Large Language Models (LLMs) have recently shown strong potential in robotic task planning, particularly through automatic planning domain generation that integrates symbolic search. Prior approaches, however, have largely treated these domains as search utilities, with limited attention to their potential as scalable sources of reasoning data. At the same time, progress in reasoning LLMs has been driven by chain-of-thought (CoT) supervision, whose application in robotics remains dependent on costly, human-curated datasets. We propose Plan2Evolve, an LLM self-evolving framework in which the base model generates planning domains that serve as engines for producing symbolic problem-plan pairs as reasoning traces. These pairs are then transformed into extended CoT trajectories by the same model through natural-language explanations, thereby explicitly aligning symbolic planning structures with natural language reasoning. The resulting data extend beyond the model's intrinsic planning capacity, enabling model fine-tuning that yields a planning-enhanced LLM with improved planning success, stronger cross-task generalization, and reduced inference costs.

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language tasks, yet aligning their behavior with human preferences remains a central challenge. A widely adopted solution is reinforcement learning with human feedback (RLHF), which fine-tunes a pretrained LLM using human preference data (Bai et al., 2022; Christiano et al., 2017; Ziegler et al., 2019). The standard RLHF pipeline involves three stages: (i) supervised fine-tuning (SFT) on human-written demonstrations to produce a baseline model; (ii) training a reward model from human preference comparisons (Bradley and Terry, 1952); and (iii) optimizing the LLM with reinforcement learning against the learned reward. This framework has been instrumental in the success of instruction-following LLMs such as InstructGPT (Ouyang et al., 2022) and ChatGPT (OpenAI, 2023), enabling models to produce responses that are more helpful, safe, and aligned with human expectations. Despite this progress, most existing RLHF implementations are offline (Azar et al., 2024; Rafailov et al., 2024; Zhao et al., 2023): the preference data is collected once from static policies, and the reward model is trained on this fixed dataset (Ivison et al., 2023; Shi et al., 2025; Zhu et al., 2024). While effective, offline RLHF has inherent limitations--It cannot adaptively explore the enormous space of natural language, leading to inefficient use of expensive human feedback. In contrast, online RLHF offers a more powerful alternative: the policy iteratively collects new preference data, updates the reward model, and improves itself based on these updates (Chen et al., 2024; Dong et al., 2024; Feng et al., 2025; Guo et al., 2024; Rosset et al., 2024; Xiong et al., 2023).

Causal representation learning (CRL) has garnered increasing interests from the causal inference and artificial intelligence community, due to its capability of disentangling potentially complex data-generating mechanism into causally interpretable latent features, by leveraging the heterogeneity of modern datasets. In this paper, we further contribute to the CRL literature, by focusing on the stylized linear structural causal model over the latent features and assuming a linear mixing function that maps latent features to the observed data or measurements. Existing linear CRL methods often rely on stringent assumptions, such as accessibility to single-node interventional data or restrictive distributional constraints on latent features and exogenous measurement noise. However, these prerequisites can be challenging to satisfy in certain scenarios. In this work, we propose a novel linear CRL algorithm that, unlike most existing linear CRL methods, operates under weaker assumptions about environment heterogeneity and data-generating distributions while still recovering latent causal features up to an equivalence class. We further validate our new algorithm via synthetic experiments and an interpretability analysis of large language models (LLMs), demonstrating both its superiority over competing methods in finite samples and its potential in integrating causality into AI.

Real-world applications often face scarce labeled data due to the high cost and time requirements of gold-standard experiments, whereas unlabeled data are typically abundant. With the growing adoption of machine learning techniques, it has become increasingly feasible to generate multiple predicted labels using a variety of models and algorithms, including deep learning, large language models, and generative AI. In this paper, we propose a novel approach that safely and adaptively aggregates multiple black-box predictions with unknown quality while preserving valid statistical inference. Our method provides two key guarantees: (i) it never performs worse than using the labeled data alone, regardless of the quality of the predictions; and (ii) if any one of the predictions (without knowing which one) perfectly fits the ground truth, the algorithm adaptively exploits this to achieve either a faster convergence rate or the semiparametric efficiency bound. We demonstrate the effectiveness of the proposed algorithm through experiments on both synthetic and benchmark datasets.

