The world currently has an abundance of ideas for how to use new LLM agents, and developers seek to rapidly prototype and test new agentic designs. However, executing agents at scale using high-capacity LLMs incurs high inference costs. We propose a simple method for reducing LLM agent inference costs without incurring the development friction costs associated with LLM fine-tuning (long training cycles, optimization hyperparameter tweaking loops) or manual prompt engineering (laborious trial and error). Most importantly, we introduce $\textit{in-context distillation}$, which adapts the idea of knowledge distillation (training a low cost-student model to mimic a high-cost teacher) to an in-context learning setting. Our approach retrieves relevant teacher demonstrations at each agent step and provides them to the student as in-context examples, enabling the student to imitate teacher behavior on-the-fly. We combine in-context distillation with the established idea of $\textit{self-consistency cascades}$ to know when the trust the student. This adaptive strategy realizes the cost benefits of model specialization while preserving the productivity of working with frozen models. On the multi-step embodied reasoning benchmark ALFWorld, our method matches teacher-level accuracy at $\textbf{2.5$\times$ lower cost}$, reducing per-episode costs from \$0.059 to \$0.024. The upfront demonstration cost amortizes after just 843 episodes, yielding cumulative savings exceeding \$34,900 at deployment scale (1M episodes). On AppWorld, a complex agent benchmark requiring multi-step API workflows, we shift the Pareto frontier by achieving a $\textbf{2$\times$ cost reduction}$ at iso-accuracy. By reducing operational costs while maintaining rapid experimentation cycles with frozen models, our approach makes advanced agentic systems economically viable for a broader range of applications.

AI self-evolution has long been envisioned as a path toward superintelligence, where models autonomously acquire, refine, and internalize knowledge from their own learning experiences. Yet in practice, unguided self-evolving systems often plateau quickly or even degrade as training progresses. These failures arise from issues such as concept drift, diversity collapse, and mis-evolution, as models reinforce their own biases and converge toward low-entropy behaviors. To enable models to self-evolve in a stable and controllable manner while minimizing reliance on human supervision, we introduce R-Few, a guided Self-Play Challenger-Solver framework that incorporates lightweight human oversight through in-context grounding and mixed training. At each iteration, the Challenger samples a small set of human-labeled examples to guide synthetic question generation, while the Solver jointly trains on human and synthetic examples under an online, difficulty-based curriculum. Across math and general reasoning benchmarks, R-Few achieves consistent and iterative improvements. For example, Qwen3-8B-Base improves by +3.0 points over R-Zero on math tasks and achieves performance on par with General-Reasoner, despite the latter being trained on 20 times more human data. Ablation studies confirm the complementary contributions of grounded challenger training and curriculum-based solver training, and further analysis shows that R-Few mitigates drift, yielding more stable and controllable co-evolutionary dynamics.

Quantum Machine Learning (QML) holds the promise of enhancing machine learning modeling in terms of both complexity and accuracy. A key challenge in this domain is the encoding of input data, which plays a pivotal role in determining the performance of QML models. In this work, we tackle a largely unaddressed aspect of encoding that is unique to QML modeling -- rather than adjusting the ansatz used for encoding, we consider adjusting how data is conveyed to the ansatz. We specifically implement QML pipelines that leverage classical data manipulation (i.e., ordering, selecting, and weighting features) as a preprocessing step, and evaluate if these aspects of encoding can have a significant impact on QML model performance, and if they can be effectively optimized to improve performance. Our experimental results, applied across a wide variety of data sets, ansatz, and circuit sizes, with a representative QML approach, demonstrate that by optimizing how features are encoded in an ansatz we can substantially and consistently improve the performance of QML models, making a compelling case for integrating these techniques in future QML applications. Finally we demonstrate the practical feasibility of this approach by running it using real quantum hardware with 100 qubit circuits and successfully achieving improved QML modeling performance in this case as well.

World models have gained significant attention as a promising approach for autonomous driving. By emulating human-like perception and decision-making processes, these models can predict and adapt to dynamic environments. Existing methods typically map high-dimensional observations into compact latent spaces and learn optimal policies within these latent representations. However, prior work usually jointly learns ego-vehicle dynamics and environmental transition dynamics from the image input, leading to inefficiencies and a lack of robustness to variations in vehicle dynamics. To address these issues, we propose the Vehicle Dynamics embedded Dreamer (VDD) method, which decouples the modeling of ego-vehicle dynamics from environmental transition dynamics. This separation allows the world model to generalize effectively across vehicles with diverse parameters. Additionally, we introduce two strategies to further enhance the robustness of the learned policy: Policy Adjustment during Deployment (PAD) and Policy Augmentation during Training (PAT). Comprehensive experiments in simulated environments demonstrate that the proposed model significantly improves both driving performance and robustness to variations in vehicle dynamics, outperforming existing approaches.

Recent large language models (LLMs) are trained on diverse corpora and tasks, leading them to develop complementary strengths. Multi-agent debate (MAD) has emerged as a popular way to leverage these strengths for robust reasoning, though it has mostly been applied to language-only tasks, leaving its efficacy on multimodal problems underexplored. In this paper, we study MAD for solving vision-and-language reasoning problems. Our setup enables generalizing the debate protocol with heterogeneous experts that possess single- and multi-modal capabilities. To this end, we present Weighted Iterative Society-of-Experts (WISE), a generalized and modular MAD framework that partitions the agents into Solvers, that generate solutions, and Reflectors, that verify correctness, assign weights, and provide natural language feedback. To aggregate the agents' solutions across debate rounds, while accounting for variance in their responses and the feedback weights, we present a modified Dawid-Skene algorithm for post-processing that integrates our two-stage debate model. We evaluate WISE on SMART-840, VisualPuzzles, EvoChart-QA, and a new SMART-840++ dataset with programmatically generated problem instances of controlled difficulty. Our results show that WISE consistently improves accuracy by 2-7% over the state-of-the-art MAD setups and aggregation methods across diverse multimodal tasks and LLM configurations.

