Large language models frequently generate harmful recommendations in personal multi-turn dialogue by ignoring user-specific safety context, exhibiting sycophantic agreement, and compromising user safety for larger group preferences. We introduce MIRROR, a modular production-focused architecture that prevents these failures through a persistent, bounded internal state that preserves personal conversational information across conversational turns. Our dual-component design inspired by Dual Process Theory separates immediate response generation (Talker) from asynchronous deliberative processing (Thinker), which synthesizes parallel reasoning threads between turns with marginal latency. On the CuRaTe personalized safety benchmark, MIRROR-augmented models achieve a 21% relative improvement (69% to 84%) across seven diverse frontier models, with open-source Llama 4 and Mistral 3 variants surpassing both GPT-4o and Claude 3.7 Sonnet at only \$0.0028 to \$0.0172 additional cost per turn, narrowing the gap between affordable open-source models to frontier systems in the safety space. The modular architecture enables flexible deployment: full internal processing for affordable models or single-component configurations for expensive systems, democratizing access to safer, personalized AI.

In enterprise settings, organizational data is segregated, siloed and carefully protected by elaborate access control frameworks. These access control structures can completely break down if an LLM fine-tuned on the siloed data serves requests, for downstream tasks, from individuals with disparate access privileges. We propose Permissioned LLMs (PermLLM), a new class of LLMs that superimpose the organizational data access control structures on query responses they generate. We formalize abstractions underpinning the means to determine whether access control enforcement happens correctly over LLM query responses. Our formalism introduces the notion of a relevant response that can be used to prove whether a PermLLM mechanism has been implemented correctly. We also introduce a novel metric, called access advantage, to empirically evaluate the efficacy of a PermLLM mechanism. We introduce three novel PermLLM mechanisms that build on Parameter Efficient Fine-Tuning to achieve the desired access control. We furthermore present two instantiations of access advantage--(i) Domain Distinguishability Index (DDI) based on Membership Inference Attacks, and (ii) Utility Gap Index (UGI) based on LLM utility evaluation. We demonstrate the efficacy of our PermLLM mechanisms through extensive experiments on five public datasets (GPQA, RCV1, SimpleQA, WMDP, and PubMedQA), in addition to evaluating the validity of DDI and UGI metrics themselves for quantifying access control in LLMs.

Reinforcement learning (RL) has been widely adopted to enhance the performance of large language models (LLMs) on Text-to-SQL tasks. However, existing methods often rely on execution-based or LLM-based Bradley-Terry reward models. The former suffers from high execution latency caused by repeated database calls, whereas the latter imposes substantial GPU memory overhead, both of which significantly hinder the efficiency and scalability of RL pipelines. To this end, we propose a novel reward model framework for RL-based Text-to-SQL named Graph-Reward-SQL, which employs the GMNScore outcome reward model. We leverage SQL graph representations to provide accurate reward signals while significantly reducing time cost and GPU memory usage. Building on this foundation, we further introduce StepRTM, a stepwise reward model that provides intermediate supervision over Common Table Expression (CTE) subqueries. This encourages both functional correctness and readability of SQL. Extensive comparative and ablation experiments on standard benchmarks, including Spider and BIRD, demonstrate that our method consistently outperforms existing reward models.

Multi-Hop Question Answering (MHQA) tasks permeate real-world applications, posing challenges in orchestrating multi-step reasoning across diverse knowledge domains. While existing approaches have been improved with iterative retrieval, they still struggle to identify and organize dynamic knowledge. To address this, we propose DualRAG, a synergistic dual-process framework that seamlessly integrates reasoning and retrieval. DualRAG operates through two tightly coupled processes: Reasoning-augmented Querying (RaQ) and progressive Knowledge Aggregation (pKA). They work in concert: as RaQ navigates the reasoning path and generates targeted queries, pKA ensures that newly acquired knowledge is systematically integrated to support coherent reasoning. This creates a virtuous cycle of knowledge enrichment and reasoning refinement. Through targeted fine-tuning, DualRAG preserves its sophisticated reasoning and retrieval capabilities even in smaller-scale models, demonstrating its versatility and core advantages across different scales. Extensive experiments demonstrate that this dual-process approach substantially improves answer accuracy and coherence, approaching, and in some cases surpassing, the performance achieved with oracle knowledge access. These results establish DualRAG as a robust and efficient solution for complex multi-hop reasoning tasks.

