Ma et al.

SWE-bench evaluates LMs on real GitHub issues; Claude 2 solves only 1.96%.

Carlos E. Jimenez, John Yang, Alexander Wettig et al.

Proposes Sheared-LLaMA, a structured pruning method that compresses a 7B model to 1.3B and 2.7B parameters, achieving performance surpassing comparable models with only 3% training cost.

Mengzhou Xia, Tianyu Gao, Zhiyuan Zeng et al.

This study analyzes RLHF's impact on LLM generalization and diversity, showing improved out-of-distribution performance but reduced output variety.

Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis et al.

Proposes Few-Shot In-Context Attack (ICA) and Defense (ICD) methods, significantly influencing LLM safety with minimal demonstrations.

Zeming Wei, Yifei Wang, Ang Li et al.

CoinSeg employs inter- and intra-class contrastive representations with multi-Gaussian modeling, achieving state-of-the-art performance on Pascal VOC and ADE20K, especially in long-term incremental scenarios.

Zekang Zhang, Guangyu Gao, Jianbo Jiao et al.

Selective Context method enhances inference efficiency by compressing context, reducing context cost by 50%.

Yucheng Li, Bo Dong, Chenghua Lin et al.

FLATTEN introduces flow-guided attention, significantly improving temporal consistency in text-to-video editing.

Yuren Cong, Mengmeng Xu, Christian Simon et al.

Introduces Conformal Decision Theory with online risk calibration, ensuring provable low-risk decisions in adversarial environments; validated in robot navigation, manufacturing, and trading.

Jordan Lekeufack, Anastasios N. Angelopoulos, Andrea Bajcsy et al.

MAGVIT-v2 introduces a large-scale, shared vocabulary tokenizer with lookup-free quantization, surpassing diffusion models in image/video generation, compression, and recognition.

Lijun Yu, José Lezama, Nitesh B. Gundavarapu et al.

LLMLingua employs a hierarchical prompt compression framework, achieving up to 20x compression while maintaining near-original task performance.

Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin et al.

FREE framework significantly reduces inference latency with synchronized parallel decoding, enhancing performance.

Sangmin Bae, Jongwoo Ko, Hwanjun Song et al.

Proposes Perturbed Maximization (PM), balancing randomness and matching quality with multi-metric optimization.

Yixuan Even Xu, Steven Jecmen, Zimeng Song et al.

NOPE evaluates object hallucination in vision-language models, finding all models score below 10% accuracy on NegP datasets.

Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya et al.

Proposes a solution-generator framework using policy gradient with structured features and entropy regularization, ensuring near-optimal solutions for combinatorial problems.

Constantine Caramanis, Dimitris Fotakis, Alkis Kalavasis et al.