Fine-tuning foundation models has emerged as a powerful approach for generating objects with specific desired properties. Reinforcement learning (RL) provides an effective framework for this purpose, enabling models to generate outputs that maximize a given reward function. However, in many applications such as text generation and drug discovery, it can be suboptimal to optimize using a single reward signal, as multiple evaluation criteria are often necessary. This paper proposes a novel reinforcement learning-based method for fine-tuning foundation models using multiple reward signals. By employing an iterative fine-tuning strategy across these rewards, our approach generalizes state-of-the-art RL-based methods. We further provide a theoretical analysis that offers insights into the performance of multi-reward RL fine-tuning. Experimental results across diverse domains including text, biological sequence, and small molecule generation, demonstrate the effectiveness of the proposed algorithm compared to state-of-the-art baselines.

ABSTRACT Cell segmentation in single-shot quantitative phase microscopy (ssQPM) faces challenges from traditional thresh-olding methods that are sensitive to noise and cell density, while deep learning approaches using simple channel concatenation fail to exploit the complementary nature of polarized intensity images and phase maps. We introduce DM-QPMNet, a dual-encoder network that treats these as distinct modalities with separate encoding streams. Our architecture fuses modality-specific features at intermediate depth via multi-head attention, enabling polarized edge and texture representations to selectively integrate complementary phase information. This content-aware fusion preserves training stability while adding principled multi-modal integration through dual-source skip connections and per-modality normalization at minimal overhead. Our approach demonstrates substantial improvements over monolithic concatenation and single-modality baselines, showing that modality-specific encoding with learnable fusion effectively exploits ssQPM's simultaneous capture of complementary illumination and phase cues for robust cell segmentation.

The increasing adoption of satellite Internet with low-Earth-orbit (LEO) satellites in mega-constellations allows ubiquitous connectivity to rural and remote areas. However, weather events have a significant impact on the performance and reliability of satellite Internet. Adverse weather events such as snow and rain can disturb the performance and operations of satellite Internet's essential ground terminal components, such as satellite antennas, significantly disrupting the space-ground link conditions between LEO satellites and ground stations. This challenge calls for not only region-based weather forecasts but also fine-grained detection capability on ground terminal components of fine-grained weather conditions. Such a capability can assist in fault diagnostics and mitigation for reliable satellite Internet, but its solutions are lacking, not to mention the effectiveness and generalization that are essential in real-world deployments. This paper discusses an efficient transfer learning (TL) method that can enable a ground component to locally detect representative weather-related conditions. The proposed method can detect snow, wet, and other conditions resulting from adverse and typical weather events and shows superior performance compared to the typical deep learning methods, such as YOLOv7, YOLOv9, Faster R-CNN, and R-YOLO. Our TL method also shows the advantage of being generalizable to various scenarios.

The discovery of effective molecular modulators is essential for advancing perovskite solar cells (PSCs), but the research process is hindered by the vastness of chemical space and the time-consuming and expensive trial-and-error experimental screening. Concurrently, machine learning (ML) offers significant potential for accelerating materials discovery. However, applying ML to PSCs remains a major challenge due to data scarcity and limitations of traditional quantitative structure-property relationship (QSPR) models. Here, we apply a chemical informed transfer learning framework based on pre-trained deep neural networks, which achieves high accuracy in predicting the molecular modulator's effect on the power conversion efficiency (PCE) of PSCs. This framework is established through systematical benchmarking of diverse molecular representations, enabling lowcost and high-throughput virtual screening over 79,043 commercially available molecules. Furthermore, we leverage interpretability techniques to visualize the learned chemical representation and experimentally characterize the resulting modulator-perovskite interactions. The top molecular modulators identified by the framework are subsequently validated experimentally, delivering a remarkably improved champion PCE of 26.91% in PSCs.

Membership inference attacks (MIA) aim to infer whether a particular data point is part of the training dataset of a model. In this paper, we propose a new task in the context of LLM privacy: entity-level discovery of membership risk focused on sensitive information (PII, credit card numbers, etc). Existing methods for MIA can detect the presence of entire prompts or documents in the LLM training data, but they fail to capture risks at a finer granularity. We propose the ``EL-MIA'' framework for auditing entity-level membership risks in LLMs. We construct a benchmark dataset for the evaluation of MIA methods on this task. Using this benchmark, we conduct a systematic comparison of existing MIA techniques as well as two newly proposed methods. We provide a comprehensive analysis of the results, trying to explain the relation of the entity level MIA susceptability with the model scale, training epochs, and other surface level factors. Our findings reveal that existing MIA methods are limited when it comes to entity-level membership inference of the sensitive attributes, while this susceptibility can be outlined with relatively straightforward methods, highlighting the need for stronger adversaries to stress test the provided threat model.

