Computed T omography (CT) plays a crucial role in clinical diagnosis, but the growing demand for CT examinations has raised concerns about diagnostic errors. While Multimodal Large Language Models (MLLMs) demonstrate promising comprehension of medical knowledge, their tendency to produce inaccurate information highlights the need for rigorous validation. However, existing medical visual question answering (VQA) benchmarks primarily focus on simple visual recognition tasks, lacking clinical relevance and failing to assess expert-level knowledge. W e introduce MedErr-CT, a novel benchmark for evaluating medical MLLMs' ability to identify and correct errors in CT reports through a VQA framework. The benchmark includes six error categories--four vision-centric errors (Omission, Insertion, Direction, Size) and two lexical error types (Unit, Typo)--and is organized into three task levels: classification, detection, and correction. Using this benchmark, we quantitatively assess the performance of state-of-the-art 3D medical MLLMs, revealing substantial variation in their capabilities across different error types. Our benchmark contributes to the development of more reliable and clinically applicable MLLMs, ultimately helping reduce diagnostic errors and improve accuracy in clinical practice.

LLMs are typically trained in high-resource languages, and tasks in lower-resourced languages tend to underperform the higher-resource language counterparts for in-context learning. Despite the large body of work on prompting settings, it is still unclear how LLMs should be adapted cross-lingually specifically for in-context learning in the low-resource target languages. We perform a comprehensive study spanning five diverse target languages, three base LLMs, and seven downstream tasks spanning over 4,100 GPU training hours (9,900+ TFLOPs) across various adaptation techniques: few-shot prompting, translate-test, fine-tuning, embedding re-initialization, and instruction fine-tuning. Our results show that the few-shot prompting and translate-test settings tend to heavily outperform the gradient-based adaptation methods. To better understand this discrepancy, we design a novel metric, Valid Output Recall (VOR), and analyze model outputs to empirically attribute the degradation of these trained models to catastrophic forgetting. To the extent of our knowledge, this is the largest study done on in-context learning for low-resource languages with respect to train compute and number of adaptation techniques considered. We make all our datasets and trained models available for public use.

Traditional mental health support systems often generate responses based solely on the user's current emotion and situations, resulting in superficial interventions that fail to address deeper emotional needs. This study introduces a novel framework by integrating spiritual wisdom from the Bhagavad Gita with advanced large language model GPT-4o to enhance emotional well-being. We present the GITes (Gita Integrated Therapy for Emotional Support) dataset, which enhances the existing ExTES mental health dataset by including 10,729 spiritually guided responses generated by GPT-4o and evaluated by domain experts. We benchmark GITes against 12 state-of-the-art LLMs, including both mental health specific and general purpose models. To evaluate spiritual relevance in generated responses beyond what conventional n-gram based metrics capture, we propose a novel Spiritual Insight metric and automate assessment via an LLM as jury framework using chain-of-thought prompting. Integrating spiritual guidance into AI driven support enhances both NLP and spiritual metrics for the best performing LLM Phi3-Mini 3.2B Instruct, achieving improvements of 122.71% in ROUGE, 126.53% in METEOR, 8.15% in BERT score, 15.92% in Spiritual Insight, 18.61% in Sufficiency and 13.22% in Relevance compared to its zero-shot counterpart. While these results reflect substantial improvements across automated empathy and spirituality metrics, further validation in real world patient populations remains a necessary step. Our findings indicate a strong potential for AI systems enriched with spiritual guidance to enhance user satisfaction and perceived support outcomes. The code and dataset will be publicly available to advance further research in this emerging area.

Large language models (LLMs) often struggle with mathematical problems that require exact computation or multi-step algebraic reasoning. Tool-integrated reasoning (TIR) offers a promising solution by leveraging external tools such as code interpreters to ensure correctness, but it introduces inference-time dependencies that hinder scalability and deployment. In this work, we propose a new paradigm for distilling tool knowledge into LLMs purely through natural language. We first construct a Solver Agent that solves math problems by interleaving planning, symbolic tool calls, and reflective reasoning. Then, using a back-translation pipeline powered by multiple LLM-based agents, we convert interleaved TIR traces into natural language reasoning traces. A Translator Agent generates explanations for individual tool calls, while a Rephrase Agent merges them into a fluent and globally coherent narrative. Empirically, we show that fine-tuning a small open-source model on these synthesized traces enables it to internalize both tool knowledge and structured reasoning patterns, yielding gains on competition-level math benchmarks without requiring tool access at inference.

Retrofitting large language models (LLMs) with new behaviors typically requires full finetuning or distillation-costly steps that must be repeated for every architecture. In this work, we introduce Command-V, a backpropagation-free behavior transfer method that copies an existing residual activation adapter from a donor model and pastes its effect into a recipient model. Command-V profiles layer activations on a small prompt set, derives linear converters between corresponding layers, and applies the donor intervention in the recipient's activation space. This process does not require access to the original training data and needs minimal compute. In three case studies-safety-refusal enhancement, jailbreak facilitation, and automatic chain-of-thought reasoning--Command-V matches or exceeds the performance of direct finetuning while using orders of magnitude less compute. Our code and data are accessible at https://github.com/GithuBarry/Command-V/.

