--We investigate consciousness-like behaviors in Large Language Models (LLMs) using the Maze T est, challenging models to navigate mazes from a first-person perspective. After synthesizing consciousness theories into 13 essential characteristics, we evaluated 12 leading LLMs across zero-shot, one-shot, and few-shot learning scenarios. Results showed reasoning-capable LLMs consistently outperforming standard versions, with Gemini 2.0 Pro achieving 52.9% Complete Path Accuracy and DeepSeek-R1 reaching 80.5% Partial Path Accuracy . The gap between these metrics indicates LLMs struggle to maintain coherent self-models throughout solutions--a fundamental consciousness aspect. While LLMs show progress in consciousness-related behaviors through reasoning mechanisms, they lack the integrated, persistent self-awareness characteristic of consciousness. The emergence of human-like capabilities in AI has been debated since the field's inception in the 1950s [1], [2]. An early case was ELIZA [3], a chatbot simulating a therapist. Though based on pattern matching, its responses were so convincing that Weizenbaum's secretary requested privacy for a "real conversation"--showing how humans can mistakenly perceive consciousness in even the simplest AI systems.

On-device running Large Language Models (LLMs) is nowadays a critical enabler towards preserving user privacy. We observe that the attention operator falls back from the special-purpose NPU to the general-purpose CPU/GPU because of quantization sensitivity in state-of-the-art frameworks. This fallback results in a degraded user experience and increased complexity in system scheduling. To this end, this paper presents shadowAttn, a system-algorithm codesigned sparse attention module with minimal reliance on CPU/GPU by only sparsely calculating the attention on a tiny portion of tokens. The key idea is to hide the overhead of estimating the important tokens with a NPU-based pilot compute. Further, shadowAttn proposes insightful techniques such as NPU compute graph bucketing, head-wise NPU-CPU/GPU pipeline and per-head fine-grained sparsity ratio to achieve high accuracy and efficiency. shadowAttn delivers the best performance with highly limited CPU/GPU resource; it requires much less CPU/GPU resource to deliver on-par performance of SoTA frameworks.

Advanced reasoning capabilities in Large Language Models (LLMs) have caused higher hallucination prevalence; yet most mitigation work focuses on after-the-fact filtering rather than shaping the queries that trigger them. We introduce QueryBandits, a bandit framework that designs rewrite strategies to maximize a reward model, that encapsulates hallucination propensity based upon the sensitivities of 17 linguistic features of the input query-and therefore, proactively steer LLMs away from generating hallucinations. Across 13 diverse QA benchmarks and 1,050 lexically perturbed queries per dataset, our top contextual QueryBandit (Thompson Sampling) achieves an 87.5% win rate over a no-rewrite baseline and also outperforms zero-shot static prompting ("paraphrase" or "expand") by 42.6% and 60.3% respectively. Therefore, we empirically substantiate the effectiveness of QueryBandits in mitigating hallucination via the intervention that takes the form of a query rewrite. Interestingly, certain static prompting strategies, which constitute a considerable number of current query rewriting literature, have a higher cumulative regret than the no-rewrite baseline, signifying that static rewrites can worsen hallucination. Moreover, we discover that the converged per-arm regression feature weight vectors substantiate that there is no single rewrite strategy optimal for all queries. In this context, guided rewriting via exploiting semantic features with QueryBandits can induce significant shifts in output behavior through forward-pass mechanisms, bypassing the need for retraining or gradient-based adaptation.

