Natural Language Processing

AI agents developed a communication system using spatial relationships, achieving over 90% accuracy in conveying relative positions of objects within a scene.

SparseLLM globally prunes large language models efficiently by decomposing the problem into manageable subproblems, achieving significant performance improvements, especially at high sparsity.

SpaceByte: A novel byte-level decoder architecture achieving near-tokenized-model performance without tokenization!

ProRec, a novel framework, bridges the binary-source semantic gap by using a binary-source encoder-decoder model and LLMs, achieving significant improvements in zero-shot binary summarization and func…

Boosting toxicity classification robustness, this paper introduces a novel bi-level optimization framework integrating crowdsourced soft-labels and GroupDRO to enhance resistance against out-of-distri…

Open-source LLMs are vulnerable to embedding space attacks, which efficiently bypass safety mechanisms and enable data extraction, even after unlearning.

SnapKV: Slashing LLM memory usage & boosting speed via smart KV cache compression!

SMOOTHIE: Label-free LLM routing achieves up to 10% accuracy gains by using a latent variable model to estimate LLM quality without labeled data.

SMALLTOLARGE (S2L) revolutionizes large language model (LLM) fine-tuning by using a small model to summarize training loss trajectories, enabling efficient data selection for larger models.

SLTrain: Sparsity+low-rank pretraining boosts LLM efficiency by up to 73% memory reduction without performance loss!

SlowFocus significantly improves fine-grained temporal understanding in video LLMs by using mixed-frequency sampling and a novel multi-frequency attention mechanism.

SlimGPT: Achieve near-optimal LLM structured pruning via Batched Greedy Pruning and Incremental Pruning Ratio, improving efficiency without sacrificing accuracy.

Self Logits Evolution Decoding (SLED) boosts LLM factuality by up to 20% without extra data or fine-tuning!

SIRIUS: A novel correction mechanism boosts the efficiency of contextually sparse LLMs for complex reasoning tasks, achieving significant latency reduction.

SimPO: a simpler, reference-free reward algorithm significantly outperforming existing offline preference optimization methods, achieving higher accuracy and efficiency in aligning LLMs with human pre…

Simplified and generalized masked diffusion models achieve state-of-the-art results in discrete data generation, surpassing previous methods in text and image modeling.

Simple masked discrete diffusion models achieve state-of-the-art language modeling results, closing the performance gap with autoregressive methods by using a novel training recipe and a Rao-Blackwell…

SILENCE, a novel lightweight system, protects user privacy in offloaded speech understanding on resource-constrained devices by selectively masking short-term audio details without impacting long-term…

Language model editing’s limitations exposed: Scaling current methods leads to knowledge loss and compromised safety, urging research into more robust techniques.

ShiftAddLLM accelerates pretrained LLMs via post-training, multiplication-less reparameterization, achieving significant memory and energy reductions with comparable or better accuracy than existing m…