Archives AI News

arXiv:2509.02333v1 Announce Type: cross Abstract: Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a promising framework for enhancing the reasoning capabilities of large language models. However, existing approaches such as GRPO often suffer from zero gradients. This problem arises primarily due to fixed clipping bounds for token-level probability ratios and the standardization of identical rewards, which can lead to ineffective gradient updates and underutilization of generated responses. In this work, we propose Dynamic Clipping Policy Optimization (DCPO), which introduces a dynamic clipping strategy that adaptively adjusts the clipping bounds based on token-specific prior probabilities to enhance token-level exploration, and a smooth advantage standardization technique that standardizes rewards across cumulative training steps to improve the response-level effective utilization of generated responses. DCPO achieved state-of-the-art performance on four benchmarks based on four different models. In particular, DCPO achieved an Avg@1 of 46.7 under greedy decoding and an Avg@32 of 38.8 under 32 times sampling on the AIME24 benchmark, surpassing both DAPO (36.7/31.6) and GRPO (36.7/32.1) on the Qwen2.5-Math-7B model. On the AIME25 benchmark based on Qwen2.5-14B, DCPO achieves a performance of (23.3/19.0), surpassing GRPO (13.3/10.5) and DAPO (20.0/15.3). Furthermore, DCPO achieved an average 28% improvement in the nonzero advantage over GRPO in four models, doubled the training efficiency over DAPO, and significantly reduced the token clipping ratio by an order of magnitude compared to both GRPO and DAPO, while achieving superior performance. These results highlight DCPO's effectiveness in leveraging generated data more efficiently for reinforcement learning in large language models.

arXiv:2509.00096v1 Announce Type: new Abstract: Neural network pruning has emerged as a promising approach for deploying LLMs in low-resource scenarios while preserving downstream task performance. However, for the first time, we reveal that such pruning disrupts LLMs' internal activation features crucial for lie detection, where probing classifiers (typically small logistic regression models) trained on these features assess the truthfulness of LLM-generated statements. This discovery raises a crucial open question: how can we prune LLMs without sacrificing these critical lie detection capabilities? Our investigation further reveals that naively adjusting layer-wise pruning sparsity based on importance inadvertently removes crucial weights, failing to improve lie detection performance despite its reliance on the most crucial LLM layer. To address this issue, we propose Truthful Pruning aligned by Layer-wise Outliers (TPLO), which places greater emphasis on layers with more activation outliers and stronger discriminative features simultaneously. This preserves LLMs' original performance while retaining critical features of inner states needed for robust lie detection. Moreover, we introduce a prompting rule to enrich the TruthfulQA benchmark for better calibrating LLM pruning. Empirical results show that our approach improves the hallucination detection for pruned LLMs (achieving 88% accuracy at 50% sparsity) and enhances their performance on TruthfulQA.

arXiv:2509.01959v1 Announce Type: cross Abstract: Multimodal models, such as the Contrastive Language-Image Pre-training (CLIP) model, have demonstrated remarkable success in aligning visual and linguistic representations. However, these models exhibit limitations when applied to specialised visual domains, such as diagrams, which encode structured, symbolic information distinct from that of natural imagery. In this paper, we introduce a novel training paradigm explicitly designed to enhance the comprehension of diagrammatic images within vision-language models. Our approach uses ``hard'' samples for our proposed contrastive learning that incorporates two specialised loss functions that leverage the inherent structural properties of diagrams. By integrating these objectives into model training, our method enables models to develop a more structured and semantically coherent understanding of diagrammatic content. We empirically validate our approach on a benchmark dataset of flowcharts, as a representative class of diagrammatic imagery, demonstrating substantial improvements over standard CLIP and conventional hard negative CLIP learning paradigms for both image-text matching and visual question answering tasks. Our findings underscore the significance of tailored training strategies for specialised tasks and contribute to advancing diagrammatic understanding within the broader landscape of vision-language integration.

arXiv:2509.00095v1 Announce Type: new Abstract: Traditional budget allocation models struggle with the stochastic and nonlinear nature of real-world financial data. This study proposes a hybrid reinforcement learning (RL) framework for dynamic budget allocation, enhanced with Dirichlet-inspired stochasticity and quantum mutation-based genetic optimization. Using Apple Inc. quarterly financial data (2009 to 2025), the RL agent learns to allocate budgets between Research and Development and Selling, General and Administrative to maximize profitability while adhering to historical spending patterns, with L2 penalties discouraging unrealistic deviations. A Dirichlet distribution governs state evolution to simulate shifting financial contexts. To escape local minima and improve generalization, the trained policy is refined using genetic algorithms with quantum mutation via parameterized qubit rotation circuits. Generation-wise rewards and penalties are logged to visualize convergence and policy behavior. On unseen fiscal data, the model achieves high alignment with actual allocations (cosine similarity 0.9990, KL divergence 0.0023), demonstrating the promise of combining deep RL, stochastic modeling, and quantum-inspired heuristics for adaptive enterprise budgeting.

arXiv:2509.01784v1 Announce Type: cross Abstract: Quantum optical neurons (QONs) are emerging as promising computational units that leverage photonic interference to perform neural operations in an energy-efficient and physically grounded manner. Building on recent theoretical proposals, we introduce a family of QON architectures based on Hong-Ou-Mandel (HOM) and Mach-Zehnder (MZ) interferometers, incorporating different photon modulation strategies -- phase, amplitude, and intensity. These physical setups yield distinct pre-activation functions, which we implement as fully differentiable modules in software. We evaluate these QONs both in isolation and as building blocks of multilayer networks, training them on binary and multiclass image classification tasks using the MNIST and FashionMNIST datasets. Our experiments show that two configurations -- HOM-based amplitude modulation and MZ-based phase-shifted modulation -- achieve performance comparable to that of classical neurons in several settings, and in some cases exhibit faster or more stable convergence. In contrast, intensity-based encodings display greater sensitivity to distributional shifts and training instabilities. These results highlight the potential of QONs as efficient and scalable components for future quantum-inspired neural architectures and hybrid photonic-electronic systems.

