Large Language Models

BoNBON alignment optimizes large language model (LLM) outputs towards human preferences using best-of-n sampling, maximizing win-rate against base models with minimal off-target impact.

BLoB: Bayesian Low-Rank Adaptation by Backpropagation enhances LLMs by jointly tuning mean and covariance of parameters during fine-tuning, improving uncertainty estimation and generalization.

BitDelta drastically shrinks fine-tuned LLMs by quantizing their weight deltas to just one bit, achieving 10x memory reduction and latency improvements without sacrificing performance.

BISCOPE: AI-generated text detection using a novel bidirectional method that outperforms existing techniques by leveraging both prediction and memorization of preceding tokens.

Bileve: a novel bi-level signature secures text provenance in LLMs against spoofing, enhancing detectability and reliability via fine-grained integrity checks and coarse-grained source tracing.

LLMs’ iterative interactions amplify subtle biases; this paper uses a Bayesian Iterated Learning framework to explain this phenomenon and offers strategies to guide LLM evolution.

Masked language models can perform in-context learning, challenging the dominance of causal models in this area.

Goldfish Loss: A novel training method for LLMs dramatically reduces memorization without impacting performance, addressing key safety, privacy, and copyright concerns.

LLM long-context ability is fundamentally limited by RoPE’s base parameter, which determines an absolute lower bound for achievable context length.

BAM! Efficiently upcycles pre-trained models into powerful Mixture-of-Experts (MoE) models, achieving state-of-the-art performance with reduced computational costs.

BAdam: A memory-efficient optimization method enabling full parameter fine-tuning of large language models using a block coordinate descent framework with Adam’s update rule, achieving comparable or s…

BackdoorAlign defends against fine-tuning-based LLM jailbreaks using a ‘backdoor trigger’ to enforce safety alignment during inference, effectively mitigating risks with minimal additional safety exam…

B’MOJO: A novel hybrid architecture for foundation models enhances transductive inference by dynamically balancing eidetic and fading memory, leading to efficient and accurate processing of long seque…

AVATAR: A novel automated framework optimizes LLM agents for effective tool usage via contrastive reasoning, significantly boosting performance on complex tasks.

AutoTimes repurposes LLMs as autoregressive time series forecasters, achieving state-of-the-art results with minimal trainable parameters and faster training/inference.

AutoSurvey automates comprehensive literature survey creation using LLMs, overcoming challenges of context limitations and knowledge constraints via a novel, efficient, and rigorously evaluated method…

AutoPSV automates process annotation for LLMs, improving reasoning by detecting confidence shifts in reasoning steps, thus efficiently enhancing model performance.

AutoMix intelligently routes queries to different-sized LLMs based on a smaller model’s self-verification, minimizing cost while maintaining performance.

LLM agents can now autonomously build environmental understanding via interactive learning, generating human-readable instruction manuals that boost task success rates.

AutoGuide: Automated generation of context-aware guidelines significantly improves LLM agent performance in unfamiliar domains.