Mixture-of-Experts (MoE) models have shown strong potential in scaling language models efficiently by activating only a small subset of experts per input. However, their widespread deployment remains limited due to the high memory overhead associated with storing all expert parameters, particularly as the number of experts increases. To address this challenge, prior works have explored expert dropping and merging strategies, yet they often suffer from performance drop at high compression ratios. In this paper, we introduce PuzzleMoE, a training-free MoE compression method that achieves both high accuracy and efficient inference through two key innovations: First, PuzzleMoE performs sparse expert merging by identifying element-wise weight redundancy and specialization. It uses a dual-mask to capture both shared and expert-specific parameters. Second, to avoid the overhead of storing binary masks and signs, PuzzleMoE introduces a bit-packed encoding scheme that reuses underutilized exponent bits, enabling efficient MoE inference on GPUs. Extensive experiments demonstrate that PuzzleMoE can compress MoE models by up to 50% while maintaining accuracy across various tasks. Specifically, it outperforms prior MoE compression methods by up to 16.7% on MMLU at 50% compression ratio, and achieves up to 1.28\times inference speedup.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an effective approach for improving the reasoning abilities of large language models (LLMs). The Group Relative Policy Optimization (GRPO) family has demonstrated strong performance in training LLMs with RLVR. However, as models train longer and scale larger, more training prompts become residual prompts, those with zero variance rewards that provide no training signal. Consequently, fewer prompts contribute to training, reducing diversity and hindering effectiveness. To fully exploit these residual prompts, we propose the Explore Residual Prompts in Policy Optimization (ERPO) framework, which encourages exploration on residual prompts and reactivates their training signals. ERPO maintains a history tracker for each prompt and adaptively increases the sampling temperature for residual prompts that previously produced all correct responses. This encourages the model to generate more diverse reasoning traces, introducing incorrect responses that revive training signals. Empirical results on the Qwen2.5 series demonstrate that ERPO consistently surpasses strong baselines across multiple mathematical reasoning benchmarks.

Modern LLM serving systems must sustain high throughput while meeting strict latency SLOs across two distinct inference phases: compute-intensive prefill and memory-bound decode phases. Existing approaches either (1) aggregate both phases on shared GPUs, leading to interference between prefill and decode phases, which degrades time-between-tokens (TBT); or (2) disaggregate the two phases across GPUs, improving latency but wasting resources through duplicated models and KV cache transfers. We present DuetServe, a unified LLM serving framework that achieves disaggregation-level isolation within a single GPU. DuetServe operates in aggregated mode by default and dynamically activates SM-level GPU spatial multiplexing when TBT degradation is predicted. Its key idea is to decouple prefill and decode execution only when needed through fine-grained, adaptive SM partitioning that provides phase isolation only when contention threatens latency service level objectives (SLOs). DuetServe integrates (1) an attention-aware roofline model to forecast iteration latency, (2) a partitioning optimizer that selects the optimal SM split to maximize throughput under TBT constraints, and (3) an interruption-free execution engine that eliminates CPU-GPU synchronization overhead. Evaluations show that DuetServe improves total throughput by up to 1.3x while maintaining low generation latency compared to state-of-the-art frameworks.

Simulation plays a key role in scaling robot learning and validating policies, but constructing simulations remains a labor-intensive process. This paper introduces ReGen, a generative simulation framework that automates simulation design via inverse design. Given a robot's behavior -- such as a motion trajectory or an objective function -- and its textual description, ReGen infers plausible scenarios and environments that could have caused the behavior. ReGen leverages large language models to synthesize scenarios by expanding a directed graph that encodes cause-and-effect relationships, relevant entities, and their properties. This structured graph is then translated into a symbolic program, which configures and executes a robot simulation environment. Our framework supports (i) augmenting simulations based on ego-agent behaviors, (ii) controllable, counterfactual scenario generation, (iii) reasoning about agent cognition and mental states, and (iv) reasoning with distinct sensing modalities, such as braking due to faulty GPS signals. We demonstrate ReGen in autonomous driving and robot manipulation tasks, generating more diverse, complex simulated environments compared to existing simulations with high success rates, and enabling controllable generation for corner cases. This approach enhances the validation of robot policies and supports data or simulation augmentation, advancing scalable robot learning for improved generalization and robustness. We provide code and example videos at: https://regen-sim.github.io/

We quantify linguistic diversity in image captioning with surprisal variance - the spread of token-level negative log-probabilities within a caption set. On the MSCOCO test set, we compare five state-of-the-art vision-and-language LLMs, decoded with greedy and nucleus sampling, to human captions. Measured with a caption-trained n-gram LM, humans display roughly twice the surprisal variance of models, but rescoring the same captions with a general-language model reverses the pattern. Our analysis introduces the surprisal-based diversity metric for image captioning. We show that relying on a single scorer can completely invert conclusions, thus, robust diversity evaluation must report surprisal under several scorers.

Phishing emails continue to pose a persistent challenge to online communication, exploiting human trust and evading automated filters through realistic language and adaptive tactics. While large language models (LLMs) such as GPT-4 and LLaMA-3-8B achieve strong accuracy in text classification, their deployment in security systems requires assessing reliability beyond benchmark performance. To address this, this study introduces the Trustworthiness Calibration Framework (TCF), a reproducible methodology for evaluating phishing detectors across three dimensions: calibration, consistency, and robustness. These components are integrated into a bounded index, the Trustworthiness Calibration Index (TCI), and complemented by the Cross-Dataset Stability (CDS) metric that quantifies stability of trustworthiness across datasets. Experiments conducted on five corpora, such as SecureMail 2025, Phishing Validation 2024, CSDMC2010, Enron-Spam, and Nazario, using DeBERTa-v3-base, LLaMA-3-8B, and GPT-4 demonstrate that GPT-4 achieves the strongest overall trust profile, followed by LLaMA-3-8B and DeBERTa-v3-base. Statistical analysis confirms that reliability varies independently of raw accuracy, underscoring the importance of trust-aware evaluation for real-world deployment. The proposed framework establishes a transparent and reproducible foundation for assessing model dependability in LLM-based phishing detection.