Recent advancements in large language models (LLMs) have significantly improved their reasoning abilities, particularly through techniques involving search and backtracking. Backtracking naturally scales test-time compute by enabling sequential, linearized exploration via long chain-of-thought (CoT) generation. However, this is not the only strategy for scaling test time-compute: parallel sampling with best-of-N selection provides an alternative that generates diverse solutions simultaneously. Despite the growing adoption of sequential search, its advantages over parallel sampling-especially under a fixed compute budget-remain poorly understood. In this paper, we systematically compare these two approaches on two challenging reasoning tasks: CountDown and Sudoku. Surprisingly, we find that sequential search underperforms parallel sampling on CountDown but outperforms it on Sudoku, suggesting that backtracking is not universally beneficial. We identify two factors that can cause backtracking to degrade performance: (1) training on fixed search traces can lock models intro suboptimal strategies, and (2) explicit CoT supervision can discourage implicit (non verbalized) reasoning. Extending our analysis to reinforcement learning (RL), we show that models with backtracking capabilities benefit significantly from RL fine-tuning, while models without backtracking see limited, mixed gains. Together, these findings challenge the assumption that backtracking universally enhances LLM reasoning, instead revealing a complex interaction between task structure, training data, model scale, and learning paradigm.

We study last-layer outlier dimensions, i.e. dimensions that display extreme activations for the majority of inputs. We show that outlier dimensions arise in many different modern language models, and trace their function back to the heuristic of constantly predicting frequent words. We further show how a model can block this heuristic when it is not contextually appropriate, by assigning a counterbalancing weight mass to the remaining dimensions, and we investigate which model parameters boost outlier dimensions and when they arise during training. We conclude that outlier dimensions are a specialized mechanism discovered by many distinct models to implement a useful token prediction heuristic.

Results on existing LLM benchmarks capture little information over the model capabilities in low-resource tasks, making it difficult to develop effective solutions in these domains. To address these challenges, we curated 14 travel-domain datasets spanning 7 common NLP tasks using anonymised data from real-world scenarios, and analysed the performance across LLMs. We report on the accuracy, scaling behaviour, and reasoning capabilities of LLMs in a variety of tasks. Our results confirm that general benchmarking results are insufficient for understanding model performance in low-resource tasks. Despite the amount of training FLOPs, out-of-the-box LLMs hit performance bottlenecks in complex, domain-specific scenarios. Furthermore, reasoning provides a more significant boost for smaller LLMs by making the model a better judge on certain tasks.

