Safeguarding vision-language models (VLMs) is a critical challenge, as existing methods often suffer from over-defense, which harms utility, or rely on shallow alignment, failing to detect complex threats that require deep reasoning. To this end, we introduce PRISM (Principled Reasoning for Integrated Safety in Multimodality), a system2-like framework that aligns VLMs by embedding a structured, safety-aware reasoning process. Our framework consists of two key components: PRISM-CoT, a dataset that teaches safety-aware chain-of-thought reasoning, and PRISM-DPO, generated via Monte Carlo Tree Search (MCTS) to further refine this reasoning through Direct Preference Optimization to help obtain a delicate safety boundary. Comprehensive evaluations demonstrate PRISM's effectiveness, achieving remarkably low attack success rates including 0.15% on JailbreakV-28K for Qwen2-VL and 90% improvement over the previous best method on VLBreak for LLaVA-1.5. PRISM also exhibits strong robustness against adaptive attacks, significantly increasing computational costs for adversaries, and generalizes effectively to out-of-distribution challenges, reducing attack success rates to just 8.70% on the challenging multi-image MIS benchmark. Remarkably, this robust defense is achieved while preserving, and in some cases enhancing, model utility. To promote reproducibility, we have made our code, data, and model weights available at https://github.com/SaFoLab-WISC/PRISM.

LLM app stores are quickly emerging as platforms that gather a wide range of intelligent applications based on LLMs, giving users many choices for content creation, coding support, education, and more. However, the current methods for ranking and recommending apps in these stores mostly rely on static metrics like user activity and favorites, which makes it hard for users to efficiently find high-quality apps. To address these challenges, we propose LaQual, an automated framework for evaluating the quality of LLM apps. LaQual consists of three main stages: first, it labels and classifies LLM apps in a hierarchical way to accurately match them to different scenarios; second, it uses static indicators, such as time-weighted user engagement and functional capability metrics, to filter out low-quality apps; and third, it conducts a dynamic, scenario-adaptive evaluation, where the LLM itself generates scenario-specific evaluation metrics, scoring rules, and tasks for a thorough quality assessment. Experiments on a popular LLM app store show that LaQual is effective. Its automated scores are highly consistent with human judgments (with Spearman's rho of 0.62 and p=0.006 in legal consulting, and rho of 0.60 and p=0.009 in travel planning). By effectively screening, LaQual can reduce the pool of candidate LLM apps by 66.7% to 81.3%. User studies further confirm that LaQual significantly outperforms baseline systems in decision confidence, comparison efficiency (with average scores of 5.45 compared to 3.30), and the perceived value of its evaluation reports (4.75 versus 2.25). Overall, these results demonstrate that LaQual offers a scalable, objective, and user-centered solution for finding and recommending high-quality LLM apps in real-world use cases.

Peer-to-peer (P2P) energy trading is becoming central to modern distribution systems as rooftop PV and home energy management systems become pervasive, yet most existing market and reinforcement learning designs emphasize efficiency or private profit and offer little real-time guidance to ensure equitable outcomes under uncertainty. To address this gap, a fairness-aware multiagent reinforcement learning framework, FairMarket-RL, is proposed in which a large language model (LLM) critic shapes bidding policies within a continuous double auction under partial observability and discrete price-quantity actions. After each trading slot, the LLM returns normalized fairness scores Fairness-to-Grid (FTG), Fairness-Between-Sellers (FBS), and Fairness-of-Pricing (FPP) that are integrated into the reward via ramped coefficients and tunable scaling, so that fairness guidance complements, rather than overwhelms, economic incentives. The environment models realistic residential load and PV profiles and enforce hard constraints on prices, physical feasibility, and policy-update stability. Across a progression of experiments from a small pilot to a larger simulated community and a mixed-asset real-world dataset, the framework shifts exchanges toward local P2P trades, lowers consumer costs relative to grid-only procurement, sustains strong fairness across participants, and preserves utility viability. Sensitivity analyses over solar availability and aggregate demand further indicate robust performance, suggesting a scalable, LLM-guided pathway to decentralized electricity markets that are economically efficient, socially equitable, and technically sound.

Large language models (LLMs) are increasingly used to simulate human decision-making, but their intrinsic biases often diverge from real human behavior--limiting their ability to reflect population-level diversity. We address this challenge with a persona-based approach that leverages individual-level behavioral data from behavioral economics to adjust model biases. Applying this method to the ultimatum game--a standard but difficult benchmark for LLMs--we observe improved alignment between simulated and empirical behavior, particularly on the responder side. While further refinement of trait representations is needed, our results demonstrate the promise of persona-conditioned LLMs for simulating human-like decision patterns at scale.

What do large language models actually model? Do they tell us something about human capacities, or are they models of the corpus we've trained them on? I give a non-deflationary defence of the latter position. Cognitive science tells us that linguistic capabilities in humans rely supralinear formats for computation. The transformer architecture, by contrast, supports at best a linear formats for processing. This argument will rely primarily on certain invariants of the computational architecture of transformers. I then suggest a positive story about what transformers are doing, focusing on Liu et al. (2022)'s intriguing speculations about shortcut automata. I conclude with why I don't think this is a terribly deflationary story. Language is not (just) a means for expressing inner state but also a kind of 'discourse machine' that lets us make new language given appropriate context. We have learned to use this technology in one way; LLMs have also learned to use it too, but via very different means.

