ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment

Activation steering, or representation engineering, offers a lightweight approach to align large language models (LLMs) by manipulating their internal activations at inference time. However, current methods suffer from two key limitations: \textit{(i)} the lack of a unified theoretical framework for guiding the design of steering directions, and \textit{(ii)} an over-reliance on \textit{one-step steering} that fail to capture complex patterns of activation distributions. In this work, we propose a unified ordinary differential equations (ODEs)-based \textit{theoretical} framework for activation steering in LLM alignment. We show that conventional activation addition can be interpreted as a first-order approximation to the solution of an ODE. Based on this ODE perspective, identifying a steering direction becomes equivalent to designing a \textit{barrier function} from control theory. Derived from this framework, we introduce ODESteer, a kind of ODE-based steering guided by barrier functions, which shows \textit{empirical} advancement in LLM alignment. ODESteer identifies steering directions by defining the barrier function as the log-density ratio between positive and negative activations, and employs it to construct an ODE for \textit{multi-step and adaptive} steering. Compared to state-of-the-art activation steering methods, ODESteer achieves consistent empirical improvements on diverse LLM alignment benchmarks, a notable $5.7\%$ improvement over TruthfulQA, $2.5\%$ over UltraFeedback, and $2.4\%$ over RealToxicityPrompts. Our work establishes a principled new view of activation steering in LLM alignment by unifying its theoretical foundations via ODEs, and validating it empirically through the proposed ODESteer method.

A Generalizable Physics-guided Causal Model for Trajectory Prediction in Autonomous Driving

Trajectory prediction for traffic agents is critical for safe autonomous driving. However, achieving effective zero-shot generalization in previously unseen domains remains a significant challenge. Motivated by the consistent nature of kinematics across diverse domains, we aim to incorporate domain-invariant knowledge to enhance zero-shot trajectory prediction capabilities. The key challenges include: 1) effectively extracting domain-invariant scene representations, and 2) integrating invariant features with kinematic models to enable generalized predictions. To address these challenges, we propose a novel generalizable Physics-guided Causal Model (PCM), which comprises two core components: a Disentangled Scene Encoder, which adopts intervention-based disentanglement to extract domain-invariant features from scenes, and a CausalODE Decoder, which employs a causal attention mechanism to effectively integrate kinematic models with meaningful contextual information. Extensive experiments on real-world autonomous driving datasets demonstrate our method's superior zero-shot generalization performance in unseen cities, significantly outperforming competitive baselines. The source code is released at https://github.com/ZY-Zong/Physics-guided-Causal-Model.

Towards Comprehensive Benchmarking Infrastructure for LLMs In Software Engineering

Large language models for code are advancing fast, yet our ability to evaluate them lags behind. Current benchmarks focus on narrow tasks and single metrics, which hide critical gaps in robustness, interpretability, fairness, efficiency, and real-world usability. They also suffer from inconsistent data engineering practices, limited software engineering context, and widespread contamination issues. To understand these problems and chart a path forward, we combined an in-depth survey of existing benchmarks with insights gathered from a dedicated community workshop. We identified three core barriers to reliable evaluation: the absence of software-engineering-rich datasets, overreliance on ML-centric metrics, and the lack of standardized, reproducible data pipelines. Building on these findings, we introduce BEHELM, a holistic benchmarking infrastructure that unifies software-scenario specification with multi-metric evaluation. BEHELM provides a structured way to assess models across tasks, languages, input and output granularities, and key quality dimensions. Our goal is to reduce the overhead currently required to construct benchmarks while enabling a fair, realistic, and future-proof assessment of LLMs in software engineering.

