Large Language Models

Rainbow Teaming: a novel black-box approach generates diverse adversarial prompts to enhance LLM robustness and safety, achieving over 90% attack success rate across various models.

LLM survey responses are systematically biased, often masking genuine model capabilities and leading to misleading alignment conclusions.

Researchers developed a query-based attack that generates adversarial prompts, fooling language models into producing harmful outputs with significantly higher success rates than previous methods, eff…

QuaRot: Revolutionizing 4-bit LLM inference with lossless quantization via rotation!

Weakly supervised strong models outperform weak models; this gain is precisely quantified by the strong model’s misfit error on weak labels.

QuanTA: Quantum-inspired Tensor Adaptation efficiently fine-tunes LLMs with high-rank updates, surpassing low-rank methods like LoRA for complex tasks while minimizing additional parameters.

QTIP: Ultra-high dimensional LLM quantization using trellis codes for faster, higher-quality inference.

QBB: A novel post-training quantization method for LLMs dramatically improves efficiency by replacing multiplications with summations, achieving state-of-the-art results with minimal accuracy loss.

PV-Tuning achieves new state-of-the-art in extreme LLM compression by going beyond traditional straight-through estimators (STE). This novel framework provides a more accurate and efficient fine-tunin…

Transformers excel at in-context learning (ICL), solving new tasks with just prompts. This paper provides a mathematical explanation, showing how transformers use multi-concept word semantics to achie…

RLHF’s overoptimization problem is mitigated by RPO, a novel algorithm that uses SFT loss as an implicit adversarial regularizer, ensuring efficient and effective LLM alignment.

IBProtector shields LLMs from harmful outputs via prompt compression, selectively preserving essential information using a trainable extractor.

LoPA: a novel parameter-efficient fine-tuning method matches state-of-the-art performance while requiring no server-side adapters, improving upon traditional prompt tuning.

LLMs’ in-context learning, though effective, exhibits surprisingly irregular decision boundaries, hindering generalization; this paper reveals this issue and proposes methods to improve smoothness via…

Pretrained transformers surprisingly learn low-dimensional nonlinear functions efficiently from few in-context examples, outperforming baseline algorithms.

Despite common belief, state-of-the-art preference learning algorithms for LLMs achieve surprisingly low ranking accuracy, highlighting significant flaws in current alignment techniques.

This paper presents a novel statistical framework for ranking LLMs using pairwise comparisons, accounting for the uncertainty introduced when using an LLM instead of human preferences. The framework …

Foundation model OOD performance prediction is reliably achieved via ensemble diversity, especially through random linear head initialization, enabling precise estimations without extensive OOD labels…

Pre-trained LLMs surprisingly use Fourier features to perform addition, with MLP layers approximating magnitude and attention layers handling modular arithmetic; this mechanism requires pre-training.

Researchers developed Policy Learning from tutorial Books (PLfB), a novel method that trains AI agents using knowledge from tutorial books instead of relying solely on real-world data.