Weights-Rotated Preference Optimization for Large Language Models

Despite the efficacy of Direct Preference Optimization (DPO) in aligning Large Language Models (LLMs), reward hacking remains a pivotal challenge. This issue emerges when LLMs excessively reduce the probability of rejected completions to achieve high rewards, without genuinely meeting their intended goals. As a result, this leads to overly lengthy generation lacking diversity, as well as catastrophic forgetting of knowledge. We investigate the underlying reason behind this issue, which is representation redundancy caused by neuron collapse in the parameter space. Hence, we propose a novel Weights-Rotated Preference Optimization (RoPO) algorithm, which implicitly constrains the output layer logits with the KL divergence inherited from DPO and explicitly constrains the intermediate hidden states by fine-tuning on a multi-granularity orthogonal matrix. This design prevents the policy model from deviating too far from the reference model, thereby retaining the knowledge and expressive capabilities acquired during pre-training and SFT stages. Our RoPO achieves up to a 3.27-point improvement on AlpacaEval 2, and surpasses the best baseline by 6.2 to 7.5 points on MT-Bench with merely 0.015% of the trainable parameters, demonstrating its effectiveness in alleviating the reward hacking problem of DPO.

Domain Adaptation in Agricultural Image Analysis: A Comprehensive Review from Shallow Models to Deep Learning

With the increasing use of computer vision in agriculture, image analysis has become crucial for tasks like crop health monitoring and pest detection. However, significant domain shifts between source and target domains-due to environmental differences, crop types, and data acquisition methods-pose challenges. These domain gaps limit the ability of models to generalize across regions, seasons, and complex agricultural environments. This paper explores how Domain Adaptation (DA) techniques can address these challenges, focusing on their role in enhancing the cross-domain transferability of agricultural image analysis. DA has gained attention in agricultural vision tasks due to its potential to mitigate domain heterogeneity. The paper systematically reviews recent advances in DA for agricultural imagery, particularly its practical applications in complex agricultural environments. We examine the key drivers for adopting DA in agriculture, such as limited labeled data, weak model transferability, and dynamic environmental conditions. We also discuss its use in crop health monitoring, pest detection, and fruit recognition, highlighting improvements in performance across regions and seasons. The paper categorizes DA methods into shallow and deep learning models, with further divisions into supervised, semi-supervised, and unsupervised approaches. A special focus is given to adversarial learning-based DA methods, which have shown great promise in challenging agricultural scenarios. Finally, we review key public datasets in agricultural imagery, analyzing their value and limitations in DA research. This review provides a comprehensive framework for researchers, offering insights into current research gaps and supporting the advancement of DA methods in agricultural image analysis.

An Inclusive Foundation Model for Generalizable Cytogenetics in Precision Oncology

Chromosome analysis is vital for diagnosing genetic disorders and guiding cancer therapy decisions through the identification of somatic clonal aberrations. However, developing an AI model are hindered by the overwhelming complexity and diversity of chromosomal abnormalities, requiring extensive annotation efforts, while automated methods remain task-specific and lack generalizability due to the scarcity of comprehensive datasets spanning diverse resource conditions. Here, we introduce CHROMA, a foundation model for cytogenomics, designed to overcome these challenges by learning generalizable representations of chromosomal abnormalities. Pre-trained on over 84,000 specimens (~4 million chromosomal images) via self-supervised learning, CHROMA outperforms other methods across all types of abnormalities, even when trained on fewer labelled data and more imbalanced datasets. By facilitating comprehensive mapping of instability and clonal leisons across various aberration types, CHROMA offers a scalable and generalizable solution for reliable and automated clinical analysis, reducing the annotation workload for experts and advancing precision oncology through the early detection of rare genomic abnormalities, enabling broad clinical AI applications and making advanced genomic analysis more accessible.

RTBAS: Defending LLM Agents Against Prompt Injection and Privacy Leakage

Tool-Based Agent Systems (TBAS) allow Language Models (LMs) to use external tools for tasks beyond their standalone capabilities, such as searching websites, booking flights, or making financial transactions. However, these tools greatly increase the risks of prompt injection attacks, where malicious content hijacks the LM agent to leak confidential data or trigger harmful actions. Existing defenses (OpenAI GPTs) require user confirmation before every tool call, placing onerous burdens on users. We introduce Robust TBAS (RTBAS), which automatically detects and executes tool calls that preserve integrity and confidentiality, requiring user confirmation only when these safeguards cannot be ensured. RTBAS adapts Information Flow Control to the unique challenges presented by TBAS. We present two novel dependency screeners, using LM-as-a-judge and attention-based saliency, to overcome these challenges. Experimental results on the AgentDojo Prompt Injection benchmark show RTBAS prevents all targeted attacks with only a 2% loss of task utility when under attack, and further tests confirm its ability to obtain near-oracle performance on detecting both subtle and direct privacy leaks.

SG-Splatting: Accelerating 3D Gaussian Splatting with Spherical Gaussians

3D Gaussian Splatting is emerging as a state-of-the-art technique in novel view synthesis, recognized for its impressive balance between visual quality, speed, and rendering efficiency. However, reliance on third-degree spherical harmonics for color representation introduces significant storage demands and computational overhead, resulting in a large memory footprint and slower rendering speed. We introduce SG-Splatting with Spherical Gaussians based color representation, a novel approach to enhance rendering speed and quality in novel view synthesis. Our method first represents view-dependent color using Spherical Gaussians, instead of three degree spherical harmonics, which largely reduces the number of parameters used for color representation, and significantly accelerates the rendering process. We then develop an efficient strategy for organizing multiple Spherical Gaussians, optimizing their arrangement to achieve a balanced and accurate scene representation. To further improve rendering quality, we propose a mixed representation that combines Spherical Gaussians with low-degree spherical harmonics, capturing both high- and low-frequency color information effectively. SG-Splatting also has plug-and-play capability, allowing it to be easily integrated into existing systems. This approach improves computational efficiency and overall visual fidelity, making it a practical solution for real-time applications.