Recent progress in reasoning-oriented Large Language Models (LLMs) has been driven by introducing Chain-of-Thought (CoT) traces, where models generate intermediate reasoning traces before producing an answer. These traces, as in DeepSeek R1, are not only used to guide inference but also serve as supervision signals for distillation into smaller models. A common but often implicit assumption is that CoT traces should be semantically meaningful and interpretable to the end user. While recent research questions the need for semantic nature of these traces, in this paper, we ask: ``\textit{Must CoT reasoning traces be interpretable to enhance LLM task performance?}" We investigate this question in the Open Book Question-Answering domain by supervised fine-tuning LLaMA and Qwen models on four types of reasoning traces: (1) DeepSeek R1 traces, (2) LLM-generated summaries of R1 traces, (3) LLM-generated post-hoc explanations of R1 traces, and (4) algorithmically generated verifiably correct traces. To quantify the trade-off between interpretability and performance, we further conduct a human-subject study with 100 participants rating the interpretability of each trace type. Our results reveal a striking mismatch: while fine-tuning on R1 traces yields the strongest performance, participants judged these traces to be the least interpretable. These findings suggest that it is useful to decouple intermediate tokens from end user interpretability.

The emergence of AI-driven web automation through Large Language Models (LLMs) offers unprecedented opportunities for optimizing digital workflows. However, deploying such systems within industry's real-world environments presents four core challenges: (1) ensuring consistent execution, (2) accurately identifying critical HTML elements, (3) meeting human-like accuracy in order to automate operations at scale and (4) the lack of comprehensive benchmarking data on internal web applications. Existing solutions are primarily tailored for well-designed, consumer-facing websites (e.g., Amazon.com, Apple.com) and fall short in addressing the complexity of poorly-designed internal web interfaces. To address these limitations, we present Cybernaut, a novel framework to ensure high execution consistency in web automation agents designed for robust enterprise use. Our contributions are threefold: (1) a Standard Operating Procedure (SOP) generator that converts user demonstrations into reliable automation instructions for linear browsing tasks, (2) a high-precision HTML DOM element recognition system tailored for the challenge of complex web interfaces, and (3) a quantitative metric to assess execution consistency. The empirical evaluation on our internal benchmark demonstrates that using our framework enables a 23.2% improvement (from 72% to 88.68%) in task execution success rate over the browser_use. Cybernaut identifies consistent execution patterns with 84.7% accuracy, enabling reliable confidence assessment and adaptive guidance during task execution in real-world systems. These results highlight Cybernaut's effectiveness in enterprise-scale web automation and lay a foundation for future advancements in web automation.

This paper investigates the integration of cognitive agents powered by Large Language Models (LLMs) within the Scaled Agile Framework (SAFe) to reinforce software project management. By deploying virtual agents in simulated software environments, this study explores their potential to fulfill fundamental roles in IT project development, thereby optimizing project outcomes through intelligent automation. Particular emphasis is placed on the adaptability of these agents to Agile methodologies and their transformative impact on decision-making, problem-solving, and collaboration dynamics. The research leverages the CogniSim ecosystem, a platform designed to simulate real-world software engineering challenges, such as aligning technical capabilities with business objectives, managing interdependencies, and maintaining project agility. Through iterative simulations, cognitive agents demonstrate advanced capabilities in task delegation, inter-agent communication, and project lifecycle management. By employing natural language processing to facilitate meaningful dialogues, these agents emulate human roles and improve the efficiency and precision of Agile practices. Key findings from this investigation highlight the ability of LLM-powered cognitive agents to deliver measurable improvements in various metrics, including task completion times, quality of deliverables, and communication coherence. These agents exhibit scalability and adaptability, ensuring their applicability across diverse and complex project environments. This study underscores the potential of integrating LLM-powered agents into Agile project management frameworks as a means of advancing software engineering practices. This integration not only refines the execution of project management tasks but also sets the stage for a paradigm shift in how teams collaborate and address emerging challenges.