With the rapid advancement of large language models (LLMs), reinforcement learning (RL) has emerged as a pivotal methodology for enhancing the reasoning capabilities of LLMs. Unlike traditional pre-training approaches, RL encompasses multiple stages: rollout, reward, and training, which necessitates collaboration among various worker types. However, current RL systems continue to grapple with substantial GPU underutilization, due to two primary factors: (1) The rollout stage dominates the overall RL process due to test-time scaling; (2) Imbalances in rollout lengths (within the same batch) result in GPU bubbles. While prior solutions like asynchronous execution and truncation offer partial relief, they may compromise training accuracy for efficiency. Our key insight stems from a previously overlooked observation: rollout responses exhibit remarkable similarity across adjacent training epochs. Based on the insight, we introduce RhymeRL, an LLM RL system designed to accelerate RL training with two key innovations. First, to enhance rollout generation, we present HistoSpec, a speculative decoding inference engine that utilizes the similarity of historical rollout token sequences to obtain accurate drafts. Second, to tackle rollout bubbles, we introduce HistoPipe, a two-tier scheduling strategy that leverages the similarity of historical rollout distributions to balance workload among rollout workers. We have evaluated RhymeRL within a real production environment, demonstrating scalability from dozens to thousands of GPUs. Experimental results demonstrate that RhymeRL achieves a 2.6x performance improvement over existing methods, without compromising accuracy or modifying the RL paradigm.

Large language models (LLMs) can potentially help with verification using proof assistants by automating proofs. However, it is unclear how effective LLMs are in this task. In this paper, we perform a case study based on two mature Rocq projects: the hs-to-coq tool and Verdi. We evaluate the effectiveness of LLMs in generating proofs by both quantitative and qualitative analysis. Our study finds that: (1) external dependencies and context in the same source file can significantly help proof generation; (2) LLMs perform great on small proofs but can also generate large proofs; (3) LLMs perform differently on different verification projects; and (4) LLMs can generate concise and smart proofs, apply classical techniques to new definitions, but can also make odd mistakes.

Visual metaphor generation is a challenging task that aims to generate an image given an input text metaphor. Inherently, it needs language understanding to bind a source concept with a target concept, in a way that preserves meaning while ensuring visual coherence. We propose a self-evaluating visual metaphor generation framework that focuses on metaphor alignment. Our self-evaluation approach combines existing metrics with our newly proposed metaphor decomposition score and a meaning alignment (MA) metric. Within this setup, we explore two novel approaches: a training-free pipeline that explicitly decomposes prompts into source-target-meaning (S-T-M) mapping for image synthesis, and a complementary training-based pipeline that improves alignment using our proposed self-evaluation reward schema, without any large-scale retraining. On the held-out test set, the training-free approach surpasses strong closed baselines (GPT-4o, Imagen) on decomposition, CLIP, and MA scores, with the training-based approach close behind. We evaluate our framework output using a user-facing study, and observed that participants preferred GPT-4o overall, while our training-free pipeline led open-source methods and edged Imagen on abstract metaphors. Our analyses show S-T-M prompting helps longer or more abstract metaphors, with closed models excelling on short, concrete cases; we also observe sensitivity to sampler settings. Overall, structured prompting and lightweight RL perform metaphor alignment well under modest compute, and remaining gaps to human preference appear driven by aesthetics and sampling.

Log schema extraction is the process of deriving human-readable templates from massive volumes of log data, which is essential yet notoriously labor-intensive. Recent studies have attempted to streamline this task by leveraging Large Language Models (LLMs) for automated schema extraction. However, existing methods invariably rely on predefined regular expressions, necessitating human domain expertise and severely limiting productivity gains. To fundamentally address this limitation, we introduce SchemaCoder, the first fully automated schema extraction framework applicable to a wide range of log file formats without requiring human customization within the flow. At its core, SchemaCoder features a novel Residual Question-Tree (Q-Tree) Boosting mechanism that iteratively refines schema extraction through targeted, adaptive queries driven by LLMs. Particularly, our method partitions logs into semantic chunks via context-bounded segmentation, selects representative patterns using embedding-based sampling, and generates schema code through hierarchical Q-Tree-driven LLM queries, iteratively refined by our textual-residual evolutionary optimizer and residual boosting. Experimental validation demonstrates SchemaCoder's superiority on the widely-used LogHub-2.0 benchmark, achieving an average improvement of 21.3% over state-of-the-arts.

When adopting the role of a teacher in learning-by-teaching environments, students often struggle to engage in knowledge-building activities, such as providing explanations and addressing misconceptions. Instead, they frequently default to knowledge-telling behaviors, where they simply dictate what they already know or what to do without deeper reflection, thereby limiting learning. Teachable agents, particularly those capable of posing persistent follow-up questions, have been shown to encourage students (tutors) to shift from knowledge-telling to knowledge-building and enhance tutor learning. Tutor learning encompasses two interrelated types of knowledge: conceptual and procedural knowledge. Research has established a bidirectional relationship between these knowledge types, where improvements in one reinforce the other. This study investigates the role of knowledge-building in mediating the bidirectional relationship between procedural and conceptual learning. Our findings revealed a stable bidirectional relationship between procedural and conceptual knowledge, with higher post-test scores observed among students who engaged in knowledge-building, regardless of their procedural and conceptual pre-test performance. This suggests that knowledge-building serves as a crucial mechanism bridging the gap between students with low prior knowledge and higher conceptual and procedural learning gain.