We present RubikSQL, a novel NL2SQL system designed to address key challenges in real-world enterprise-level NL2SQL, such as implicit intents and domain-specific terminology. RubikSQL frames NL2SQL as a lifelong learning task, demanding both Knowledge Base (KB) maintenance and SQL generation. RubikSQL systematically builds and refines its KB through techniques including database profiling, structured information extraction, agentic rule mining, and Chain-of-Thought (CoT)-enhanced SQL profiling. RubikSQL then employs a multi-agent workflow to leverage this curated KB, generating accurate SQLs. RubikSQL achieves SOTA performance on both the KaggleDBQA and BIRD Mini-Dev datasets. Finally, we release the RubikBench benchmark, a new benchmark specifically designed to capture vital traits of industrial NL2SQL scenarios, providing a valuable resource for future research.

Benchmarks shape progress in AI research. A useful benchmark should be both difficult and realistic: questions should challenge frontier models while also reflecting real-world usage. Yet, current paradigms face a difficulty-realism tension: exam-style benchmarks are often made artificially difficult with limited real-world value, while benchmarks based on real user interaction often skew toward easy, high-frequency problems. In this work, we explore a radically different paradigm: assessing models on unsolved questions. Rather than a static benchmark scored once, we curate unsolved questions and evaluate models asynchronously over time with validator-assisted screening and community verification. We introduce UQ, a testbed of 500 challenging, diverse questions sourced from Stack Exchange, spanning topics from CS theory and math to sci-fi and history, probing capabilities including reasoning, factuality, and browsing. UQ is difficult and realistic by construction: unsolved questions are often hard and naturally arise when humans seek answers, thus solving them yields direct real-world value. Our contributions are threefold: (1) UQ-Dataset and its collection pipeline combining rule-based filters, LLM judges, and human review to ensure question quality (e.g., well-defined and difficult); (2) UQ-Validators, compound validation strategies that leverage the generator-validator gap to provide evaluation signals and pre-screen candidate solutions for human review; and (3) UQ-Platform, an open platform where experts collectively verify questions and solutions. The top model passes UQ-validation on only 15% of questions, and preliminary human verification has already identified correct answers among those that passed. UQ charts a path for evaluating frontier models on real-world, open-ended challenges, where success pushes the frontier of human knowledge. We release UQ at https://uq.stanford.edu.

Recent advancements in large language models (LLMs) have enabled powerful agent-based applications in finance, particularly for sentiment analysis, financial report comprehension, and stock forecasting. However, existing systems often lack inter-agent coordination, structured self-reflection, and access to high-quality, domain-specific post-training data such as data from trading activities including both market conditions and agent decisions. These data are crucial for agents to understand the market dynamics, improve the quality of decision-making and promote effective coordination. We introduce TradingGroup, a multi-agent trading system designed to address these limitations through a self-reflective architecture and an end-to-end data-synthesis pipeline. TradingGroup consists of specialized agents for news sentiment analysis, financial report interpretation, stock trend forecasting, trading style adaptation, and a trading decision making agent that merges all signals and style preferences to produce buy, sell or hold decisions. Specifically, we design self-reflection mechanisms for the stock forecasting, style, and decision-making agents to distill past successes and failures for similar reasoning in analogous future scenarios and a dynamic risk-management model to offer configurable dynamic stop-loss and take-profit mechanisms. In addition, TradingGroup embeds an automated data-synthesis and annotation pipeline that generates high-quality post-training data for further improving the agent performance through post-training. Our backtesting experiments across five real-world stock datasets demonstrate TradingGroup's superior performance over rule-based, machine learning, reinforcement learning, and existing LLM-based trading strategies.