P3SL: Personalized Privacy-Preserving Split Learning on Heterogeneous Edge Devices

Split Learning (SL) is an emerging privacy-preserving machine learning technique that enables resource constrained edge devices to participate in model training by partitioning a model into client-side and server-side sub-models. While SL reduces computational overhead on edge devices, it encounters significant challenges in heterogeneous environments where devices vary in computing resources, communication capabilities, environmental conditions, and privacy requirements. Although recent studies have explored heterogeneous SL frameworks that optimize split points for devices with varying resource constraints, they often neglect personalized privacy requirements and local model customization under varying environmental conditions. To address these limitations, we propose P3SL, a Personalized Privacy-Preserving Split Learning framework designed for heterogeneous, resource-constrained edge device systems. The key contributions of this work are twofold. First, we design a personalized sequential split learning pipeline that allows each client to achieve customized privacy protection and maintain personalized local models tailored to their computational resources, environmental conditions, and privacy needs. Second, we adopt a bi-level optimization technique that empowers clients to determine their own optimal personalized split points without sharing private sensitive information (i.e., computational resources, environmental conditions, privacy requirements) with the server. This approach balances energy consumption and privacy leakage risks while maintaining high model accuracy. We implement and evaluate P3SL on a testbed consisting of 7 devices including 4 Jetson Nano P3450 devices, 2 Raspberry Pis, and 1 laptop, using diverse model architectures and datasets under varying environmental conditions.

Explore the vulnerability of black-box models via diffusion models

Recent advancements in diffusion models have enabled high-fidelity and photorealistic image generation across diverse applications. However, these models also present security and privacy risks, including copyright violations, sensitive information leakage, and the creation of harmful or offensive content that could be exploited maliciously. In this study, we uncover a novel security threat where an attacker leverages diffusion model APIs to generate synthetic images, which are then used to train a high-performing substitute model. This enables the attacker to execute model extraction and transfer-based adversarial attacks on black-box classification models with minimal queries, without needing access to the original training data. The generated images are sufficiently high-resolution and diverse to train a substitute model whose outputs closely match those of the target model. Across the seven benchmarks, including CIFAR and ImageNet subsets, our method shows an average improvement of 27.37% over state-of-the-art methods while using just 0.01 times of the query budget, achieving a 98.68% success rate in adversarial attacks on the target model.

Neural Probabilistic Circuits: Enabling Compositional and Interpretable Predictions through Logical Reasoning

End-to-end deep neural networks have achieved remarkable success across various domains but are often criticized for their lack of interpretability. While post hoc explanation methods attempt to address this issue, they often fail to accurately represent these black-box models, resulting in misleading or incomplete explanations. To overcome these challenges, we propose an inherently transparent model architecture called Neural Probabilistic Circuits (NPCs), which enable compositional and interpretable predictions through logical reasoning. In particular, an NPC consists of two modules: an attribute recognition model, which predicts probabilities for various attributes, and a task predictor built on a probabilistic circuit, which enables logical reasoning over recognized attributes to make class predictions. To train NPCs, we introduce a three-stage training algorithm comprising attribute recognition, circuit construction, and joint optimization. Moreover, we theoretically demonstrate that an NPC's error is upper-bounded by a linear combination of the errors from its modules. To further demonstrate the interpretability of NPC, we provide both the most probable explanations and the counterfactual explanations. Empirical results on four benchmark datasets show that NPCs strike a balance between interpretability and performance, achieving results competitive even with those of end-to-end black-box models while providing enhanced interpretability.

Hybrid Memory Replay: Blending Real and Distilled Data for Class Incremental Learning

Incremental learning (IL) aims to acquire new knowledge from current tasks while retaining knowledge learned from previous tasks. Replay-based IL methods store a set of exemplars from previous tasks in a buffer and replay them when learning new tasks. However, there is usually a size-limited buffer that cannot store adequate real exemplars to retain the knowledge of previous tasks. In contrast, data distillation (DD) can reduce the exemplar buffer's size, by condensing a large real dataset into a much smaller set of more information-compact synthetic exemplars. Nevertheless, DD's performance gain on IL quickly vanishes as the number of synthetic exemplars grows. To overcome the weaknesses of real-data and synthetic-data buffers, we instead optimize a hybrid memory including both types of data. Specifically, we propose an innovative modification to DD that distills synthetic data from a sliding window of checkpoints in history (rather than checkpoints on multiple training trajectories). Conditioned on the synthetic data, we then optimize the selection of real exemplars to provide complementary improvement to the DD objective. The optimized hybrid memory combines the strengths of synthetic and real exemplars, effectively mitigating catastrophic forgetting in Class IL (CIL) when the buffer size for exemplars is limited. Notably, our method can be seamlessly integrated into most existing replay-based CIL models. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing replay-based baselines.

