Automated model discovery is the process of automatically searching and identifying the most appropriate model for a given dataset over a large combinatorial search space. Existing approaches, however, often face challenges in balancing the capture of fine-grained details with ensuring generalizability beyond training data regimes with a reasonable model complexity. In this paper, we present a multi-modal \& multi-step pipeline for effective automated model discovery. Our approach leverages two vision-language-based modules (VLM), AnalyzerVLM and EvaluatorVLM, for effective model proposal and evaluation in an agentic way. AnalyzerVLM autonomously plans and executes multi-step analyses to propose effective candidate models. EvaluatorVLM assesses the candidate models both quantitatively and perceptually, regarding the fitness for local details and the generalibility for overall trends. Our results demonstrate that our pipeline effectively discovers models that capture fine details and ensure strong generalizability. Additionally, extensive ablation studies show that both multi-modality and multi-step reasoning play crucial roles in discovering favorable models.

Large language model (LLM) inference often suffers from high latency, particularly in resource-constrained environments such as on-device or edge deployments. To address this challenge, we present StorInfer, a novel storage-assisted LLM inference system that accelerates response time by precomputing and storing predictable query-response pairs offline. When a user query semantically matches a precomputed query, StorInfer bypasses expensive GPU inference and instantly returns the stored response, significantly reducing latency and compute costs. To maximize coverage and effectiveness, StorInfer employs an LLM-driven generator that adaptively produces diverse and deduplicated queries based on a given knowledge base. This is achieved via two techniques: adaptive query masking, which prevents regeneration of similar queries, and adaptive sampling, which dynamically tunes generation parameters to promote semantic diversity. The resulting query-response pairs are embedded and indexed using a disk-backed vector database to enable fast, similarity-based retrieval at runtime. Using this approach, we generated 150K unique precomputed pairs (taking up to 830 MB of storage space), achieving up to 17.3% latency reduction with no loss in response quality. Our evaluation across multiple QA datasets demonstrates the practicality and scalability of storage-assisted inference, especially in scenarios with predictable query distributions. StorInfer highlights a promising direction in leveraging storage as a primary enabler for efficient, low-latency LLM deployment.

Since no metrics are available to evaluate specific aspects of a text, such as its personalization quality, the researchers often rely solely on large language models to meta-evaluate such texts. Due to internal biases of individual language models, it is recommended to use multiple of them for combined evaluation, which directly increases costs of such meta-evaluation. In this paper, a computationally efficient method for evaluation of personalization quality of a given text (generated by a language model) is introduced, called PerQ. A case study of comparison of generation capabilities of large and small language models shows the usability of the proposed metric in research, effectively reducing the waste of resources.

Humans often encounter role conflicts -- social dilemmas where the expectations of multiple roles clash and cannot be simultaneously fulfilled. As large language models (LLMs) become increasingly influential in human decision-making, understanding how they behave in complex social situations is essential. While previous research has evaluated LLMs' social abilities in contexts with predefined correct answers, role conflicts represent inherently ambiguous social dilemmas that require contextual sensitivity: the ability to recognize and appropriately weigh situational cues that can fundamentally alter decision priorities. To address this gap, we introduce RoleConflictBench, a novel benchmark designed to evaluate LLMs' contextual sensitivity in complex social dilemmas. Our benchmark employs a three-stage pipeline to generate over 13K realistic role conflict scenarios across 65 roles, systematically varying their associated expectations (i.e., their responsibilities and obligations) and situational urgency levels. By analyzing model choices across 10 different LLMs, we find that while LLMs show some capacity to respond to these contextual cues, this sensitivity is insufficient. Instead, their decisions are predominantly governed by a powerful, inherent bias related to social roles rather than situational information. Our analysis quantifies these biases, revealing a dominant preference for roles within the Family and Occupation domains, as well as a clear prioritization of male roles and Abrahamic religions across most evaluatee models.

Embodied agents powered by large language models (LLMs) inherit advanced planning capabilities; however, their direct interaction with the physical world exposes them to safety vulnerabilities. In this work, we identify four key reasoning stages where hazards may arise: Task Understanding, Environment Perception, High-Level Plan Generation, and Low-Level Action Generation. We further formalize three orthogonal safety constraint types (Factual, Causal, and Temporal) to systematically characterize potential safety violations. Building on this risk model, we present SafeMindBench, a multimodal benchmark with 5,558 samples spanning four task categories (Instr-Risk, Env-Risk, Order-Fix, Req-Align) across high-risk scenarios such as sabotage, harm, privacy, and illegal behavior. Extensive experiments on SafeMindBench reveal that leading LLMs (e.g., GPT-4o) and widely used embodied agents remain susceptible to safety-critical failures. To address this challenge, we introduce SafeMindAgent, a modular Planner-Executor architecture integrated with three cascaded safety modules, which incorporate safety constraints into the reasoning process. Results show that SafeMindAgent significantly improves safety rate over strong baselines while maintaining comparable task completion. Together, SafeMindBench and SafeMindAgent provide both a rigorous evaluation suite and a practical solution that advance the systematic study and mitigation of safety risks in embodied LLM agents.