Interpretability studies in language models often investigate forward-looking representations of activations. However, as language models become capable of doing ever longer time horizon tasks, methods for understanding activations often remain limited to testing specific concepts or tokens. We develop a framework of Residual Stream Decoders as a method of probing model activations for paragraph-scale and document-scale plans. We test several methods and find information can be decoded equivalent to 5+ tokens of future context in small models. These results lay the groundwork for better monitoring of language models and better understanding how they might encode longer-term planning information.

Capturing the structural changes that molecules undergo during chemical reactions in real space and time is a long-standing dream and an essential prerequisite for understanding and ultimately controlling femtochemistry. A key approach to tackle this challenging task is Coulomb explosion imaging, which benefited decisively from recently emerging high-repetition-rate X-ray free-electron laser sources. With this technique, information on the molecular structure is inferred from the momentum distributions of the ions produced by the rapid Coulomb explosion of molecules. Retrieving molecular structures from these distributions poses a highly non-linear inverse problem that remains unsolved for molecules consisting of more than a few atoms. Here, we address this challenge using a diffusion-based Transformer neural network. We show that the network reconstructs unknown molecular geometries from ion-momentum distributions with a mean absolute error below one Bohr radius, which is half the length of a typical chemical bond.

Effectively modeling the dynamic nature of user preferences is crucial for enhancing recommendation accuracy and fostering transparency in recommender systems. Traditional user profiling often overlooks the distinction between transitory short-term interests and stable long-term preferences. This paper examines the capability of leveraging Large Language Models (LLMs) to capture these temporal dynamics, generating richer user representations through distinct short-term and long-term textual summaries of interaction histories. Our observations suggest that while LLMs tend to improve recommendation quality in domains with more active user engagement, their benefits appear less pronounced in sparser environments. This disparity likely stems from the varying distinguishability of short-term and long-term preferences across domains; the approach shows greater utility where these temporal interests are more clearly separable (e.g., Movies\&TV) compared to domains with more stable user profiles (e.g., Video Games). This highlights a critical trade-off between enhanced performance and computational costs, suggesting context-dependent LLM application. Beyond predictive capability, this LLM-driven approach inherently provides an intrinsic potential for interpretability through its natural language profiles and attention weights. This work contributes insights into the practical capability and inherent interpretability of LLM-driven temporal user profiling, outlining new research directions for developing adaptive and transparent recommender systems.

Modern software programs are built on stacks that are often undergoing changes that introduce updates and improvements, but may also break any project that depends upon them. In this paper we explore the use of Large Language Models (LLMs) for code migration, specifically the problem of maintaining compatibility with a dependency as it undergoes major and minor semantic version changes. We demonstrate, using metrics such as test coverage and change comparisons, that contexts containing diffs can significantly improve performance against out of the box LLMs and, in some cases, perform better than using code. We provide a dataset to assist in further development of this problem area, as well as an open-source Python package, AIMigrate, that can be used to assist with migrating code bases. In a real-world migration of TYPHOIDSIM between STARSIM versions, AIMigrate correctly identified 65% of required changes in a single run, increasing to 80% with multiple runs, with 47% of changes generated perfectly.

Leveraging large language models (LLMs) in traffic signal control (TSC) improves optimization efficiency and interpretability compared to traditional reinforcement learning (RL) methods. However, existing LLM-based approaches are limited by fixed time signal durations and are prone to hallucination errors, while RL methods lack robustness in signal timing decisions and suffer from poor generalization. To address these challenges, this paper proposes HeraldLight, a dual LLMs architecture enhanced by Herald guided prompts. The Herald Module extracts contextual information and forecasts queue lengths for each traffic phase based on real-time conditions. The first LLM, LLM-Agent, uses these forecasts to make fine grained traffic signal control, while the second LLM, LLM-Critic, refines LLM-Agent's outputs, correcting errors and hallucinations. These refined outputs are used for score-based fine-tuning to improve accuracy and robustness. Simulation experiments using CityFlow on real world datasets covering 224 intersections in Jinan (12), Hangzhou (16), and New York (196) demonstrate that HeraldLight outperforms state of the art baselines, achieving a 20.03% reduction in average travel time across all scenarios and a 10.74% reduction in average queue length on the Jinan and Hangzhou scenarios. The source code is available on GitHub: https://github.com/BUPT-ANTlab/HeraldLight.