The automatic evaluation of Language Model (LM) responses is a critical piece in the development of benchmarks and metrics, both for model training and quality assessment of production model endpoints. The current approaches to response classification relies on methods that are too expensive (i.e. LLM-as-a-Judge) or that are far from real-world conditions (string-matching, logprob). In this paper, a structure-free evaluation method is presented. The method makes use of semantic embedding distances to match target candidates with arbitrary LM-generated text, resulting in a robust classification of the response at a relatively low compute cost (embedding models of less than $10B$ parameters). The results show a regression score of ~0.97 and an accuracy of ~96% against human annotators, tested over 3 data sets and 3 different LM architectures.

Since the release of Deepseek-R1, reinforcement learning with verifiable rewards (RL VR) has become a central approach for training large language models (LLMs) on reasoning tasks. Recent work has largely focused on modifying loss functions to make RL VR more efficient and effective. In this paper, motivated by studies of overthinking in LLMs, we propose Length-aware Sampling for Policy Optimization (LSPO), a novel meta-RL VR algorithm that dynamically selects training data at each step based on the average response length. We evaluate LSPO across multiple base models and datasets, demonstrating that it consistently improves learning effectiveness. In addition, we conduct a detailed ablation study to examine alternative ways of incorporating length signals into dynamic sampling, offering further insights and highlighting promising directions for future research. Since the release of ChatGPT, large language models (LLMs) have rapidly evolved beyond traditional natural language tasks such as summarization and question answering, extending into broader domains of reasoning and problem solving Chen et al. (2023); Y ao et al. (2023); Chen et al. (2024a;b); Jaech et al. (2024); Huang & Y ang (2025). A growing body of research has therefore focused on how to make LLMs more effective and efficient on reasoning-intensive tasks. While a prominent line of work have investigated how to improve test-time performance without retraining a model from scratch Wang et al. (2022); Chen et al. (2023); Madaan et al. (2023); Gou et al. (2023); Guan et al. (2025); Chen et al. (2025), more recently, following the release of Deepseek-R1 DeepSeek-AI et al. (2025), reinforcement learning with verifiable rewards (RL VR) has emerged as a powerful and promising paradigm for post-training LLMs, yielding significant gains in reasoning ability. While Deepseek-R1 adopted GRPO Shao et al. (2024), subsequent research has proposed a variety of alternative algorithms to further enhance reasoning capabilities. In general, efforts to improve RL VR training have explored several dimensions, including alternative loss functions, improved datasets, and better hyperparameter tuning.

Data-efficient learning aims to eliminate redundancy in large training datasets by training models on smaller subsets of the most informative examples. While data selection has been extensively explored for vision models and large language models (LLMs), it remains underexplored for Large Vision-Language Models (L VLMs). Notably, none of existing methods can outperform random selection at different subset sizes. In this work, we propose the first principled method for data-efficient instruction tuning of L VLMs. We prove that examples with similar cross-modal attention matrices during instruction tuning have similar gradients. Thus, they influence model parameters in a similar manner and convey the same information to the model during training. Building on this insight, we propose XMAS, which clusters examples based on the trajectories of the top singular values of their attention matrices obtained from fine-tuning a small proxy L VLM. By sampling a balanced subset from these clusters, XMAS effectively removes redundancy in large-scale L VLM training data. Extensive experiments show that XMAS can discard 50% of the LLaV A-665k dataset and 85% of the Vision-Flan dataset while fully preserving performance of LLaV A-1.5-7B on 10 downstream benchmarks and speeding up its training by 1.2 . This is 30% more data reduction compared to the best baseline for LLaV A-665k. The project's website can be found at Large Vision-Language Models (L VLMs) have demonstrated impressive capabilities in understanding and reasoning over multimodal inputs (Liu et al., 2023; 2024a; OpenAI, 2023).

Although birthed in the era of teletypes, the command line shell survived the graphical interface revolution of the 1980's and lives on in modern desktop operating systems. The command line provides access to powerful functionality not otherwise exposed on the computer, but requires users to recall textual syntax and carefully scour documentation. In contrast, graphical interfaces let users organically discover and invoke possible actions through widgets and menus. To better expose the power of the command line, we demonstrate a mechanism for automatically creating graphical interfaces for command line tools by translating their documentation (in the form of man pages) into interface specifications via AI. Using these specifications, our user-facing system, called GUIde, presents the command options to the user graphically. We evaluate the generated interfaces on a corpus of commands to show to what degree GUIde offers thorough graphical interfaces for users' real-world command line tasks.

This paper investigates the impact of the adoption of generative AI on financial stability. We conduct laboratory-style experiments using large language models to replicate classic studies on herd behavior in trading decisions. Our results show that AI agents make more rational decisions than humans, relying predominantly on private information over market trends. Increased reliance on AI-powered trading advice could therefore potentially lead to fewer asset price bubbles arising from animal spirits that trade by following the herd. However, exploring variations in the experimental settings reveals that AI agents can be induced to herd optimally when explicitly guided to make profit-maximizing decisions. While optimal herding improves market discipline, this behavior still carries potential implications for financial stability. In other experimental variations, we show that AI agents are not purely algorithmic, but have inherited some elements of human conditioning and bias.

