Multi-token prediction has emerged as a promising objective for improving language model pretraining, but its benefits have not consistently generalized to other settings such as fine-tuning. In this paper, we propose MuToR, a simple and effective approach to multi-token prediction that interleaves learnable register tokens into the input sequence, each tasked with predicting future targets. Compared to existing methods, MuToR offers several key advantages: it introduces only a negligible number of additional parameters, requires no architectural changes--ensuring compatibility with off-the-shelf pretrained language models--and remains aligned with the next-token pretraining objective, making it especially well-suited for supervised fine-tuning. Moreover, it naturally supports scalable prediction horizons. We demonstrate the effectiveness and versatility of MuToR across a range of use cases, including supervised fine-tuning, parameter-efficient fine-tuning (PEFT), and pretraining, on challenging generative tasks in both language and vision domains. Our code will be available at: https://github.com/nasosger/MuToR.

Code security and usability are both essential for various coding assistant applications driven by large language models (LLMs). Current code security benchmarks focus solely on single evaluation task and paradigm, such as code completion and generation, lacking comprehensive assessment across dimensions like secure code generation, vulnerability repair and discrimination. In this paper, we first propose CoV-Eval, a multi-task benchmark covering various tasks such as code completion, vulnerability repair, vulnerability detection and classification, for comprehensive evaluation of LLM code security. Besides, we developed VC-Judge, an improved judgment model that aligns closely with human experts and can review LLM-generated programs for vulnerabilities in a more efficient and reliable way. We conduct a comprehensive evaluation of 20 proprietary and open-source LLMs. Overall, while most LLMs identify vulnerable codes well, they still tend to generate insecure codes and struggle with recognizing specific vulnerability types and performing repairs. Extensive experiments and qualitative analyses reveal key challenges and optimization directions, offering insights for future research in LLM code security.

The emergence of unified multimodal understanding and generation models is rapidly attracting attention because of their ability to enhance instruction-following capabilities while minimizing model redundancy. However, there is a lack of a unified evaluation framework for these models, which would enable an elegant, simplified, and overall evaluation. Current models conduct evaluations on multiple task-specific benchmarks, but there are significant limitations, such as the lack of overall results, errors from extra evaluation models, reliance on extensive labeled images, benchmarks that lack diversity, and metrics with limited capacity for instruction-following evaluation. To tackle these challenges, we introduce UniEval, the first evaluation framework designed for unified multimodal models without extra models, images, or annotations. This facilitates a simplified and unified evaluation process. The UniEval framework contains a holistic benchmark, UniBench (supports both unified and visual generation models), along with the corresponding UniScore metric. UniBench includes 81 fine-grained tags contributing to high diversity. Experimental results indicate that UniBench is more challenging than existing benchmarks, and UniScore aligns closely with human evaluations, surpassing current metrics. Moreover, we extensively evaluated SoTA unified and visual generation models, uncovering new insights into Univeral's unique values.

It is commonly believed that scaling language models should commit a significant space or time cost, by increasing the parameters (parameter scaling) or output tokens (inference-time scaling). We introduce the third and more inference-efficient scaling paradigm: increasing the model's parallel computation during both training and inference time. We apply $P$ diverse and learnable transformations to the input, execute forward passes of the model in parallel, and dynamically aggregate the $P$ outputs. This method, namely parallel scaling (ParScale), scales parallel computation by reusing existing parameters and can be applied to any model structure, optimization procedure, data, or task. We theoretically propose a new scaling law and validate it through large-scale pre-training, which shows that a model with $P$ parallel streams is similar to scaling the parameters by $O(\log P)$ while showing superior inference efficiency. For example, ParScale can use up to 22$\times$ less memory increase and 6$\times$ less latency increase compared to parameter scaling that achieves the same performance improvement. It can also recycle an off-the-shelf pre-trained model into a parallelly scaled one by post-training on a small amount of tokens, further reducing the training budget. The new scaling law we discovered potentially facilitates the deployment of more powerful models in low-resource scenarios, and provides an alternative perspective for the role of computation in machine learning.

Vision language models (VLMs) are AI systems paired with both language and vision encoders to process multimodal input. They are capable of performing complex semantic tasks such as automatic captioning, but it remains an open question about how well they comprehend the visuospatial properties of scenes depicted in the images they process. We argue that descriptions of virtual objects -- objects that are not visually represented in an image -- can help test scene comprehension in these AI systems. For example, an image that depicts a person standing under a tree can be paired with the following prompt: imagine that a kite is stuck in the tree. VLMs that comprehend the scene should update their representations and reason sensibly about the spatial relations between all three objects. We describe systematic evaluations of state-of-the-art VLMs and show that their ability to process virtual objects is inadequate.

