Large Language Models

CRITICEVAL: A new benchmark reliably evaluates LLMs’ ability to identify and correct flaws in their responses, addressing limitations of existing methods by offering comprehensive and reliable evaluat…

LLM-based text embeddings are powerful but lack interpretability. This paper introduces QA-Emb, a novel method that uses an LLM to answer yes/no questions about a text, thereby producing an interpreta…

CorDA: Context-oriented weight decomposition enhances large language model fine-tuning by integrating task context, improving efficiency and mitigating catastrophic forgetting.

LLMs struggle to cooperate sustainably; GOVSIM reveals this, showing communication and ‘universalization’ reasoning improve outcomes.

Context-Aware Testing (CAT) revolutionizes ML model testing by using contextual information to identify relevant failures, surpassing traditional data-only methods.

ConStat: Exposing hidden LLM contamination!

LLMs surprisingly infer censored knowledge from implicit training data hints, posing safety challenges.

LLMs regulate uncertainty via specialized ’entropy’ and ’token frequency’ neurons, impacting prediction confidence without directly altering logits.

CALDERA: a new post-training LLM compression algorithm achieving state-of-the-art zero-shot performance using low-rank, low-precision decomposition.

LLM-directed compositional 3D-aware video generation (C3V) achieves high-fidelity video generation with diverse motion and flexible concept control by decomposing prompts, generating 3D concepts, and …

MINITRON: Efficiently creating smaller, high-performing LLMs via pruning & distillation, slashing training costs by up to 40x!

CoMERA achieves 2-3x faster AI model training via rank-adaptive tensor optimization, significantly improving both computing and memory efficiency.

COLD framework rigorously evaluates LLMs’ causal reasoning in everyday scenarios using 9 million causal queries derived from human-generated scripts of daily activities.

CORY: a novel multi-agent RL framework boosts LLM fine-tuning!

New program synthesis method, REX, leverages Thompson Sampling to balance exploration and exploitation in iterative LLM code refinement, solving more problems with fewer model calls.

Language models struggle to learn facts; this study reveals they prioritize word co-occurrence over true factual associations, proposing new training strategies for improved factual knowledge generali…

CLUES: Collaborative learning selects high-quality private data for LLM fine-tuning via training dynamics, significantly boosting performance in diverse domains.

CherryQ, a novel quantization method, leverages parameter heterogeneity in LLMs to achieve superior performance by selectively quantizing less critical parameters while preserving essential ones.

Chat-Scene: Bridging 3D scenes and LLMs using object identifiers for efficient, object-level interaction and improved scene comprehension.

LLMs can reason effectively without prompting by simply adjusting the decoding process to reveal inherent chain-of-thought paths.