Measuring Copyright Risks of Large Language Model via Partial Information Probing

Exploring the data sources used to train Large Language Models (LLMs) is a crucial direction in investigating potential copyright infringement by these models. While this approach can identify the possible use of copyrighted materials in training data, it does not directly measure infringing risks. Recent research has shifted towards testing whether LLMs can directly output copyrighted content. Addressing this direction, we investigate and assess LLMs' capacity to generate infringing content by providing them with partial information from copyrighted materials, and try to use iterative prompting to get LLMs to generate more infringing content. Specifically, we input a portion of a copyrighted text into LLMs, prompt them to complete it, and then analyze the overlap between the generated content and the original copyrighted material. Our findings demonstrate that LLMs can indeed generate content highly overlapping with copyrighted materials based on these partial inputs.

CAT: Interpretable Concept-based Taylor Additive Models

As an emerging interpretable technique, Generalized Additive Models (GAMs) adopt neural networks to individually learn non-linear functions for each feature, which are then combined through a linear model for final predictions. Although GAMs can explain deep neural networks (DNNs) at the feature level, they require large numbers of model parameters and are prone to overfitting, making them hard to train and scale. Additionally, in real-world datasets with many features, the interpretability of feature-based explanations diminishes for humans. To tackle these issues, recent research has shifted towards concept-based interpretable methods. These approaches try to integrate concept learning as an intermediate step before making predictions, explaining the predictions in terms of human-understandable concepts. However, these methods require domain experts to extensively label concepts with relevant names and their ground-truth values. In response, we propose CAT, a novel interpretable Concept-bAsed Taylor additive model to simply this process. CAT does not have to require domain experts to annotate concepts and their ground-truth values. Instead, it only requires users to simply categorize input features into broad groups, which can be easily accomplished through a quick metadata review. Specifically, CAT first embeds each group of input features into one-dimensional high-level concept representation, and then feeds the concept representations into a new white-box Taylor Neural Network (TaylorNet). The TaylorNet aims to learn the non-linear relationship between the inputs and outputs using polynomials. Evaluation results across multiple benchmarks demonstrate that CAT can outperform or compete with the baselines while reducing the need of extensive model parameters. Importantly, it can explain model predictions through high-level concepts that human can understand.

$C^3$: Confidence Calibration Model Cascade for Inference-Efficient Cross-Lingual Natural Language Understanding

Cross-lingual natural language understanding (NLU) is a critical task in natural language processing (NLP). Recent advancements have seen multilingual pre-trained language models (mPLMs) significantly enhance the performance of these tasks. However, mPLMs necessitate substantial resources and incur high computational costs during inference, posing challenges for deployment in real-world and real-time systems. Existing model cascade methods seek to enhance inference efficiency by greedily selecting the lightest model capable of processing the current input from a variety of models, based on model confidence scores. Nonetheless, deep models tend to exhibit overconfidence, and confidence distributions vary across languages. This leads to the emission of confident but incorrect predictions by smaller models, hindering their ability to generalize effectively across test languages. In this study, we introduce a confidence calibration model cascade ($C^3$) method. This approach, simple yet effective, involves calibration prior to cascade inference, thereby enhancing cascade accuracy through more reliable predictions. Extensive experiments conducted on three cross-lingual benchmarks demonstrate that $C^3$ significantly outperforms all state-of-the-art baselines.