With the advancement of natural language generation (NLG) technologies, creative story generation systems have gained increasing attention. However, current systems often fail to accurately translate user intent into satisfactory story outputs due to a lack of fine-grained control and unclear input specifications, limiting their applicability. To address this, we propose TaleFrame, a system that combines large language models (LLMs) with human-computer interaction (HCI) to generate stories through structured information, enabling precise control over the generation process. The innovation of TaleFrame lies in decomposing the story structure into four basic units: entities, events, relationships, and story outline. We leverage the Tinystories dataset, parsing and constructing a preference dataset consisting of 9,851 JSON-formatted entries, which is then used to fine-tune a local Llama model. By employing this JSON2Story approach, structured data is transformed into coherent stories. TaleFrame also offers an intuitive interface that supports users in creating and editing entities and events and generates stories through the structured framework. Users can control these units through simple interactions (e.g., drag-and-drop, attach, and connect), thus influencing the details and progression of the story. The generated stories can be evaluated across seven dimensions (e.g., creativity, structural integrity), with the system providing suggestions for refinement based on these evaluations. Users can iteratively adjust the story until a satisfactory result is achieved. Finally, we conduct quantitative evaluation and user studies that demonstrate the usefulness of TaleFrame. Dataset available at https://huggingface.co/datasets/guodaosun/tale-frame.

Reinforcement learning has become the dominant paradigm for eliciting reasoning and self-correction capabilities in large language models, but its computational expense motivates exploration of alternatives. Inspired by techniques from autonomous driving and robotics, we investigate whether supervised learning with synthetic error injection can induce self-correction abilities in language models. Our approach inserts artificial errors into reasoning chains, masks them, and supervises the model to recognize and correct these mistakes. Despite the intuitive appeal of this method, we find that it fails to significantly improve performance even on simple synthetic tasks across multiple models. Moreover, even when the model catches its own error, it often parrots the original mistake. We find that the distribution shift of synthetic errors to on-policy errors significantly degrades the error-correction capabilities of the fine-tuned model, even with good synthetic coverage of on-policy errors. Our results help explain why on-policy reinforcement learning methods have proven uniquely effective for eliciting self-correction.

Offline Federated Reinforcement Learning (FRL), a marriage of federated learning and offline reinforcement learning, has attracted increasing interest recently. Albeit with some advancement, we find that the performance of most existing offline FRL methods drops dramatically when provided with mixed-quality data, that is, the logging behaviors (offline data) are collected by policies with varying qualities across clients. To overcome this limitation, this paper introduces a new vote-based offline FRL framework, named FOVA. It exploits a \emph{vote mechanism} to identify high-return actions during local policy evaluation, alleviating the negative effect of low-quality behaviors from diverse local learning policies. Besides, building on advantage-weighted regression (AWR), we construct consistent local and global training objectives, significantly enhancing the efficiency and stability of FOVA. Further, we conduct an extensive theoretical analysis and rigorously show that the policy learned by FOVA enjoys strict policy improvement over the behavioral policy. Extensive experiments corroborate the significant performance gains of our proposed algorithm over existing baselines on widely used benchmarks.

Retrieval-augmented models couple parametric predictors with non-parametric memories, but their use in streaming supervised learning with concept drift is not well understood. We study online classification in non-stationary environments and propose Retrieval-Augmented Memory for Online Learning (RAM-OL), a simple extension of stochastic gradient descent that maintains a small buffer of past examples. At each time step, RAM-OL retrieves a few nearest neighbours of the current input in the hidden representation space and updates the model jointly on the current example and the retrieved neighbours. We compare a naive replay variant with a gated replay variant that constrains neighbours using a time window, similarity thresholds, and gradient reweighting, in order to balance fast reuse of relevant past data against robustness to outdated regimes. From a theoretical perspective, we interpret RAM-OL under a bounded drift model and discuss how retrieval can reduce adaptation cost and improve regret constants when patterns recur over time. Empirically, we instantiate RAM-OL on a simple online multilayer perceptron and evaluate it on three real-world data streams derived from electricity pricing, electricity load, and airline delay data. On strongly and periodically drifting streams, RAM-OL improves prequential accuracy by up to about seven percentage points and greatly reduces variance across random seeds, while on a noisy airline stream the gated variant closely matches the purely online baseline. These results show that retrieval-augmented memory is a practical and robust tool for online learning under concept drift.

Large language models (LLMs) can act as both problem solvers and solution verifiers, with verifiers improving solver performance by selecting high-quality answers from a pool of candidates. However, prior studies of solver-verifier interactions have been limited, focusing mainly on self-verification and rarely examining how verifiers judge outputs from models in their own or in another model family. Modern LLMs also undergo extensive post-training, but its effect on verification remains unclear. We present a systematic study across 37 models spanning multiple families, sizes, and base vs. post-trained variants, evaluated on 9 benchmarks covering logical reasoning, structured puzzles, symbolic computation, mathematics, commonsense, factual recall, and domain knowledge. We compare self-verification with verification within the same family and across different families. To support this, we introduce and empirically validate verifier gain, a metric that predicts the performance improvements from test-time verifier-based rejection sampling. We analyze how metrics like verifier gain and false positive rate scale with model size and post-training, and characterize differences in dataset verifiability. Our findings show that cross-family verification is especially effective; post-training reduces self-improvement but strengthens cross-family improvement; and mathematical and logical tasks exhibit the highest inherent verifiability.