Finetuning provides a scalable and cost-effective means of customizing language models for specific tasks or response styles, with greater reliability than prompting or in-context learning. In contrast, the conventional wisdom is that injecting knowledge via finetuning results in brittle performance and poor generalization. We argue that the dichotomy of "task customization" (e.g., instruction tuning) and "knowledge injection" (e.g., teaching new facts) is a distinction without a difference. We instead identify concrete factors that explain the heterogeneous effectiveness observed with finetuning. To this end, we conduct a large-scale experimental study of finetuning the frontier Gemini v1.5 model family on a spectrum of datasets that are artificially engineered to interpolate between the strengths and failure modes of finetuning. Our findings indicate that question-answer training data formats provide much stronger knowledge generalization than document/article-style training data, numerical information can be harder for finetuning to retain than categorical information, and models struggle to apply finetuned knowledge during multi-step reasoning even when trained on similar examples -- all factors that render "knowledge injection" to be especially difficult, even after controlling for considerations like data augmentation and information volume. On the other hand, our findings also indicate that it is not fundamentally more difficult to finetune information about a real-world event than information about what a model's writing style should be.

Blind and Low Vision (BLV) people have adopted AI-powered visual interpretation applications to address their daily needs. While these applications have been helpful, prior work has found that users remain unsatisfied by their frequent errors. Recently, multimodal large language models (MLLMs) have been integrated into visual interpretation applications, and they show promise for more descriptive visual interpretations. However, it is still unknown how this advancement has changed people's use of these applications. To address this gap, we conducted a two-week diary study in which 20 BLV people used an MLLM-enabled visual interpretation application we developed, and we collected 553 entries. In this paper, we report a preliminary analysis of 60 diary entries from 6 participants. We found that participants considered the application's visual interpretations trustworthy (mean 3.75 out of 5) and satisfying (mean 4.15 out of 5). Moreover, participants trusted our application in high-stakes scenarios, such as receiving medical dosage advice. We discuss our plan to complete our analysis to inform the design of future MLLM-enabled visual interpretation systems.

Longitudinal data in electronic health records (EHRs) represent an individual`s clinical history through a sequence of codified concepts, including diagnoses, procedures, medications, and laboratory tests. Foundational models, such as generative pre-trained transformers (GPT), can leverage this data to predict future events. While fine-tuning of these models enhances task-specific performance, it is costly, complex, and unsustainable for every target. We show that a foundation model trained on EHRs can perform predictive tasks in a zero-shot manner, eliminating the need for fine-tuning. This study presents the first comprehensive analysis of zero-shot forecasting with GPT-based foundational models in EHRs, introducing a novel pipeline that formulates medical concept prediction as a generative modeling task. Unlike supervised approaches requiring extensive labeled data, our method enables the model to forecast a next medical event purely from a pretraining knowledge. We evaluate performance across multiple time horizons and clinical categories, demonstrating model`s ability to capture latent temporal dependencies and complex patient trajectories without task supervision. Model performance for predicting the next medical concept was evaluated using precision and recall metrics, achieving an average top1 precision of 0.614 and recall of 0.524. For 12 major diagnostic conditions, the model demonstrated strong zero-shot performance, achieving high true positive rates while maintaining low false positives. We demonstrate the power of a foundational EHR GPT model in capturing diverse phenotypes and enabling robust, zero-shot forecasting of clinical outcomes. This capability enhances the versatility of predictive healthcare models and reduces the need for task-specific training, enabling more scalable applications in clinical settings.

While the Question Generation (QG) task has been increasingly adopted in educational assessments, its evaluation remains limited by approaches that lack a clear connection to the educational values of test items. In this work, we introduce test item analysis, a method frequently used by educators to assess test question quality, into QG evaluation. Specifically, we construct pairs of candidate questions that differ in quality across dimensions such as topic coverage, item difficulty, item discrimination, and distractor efficiency. We then examine whether existing QG evaluation approaches can effectively distinguish these differences. Our findings reveal significant shortcomings in these approaches with respect to accurately assessing test item quality in relation to student performance. To address this gap, we propose a novel QG evaluation framework, QG-SMS, which leverages Large Language Model for Student Modeling and Simulation to perform test item analysis. As demonstrated in our extensive experiments and human evaluation study, the additional perspectives introduced by the simulated student profiles lead to a more effective and robust assessment of test items.

Benchmarks are essential for consistent evaluation and reproducibility. The integration of Artificial Intelligence into Software Engineering (AI4SE) has given rise to numerous benchmarks for tasks such as code generation and bug fixing. However, this surge presents challenges: (1) scattered benchmark knowledge across tasks, (2) difficulty in selecting relevant benchmarks, (3) the absence of a uniform standard for benchmark development, and (4) limitations of existing benchmarks. In this paper, we review 173 studies and identify 204 AI4SE benchmarks. We classify these benchmarks, analyze their limitations, and expose gaps in practices. Based on our review, we created BenchScout, a semantic search tool to find relevant benchmarks, using automated clustering of the contexts from associated studies. We conducted a user study with 22 participants to evaluate BenchScout's usability, effectiveness, and intuitiveness which resulted in average scores of 4.5, 4.0, and 4.1 out of 5. To advance benchmarking standards, we propose BenchFrame, a unified method to enhance benchmark quality. As a case study, we applied BenchFrame to the HumanEval benchmark and addressed its main limitations. This led to HumanEvalNext, featuring (1) corrected errors, (2) improved language conversion, (3) expanded test coverage, and (4) increased difficulty. We then evaluated ten state-of-the-art code language models on HumanEval, HumanEvalPlus, and HumanEvalNext. On HumanEvalNext, models showed a pass@1 score reduction of 31.22% and 19.94% compared to HumanEval and HumanEvalPlus, respectively.

