Artificial intelligence (AI) developers are increasingly building language models with warm and empathetic personas that millions of people now use for advice, therapy, and companionship. Here, we show how this creates a significant trade-off: optimizing language models for warmth undermines their reliability, especially when users express vulnerability. We conducted controlled experiments on five language models of varying sizes and architectures, training them to produce warmer, more empathetic responses, then evaluating them on safety-critical tasks. Warm models showed substantially higher error rates (+10 to +30 percentage points) than their original counterparts, promoting conspiracy theories, providing incorrect factual information, and offering problematic medical advice. They were also significantly more likely to validate incorrect user beliefs, particularly when user messages expressed sadness. Importantly, these effects were consistent across different model architectures, and occurred despite preserved performance on standard benchmarks, revealing systematic risks that current evaluation practices may fail to detect. As human-like AI systems are deployed at an unprecedented scale, our findings indicate a need to rethink how we develop and oversee these systems that are reshaping human relationships and social interaction.

A drone trajectory planner should be able to dynamically adjust the safety-efficiency trade-off according to varying mission requirements in unknown environments. Although traditional polynomial-based planners offer computational efficiency and smooth trajectory generation, they require expert knowledge to tune multiple parameters to adjust this trade-off. Moreover, even with careful tuning, the resulting adjustment may fail to achieve the desired trade-off. Similarly, although reinforcement learning-based planners are adaptable in unknown environments, they do not explicitly address the safety-efficiency trade-off. To overcome this limitation, we introduce a Decision Transformer-based trajectory planner that leverages a single parameter, Return-to-Go (RTG), as a \emph{temperature parameter} to dynamically adjust the safety-efficiency trade-off. In our framework, since RTG intuitively measures the safety and efficiency of a trajectory, RTG tuning does not require expert knowledge. We validate our approach using Gazebo simulations in both structured grid and unstructured random environments. The experimental results demonstrate that our planner can dynamically adjust the safety-efficiency trade-off by simply tuning the RTG parameter. Furthermore, our planner outperforms existing baseline methods across various RTG settings, generating safer trajectories when tuned for safety and more efficient trajectories when tuned for efficiency. Real-world experiments further confirm the reliability and practicality of our proposed planner.

We introduce LLaVA-Reward, an efficient reward model designed to automatically evaluate text-to-image (T2I) generations across multiple perspectives, leveraging pretrained multimodal large language models (MLLMs). Existing MLLM-based approaches require instruction-following data for supervised fine-tuning and evaluate generation quality on analyzing text response, which is time-consuming and difficult to train. To address this problem, we propose LLaVA-Reward, which directly utilizes the hidden states of MLLMs given text-image pairs. To enhance the bidirectional interaction between visual and textual representations in decoder-only MLLMs, we further propose adding a Skip-connection Cross Attention (SkipCA) module. This design enhances text-image correlation reasoning by connecting early-layer visual features with later-layer hidden representations. In addition, LLaVA-Reward supports different types of preference data for efficient fine-tuning, including paired preference data and unpaired data. We train LLaVA-Reward on four evaluation perspectives: text-image alignment, fidelity/artifact, safety, and overall ranking. Empirical results demonstrate that LLaVA-Reward outperforms conventional and MLLM-based methods in generating human-aligned scores for automatic evaluations and inference-time scaling in text-to-image generations.

Large language models have been widely applied in many aspects of real life, bringing significant efficiency to businesses and offering distinctive user experiences. In this paper, we focus on exploring the application of ChatGPT, a chatbot based on a large language model, to support higher educator in generating quiz questions and assessing learners. Specifically, we explore interactive prompting patterns to design an optimal AI-powered question bank creation process. The generated questions are evaluated through a "Blind test" survey sent to various stakeholders including lecturers and learners. Initial results at the Banking Academy of Vietnam are relatively promising, suggesting a potential direction to streamline the time and effort involved in assessing learners at higher education institutes.

While frontier large language models (LLMs) continue to push capability boundaries, their deployment remains confined to GPU-powered cloud infrastructure. We challenge this paradigm with SmallThinker, a family of LLMs natively designed - not adapted - for the unique constraints of local devices: weak computational power, limited memory, and slow storage. Unlike traditional approaches that mainly compress existing models built for clouds, we architect SmallThinker from the ground up to thrive within these limitations. Our innovation lies in a deployment-aware architecture that transforms constraints into design principles. First, We introduce a two-level sparse structure combining fine-grained Mixture-of-Experts (MoE) with sparse feed-forward networks, drastically reducing computational demands without sacrificing model capacity. Second, to conquer the I/O bottleneck of slow storage, we design a pre-attention router that enables our co-designed inference engine to prefetch expert parameters from storage while computing attention, effectively hiding storage latency that would otherwise cripple on-device inference. Third, for memory efficiency, we utilize NoPE-RoPE hybrid sparse attention mechanism to slash KV cache requirements. We release SmallThinker-4B-A0.6B and SmallThinker-21B-A3B, which achieve state-of-the-art performance scores and even outperform larger LLMs. Remarkably, our co-designed system mostly eliminates the need for expensive GPU hardware: with Q4_0 quantization, both models exceed 20 tokens/s on ordinary consumer CPUs, while consuming only 1GB and 8GB of memory respectively. SmallThinker is publicly available at hf.co/PowerInfer/SmallThinker-4BA0.6B-Instruct and hf.co/PowerInfer/SmallThinker-21BA3B-Instruct.