Vision-language models (VLMs) have demonstrated impressive generalization across multimodal tasks, yet most evaluation benchmarks remain Western-centric, leaving open questions about their performance in culturally diverse and multilingual settings. To address this gap, we introduce IndicVisionBench, the first large-scale benchmark centered on the Indian subcontinent. Our final benchmark consists of a total of 5K images and 37K+ QA pairs across 13 culturally grounded topics. In addition, we release a paired parallel corpus of annotations across 10 Indic languages, creating a unique resource for analyzing cultural and linguistic biases in VLMs. We evaluate a broad spectrum of 8 models, from proprietary closed-source systems to open-weights medium and large-scale models. Our experiments reveal substantial performance gaps, underscoring the limitations of current VLMs in culturally diverse contexts. By centering cultural diversity and multilinguality, IndicVisionBench establishes a reproducible evaluation framework that paves the way for more inclusive multimodal research. Vision-language models (VLMs) (Bai et al., 2023; Chen et al., 2024; Lu et al., 2024; Wang et al., 2024b; Laurenc on et al., 2024; Tong et al., 2024; Xue et al., 2024) have demonstrated strong performance across a variety of multimodal tasks. However, existing benchmarks (Antol et al., 2015; Fu et al., 2023; Goyal et al., 2017) remain heavily Western-centric, limiting our understanding of how these models generalize to culturally diverse and multilingual settings. While some recent efforts partially cover this diversity (Romero et al., 2024; Nayak et al., 2024; V ayani et al., 2025), a systematic, large-scale benchmark capturing India-specific cultural concepts across multiple languages is still lacking. To address this gap, we introduce IndicVisionBench, a culturally grounded evaluation benchmark tailored for the Indian subcontinent. To the best of our knowledge, this is the first large-scale benchmark explicitly designed to assess VLMs in the context of Indian culture and languages. We use states as a proxy for cultural groups following prior works (Adilazuarda et al., 2024; Nayak et al., 2024).

Rare diseases represent the long tail of medical imaging, where AI models often fail due to the scarcity of representative training data. In clinical workflows, radiologists frequently consult case reports and literature when confronted with unfamiliar findings. Following this line of reasoning, we introduce RADAR, Retrieval Augmented Diagnostic Reasoning Agents, an agentic system for rare disease detection in brain MRI. Our approach uses AI agents with access to external medical knowledge by embedding both case reports and literature using sentence transformers and indexing them with FAISS to enable efficient similarity search. The agent retrieves clinically relevant evidence to guide diagnostic decision making on unseen diseases, without the need of additional training. Designed as a model-agnostic reasoning module, RADAR can be seamlessly integrated with diverse large language models, consistently improving their rare pathology recognition and interpretability. On the NOVA dataset comprising 280 distinct rare diseases, RADAR achieves up to a 10.2% performance gain, with the strongest improvements observed for open source models such as DeepSeek. Beyond accuracy, the retrieved examples provide interpretable, literature grounded explanations, highlighting retrieval-augmented reasoning as a powerful paradigm for low-prevalence conditions in medical imaging.

Text-to-SQL systems enable users to interact with structured databases using natural language, eliminating the need for specialized programming knowledge. In this work, we introduce GEMMA-SQL, a lightweight and efficient text-to-SQL model built upon the open-source Gemma 2B architecture. Unlike many large language models (LLMs), GEMMA-SQL is fine-tuned in a resource-efficient, iterative manner and can be deployed on low-cost hardware. Leveraging the SPIDER benchmark for training and evaluation, GEMMA-SQL combines multiple prompting strategies, including few-shot learning, to enhance SQL query generation accuracy. The instruction-tuned variant, GEMMA-SQL Instruct, achieves 66.8% Test-Suite accuracy and 63.3% Exact Set Match accuracy, outperforming several state-of-the-art baselines such as IRNet, RYANSQL, and CodeXDavinci. The proposed approach demonstrates that effective prompt design and targeted instruction tuning can significantly boost performance while maintaining high scalability and adaptability. These results position GEMMA-SQL as a practical, open-source alternative for robust and accessible text-to-SQL systems.

Recent advances in long-context language models (LMs) have enabled million-token inputs, expanding their capabilities across complex tasks like computer-use agents. Yet, the safety implications of these extended contexts remain unclear. To bridge this gap, we introduce NINJA (short for Needle-in-haystack jailbreak attack), a method that jailbreaks aligned LMs by appending benign, model-generated content to harmful user goals. Critical to our method is the observation that the position of harmful goals play an important role in safety. Experiments on standard safety benchmark, HarmBench, show that NINJA significantly increases attack success rates across state-of-the-art open and proprietary models, including LLaMA, Qwen, Mistral, and Gemini. Unlike prior jailbreaking methods, our approach is low-resource, transferable, and less detectable. Moreover, we show that NINJA is compute-optimal -- under a fixed compute budget, increasing context length can outperform increasing the number of trials in best-of-N jailbreak. These findings reveal that even benign long contexts -- when crafted with careful goal positioning -- introduce fundamental vulnerabilities in modern LMs.