Introducing anchor-based attention masks that optimize self-attention in LLMs by controlling how tokens access information across sequences.

Explore the inner workings of decoder-only transformers in LLMs, focusing on self-attention networks and the challenges of keys/values caching during inference.

Discover how our research extends in-context learning by introducing anchor tokens for sequence compression in LLMs.

Learn how anchor-based LLMs compress sequence information to reduce memory demand and boost inference speed.

Anchor-based LLMs show 3.5x speed gains, improve model scalability, and optimize real-time inference by reducing key-value cache lengths in translation tasks.

Learn how Anchor-based LLMs with AnSAN reduce memory use by 99% and boost inference speed by 3.5x.

This analysis explores the flexibility and memory efficiency of AnLLM models in the German-to-English translation task.

The results in this section highlight how AnLLM-EP and AnLLM-AC models achieve strong accuracy across various benchmarks.

Explore the training and testing data settings for AnLLM models, including examples of tokenization and evaluation methods using HuggingFace datasets.

Explore how human-in-the-loop preferences refine reward functions in tasks like humanoid running and jumping.

Explore how In-Context Preference Learning (ICPL) progressively refined reward functions in humanoid tasks using proxy human preferences.

Discover the key environment details, task descriptions, and metrics for 9 tasks in IsaacGym, as outlined in this paper.

Learn how disagreement sampling and PrefPPO optimize reward learning in reinforcement learning.

Full Prompts and ICPL Details for study Few-shot in-context preference learning with LLMs

ICPL enhances reinforcement learning by integrating LLMs and human preferences for efficient reward function synthesis.

A. Appendix

ICPL integrates LLMs with human preferences to iteratively synthesize reward functions, offering an efficient, feedback-driven approach to RL reward design.

Explore how ICPL builds on foundational works like EUREKA to redefine reward design in reinforcement learning.

Explore how ICPL builds on foundational works like EUREKA to redefine reward design in reinforcement learning.

Discover ICPL, a novel approach that leverages Large Language Models to enhance reward learning efficiency in reinforcement learning.