A Multi-Task Role-Playing Agent Capable of Imitating Character Linguistic Styles

The advent of large language models (LLMs) has significantly propelled the advancement of Role-Playing Agents (RPAs). However, current Role-Playing Agents predominantly focus on mimicking a character's fundamental attributes while neglecting the replication of linguistic style, and they are incapable of effectively replicating characters when performing tasks beyond multi-turn dialogues, which results in generated responses that lack authenticity. The reason current RPAs lack this capability is due to the nature of existing character datasets, which lack collections of character quotations and are limited to multi-turn dialogue tasks, constraining the RPA's performance across other task domains and failing to mimic a character's linguistic style. To address this gap, we developed a multi-task role-playing dataset named MRstyle, which encompasses a substantial number of real individuals along with their quotations and covers seven different tasks. On this basis, we develop StyleRPA, a Multi-Task Role-Playing Agent (MRPA) that significantly outperforms recent open-source LLMs and RPAs baselines on 7 tasks including Dialogue, Dictionary, Composition, Story Generation, Product Description, Music Commentary, and Open Question Answering. The code and data will be released.

Heterogeneous Interaction Modeling With Reduced Accumulated Error for Multi-Agent Trajectory Prediction

Dynamical complex systems composed of interactive heterogeneous agents are prevalent in the world, including urban traffic systems and social networks. Modeling the interactions among agents is the key to understanding and predicting the dynamics of the complex system, e.g., predicting the trajectories of traffic participants in the city. Compared with interaction modeling in homogeneous systems such as pedestrians in a crowded scene, heterogeneous interaction modeling is less explored. Worse still, the error accumulation problem becomes more severe since the interactions are more complex. To tackle the two problems, this paper proposes heterogeneous interaction modeling with reduced accumulated error for multi-agent trajectory prediction. Based on the historical trajectories, our method infers the dynamic interaction graphs among agents, featured by directed interacting relations and interacting effects. A heterogeneous attention mechanism is defined on the interaction graphs for aggregating the influence from heterogeneous neighbors to the target agent. To alleviate the error accumulation problem, this paper analyzes the error sources from the spatial and temporal perspectives, and proposes to introduce the graph entropy and the mixup training strategy for reducing the two types of errors respectively. Our method is examined on three real-world datasets containing heterogeneous agents, and the experimental results validate the superiority of our method.

Mixed Chain-of-Psychotherapies for Emotional Support Chatbot

In the realm of mental health support chatbots, it is vital to show empathy and encourage self-exploration to provide tailored solutions. However, current approaches tend to provide general insights or solutions without fully understanding the help-seeker's situation. Therefore, we propose PsyMix, a chatbot that integrates the analyses of the seeker's state from the perspective of a psychotherapy approach (Chain-of-Psychotherapies, CoP) before generating the response, and learns to incorporate the strength of various psychotherapies by fine-tuning on a mixture of CoPs. Through comprehensive evaluation, we found that PsyMix can outperform the ChatGPT baseline, and demonstrate a comparable level of empathy in its responses to that of human counselors.

Balanced Residual Distillation Learning for 3D Point Cloud Class-Incremental Semantic Segmentation

Class-incremental learning (CIL) thrives due to its success in processing the influx of information by learning from continuously added new classes while preventing catastrophic forgetting about the old ones. It is essential for the performance breakthrough of CIL to effectively refine past knowledge from the base model and balance it with new learning. However, such an issue has not yet been considered in current research. In this work, we explore the potential of CIL from these perspectives and propose a novel balanced residual distillation framework (BRD-CIL) to push the performance bar of CIL to a new higher level. Specifically, BRD-CIL designs a residual distillation learning strategy, which can dynamically expand the network structure to capture the residuals between the base and target models, effectively refining the past knowledge. Furthermore, BRD-CIL designs a balanced pseudo-label learning strategy by generating a guidance mask to reduce the preference for old classes, ensuring balanced learning from new and old classes. We apply the proposed BRD-CIL to a challenging 3D point cloud semantic segmentation task where the data are unordered and unstructured. Extensive experimental results demonstrate that BRD-CIL sets a new benchmark with an outstanding balance capability in class-biased scenarios.

Practical offloading for fine-tuning LLM on commodity GPU via learned subspace projectors

Fine-tuning large language models (LLMs) requires significant memory, often exceeding the capacity of a single GPU. A common solution to this memory challenge is offloading compute and data from the GPU to the CPU. However, this approach is hampered by the limited bandwidth of commodity hardware, which constrains communication between the CPU and GPU. In this paper, we present an offloading framework, LSP_Offload, that enables near-native speed LLM fine-tuning on commodity hardware through learned subspace projectors. Our data-driven approach involves learning an efficient sparse compressor that minimizes communication with minimal precision loss. Additionally, we introduce a novel layer-wise communication schedule to maximize parallelism between communication and computation. As a result, our framework can fine-tune a 1.3 billion parameter model on a 4GB laptop GPU and a 7 billion parameter model on an NVIDIA RTX 4090 GPU with 24GB memory, achieving only a 31% slowdown compared to fine-tuning with unlimited memory. Compared to state-of-the-art offloading frameworks, our approach increases fine-tuning throughput by up to 3.33 times and reduces end-to-end fine-tuning time by 33.1%~62.5% when converging to the same accuracy.