🏢 AIRI

LLMs’ reasoning is decoded via sparse autoencoders, revealing key features that, when steered, enhance performance. First mechanistic account of reasoning in LLMs!

Efficient image representation via adaptive token reduction.

LLMs use punctuation in context memory, surprisingly boosting performance by using seemingly trivial tokens.

Packing new knowledge into LoRA adapters can harm LLMs! A delicate balance is needed to prevent performance decline.

LLMs can losslessly compress 1568 tokens into a single vector, surpassing prior methods by two orders of magnitude.

MIKASA, a new benchmark for memory-intensive reinforcement learning, provides a unified framework for evaluating memory capabilities in diverse scenarios, including complex robotic manipulation tasks.

SynthDetoxM generates high-quality multilingual parallel data for text detoxification using LLMs, outperforming existing datasets and models in few-shot settings.

SRMT: Shared Recurrent Memory Transformer boosts multi-agent coordination by implicitly sharing information via a global memory, significantly outperforming baselines in complex pathfinding tasks.

3DGraphLLM boosts 3D scene understanding by cleverly merging semantic graphs and LLMs, enabling more accurate scene descriptions and outperforming existing methods.