Human learning relies on specialization--distinct cognitive mechanisms working together to enable rapid learning. In contrast, most modern neural networks rely on a single mechanism: gradient descent over an objective function. This raises the question: might human learners' relatively rapid learning from just tens of examples instead of tens of thousands in data-driven deep learning arise from our ability to use multiple specialized mechanisms of learning in combination? We investigate this question through an ablation analysis of inductive human learning simulations in online tutoring environments. Comparing reinforcement learning to a more data-efficient 3-mechanism symbolic rule induction approach, we find that decomposing learning into multiple distinct mechanisms significantly improves data efficiency, bringing it in line with human learning. Furthermore, we show that this decomposition has a greater impact on efficiency than the distinction between symbolic and subsymbolic learning alone. Efforts to align data-driven machine learning with human learning often overlook the stark difference in learning efficiency. Our findings suggest that integrating multiple specialized learning mechanisms may be key to bridging this gap. A key idea within the learning sciences, popularized by Anderson's ACT -R theory (2013) and expanded upon by others (Koedinger, Corbett, & Perfetti, 2012), is that human performance is enabled by independent knowledge components--individual facts, skills, or principles--that must be understood and retained to exhibit mastery of higher-level capabilities.

Plain language summaries (PLSs) are essential for facilitating effective communication between clinicians and patients by making complex medical information easier for laypeople to understand and act upon. Large language models (LLMs) have recently shown promise in automating PLS generation, but their effectiveness in supporting health information comprehension remains unclear. Prior evaluations have generally relied on automated scores that do not measure understandability directly, or subjective Likert-scale ratings from convenience samples with limited generalizability. To address these gaps, we conducted a large-scale crowdsourced evaluation of LLM-generated PLSs using Amazon Mechanical Turk with 150 participants. We assessed PLS quality through subjective Likert-scale ratings focusing on simplicity, informativeness, coherence, and faithfulness; and objective multiple-choice comprehension and recall measures of reader understanding. Additionally, we examined the alignment between 10 automated evaluation metrics and human judgments. Our findings indicate that while LLMs can generate PLSs that appear indistinguishable from human-written ones in subjective evaluations, human-written PLSs lead to significantly better comprehension. Furthermore, automated evaluation metrics fail to reflect human judgment, calling into question their suitability for evaluating PLSs. This is the first study to systematically evaluate LLM-generated PLSs based on both reader preferences and comprehension outcomes. Our findings highlight the need for evaluation frameworks that move beyond surface-level quality and for generation methods that explicitly optimize for layperson comprehension.

With the advent of neural language models, the performance of code generation has been significantly boosted. However, the problem of repetitions during the generation process continues to linger. Previous work has primarily focused on content repetition, which is merely a fraction of the broader repetition problem in code generation. A more prevalent and challenging problem is structural repetition. In structural repetition, the repeated code appears in various patterns but possesses a fixed structure, which can be inherently reflected in grammar. In this paper, we formally define structural repetition and propose an efficient decoding approach called RPG, which stands for Repetition Penalization based on Grammar, to alleviate the repetition problems in code generation for LLMs. Specifically, RPG first leverages grammar rules to identify repetition problems during code generation, and then strategically decays the likelihood of critical tokens that contribute to repetitions, thereby mitigating them in code generation. To facilitate this study, we construct a new dataset CodeRepetEval to comprehensively evaluate approaches for mitigating the repetition problems in code generation. Extensive experimental results demonstrate that RPG substantially outperforms the best-performing baselines on CodeRepetEval dataset as well as HumanEval and MBPP benchmarks, effectively reducing repetitions and enhancing the quality of generated code.

A recent area of increasing research is the use of Large Language Models (LLMs) in penetration testing, which promises to reduce costs and thus allow for higher frequency. We conduct a review of related work, identifying best practices and common evaluation issues. We then present AutoPentest, an application for performing black-box penetration tests with a high degree of autonomy. AutoPentest is based on the LLM GPT-4o from OpenAI and the LLM agent framework LangChain. It can perform complex multi-step tasks, augmented by external tools and knowledge bases. We conduct a study on three capture-the-flag style Hack The Box (HTB) machines, comparing our implementation AutoPentest with the baseline approach of manually using the ChatGPT-4o user interface. Both approaches are able to complete 15-25 % of the subtasks on the HTB machines, with AutoPentest slightly outperforming ChatGPT. We measure a total cost of \$96.20 US when using AutoPentest across all experiments, while a one-month subscription to ChatGPT Plus costs \$20. The results show that further implementation efforts and the use of more powerful LLMs released in the future are likely to make this a viable part of vulnerability management.

Transformer models have revolutionized AI, powering applications like content generation and sentiment analysis. However, their deployment in Machine Learning as a Service (MLaaS) raises significant privacy concerns, primarily due to the centralized processing of sensitive user data. Private Transformer Inference (PTI) offers a solution by utilizing cryptographic techniques such as secure multi-party computation and homomorphic encryption, enabling inference while preserving both user data and model privacy. This paper reviews recent PTI advancements, highlighting state-of-the-art solutions and challenges. We also introduce a structured taxonomy and evaluation framework for PTI, focusing on balancing resource efficiency with privacy and bridging the gap between high-performance inference and data privacy.