Large Language Models

Richelieu: a self-evolving LLM-based AI agent masters Diplomacy, a complex game requiring strategic planning and negotiation, without human data, by integrating self-play for continuous improvement.

Reverse the forget-retain objectives for efficient LLM unlearning!

PaRO boosts LLM training speed by up to 266% through refined model state partitioning and optimized communication.

Researchers propose Adversarial Compression Ratio (ACR) to assess LLM memorization, offering an adversarial, flexible, and computationally efficient method for monitoring data misuse and compliance.

ReST-MCTS*: A novel LLM self-training method using process reward guided tree search, outperforming existing methods by generating higher-quality reasoning traces for improved model accuracy.

New research resolves discrepancies in language model scaling laws, revealing three key factors driving the differences and improving accuracy in predicting optimal model size based on compute budget.

Boost LLM reliability by adding redundancy! This paper uses a communication theory framework to show that generating multiple LLM outputs and reranking them significantly reduces errors, even with imp…

This research unveils a compute-optimal recipe for fine-tuning language models into high-quality text embedding models, offering practical guidance and scaling laws for resource-constrained settings.

RepNoise: a novel defense against harmful fine-tuning of LLMs by removing information about harmful representations, generalizing across different harmful tasks, and maintaining LLM capabilities.

ReMoDetect leverages reward models to identify and classify LLM-generated text. By using continual preference fine-tuning and incorporating human/LLM mixed text, ReMoDetect achieves state-of-the-art p…

POAD enhances LLM agents by decomposing language agent optimization to the token level, achieving finer-grained credit assignment and improved learning efficiency and generalization.

Regularizing hidden states improves reward model generalization in RLHF for LLMs, boosting accuracy and mitigating over-optimization.

ReFT: Revolutionizing language model finetuning by directly manipulating hidden representations, achieving superior efficiency and performance compared to existing methods.

COPPER enhances LLM-based multi-agent collaboration via a self-reflection mechanism and counterfactual PPO. It improves reflection quality, alleviates credit assignment issues, and shows strong perfo…

Reference Trustable Decoding (RTD) revolutionizes large language model adaptation by offering a training-free method, enabling efficient and cost-effective task adaptation without parameter adjustment…

Cross-Layer Attention (CLA) shrinks Transformer Key-Value cache 2x, improving LLMs’ memory efficiency without accuracy loss.

RISE: Recursive Introspection teaches LLMs to iteratively improve their responses, enabling self-correction and enhanced performance on challenging reasoning tasks.

Overparameterized neural networks surprisingly recover from catastrophic interference when trained cyclically on repeated data sequences, exhibiting anticipatory knowledge reactivation.

Boosting large language model performance after knowledge editing: A new method (D4S) minimizes model damage by regulating the explosive growth of parameter layers, enabling multiple effective edits.

New surrogate loss functions for learning-to-defer achieve Bayes-consistency, realizable H-consistency, and H-consistency bounds simultaneously, resolving open questions and improving L2D performance.