SACHIN KUMARDART-Math: Difficulty-Aware Rejection Tuning of LLMs for better Mathematical Problem-SolvingPrevious works usually synthesize data from proprietary models to augment existing datasets, followed by instruction tuning to achieve…5d ago5d ago
SACHIN KUMARDecoupled Refusal Training for improving Safety in LLMsIn this paper [1], authors introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to…Jul 161Jul 161
SACHIN KUMARSpeculative RAG: enhancing RAG with multiple drafts generation and verificationRecent RAG advancements focus on improving retrieval outcomes through iterative LLM refinement or self-critique capabilities acquired…Jul 14Jul 14
SACHIN KUMARLookback-lens: Detect and Mitigate Hallucinations in LLMs with Attention MapsWhen summarizing articles or answering questions for a given passage, LLMs can hallucinate details and respond with inaccurate or…Jul 10Jul 10
SACHIN KUMARAutoDetect: Framework for Automated Weakness Detection in LLMs across various tasksLLMs do show superiority in accomplishing certain tasks, but still exhibit significant but subtle weaknesses, such as mistakes in…Jun 26Jun 26
SACHIN KUMARInstruction Pre-Training: using instruction-response pairs to pre-train LLMsSupervised multitask learning holds significant promise, as scaling it in the post-training stage trends towards better generalization.Jun 23Jun 23
SACHIN KUMARGraphReader: a graph based Agent to enhance long-context abilities of LLMsLong-context capabilities in LLMs helps in tackling complex and long-input tasks, but that was hindered by challenges that persists in…Jun 23Jun 23
SACHIN KUMAREvaluating Alignment and Vulnerabilities in LLMs-as-JudgesLLM-as-a-judge has emerged as an approach for evaluating large language models (LLMs), but still has many open questions about the…Jun 20Jun 20
SACHIN KUMARMeta-Reasoning Prompting : efficient system prompting method for LLMs inspired by human…Traditional in-context learning-based reasoning techniques, such as Tree-of-Thoughts, show promise but lack consistent state-of-the-art…Jun 18Jun 18
SACHIN KUMARMitigating Memorization in Generative LLMs to prevent training data leaks in responsesLLMs can memorize and repeat their training data, causing privacy and copyright risks. To mitigate memorization, authors of this paper[1]…Jun 18Jun 18