arXiv · · Significant research
Researchers at MBZUAI proposed CopyGen, a lightweight, copy-constrained plugin designed to improve the exact reproduction of text spans by large language models (LLMs). CopyGen addresses the challenge of LLMs failing to precisely copy information from context by decomposing the process into distinct decisions for when to copy, how far, and when to stop, and adaptively adjusting logits. Evaluated across three LLM backbones and four benchmark tasks, CopyGen achieved a 35% relative improvement in average exact match without fine-tuning and further accelerated generation. Why it matters: This research from a leading regional AI institution provides a practical solution to enhance LLM reliability for critical applications requiring precise text reproduction, such as code generation or contract analysis.
LLM · CopyGen · Decoding · Span Reproduction · MBZUAI
Get the weekly digest
Top AI stories from the GCC region, every week.
arXiv ·
This paper introduces DetectLLM-LRR and DetectLLM-NPR, two novel zero-shot methods for detecting machine-generated text using log rank information. Experiments across three datasets and seven language models demonstrate improvements of up to 3.9 AUROC points over state-of-the-art methods. The code and data for both methods are available on Github.
MBZUAI ·
MBZUAI researchers identified 'self-referencing causal cycles' in LLM training data that can mitigate the 'reversal curse,' where LLMs struggle with information presented in reverse order. The study, to be presented at ACL, explains that the transformer architecture's unidirectional token generation causes this issue. By leveraging the repetitive nature of information in training texts, the team developed an efficient solution to improve LLM performance. Why it matters: Overcoming the reversal curse can significantly enhance LLM accuracy and reliability, especially in tasks requiring bidirectional reasoning and understanding of context.
arXiv ·
A new method is proposed to reduce the verbosity of LLMs in step-by-step reasoning by retaining moderately easy problems during Reinforcement Learning with Verifiable Rewards (RLVR) training. This approach acts as an implicit length regularizer, preventing the model from excessively increasing output length on harder problems. Experiments using Qwen3-4B-Thinking-2507 show the model achieves baseline accuracy with nearly twice shorter solutions.
MBZUAI ·
MBZUAI researchers created Open CaptchaWorld, a new benchmark to test AI agents on solving CAPTCHAs. The benchmark includes 20 modern CAPTCHA types that require perception, reasoning, and interactive actions within a browser. While humans achieve 93.3% accuracy, the best AI agent only reaches 40% on the benchmark. Why it matters: This research highlights a critical gap in current AI agent capabilities, as CAPTCHAs are gatekeepers to high-value web actions like e-commerce and secure logins.