Prompt injection threatens novel applications that emerge from adapting LLMs for various user tasks. The newly developed LLM-based software applications become more ubiquitous and diverse. However, the threat of prompt injection attacks undermines the security of these systems as the mitigation and defenses against them, proposed so far, are insufficient. We investigated the capabilities of early prompt injection detection systems, focusing specifically on the detection performance of techniques implemented in various open-source solutions. These solutions are supposed to detect certain types of prompt injection attacks, including the prompt leak. In prompt leakage attacks, an attacker maliciously manipulates the LLM into outputting its system instructions, violating the system's confidentiality. Our study presents analyzes of distinct prompt leakage detection techniques, and a comparative analysis of several detection solutions, which implement those techniques. We identify the strengths and weaknesses of these techniques and elaborate on their optimal configuration and usage in high-stake deployments. In one of the first studies on existing prompt leak detection solutions, we compared the performances of LLM Guard, Vigil, and Rebuff. We concluded that the implementations of canary word checks in Vigil and Rebuff were not effective at detecting prompt leak attacks, and we proposed improvements for them. We also found an evasion weakness in Rebuff's secondary model-based technique and proposed a mitigation. Then, the result of the comparison of LLM Guard, Vigil, and Rebuff at their peak performance revealed that Vigil is optimal for cases when minimal false positive rate is required, and Rebuff is the most optimal for average needs.

Large language models (LLMs) are known to perform well on language tasks, but struggle with reasoning tasks. This paper explores the ability of LLMs to play the 2D puzzle game Baba is You, in which players manipulate rules by rearranging text blocks that define object properties. Given that this rule-manipulation relies on language abilities and reasoning, it is a compelling challenge for LLMs. Six LLMs are evaluated using different prompt types, including (1) simple, (2) rule-extended and (3) action-extended prompts. In addition, two models (Mistral, OLMo) are finetuned using textual and structural data from the game. Results show that while larger models (particularly GPT-4o) perform better in reasoning and puzzle solving, smaller unadapted models struggle to recognize game mechanics or apply rule changes. Finetuning improves the ability to analyze the game levels, but does not significantly improve solution formulation. We conclude that even for state-of-the-art and finetuned LLMs, reasoning about dynamic rule changes is difficult (specifically, understanding the use-mention distinction). The results provide insights into the applicability of LLMs to complex problem-solving tasks and highlight the suitability of games with dynamically changing rules for testing reasoning and reflection by LLMs.

Matching job titles is a highly relevant task in the computational job market domain, as it improves e.g., automatic candidate matching, career path prediction, and job market analysis. Furthermore, aligning job titles to job skills can be considered an extension to this task, with similar relevance for the same downstream tasks. In this report, we outline NLPnorth's submission to TalentCLEF 2025, which includes both of these tasks: Multilingual Job Title Matching, and Job Title-Based Skill Prediction. For both tasks we compare (fine-tuned) classification-based, (fine-tuned) contrastive-based, and prompting methods. We observe that for Task A, our prompting approach performs best with an average of 0.492 mean average precision (MAP) on test data, averaged over English, Spanish, and German. For Task B, we obtain an MAP of 0.290 on test data with our fine-tuned classification-based approach. Additionally, we made use of extra data by pulling all the language-specific titles and corresponding \emph{descriptions} from ESCO for each job and skill. Overall, we find that the largest multilingual language models perform best for both tasks. Per the provisional results and only counting the unique teams, the ranking on Task A is 5$^{\text{th}}$/20 and for Task B 3$^{\text{rd}}$/14.

The recent work by Shojaee et al. (2025), titled The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity, presents a compelling empirical finding, a reasoning cliff, where the performance of Large Reasoning Models (LRMs) collapses beyond a specific complexity threshold, which the authors posit as an intrinsic scaling limitation of Chain-of-Thought (CoT) reasoning. This commentary, while acknowledging the study's methodological rigor, contends that this conclusion is confounded by experimental artifacts. We argue that the observed failure is not evidence of a fundamental cognitive boundary, but rather a predictable outcome of system-level constraints in the static, text-only evaluation paradigm, including tool use restrictions, context window recall issues, the absence of crucial cognitive baselines, inadequate statistical reporting, and output generation limits. We reframe this performance collapse through the lens of an agentic gap, asserting that the models are not failing at reasoning, but at execution within a profoundly restrictive interface. We empirically substantiate this critique by demonstrating a striking reversal. A model, initially declaring a puzzle impossible when confined to text-only generation, now employs agentic tools to not only solve it but also master variations of complexity far beyond the reasoning cliff it previously failed to surmount. Additionally, our empirical analysis of tool-enabled models like o4-mini and GPT-4o reveals a hierarchy of agentic reasoning, from simple procedural execution to complex meta-cognitive self-correction, which has significant implications for how we define and measure machine intelligence. The illusion of thinking attributed to LRMs is less a reasoning deficit and more a consequence of an otherwise capable mind lacking the tools for action.

We propose Chain-of-Experts (CoE), a new Mixture-of-Experts (MoE) architecture that introduces sequential expert communication within each layer. Unlike traditional MoE models, where experts operate independently in parallel, CoE processes tokens iteratively across a chain of experts inside a layer. To support dynamic expert selection across iterations, CoE employs a dedicated router at each iteration step within a layer. This design allows tokens to re-evaluate and select different experts during each iteration, rather than being statically assigned. As a result, CoE introduces a flexible routing mechanism that increases the diversity of expert combinations and enriches the model's representational capacity. CoE demonstrates improved performance under fixed compute: on math reasoning tasks, it reduces validation loss from 1.20 to 1.12 compared to a standard MoE. Beyond performance, CoE offers a new scaling axis: depth through expert iteration, which complements conventional width/depth scaling. For example, using 2x iterations matches the performance of 3x expert selections (in width), while reducing memory usage by 17.6-42% relative to other scaling strategies. Our analysis reveals that CoE's benefits stem from its iterative residual structure and enhanced expert specialization empowered by iterative routing, which together unlock more expressive representations. Code is available at https://github.com/ZihanWang314/coe.