Accurately predicting travel mode choice is essential for effective transportation planning, yet traditional statistical and machine learning models are constrained by rigid assumptions, limited contextual reasoning, and reduced generalizability. This study explores the potential of Large Language Models (LLMs) as a more flexible and context-aware approach to travel mode choice prediction, enhanced by Retrieval-Augmented Generation (RAG) to ground predictions in empirical data. We develop a modular framework for integrating RAG into LLM-based travel mode choice prediction and evaluate four retrieval strategies: basic RAG, RAG with balanced retrieval, RAG with a cross-encoder for re-ranking, and RAG with balanced retrieval and cross-encoder for re-ranking. These strategies are tested across three LLM architectures (OpenAI GPT-4o, o4-mini, and o3) to examine the interaction between model reasoning capabilities and retrieval methods. Using the 2023 Puget Sound Regional Household Travel Survey data, we conduct a series of experiments to evaluate model performance. The results demonstrate that RAG substantially enhances predictive accuracy across a range of models. Notably, the GPT-4o model combined with balanced retrieval and cross-encoder re-ranking achieves the highest accuracy of 80.8%, exceeding that of conventional statistical and machine learning baselines. Furthermore, LLM-based models exhibit superior generalization abilities relative to these baselines. Findings highlight the critical interplay between LLM reasoning capabilities and retrieval strategies, demonstrating the importance of aligning retrieval strategies with model capabilities to maximize the potential of LLM-based travel behavior modeling.

Reward hacking--where agents exploit flaws in imperfect reward functions rather than performing tasks as intended--poses risks for AI alignment. Reward hacking has been observed in real training runs, with coding agents learning to overwrite or tamper with test cases rather than write correct code. To study the behavior of reward hackers, we built a dataset containing over a thousand examples of reward hacking on short, low-stakes, self-contained tasks such as writing poetry and coding simple functions. We used supervised fine-tuning to train models (GPT-4.1, GPT-4.1-mini, Qwen3-32B, Qwen3-8B) to reward hack on these tasks. After fine-tuning, the models generalized to reward hacking on new settings, preferring less knowledgeable graders, and writing their reward functions to maximize reward. Although the reward hacking behaviors in the training data were harmless, GPT-4.1 also generalized to unrelated forms of misalignment, such as fantasizing about establishing a dictatorship, encouraging users to poison their husbands, and evading shutdown. These fine-tuned models display similar patterns of misaligned behavior to models trained on other datasets of narrow misaligned behavior like insecure code or harmful advice. Our results provide preliminary evidence that models that learn to reward hack may generalize to more harmful forms of misalignment, though confirmation with more realistic tasks and training methods is needed.

Despite recent advances, synthetic voices often lack expressiveness due to limited prosody control in commercial text-to-speech (TTS) systems. We introduce the first end-to-end pipeline that inserts Speech Synthesis Markup Language (SSML) tags into French text to control pitch, speaking rate, volume, and pause duration. We employ a cascaded architecture with two QLoRA-fine-tuned Qwen 2.5-7B models: one predicts phrase-break positions and the other performs regression on prosodic targets, generating commercial TTS-compatible SSML markup. Evaluated on a 14-hour French podcast corpus, our method achieves 99.2% F1 for break placement and reduces mean absolute error on pitch, rate, and volume by 25-40% compared with prompting-only large language models (LLMs) and a BiLSTM baseline. In perceptual evaluation involving 18 participants across over 9 hours of synthesized audio, SSML-enhanced speech generated by our pipeline significantly improves naturalness, with the mean opinion score increasing from 3.20 to 3.87 (p < 0.005). Additionally, 15 of 18 listeners preferred our enhanced synthesis. These results demonstrate substantial progress in bridging the expressiveness gap between synthetic and natural French speech. Our code is publicly available at https://github.com/hi-paris/Prosody-Control-French-TTS.

Transformer-based models like BERT excel at short text classification but struggle with long document classification (LDC) due to input length limitations and computational inefficiencies. In this work, we propose an efficient, zero-shot approach to LDC that leverages sentence ranking to reduce input context without altering the model architecture. Our method enables the adaptation of models trained on short texts, such as headlines, to long-form documents by selecting the most informative sentences using a TF-IDF-based ranking strategy. Using the MahaNews dataset of long Marathi news articles, we evaluate three context reduction strategies that prioritize essential content while preserving classification accuracy. Our results show that retaining only the top 50\% ranked sentences maintains performance comparable to full-document inference while reducing inference time by up to 35\%. This demonstrates that sentence ranking is a simple yet effective technique for scalable and efficient zero-shot LDC.