Enabling robots to execute long-horizon manipulation tasks from free-form language instructions remains a fundamental challenge in embodied AI. While vision-language models (VLMs) have shown promise as high-level planners, their deployment in the real world is hindered by two gaps: (i) the scarcity of large-scale, sequential manipulation data that couples natural language with multi-step action plans, and (ii) the absence of dense, interpretable rewards for fine-tuning VLMs on planning objectives. To address these issues, we propose REVER, a framework that empowers VLMs to generate and validate long-horizon manipulation plans from natural language instructions in real-world scenarios. Under REVER we train and release RoboFarseer, a VLM incentivized to emit chain-of-thought that perform temporal and spatial reasoning, ensuring physically plausible and logically coherent plans. To obtain training data, we leverage the Universal Manipulation Interface framework to capture hardware-agnostic demonstrations of atomic skills. An automated annotation engine converts each demonstration into vision-instruction-plan triplet. We introduce a verifiable reward that scores the generated plan by its ordered bipartite matching overlap with the ground-truth skill sequence. At run time, the fine-tuned VLM functions both as a planner and as a monitor, verifying step-wise completion. RoboFarseer matches or exceeds the performance of proprietary models that are orders of magnitude larger, while on open-ended planning it surpasses the best baseline by more than 40%. In real-world, long-horizon tasks, the complete system boosts overall success by roughly 60% compared with the same low-level controller without the planner. We will open-source both the dataset and the trained model upon publication.

Large language models (LLMs), despite being safety-aligned, exhibit brittle refusal behaviors that can be circumvented by simple linguistic changes. As tense jailbreaking demonstrates that models refusing harmful requests often comply when rephrased in past tense, a critical generalization gap is revealed in current alignment methods whose underlying mechanisms are poorly understood. In this work, we introduce Activation-Scaling Guard (ASGuard), an insightful, mechanistically-informed framework that surgically mitigates this specific vulnerability. For the first step, we use circuit analysis to identify the specific attention heads causally linked to the targeted jailbreaking, the tense-changing attack. Second, we train a precise, channel-wise scaling vector to recalibrate the activation of tense vulnerable heads. Lastly, we apply it into a "preventative fine-tuning", forcing the model to learn a more robust refusal mechanism. Across three LLMs, ASGuard effectively reduces the attack success rate of targeted jailbreaking while preserving general capabilities and minimizing over refusal, achieving a Pareto-optimal balance between safety and utility. Our findings underscore how adversarial suffixes suppress the propagation of the refusal-mediating direction, based on mechanistic analysis. Furthermore, our work showcases how a deep understanding of model internals can be leveraged to develop practical, efficient, and targeted methods for adjusting model behavior, charting a course for more reliable and interpretable AI safety.

While large reasoning models trained with critic-free reinforcement learning and verifiable rewards (RL VR) represent the state-of-the-art, their practical utility is hampered by "overthinking", a critical issue where models generate excessively long reasoning paths without any performance benefit. Existing solutions that penalize length often fail, inducing performance degradation due to a fundamental misalignment between trajectory-level rewards and token-level optimization. CS, built on our theoretical discovery of two previously unaddressed flaws in current length rewards: (1) the erroneous penalization of essential exploratory tokens and (2) the inadvertent rewarding of partial redundancy. Our framework's innovations include (i) a first-of-its-kind decoupled token-level reward mechanism that surgically distinguishes and penalizes redundant tokens, and (ii) a novel curriculum batch scheduling strategy to master the efficiency-efficacy equilibrium. CS can achieve a dramatic reduction in reasoning tokens by over 50% across seven benchmarks while simultaneously maintaining or even improving performance. It demonstrates conclusively that substantial gains in reasoning efficiency can be achieved without compromising a model's underlying reasoning power.Figure 1: Left: Two major flaws of prior practice apply sequence-level length reward without control of training data. CS improves pass@1 of base models while reducing 60% token costs compared to the base model across 7 benchmarks. Experimental details are presented in Appendix E.5. Recent large reasoning models (LRM; Guo et al. (2025); OpenAI (2025); Qwen (2025)) trained with critic-free reinforcement learning (RL) algorithms, such as GRPO (Shao et al., 2024), DAPO (Y u et al., 2025), and REINFORCE++ (Hu et al., 2025a), have demonstrated impressive reasoning capabilities through verifiable outcome rewards.

In this study, we focus on the automatic evaluation of long and detailed image captions generated by multimodal Large Language Models (MLLMs). Most existing automatic evaluation metrics for image captioning are primarily designed for short captions and are not suitable for evaluating long captions. Moreover, recent LLM-as-a-Judge approaches suffer from slow inference due to their reliance on autoregressive inference and early fusion of visual information. To address these limitations, we propose VELA, an automatic evaluation metric for long captions developed within a novel LLM-Hybrid-as-a-Judge framework. Furthermore, we propose LongCap-Arena, a benchmark specifically designed for evaluating metrics for long captions. This benchmark comprises 7,805 images, the corresponding human-provided long reference captions and long candidate captions, and 32,246 human judgments from three distinct perspectives: Descriptiveness, Relevance, and Fluency. We demonstrated that VELA outperformed existing metrics and achieved superhuman performance on LongCap-Arena.

Recent advances in question answering have led to substantial progress in tasks such as multi-hop reasoning. However, global sensemaking-answering questions by synthesizing information from an entire corpus remains a significant challenge. A prior graph-based approach to global sensemaking lacks retrieval mechanisms, topic specificity, and incurs high inference costs. To address these limitations, we propose ReTAG, a Retrieval-Enhanced, Topic-Augmented Graph framework that constructs topic-specific subgraphs and retrieves the relevant summaries for response generation. Experiments show that ReTAG improves response quality while significantly reducing inference time compared to the baseline. Our code is available at https://github.com/bykimby/retag.