Abstract--Genetic programming (GP) has demonstrated strong effectiveness in evolving tree-structured heuristics for complex optimization problems. Y et, in dynamic and large-scale scenarios, the most effective heuristics are often highly complex, hindering interpretability, slowing convergence, and limiting transferability across tasks. T o address these challenges, we present EvoSpeak, a novel framework that integrates GP with large language models (LLMs) to enhance the efficiency, transparency, and adaptability of heuristic evolution. EvoSpeak learns from high-quality GP heuristics, extracts knowledge, and leverages this knowledge to (i) generate warm-start populations that accelerate convergence, (ii) translate opaque GP trees into concise natural-language explanations that foster interpretability and trust, and (iii) enable knowledge transfer and preference-aware heuristic generation across related tasks. We verify the effectiveness of EvoSpeak through extensive experiments on dynamic flexible job shop scheduling (DFJSS), under both single-and multi-objective formulations. The results demonstrate that EvoSpeak produces more effective heuristics, improves evolutionary efficiency, and delivers human-readable reports that enhance usability. EURISTICS are indispensable tools for solving complex decision-making and optimization problems, with applications spanning scheduling [1], routing [2], and resource allocation [3]. They are designed to provide adaptive, domain-specific solutions that balance solution quality and computational efficiency, enabling practitioners to make near-optimal decisions in real time. Among the diverse methodologies for heuristic design, Genetic Programming (GP) [4] has emerged as a particularly powerful paradigm, capable of evolving interpretable symbolic rules that adapt to different problem structures [5]. GP-generated heuristics often rival, and sometimes surpass, learning-based methods such as neural combinatorial optimization [6], especially in terms of transparency and adaptability. Meng Xu is with the Singapore Institute of Manufacturing Technology, Agency for Science, Technology and Research, Singapore (e-mail: xu_meng@simtech.a-star.edu.sg). Jiao Liu is with the College of Computing & Data Science, Nanyang Technological University, Singapore (e-mail: jiao.liu@ntu.edu.sg). Y ew Soon Ong is with the College of Computing and Data Science, Nanyang Technological University, and the Centre for Frontier AI Research, Institute of High Performance Computing, Agency for Science, Technology and Research, Singapore (e-mail: asysong@ntu.edu.sg). Despite these advantages, the practical deployment of GPevolved heuristics faces two persistent challenges: complexity and transferability.

Advanced Planning and Scheduling (APS) systems have become indispensable for modern manufacturing operations, enabling optimized resource allocation and production efficiency in increasingly complex and dynamic environments. While algorithms for solving abstracted scheduling problems have been extensively investigated, the critical prerequisite of specifying manufacturing requirements into formal constraints remains manual and labor-intensive. Although recent advances of generative models, particularly Large Language Models (LLMs), show promise in automating constraint specification from heterogeneous raw manufacturing data, their direct application faces challenges due to natural language ambiguity, non-deterministic outputs, and limited domain-specific knowledge. This paper presents a constraint-centric architecture that regulates LLMs to perform reliable automated constraint specification for production scheduling. The architecture defines a hierarchical structural space organized across three levels, implemented through domain-specific representation to ensure precision and reliability while maintaining flexibility. Furthermore, an automated production scenario adaptation algorithm is designed and deployed to efficiently customize the architecture for specific manufacturing configurations. Experimental results demonstrate that the proposed approach successfully balances the generative capabilities of LLMs with the reliability requirements of manufacturing systems, significantly outperforming pure LLM-based approaches in constraint specification tasks.

As vision-language models (VLMs) gain prominence, their multimodal interfaces also introduce new safety vulnerabilities, making the safety evaluation challenging and critical. Existing red-teaming efforts are either restricted to a narrow set of adversarial patterns or depend heavily on manual engineering, lacking scalable exploration of emerging real-world VLM vulnerabilities. To bridge this gap, we propose ARMs, an adaptive red-teaming agent that systematically conducts comprehensive risk assessments for VLMs. Given a target harmful behavior or risk definition, ARMs automatically optimizes diverse red-teaming strategies with reasoning-enhanced multi-step orchestration, to effectively elicit harmful outputs from target VLMs. We propose 11 novel multimodal attack strategies, covering diverse adversarial patterns of VLMs (e.g., reasoning hijacking, contextual cloaking), and integrate 17 red-teaming algorithms into ARMs via model context protocol (MCP). To balance the diversity and effectiveness of the attack, we design a layered memory with an epsilon-greedy attack exploration algorithm. Extensive experiments on instance- and policy-based benchmarks show that ARMs achieves SOTA attack success rates, exceeding baselines by an average of 52.1% and surpassing 90% on Claude-4-Sonnet. We show that the diversity of red-teaming instances generated by ARMs is significantly higher, revealing emerging vulnerabilities in VLMs. Leveraging ARMs, we construct ARMs-Bench, a large-scale multimodal safety dataset comprising over 30K red-teaming instances spanning 51 diverse risk categories, grounded in both real-world multimodal threats and regulatory risks. Safety fine-tuning with ARMs-Bench substantially improves the robustness of VLMs while preserving their general utility, providing actionable guidance to improve multimodal safety alignment against emerging threats.