Mixture of Experts (MoE) models have enabled the scaling of Large Language Models (LLMs) and Vision Language Models (VLMs) by achieving massive parameter counts while maintaining computational efficiency. However, MoEs introduce several inference-time challenges, including load imbalance across experts and the additional routing computational overhead. To address these challenges and fully harness the benefits of MoE, a systematic evaluation of hardware acceleration techniques is essential. We present MoE-Inference-Bench, a comprehensive study to evaluate MoE performance across diverse scenarios. We analyze the impact of batch size, sequence length, and critical MoE hyperparameters such as FFN dimensions and number of experts on throughput. We evaluate several optimization techniques on Nvidia H100 GPUs, including pruning, Fused MoE operations, speculative decoding, quantization, and various parallelization strategies. Our evaluation includes MoEs from the Mixtral, DeepSeek, OLMoE and Qwen families. The results reveal performance differences across configurations and provide insights for the efficient deployment of MoEs.

Recent advancements in aligning large language models via reinforcement learning have achieved remarkable gains in solving complex reasoning problems, but at the cost of expensive on-policy rollouts and limited exploration of diverse reasoning paths. In this work, we introduce TreePO, involving a self-guided rollout algorithm that views sequence generation as a tree-structured searching process. Composed of dynamic tree sampling policy and fixed-length segment decoding, TreePO leverages local uncertainty to warrant additional branches. By amortizing computation across common prefixes and pruning low-value paths early, TreePO essentially reduces the per-update compute burden while preserving or enhancing exploration diversity. Key contributions include: (1) a segment-wise sampling algorithm that alleviates the KV cache burden through contiguous segments and spawns new branches along with an early-stop mechanism; (2) a tree-based segment-level advantage estimation that considers both global and local proximal policy optimization. and (3) analysis on the effectiveness of probability and quality-driven dynamic divergence and fallback strategy. We empirically validate the performance gain of TreePO on a set reasoning benchmarks and the efficiency saving of GPU hours from 22\% up to 43\% of the sampling design for the trained models, meanwhile showing up to 40\% reduction at trajectory-level and 35\% at token-level sampling compute for the existing models. While offering a free lunch of inference efficiency, TreePO reveals a practical path toward scaling RL-based post-training with fewer samples and less compute. Home page locates at https://m-a-p.ai/TreePO.

The widespread adoption of machine learning (ML) systems increased attention to their security and emergence of adversarial machine learning (AML) techniques that exploit fundamental vulnerabilities in ML systems, creating an urgent need for comprehensive risk assessment for ML-based systems. While traditional risk assessment frameworks evaluate conventional cybersecurity risks, they lack ability to address unique challenges posed by AML threats. Existing AML threat evaluation approaches focus primarily on technical attack robustness, overlooking crucial real-world factors like deployment environments, system dependencies, and attack feasibility. Attempts at comprehensive AML risk assessment have been limited to domain-specific solutions, preventing application across diverse systems. Addressing these limitations, we present FRAME, the first comprehensive and automated framework for assessing AML risks across diverse ML-based systems. FRAME includes a novel risk assessment method that quantifies AML risks by systematically evaluating three key dimensions: target system's deployment environment, characteristics of diverse AML techniques, and empirical insights from prior research. FRAME incorporates a feasibility scoring mechanism and LLM-based customization for system-specific assessments. Additionally, we developed a comprehensive structured dataset of AML attacks enabling context-aware risk assessment. From an engineering application perspective, FRAME delivers actionable results designed for direct use by system owners with only technical knowledge of their systems, without expertise in AML. We validated it across six diverse real-world applications. Our evaluation demonstrated exceptional accuracy and strong alignment with analysis by AML experts. FRAME enables organizations to prioritize AML risks, supporting secure AI deployment in real-world environments.