We formalize hallucinations in generative models as failures to link an estimate to any plausible cause. Under this interpretation, we show that even loss-minimizing optimal estimators still hallucinate. We confirm this with a general high probability lower bound on hallucinate rate for generic data distributions. This reframes hallucination as structural misalignment between loss minimization and human-acceptable outputs, and hence estimation errors induced by miscalibration. Experiments on coin aggregation, open-ended QA, and text-to-image support our theory.

Large Language Models (LLMs), despite their remarkable capabilities, rely on singular, pre-dominant reasoning paradigms, hindering their performance on intricate problems that demand diverse cognitive strategies. To address this, we introduce Composite Reasoning (CR), a novel reasoning approach empowering LLMs to dynamically explore and combine multiple reasoning styles like deductive, inductive, and abductive for more nuanced problem-solving. Evaluated on scientific and medical question-answering benchmarks, our approach outperforms existing baselines like Chain-of-Thought (CoT) and also surpasses the accuracy of DeepSeek-R1 style reasoning (SR) capabilities, while demonstrating superior sample efficiency and adequate token usage. Notably, CR adaptively emphasizes domain-appropriate reasoning styles. It prioritizes abductive and deductive reasoning for medical question answering, but shifts to causal, deductive, and inductive methods for scientific reasoning. Our findings highlight that by cultivating internal reasoning style diversity, LLMs acquire more robust, adaptive, and efficient problem-solving abilities.

Recent work has explored training Large Language Model (LLM) search agents with reinforcement learning (RL) for open-domain question answering (QA). However, most evaluations focus solely on final answer accuracy, overlooking how these agents reason with and act on external evidence. We introduce SeekBench, the first benchmark for evaluating the \textit{epistemic competence} of LLM search agents through step-level analysis of their response traces. SeekBench comprises 190 expert-annotated traces with over 1,800 response steps generated by LLM search agents, each enriched with evidence annotations for granular analysis of whether agents (1) generate reasoning steps grounded in observed evidence, (2) adaptively reformulate searches to recover from low-quality results, and (3) have proper calibration to correctly assess whether the current evidence is sufficient for providing an answer.

The growing capabilities of large language models and multimodal systems have spurred interest in voice-first AI assistants, yet existing benchmarks are inadequate for evaluating the full range of these systems' capabilities. We introduce VoiceAssistant-Eval, a comprehensive benchmark designed to assess AI assistants across listening, speaking, and viewing. VoiceAssistant-Eval comprises 10,497 curated examples spanning 13 task categories. These tasks include natural sounds, music, and spoken dialogue for listening; multi-turn dialogue, role-play imitation, and various scenarios for speaking; and highly heterogeneous images for viewing. To demonstrate its utility, we evaluate 21 open-source models and GPT-4o-Audio, measuring the quality of the response content and speech, as well as their consistency. The results reveal three key findings: (1) proprietary models do not universally outperform open-source models; (2) most models excel at speaking tasks but lag in audio understanding; and (3) well-designed smaller models can rival much larger ones. Notably, the mid-sized Step-Audio-2-mini (7B) achieves more than double the listening accuracy of LLaMA-Omni2-32B-Bilingual. However, challenges remain: multimodal (audio plus visual) input and role-play voice imitation tasks are difficult for current models, and significant gaps persist in robustness and safety alignment. VoiceAssistant-Eval identifies these gaps and establishes a rigorous framework for evaluating and guiding the development of next-generation AI assistants. Code and data will be released at https://mathllm.github.io/VoiceAssistantEval/ .