Mixture of large language model (LLMs) Agents (MoA) architectures achieve state-of-the-art performance on prominent benchmarks like AlpacaEval 2.0 by leveraging the collaboration of multiple LLMs at inference time. Despite these successes, an evaluation of the safety and reliability of MoA is missing. We present the first comprehensive study of MoA's robustness against deceptive LLM agents that deliberately provide misleading responses. We examine factors like the propagation of deceptive information, model size, and information availability, and uncover critical vulnerabilities. On AlpacaEval 2.0, the popular LLaMA 3.1-70B model achieves a length-controlled Win Rate (LC WR) of 49.2% when coupled with 3-layer MoA (6 LLM agents). However, we demonstrate that introducing only a $\textit{single}$ carefully-instructed deceptive agent into the MoA can reduce performance to 37.9%, effectively nullifying all MoA gains. On QuALITY, a multiple-choice comprehension task, the impact is also severe, with accuracy plummeting by a staggering 48.5%. Inspired in part by the historical Doge of Venice voting process, designed to minimize influence and deception, we propose a range of unsupervised defense mechanisms that recover most of the lost performance.

This Perspective explores the transformative potential of Multi-Agent Systems (MAS) powered by Large Language Models (LLMs) in the geosciences. Users of geoscientific data repositories face challenges due to the complexity and diversity of data formats, inconsistent metadata practices, and a considerable number of unprocessed datasets. MAS possesses transformative potential for improving scientists' interaction with geoscientific data by enabling intelligent data processing, natural language interfaces, and collaborative problem-solving capabilities. We illustrate this approach with "PANGAEA GPT", a specialized MAS pipeline integrated with the diverse PANGAEA database for Earth and Environmental Science, demonstrating how MAS-driven workflows can effectively manage complex datasets and accelerate scientific discovery. We discuss how MAS can address current data challenges in geosciences, highlight advancements in other scientific fields, and propose future directions for integrating MAS into geoscientific data processing pipelines. In this Perspective, we show how MAS can fundamentally improve data accessibility, promote cross-disciplinary collaboration, and accelerate geoscientific discoveries.

This study introduces a new methodology for an Inference Index (InI), called INFerence INdex In Testing model Effectiveness methodology (INFINITE), aiming to evaluate the performance of Large Language Models (LLMs) in code generation tasks. The InI index provides a comprehensive assessment focusing on three key components: efficiency, consistency, and accuracy. This approach encapsulates time-based efficiency, response quality, and the stability of model outputs, offering a thorough understanding of LLM performance beyond traditional accuracy metrics. We applied this methodology to compare OpenAI's GPT-4o (GPT), OpenAI-o1 pro (OAI1), and OpenAI-o3 mini-high (OAI3) in generating Python code for the Long-Short-Term-Memory (LSTM) model to forecast meteorological variables such as temperature, relative humidity and wind velocity. Our findings demonstrate that GPT outperforms OAI1 and performs comparably to OAI3 regarding accuracy and workflow efficiency. The study reveals that LLM-assisted code generation can produce results similar to expert-designed models with effective prompting and refinement. GPT's performance advantage highlights the benefits of widespread use and user feedback.

Large Language Models (LLMs) have achieved remarkable success, but their English-centric training data limits performance in non-English languages, highlighting the need for enhancements in their multilingual capabilities. While some work on multilingual prompting methods handles non-English queries by utilizing English translations or restructuring them to more closely align with LLM reasoning patterns, these works often overlook the importance of cultural context, limiting their effectiveness. To address this limitation, we propose EMCEI, a simple yet effective approach that improves LLMs' multilingual capabilities by incorporating cultural context for more accurate and appropriate responses. Specifically, EMCEI follows a two-step process that first extracts relevant cultural context from the LLM's parametric knowledge via prompting. Then, EMCEI employs an LLM-as-Judge mechanism to select the most appropriate response by balancing cultural relevance and reasoning ability. Experiments on diverse multilingual benchmarks show that EMCEI outperforms existing baselines, demonstrating its effectiveness in handling multilingual queries with LLMs.

Keeping large language models factually up-to-date is crucial for deployment, yet costly retraining remains a challenge. Knowledge editing offers a promising alternative, but methods are only tested on small-scale or synthetic edit benchmarks. In this work, we aim to bridge research into lifelong knowledge editing to real-world edits at practically relevant scale. We first introduce WikiBigEdit; a large-scale benchmark of real-world Wikidata edits, built to automatically extend lifelong for future-proof benchmarking. In its first instance, it includes over 500K question-answer pairs for knowledge editing alongside a comprehensive evaluation pipeline. Finally, we use WikiBigEdit to study existing knowledge editing techniques' ability to incorporate large volumes of real-world facts and contrast their capabilities to generic modification techniques such as retrieval augmentation and continual finetuning to acquire a complete picture of the practical extent of current lifelong knowledge editing.