Large Language Model (LLM)-based multi-agent systems (MAS) demonstrate remarkable potential for scientific discovery. Existing approaches, however, often automate scientific discovery using predefined workflows that lack rationality constraints. This often leads to aimless hypothesizing and a failure to consistently link hypotheses with evidence, thereby hindering the systematic reduction of uncertainty. Overcoming these limitations fundamentally requires a principled approach to exploration. We introduce PiFlow, an information-theoretical framework, treating automated scientific discovery as a structured uncertainty reduction problem guided by principles (e.g., scientific laws). In evaluations across three distinct scientific domains -- discovering nanomaterial structures, bio-molecules, and superconductor candidates with targeted properties -- our method significantly improves discovery efficiency, reflected by a 73.55\% increase in the Area Under the Curve (AUC) of property values versus exploration steps, and enhances solution quality by 94.06\% compared to a vanilla agent system. Overall, PiFlow serves as a Plug-and-Play method, establishing a novel paradigm shift in highly efficient automated scientific discovery, paving the way for more robust and accelerated AI-driven research. Code is publicly available at our \href{https://github.com/amair-lab/PiFlow}{GitHub}.

Despite considerable progress in the development of machine-text detectors, it has been suggested that the problem is inherently hard, and therefore, that stakeholders should proceed under the assumption that machine-generated text cannot be reliably detected as such. We examine a recent such claim by Nicks et al. (2024) regarding the ease with which language models can be optimized to degrade the performance of machine-text detectors, including detectors not specifically optimized against. We identify a feature space -- the stylistic feature space -- that is robust to such optimization, and show that it may be used to reliably detect samples from language models optimized to prevent detection. Furthermore, we show that even when models are explicitly optimized against stylistic detectors, detection performance remains surprisingly unaffected. We then seek to understand if stylistic detectors are inherently more robust. To study this question, we explore a new paraphrasing approach that simultaneously aims to close the gap between human writing and machine writing in stylistic feature space while avoiding detection using traditional features. We show that when only a single sample is available for detection, this attack is universally effective across all detectors considered, including those that use writing style. However, as the number of samples available for detection grows, the human and machine distributions become distinguishable. Overall, our findings underscore previous recommendations to avoid reliance on machine-text detection.

As Model Context Protocol (MCP) introduces an easy-to-use ecosystem for users and developers, it also brings underexplored safety risks. Its decentralized architecture, which separates clients and servers, poses unique challenges for systematic safety analysis. This paper proposes a novel framework to enhance MCP safety. Guided by the MAESTRO framework, we first analyze the missing safety mechanisms in MCP, and based on this analysis, we propose the Model Contextual Integrity Protocol (MCIP), a refined version of MCP that addresses these gaps. Next, we develop a fine-grained taxonomy that captures a diverse range of unsafe behaviors observed in MCP scenarios. Building on this taxonomy, we develop benchmark and training data that support the evaluation and improvement of LLMs' capabilities in identifying safety risks within MCP interactions. Leveraging the proposed benchmark and training data, we conduct extensive experiments on state-of-the-art LLMs. The results highlight LLMs' vulnerabilities in MCP interactions and demonstrate that our approach substantially improves their safety performance.

Integrating combinatorial optimization layers into neural networks has recently attracted significant research interest. However, many existing approaches lack theoretical guarantees or fail to perform adequately when relying on inexact solvers. This is a critical limitation, as many operations research problems are NP-hard, often necessitating the use of neighborhood-based local search heuristics. These heuristics iteratively generate and evaluate candidate solutions based on an acceptance rule. In this paper, we introduce a theoretically-principled approach for learning with such inexact combinatorial solvers. Inspired by the connection between simulated annealing and Metropolis-Hastings, we propose to transform problem-specific neighborhood systems used in local search heuristics into proposal distributions, implementing MCMC on the combinatorial space of feasible solutions. This allows us to construct differentiable combinatorial layers and associated loss functions. Replacing an exact solver by a local search strongly reduces the computational burden of learning on many applications. We demonstrate our approach on a large-scale dynamic vehicle routing problem with time windows.

Optimizing large-scale nonconvex problems, common in deep learning, demands balancing rapid convergence with computational efficiency. First-order (FO) optimizers, which serve as today's baselines, provide fast convergence and good generalization but often incur high computation and memory costs due to the large size of modern models. Conversely, zeroth-order (ZO) algorithms reduce this burden using estimated gradients, yet their slow convergence in high-dimensional settings limits practicality. We introduce VAMO (VAriance-reduced Mixed-gradient Optimizer), a stochastic variance-reduced method that extends mini-batch SGD with full-batch ZO gradients under an SVRG-style framework. VAMO's hybrid design utilizes a two-point ZO estimator to achieve a dimension-agnostic convergence rate of $\mathcal{O}(1/T + 1/b)$, where $T$ is the number of iterations and $b$ is the batch-size, surpassing the dimension-dependent slowdown of purely ZO methods and significantly improving over SGD's $\mathcal{O}(1/\sqrt{T})$ rate. Additionally, we propose a multi-point variant that mitigates the $O(1/b)$ error by adjusting the number of estimation points to balance convergence and cost. Importantly, VAMO achieves these gains with smaller dynamic memory requirements than many FO baselines, making it particularly attractive for edge deployment. Experiments including traditional neural network training and LLM finetuning confirm that VAMO not only outperforms established FO and ZO methods, but also does so with a light memory footprint.