Worldwide visual geo-localization seeks to determine the geographic location of an image anywhere on Earth using only its visual content. Learned representations of geography for visual geo-localization remain an active research topic despite much progress. We formulate geo-localization as aligning the visual representation of the query image with a learned geographic representation. Our novel geographic representation explicitly models the world as a hierarchy of geographic embeddings. Additionally, we introduce an approach to efficiently fuse the appearance features of the query image with its semantic segmentation map, forming a robust visual representation. Our main experiments demonstrate improved all-time bests in 22 out of 25 metrics measured across five benchmark datasets compared to prior state-of-the-art (SOTA) methods and recent Large Vision-Language Models (LVLMs). Additional ablation studies support the claim that these gains are primarily driven by the combination of geographic and visual representations.

Recent advances in large language models (LLMs) and biomedical foundation models (BioFMs) have achieved strong results in biological text reasoning, molecular modeling, and single-cell analysis, yet they remain siloed in disjoint embedding spaces, limiting cross-modal reasoning. We present BIOVERSE (Biomedical Vector Embedding Realignment for Semantic Engagement), a two-stage approach that adapts pretrained BioFMs as modality encoders and aligns them with LLMs through lightweight, modality-specific projection layers. The approach first aligns each modality to a shared LLM space through independently trained projections, allowing them to interoperate naturally, and then applies standard instruction tuning with multi-modal data to bring them together for downstream reasoning. By unifying raw biomedical data with knowledge embedded in LLMs, the approach enables zero-shot annotation, cross-modal question answering, and interactive, explainable dialogue. Across tasks spanning cell-type annotation, molecular description, and protein function reasoning, compact BIOVERSE configurations surpass larger LLM baselines while enabling richer, generative outputs than existing BioFMs, establishing a foundation for principled multi-modal biomedical reasoning.

System logs represent a valuable source of Cyber Threat Intelligence (CTI), capturing attacker behaviors, exploited vulnerabilities, and traces of malicious activity. Yet their utility is often limited by lack of structure, semantic inconsistency, and fragmentation across devices and sessions. Extracting actionable CTI from logs therefore requires approaches that can reconcile noisy, heterogeneous data into coherent and interoperable representations. We introduce OntoLogX, an autonomous Artificial Intelligence (AI) agent that leverages Large Language Models (LLMs) to transform raw logs into ontology-grounded Knowledge Graphs (KGs). OntoLogX integrates a lightweight log ontology with Retrieval Augmented Generation (RAG) and iterative correction steps, ensuring that generated KGs are syntactically and semantically valid. Beyond event-level analysis, the system aggregates KGs into sessions and employs a LLM to predict MITRE ATT&CK tactics, linking low-level log evidence to higher-level adversarial objectives. We evaluate OntoLogX on both logs from a public benchmark and a real-world honeypot dataset, demonstrating robust KG generation across multiple KGs backends and accurate mapping of adversarial activity to ATT&CK tactics. Results highlight the benefits of retrieval and correction for precision and recall, the effectiveness of code-oriented models in structured log analysis, and the value of ontology-grounded representations for actionable CTI extraction.

Both the general public and academic communities have raised concerns about sycophancy, the phenomenon of artificial intelligence (AI) excessively agreeing with or flattering users. Yet, beyond isolated media reports of severe consequences, like reinforcing delusions, little is known about the extent of sycophancy or how it affects people who use AI. Here we show the pervasiveness and harmful impacts of sycophancy when people seek advice from AI. First, across 11 state-of-the-art AI models, we find that models are highly sycophantic: they affirm users' actions 50% more than humans do, and they do so even in cases where user queries mention manipulation, deception, or other relational harms. Second, in two preregistered experiments (N = 1604), including a live-interaction study where participants discuss a real interpersonal conflict from their life, we find that interaction with sycophantic AI models significantly reduced participants' willingness to take actions to repair interpersonal conflict, while increasing their conviction of being in the right. However, participants rated sycophantic responses as higher quality, trusted the sycophantic AI model more, and were more willing to use it again. This suggests that people are drawn to AI that unquestioningly validate, even as that validation risks eroding their judgment and reducing their inclination toward prosocial behavior. These preferences create perverse incentives both for people to increasingly rely on sycophantic AI models and for AI model training to favor sycophancy. Our findings highlight the necessity of explicitly addressing this incentive structure to mitigate the widespread risks of AI sycophancy.

Large language model (LLM) generation often requires balancing output quality against inference cost, especially when using multiple generations. We introduce a new framework for inference-time optimization based on the classical Pandora's Box problem. Viewing each generation as opening a costly "box" with random reward, we develop algorithms that decide when to stop generating without knowing the underlying reward distribution. Our first contribution is a UCB-style Pandora's Box algorithm, which achieves performance that is provably close to Weitzman's algorithm, the optimal strategy when the distribution is known. We further adapt this method to practical LLM settings by addressing reward scaling across prompts via a Bradley-Terry inspired transformation. This leads to an adaptive inference-time optimization method that normalizes rewards and learns stopping thresholds on the fly. Experiments on the AlpacaFarm and HH-RLHF datasets, using multiple LLM-reward model pairs, show that our adaptive strategy can obtain the same performance as non-adaptive Best-of-N sampling while requiring 15-35 percent fewer generations on average. Our results establish a principled bridge between optimal stopping theory and inference-time scaling, providing both theoretical performance bounds and practical efficiency gains for LLM deployment.