Many existing studies have achieved significant improvements in the reasoning capabilities of large language models (LLMs) through reinforcement learning with verifiable rewards (RLVR), while the enhancement of reasoning abilities in small language models (SLMs) has not yet been sufficiently explored. Combining distilled data from larger models with RLVR on small models themselves is a natural approach, but it still faces various challenges and issues. Therefore, we propose \textit{\underline{R}}ecall-\textit{\underline{E}}xtend \textit{\underline{D}}ynamics(RED): Enhancing Small Language Models through Controlled Exploration and Refined Offline Integration. In this paper, we explore the perspective of varying exploration spaces, balancing offline distillation with online reinforcement learning. Simultaneously, we specifically design and optimize for the insertion problem within offline data. By monitoring the ratio of entropy changes in the model concerning offline and online data, we regulate the weight of offline-SFT, thereby addressing the issues of insufficient exploration space in small models and the redundancy and complexity during the distillation process. Furthermore, to tackle the distribution discrepancies between offline data and the current policy, we design a sample-accuracy-based policy shift mechanism that dynamically chooses between imitating offline distilled data and learning from its own policy.

Transformer architecture gradually dominates the LLM field. Recent advances in training optimization for Transformer-based large language models (LLMs) primarily focus on architectural modifications or optimizer adjustments. However, these approaches lack systematic optimization of weight patterns during training. Weight pattern refers to the distribution and relative magnitudes of weight parameters in a neural network. To address this issue, we propose a Weight Scaling method called WISCA to enhance training efficiency and model quality by strategically improving neural network weight patterns without changing network structures. By rescaling weights while preserving model outputs, WISCA indirectly optimizes the model's training trajectory. Experiments demonstrate that WISCA significantly improves convergence quality (measured by generalization capability and loss reduction), particularly in LLMs with Grouped Query Attention (GQA) architectures and LoRA fine-tuning tasks. Empirical results show 5.6% average improvement on zero-shot validation tasks and 2.12% average reduction in training perplexity across multiple architectures.

Medical report interpretation plays a crucial role in healthcare, enabling both patient-facing explanations and effective information flow across clinical systems. While recent vision-language models (VLMs) and large language models (LLMs) have demonstrated general document understanding capabilities, there remains a lack of standardized benchmarks to assess structured interpretation quality in medical reports. We introduce MedRepBench, a comprehensive benchmark built from 1,900 de-identified real-world Chinese medical reports spanning diverse departments, patient demographics, and acquisition formats. The benchmark is designed primarily to evaluate end-to-end VLMs for structured medical report understanding. To enable controlled comparisons, we also include a text-only evaluation setting using high-quality OCR outputs combined with LLMs, allowing us to estimate the upper-bound performance when character recognition errors are minimized. Our evaluation framework supports two complementary protocols: (1) an objective evaluation measuring field-level recall of structured clinical items, and (2) an automated subjective evaluation using a powerful LLM as a scoring agent to assess factuality, interpretability, and reasoning quality. Based on the objective metric, we further design a reward function and apply Group Relative Policy Optimization (GRPO) to improve a mid-scale VLM, achieving up to 6% recall gain. We also observe that the OCR+LLM pipeline, despite strong performance, suffers from layout-blindness and latency issues, motivating further progress toward robust, fully vision-based report understanding. We will release the dataset and evaluation toolkit upon acceptance.

Artificial Intelligence (AI) is increasingly used in hiring, with large language models (LLMs) having the potential to influence or even make hiring decisions. However, this raises pressing concerns about bias, fairness, and trust, particularly across diverse cultural contexts. Despite their growing role, few studies have systematically examined the potential biases in AI-driven hiring evaluation across cultures. In this study, we conduct a systematic analysis of how LLMs assess job interviews across cultural and identity dimensions. Using two datasets of interview transcripts, 100 from UK and 100 from Indian job seekers, we first examine cross-cultural differences in LLM-generated scores for hirability and related traits. Indian transcripts receive consistently lower scores than UK transcripts, even when they were anonymized, with disparities linked to linguistic features such as sentence complexity and lexical diversity. We then perform controlled identity substitutions (varying names by gender, caste, and region) within the Indian dataset to test for name-based bias. These substitutions do not yield statistically significant effects, indicating that names alone, when isolated from other contextual signals, may not influence LLM evaluations. Our findings underscore the importance of evaluating both linguistic and social dimensions in LLM-driven evaluations and highlight the need for culturally sensitive design and accountability in AI-assisted hiring.