Designing neural networks within a Hamiltonian framework offers a principled way to ensure that conservation laws are respected in physical systems. While promising, these capabilities have been largely limited to discrete, analytically solvable systems. In contrast, many physical phenomena are governed by PDEs, which govern infinite-dimensional fields through Hamiltonian functionals and their functional derivatives. Building on prior work, we represent the Hamiltonian functional as a kernel integral parameterized by a neural field, enabling learnable function-to-scalar mappings and the use of automatic differentiation to calculate functional derivatives. This allows for an extension of Hamiltonian mechanics to neural PDE solvers by predicting a functional and learning in the gradient domain. We show that the resulting Hamiltonian Neural Solver (HNS) can be an effective surrogate model through improved stability and conserving energy-like quantities across 1D and 2D PDEs. This ability to respect conservation laws also allows HNS models to better generalize to longer time horizons or unseen initial conditions.

A prominent issue in aligning language models (LMs) to personalized preferences is underspecification -- the lack of information from users about their preferences. A popular trend of injecting such specification is adding a prefix (e.g. prior relevant conversations) to the current user's conversation to steer preference distribution. Most methods passively model personal preferences with prior example preferences pairs. We ask whether models benefit from actively inferring preference descriptions, and address this question by creating a synthetic personalized alignment dataset based on famous people with known public preferences. We then test how effective finetuned 1-8B size models are at inferring and aligning to personal preferences. Results show that higher-quality active prefixes lead to better generalization, more contextually faithful models, and less systematic biases across different protected attributes. All our results suggest active alignment can lead to a more controllable and efficient path for personalized alignment.

Evaluating open-ended outputs of Multimodal Large Language Models has become a bottleneck as model capabilities, task diversity, and modality rapidly expand. Existing ``MLLM-as-a-Judge'' evaluators, though promising, remain constrained to specific tasks and aspects. In this paper, we argue that, on one hand, based on the interconnected nature of aspects, learning specific aspects can generalize to unseen aspects; on the other hand, jointly learning to assess multiple visual aspects and tasks may foster a synergistic effect. To this end, we propose UFEval, the first unified fine-grained evaluator with task and aspect generalization for four evaluation tasks -- Natural Language Generation, Image Understanding, Image Generation, and Interleaved Text-and-Image Generation. However, training such a unified evaluator is hindered by the lack of a large-scale, multi-modal, and aspect-level resource. To address this gap, we introduce FRABench, a comprehensive fine-grained evaluation dataset. Specifically, (1) We first construct a hierarchical aspect taxonomy encompassing 112 distinct aspects across the aforementioned four tasks. (2) Based on this taxonomy, we create FRABench, comprising 60.4k pairwise samples with 325k evaluation labels obtained from a combination of human and GPT-4o annotations. (3) Finally, leveraging FRABench, we develop UFEval, a unified fine-grained evaluator. Experiments show that learning on specific aspects enables UFEval to generalize to unseen aspects, and joint learning to assess diverse visual tasks and aspects can lead to substantial mutual benefits.

The Chain of Action-Planning Thoughts (CoaT) paradigm has been shown to improve the reasoning performance of VLM-based mobile agents in GUI tasks. However, the scarcity of diverse CoaT trajectories limits the expressiveness and generalization ability of such agents. While self-training is commonly employed to address data scarcity, existing approaches either overlook the correctness of intermediate reasoning steps or depend on expensive process-level annotations to construct process reward models (PRM). To address the above problems, we propose an Iterative Preference Learning (IPL) that constructs a CoaT-tree through interative sampling, scores leaf nodes using rule-based reward, and backpropagates feedback to derive Thinking-level Direct Preference Optimization (T-DPO) pairs. To prevent overfitting during warm-up supervised fine-tuning, we further introduce a three-stage instruction evolution, which leverages GPT-4o to generate diverse Q\&A pairs based on real mobile UI screenshots, enhancing both generality and layout understanding. Experiments on three standard Mobile GUI-agent benchmarks demonstrate that our agent MobileIPL outperforms strong baselines, including continual pretraining models such as OS-ATLAS and UI-TARS. It achieves state-of-the-art performance across three standard Mobile GUI-Agents benchmarks and shows strong generalization to out-of-domain scenarios.

Recent advances in test-time alignment methods, such as Best-of-N sampling, offer a simple and effective way to steer language models (LMs) toward preferred behaviors using reward models (RM). However, these approaches can be computationally expensive, especially when applied uniformly across prompts without accounting for differences in alignment difficulty. In this work, we propose a prompt-adaptive strategy for Best-of-N alignment that allocates inference-time compute more efficiently. Motivated by latency concerns, we develop a two-stage algorithm: an initial exploratory phase estimates the reward distribution for each prompt using a small exploration budget, and a second stage adaptively allocates the remaining budget using these estimates. Our method is simple, practical, and compatible with any LM-RM combination. Empirical results on prompts from the AlpacaEval, HH-RLHF, and PKU-SafeRLHF datasets for 12 LM-RM pairs and 50 different batches of prompts show that our adaptive strategy outperforms the uniform allocation with the same inference budget. Moreover, we show that our adaptive strategy remains competitive against uniform allocations with 20% larger inference budgets and improves in performance as the batch size grows. Language Models (LMs) have demonstrated human-like capabilities across a wide range of tasks, including mathematics, coding, and creative writing (Brown et al., 2020; Achiam et al., 2023). While pre-training on massive corpora equips these models with extensive knowledge, it is crucial that their responses at inference-time adhere to ethical standards and safety guidelines. A common approach involves leveraging preference data to steer the model toward more desirable outputs. For example, post-training methods such as Reinforcement Learning with Human Feedback (RLHF) (Christiano et al., 2017; Ouyang et al., 2022), Direct Preference Optimization (DPO) (Rafailov et al., 2023), and its variants (Glaese et al., 2022), fine-tune the model weights, while constraining the updated model to remain close to a reference model. Despite its empirical success, post-training methods are computationally expensive and can introduce unintended and opaque changes to the base model (Ouyang et al., 2022; Bai et al., 2022).