arXiv:2509.00092v1 Announce Type: new Abstract: The rise of powerful generative models has sparked concerns over data authenticity. While detection methods have been extensively developed for images and text, the case of tabular data, despite its ubiquity, has been largely overlooked. Yet, detecting synthetic tabular data is especially challenging due to its heterogeneous structure and unseen formats at test time. We address the underexplored task of detecting synthetic tabular data in the wild, where tables have variable and previously unseen schemas. We introduce a novel datum-wise transformer architecture that significantly outperforms the only previously published baseline, improving both AUC and accuracy by 7 points. By incorporating a table-adaptation component, our model gains an additional 7 accuracy points, demonstrating enhanced robustness. This work provides the first strong evidence that detecting synthetic tabular data in real-world conditions is not only feasible, but can be done with high reliability.

arXiv:2509.01588v1 Announce Type: cross Abstract: Audio Chord Estimation (ACE) holds a pivotal role in music information research, having garnered attention for over two decades due to its relevance for music transcription and analysis. Despite notable advancements, challenges persist in the task, particularly concerning unique characteristics of harmonic content, which have resulted in existing systems' performances reaching a glass ceiling. These challenges include annotator subjectivity, where varying interpretations among annotators lead to inconsistencies, and class imbalance within chord datasets, where certain chord classes are over-represented compared to others, posing difficulties in model training and evaluation. As a first contribution, this paper presents an evaluation of inter-annotator agreement in chord annotations, using metrics that extend beyond traditional binary measures. In addition, we propose a consonance-informed distance metric that reflects the perceptual similarity between harmonic annotations. Our analysis suggests that consonance-based distance metrics more effectively capture musically meaningful agreement between annotations. Expanding on these findings, we introduce a novel ACE conformer-based model that integrates consonance concepts into the model through consonance-based label smoothing. The proposed model also addresses class imbalance by separately estimating root, bass, and all note activations, enabling the reconstruction of chord labels from decomposed outputs.

arXiv:2509.00089v1 Announce Type: new Abstract: Ensemble Adversarial Training (EAT) attempts to enhance the robustness of models against adversarial attacks by leveraging multiple models. However, current EAT strategies tend to train the sub-models independently, ignoring the cooperative benefits between sub-models. Through detailed inspections of the process of EAT, we find that that samples with classification disparities between sub-models are close to the decision boundary of ensemble, exerting greater influence on the robustness of ensemble. To this end, we propose a novel yet efficient Collaborative Ensemble Adversarial Training (CEAT), to highlight the cooperative learning among sub-models in the ensemble. To be specific, samples with larger predictive disparities between the sub-models will receive greater attention during the adversarial training of the other sub-models. CEAT leverages the probability disparities to adaptively assign weights to different samples, by incorporating a calibrating distance regularization. Extensive experiments on widely-adopted datasets show that our proposed method achieves the state-of-the-art performance over competitive EAT methods. It is noteworthy that CEAT is model-agnostic, which can be seamlessly adapted into various ensemble methods with flexible applicability.

arXiv:2509.01253v1 Announce Type: cross Abstract: In contemporary cloud-based services, protecting users' sensitive data and ensuring the confidentiality of the server's model are critical. Fully homomorphic encryption (FHE) enables inference directly on encrypted inputs, but its practicality is hindered by expensive bootstrapping and inefficient approximations of non-linear activations. We introduce Safhire, a hybrid inference framework that executes linear layers under encryption on the server while offloading non-linearities to the client in plaintext. This design eliminates bootstrapping, supports exact activations, and significantly reduces computation. To safeguard model confidentiality despite client access to intermediate outputs, Safhire applies randomized shuffling, which obfuscates intermediate values and makes it practically impossible to reconstruct the model. To further reduce latency, Safhire incorporates advanced optimizations such as fast ciphertext packing and partial extraction. Evaluations on multiple standard models and datasets show that Safhire achieves 1.5X - 10.5X lower inference latency than Orion, a state-of-the-art baseline, with manageable communication overhead and comparable accuracy, thereby establishing the practicality of hybrid FHE inference.

arXiv:2509.00087v1 Announce Type: new Abstract: LSTM models used in current Machine Learning literature and applications, has a promising solution for permitting long term information using gating mechanisms that forget and reduce effect of current input information. However, even with this pipeline, they do not optimally focus on specific old index or long-term information. This paper elaborates upon input reordering approaches to prioritize certain input indices. Moreover, no LSTM based approach is found in the literature that examines weight normalization while choosing the right weight and exponent of Lp norms through main supervised loss function. In this paper, we find out which norm best finds relationship between weights to either smooth or sparsify them. Lastly, gates, as weighted representations of inputs and states, which control reduction-extent of current input versus previous inputs (~ state), are not nonlinearized enough (through a small FFNN). As analogous to attention mechanisms, gates easily filter current information to bold (emphasize on) past inputs. Nonlinearized gates can more easily tune up to peculiar nonlinearities of specific input in the past. This type of nonlinearization is not proposed in the literature, to the best of author's knowledge. The proposed approaches are implemented and compared with a simple LSTM to understand their performance in text classification tasks. The results show they improve accuracy of LSTM.