Large Language Models

Boost LLM performance with INSTRUCTION MODELLING: a simple yet effective instruction tuning method that improves model outputs by over 100% in some cases by applying loss to both instructions and outp…

Instance-adaptive prompting significantly improves zero-shot Chain-of-Thought reasoning in LLMs by dynamically selecting prompts tailored to each instance, leading to consistent performance gains acro…

Transformer model initialization dramatically affects whether it reasons or memorizes, impacting performance on compositional tasks.

InfoRE: A novel method improving large language models’ reasoning by reorganizing information to highlight logical relationships, resulting in a 4% average accuracy boost across various tasks.

InfoRM tackles reward hacking in RLHF using an information-theoretic approach, enhancing generalizability and enabling overoptimization detection.

InfLLM: Training-free long-context extrapolation for LLMs via efficient context memory.

Injecting watermarks into LLM outputs while speeding up generation is impossible; this paper proves this trade-off and offers methods prioritizing either watermark strength or speed.

Restricted models often outperform full-featured models when training data is limited. This paper introduces Induced Model Matching (IMM), a novel technique that uses a restricted model as a guide to…

INDICT, a novel framework, empowers LLMs with internal dialogues of critiques to enhance code generation, prioritizing both safety and helpfulness, resulting in +10% absolute improvement across variou…

Transformers learn contextual information for generalization to unseen examples and tasks, even with limited training data, converging linearly to a global minimum.

This paper introduces inner and momentum optimization to enhance the state vector for in-context learning, improving performance and scalability in LLMs.

Linear Transformer Blocks (LTBs) achieve near-optimal in-context learning (ICL) for linear regression by effectively implementing one-step gradient descent with learnable initialization, a significant…

Gated Sparse Autoencoders (GSAEs) achieve Pareto improvement over baseline SAEs for unsupervised feature discovery in language models, resolving the shrinkage bias of L1 penalty by separating feature …

Context-aware preference modeling improves language model alignment by resolving ambiguity through a two-step process: context selection followed by context-specific preference evaluation. The approa…

Researchers developed novel methods to improve the generation of adversarial examples against safety-aligned LLMs, achieving significantly higher attack success rates compared to existing techniques.

Improved few-shot jailbreaking techniques efficiently circumvent aligned language models and their defenses, achieving high success rates even against advanced protection methods.

Boosting deep learning generalization, this work unveils SAM’s implicit regularization using ‘balancedness’, a new metric. A resource-efficient variant, BAR, achieves 95% computational savings with i…

Researchers reveal implicit optimization biases in next-token prediction for language models, showing how gradient descent selects solutions based on data sparsity and a novel margin concept, impactin…

This study presents a novel Inverse Reinforcement Learning (IRL) approach for fine-tuning large language models, offering improved performance and generation diversity compared to standard methods.

IDGen synthesizes LLM evaluation prompts using Item Discrimination theory, creating a more challenging and discriminative dataset than previous methods.