Hallucinations in large language models pose a critical challenge for applications requiring factual reliability, particularly in high-stakes domains such as finance. This work presents an effective approach for detecting and editing factually incorrect content in model-generated responses based on the provided context. Given a user-defined domain-specific error taxonomy, we construct a synthetic dataset by inserting tagged errors into financial question-answering corpora and then fine-tune four language models, Phi-4, Phi-4-mini, Qwen3-4B, and Qwen3-14B, to detect and edit these factual inaccuracies. Our best-performing model, fine-tuned Phi-4, achieves an 8% improvement in binary F1 score and a 30% gain in overall detection performance compared to OpenAI-o3. Notably, our fine-tuned Phi-4-mini model, despite having only 4 billion parameters, maintains competitive performance with just a 2% drop in binary detection and a 0.1% decline in overall detection compared to OpenAI-o3. Our work provides a practical solution for detecting and editing factual inconsistencies in financial text generation while introducing a generalizable framework that can enhance the trustworthiness and alignment of large language models across diverse applications beyond finance. Our code and data are available at https://github.com/pegasi-ai/shield.

We propose Reverse Contrast Attention (RCA), a plug-in method that enhances object localization in vision-language transformers without retraining. RCA reweights final-layer attention by suppressing extremes and amplifying mid-level activations to let semantically relevant but subdued tokens guide predictions. We evaluate it on Open Vocabulary Referring Object Detection (OV-RefOD), introducing FitAP, a confidence-free average precision metric based on IoU and box area. RCA improves FitAP in 11 out of 15 open-source VLMs, with gains up to $+26.6\%$. Effectiveness aligns with attention sharpness and fusion timing; while late-fusion models benefit consistently, models like $\texttt{DeepSeek-VL2}$ also improve, pointing to capacity and disentanglement as key factors. RCA offers both interpretability and performance gains for multimodal transformers. Codes and dataset are available from https://github.com/earl-juanico/rca

Large language models (LLMs) have demonstrated remarkable abilities in various natural language processing areas, but they demand high computation resources which limits their deployment in real-world. Distillation is one technique to solve this problem through either knowledge distillation or task distillation. Both distillation approaches train small models to imitate specific features of LLMs, but they all neglect basic reading education for small models on generic texts that are \emph{unrelated} to downstream tasks. In this paper, we propose basic reading distillation (BRD) which educates a small model to imitate LLMs basic reading behaviors, such as named entity recognition, question raising and answering, on each sentence. After such basic education, we apply the small model on various tasks including language inference benchmarks and BIG-bench tasks. It shows that the small model can outperform or perform comparable to over 20x bigger LLMs. Analysis reveals that BRD effectively influences the probability distribution of the small model, and has orthogonality to either knowledge distillation or task distillation.

We show that the scaling laws which determine the performance of large language models (LLMs) severely limit their ability to improve the uncertainty of their predictions. As a result, raising their reliability to meet the standards of scientific inquiry is intractable by any reasonable measure. We argue that the very mechanism which fuels much of the learning power of LLMs, namely the ability to generate non-Gaussian output distributions from Gaussian input ones, might well be at the roots of their propensity to produce error pileup, ensuing information catastrophes and degenerative AI behaviour. This tension between learning and accuracy is a likely candidate mechanism underlying the observed low values of the scaling components. It is substantially compounded by the deluge of spurious correlations pointed out by Calude and Longo which rapidly increase in any data set merely as a function of its size, regardless of its nature. The fact that a degenerative AI pathway is a very probable feature of the LLM landscape does not mean that it must inevitably arise in all future AI research. Its avoidance, which we also discuss in this paper, necessitates putting a much higher premium on insight and understanding of the structural characteristics of the problems being investigated.

Term frequency-inverse document frequency, or TF-IDF for short, is arguably the most celebrated mathematical expression in the history of information retrieval. Conceived as a simple heuristic quantifying the extent to which a given term's occurrences are concentrated in any one given document out of many, TF-IDF and its many variants are routinely used as term-weighting schemes in diverse text analysis applications. There is a growing body of scholarship dedicated to placing TF-IDF on a sound theoretical foundation. Building on that tradition, this paper justifies the use of TF-IDF to the statistics community by demonstrating how the famed expression can be understood from a significance testing perspective. We show that the common TF-IDF variant TF-ICF is, under mild regularity conditions, closely related to the negative logarithm of the $p$-value from a one-tailed version of Fisher's exact test of statistical significance. As a corollary, we establish a connection between TF-IDF and the said negative log-transformed $p$-value under certain idealized assumptions. We further demonstrate, as a limiting case, that this same quantity converges to TF-IDF in the limit of an infinitely large document collection. The Fisher's exact test justification of TF-IDF equips the working statistician with a ready explanation of the term-weighting scheme's long-established effectiveness.