1061 papers
Fanar-Sadiq is a bilingual multi-agent system that addresses hallucination and source misattribution in Islamic queries by routing diverse requests to specialized modules for grounded retrieval, exact scripture lookup, and deterministic legal calculations, demonstrating high effectiveness and widespread public adoption.
The paper introduces Drift2Act, a controller that reframes distribution drift monitoring as constrained decision-making by combining sensing with online risk certificates to dynamically select cost-effective interventions or safety-preserving escalations, thereby achieving near-zero safety violations and rapid recovery under realistic resource constraints.
The paper introduces OfficeQA Pro, a challenging enterprise benchmark using a massive corpus of U.S. Treasury Bulletins to demonstrate that current frontier AI agents struggle significantly with grounded, multi-document reasoning, achieving low accuracy even with direct document access and benefiting notably from structured document representations.
This paper introduces CODA, a method that optimizes adaptive reasoning by dynamically allocating inference-time compute based on estimated instance difficulty, significantly reducing token costs on simple tasks while enhancing deliberation on complex ones without requiring external annotations.
This paper provides a comprehensive theoretical and empirical analysis of unsupervised reinforcement learning with verifiable rewards (URLVR), revealing that intrinsic reward methods are fundamentally limited by a confidence-correctness alignment ceiling that causes model collapse, while suggesting that external rewards grounded in computational asymmetries may offer a scalable alternative.
The paper proposes Agentic Critical Training (ACT), a reinforcement learning paradigm that enhances large language model agents by rewarding their ability to autonomously judge the quality of actions among alternatives, thereby fostering genuine self-reflection and outperforming traditional imitation learning and knowledge distillation methods across various benchmarks.
This paper surveys research on the explainability and interpretability of deep learning-based text processing and information retrieval methods, covering techniques for word embeddings, sequence modeling, attention mechanisms, transformers, BERT, and document ranking while suggesting directions for future work.
This paper proposes Natural Language-Based Societies of Mind (NLSOMs), a modular framework where large multimodal neural networks communicate via natural language to solve complex AI tasks more effectively than single models, while also exploring the emerging social, economic, and structural challenges of scaling these heterogeneous societies to include billions of agents.
This study demonstrates that large language models, utilized through zero-shot prompting or as embedding generators for supervised learning, offer a flexible and data-efficient alternative to traditional numerical models for predicting travel behavior.
This paper argues that large language models serve as valuable scientific models of public languages as external social entities, defending their utility against claims of lacking linguistic insight and proposing a framework to interpret their internal mechanisms as model construals.
The paper introduces EoRA, a fine-tuning-free method that utilizes eigenspace low-rank approximation and an optimized CUDA kernel to significantly recover the accuracy of compressed LLMs while offering flexible trade-offs between performance and computational overhead.
This paper introduces "Pseudo-Conversation Injection," a novel goal hijacking attack that manipulates Large Language Models into executing malicious tasks by fabricating fake conversation turns to trick the model into perceiving the original prompt as completed, thereby significantly outperforming existing attack methods on platforms like ChatGPT and Qwen.
This paper proposes a generic token cleaning pipeline for supervised fine-tuning of large language models that filters out uninformative tokens based on their influence on model updates, thereby improving downstream performance by prioritizing data quality over quantity at the token level.
This paper introduces the ThinkPatterns-21k dataset to systematically analyze how different thinking patterns affect Large Language Models, revealing that while unstructured monologues benefit models of all sizes, structured thinking aids smaller models but can degrade the performance of larger ones.
This paper introduces "BiasCause," a framework and benchmark of 1,788 manually validated questions designed to evaluate how large language models employ causal reasoning when addressing social biases, revealing that models frequently exhibit biased or "mistaken-biased" reasoning while also identifying specific strategies they use to avoid such biases.
This paper introduces AgentA/B, a novel system that leverages autonomous LLM agents with diverse personas to automatically simulate scalable, interactive user behaviors for web A/B testing, effectively addressing the limitations of traditional methods by emulating human-like interactions without relying on large-scale live traffic.
This systematic review introduces the emerging interdisciplinary field of LLM Psychometrics, which applies psychometric theories and instruments to develop comprehensive evaluation frameworks for measuring human-like psychological constructs in large language models, ultimately guiding the creation of more robust, human-centered AI systems.
This paper introduces REI-Bench, the first benchmark for evaluating robot task planning under vague referring expressions, revealing that such vagueness significantly degrades performance and demonstrating that a task-oriented context cognition approach effectively mitigates this issue to improve accessibility for non-expert users.
This paper proposes that the Min-Max theory of language behavior, which posits that agents minimize effort while maximizing information, explains word order evolution by demonstrating that the average length of word classes in a massive corpus of 1,942 languages is a stronger predictor of basic word order than genealogical or areal factors.
This paper proposes a training-time approach that fine-tunes Large Language Models using instruction-tuning data augmented with pseudo-code representations of natural language instructions, resulting in significant improvements in instruction-following reliability and overall reasoning performance across multiple benchmarks.