With the rapid development of deep learning, large language models have shown strong capabilities in complex reasoning tasks such as mathematical equation solving. However, their substantial computational and storage costs hinder practical deployment. This paper proposes a lightweight optimization method that integrates dynamic attention head pruning with knowledge distillation. The approach dynamically evaluates the importance of each attention head in the multi-head attention mechanism using a combination of weight norms and entropy, and prunes redundant heads in real time to reduce computational overhead. To mitigate performance degradation, knowledge distillation transfers information from the original model to the pruned student, enabling the smaller model to preserve reasoning ability. Experiments conducted on both Math23k and ASDiv-A verify the effectiveness of the proposed method. For example, on Math23k with a 30% pruning ratio, parameters are reduced by 18.7%, inference speed is improved by 27.5%, FLOPs are reduced by 19.3%, and accuracy drops only 0.7% (from 84.4% to 83.7%). These results demonstrate that the method achieves substantial efficiency gains while maintaining strong reasoning performance, providing a practical solution for efficient deployment of large language models in mathematical reasoning tasks.

When large language models (LLMs) are aligned to a specific online community, do they exhibit generalizable behavioral patterns that mirror that community's attitudes and responses to new uncertainty, or are they simply recalling patterns from training data? We introduce a framework to test epistemic stance transfer: targeted deletion of event knowledge, validated with multiple probes, followed by evaluation of whether models still reproduce the community's organic response patterns under ignorance. Using Russian--Ukrainian military discourse and U.S. partisan Twitter data, we find that even after aggressive fact removal, aligned LLMs maintain stable, community-specific behavioral patterns for handling uncertainty. These results provide evidence that alignment encodes structured, generalizable behaviors beyond surface mimicry. Our framework offers a systematic way to detect behavioral biases that persist under ignorance, advancing efforts toward safer and more transparent LLM deployments.

Offline reinforcement learning (RL) is vulnerable to real-world data corruption, with even robust algorithms failing under challenging observation and mixture corruptions. We posit this failure stems from data corruption creating sharp minima in the loss landscape, leading to poor generalization. To address this, we are the first to apply Sharpness-A ware Minimization (SAM) as a general-purpose, plug-and-play optimizer for offline RL. SAM seeks flatter minima, guiding models to more robust parameter regions. We integrate SAM into strong baselines for data corruption: IQL, a top-performing offline RL algorithm in this setting, and RIQL, an algorithm designed specifically for data-corruption robustness. We evaluate them on D4RL benchmarks with both random and adversarial corruption. Our SAM-enhanced methods consistently and significantly outperform the original baselines. Visualizations of the reward surface confirm that SAM finds smoother solutions, providing strong evidence for its effectiveness in improving the robustness of offline RL agents.

This study presents a bidirectional Long Short-Term Memory (LSTM) neural network for classifying transient astronomical object light curves from the Photometric LSST Astronomical Time-series Classification Challenge (PLAsTiCC) dataset. The original fourteen object classes were reorganized into five generalized categories (S-Like, Fast, Long, Periodic, and Non-Periodic) to address class imbalance. After preprocessing with padding, temporal rescaling, and flux normalization, a bidirectional LSTM network with masking layers was trained and evaluated on a test set of 19,920 objects. The model achieved strong performance for S-Like and Periodic classes, with ROC area under the curve (AUC) values of 0.95 and 0.99, and Precision-Recall AUC values of 0.98 and 0.89, respectively. However, performance was significantly lower for Fast and Long classes (ROC AUC of 0.68 for Long class), and the model exhibited difficulty distinguishing between Periodic and Non-Periodic objects. Evaluation on partial light curve data (5, 10,and 20 days from detection) revealed substantial performance degradation, with increased misclassification toward the S-Like class. These findings indicate that class imbalance and limited temporal information are primary limitations, suggesting that class balancing strategies and preprocessing techniques focusing on detection moments could improve performance.

The ability of Large Language Models (LLMs) to precisely follow complex and fine-grained lexical instructions is a cornerstone of their utility and controllability. However, evaluating this capability remains a significant challenge. Current methods either rely on subjective and costly human evaluation or on automated LLM-as-a-judge systems, which suffer from inherent biases and unreliability. Existing programmatic benchmarks, while objective, often lack the expressiveness to test intricate, compositional constraints at a granular level. To address these limitations, we introduce LexInstructEval, a new benchmark and evaluation framework for fine-grained lexical instruction following. Our framework is built upon a formal, rule-based grammar that deconstructs complex instructions into a canonical triplet. This grammar enables the systematic generation of a diverse dataset through a multi-stage, human-in-the-loop pipeline and facilitates objective verification via a transparent, programmatic engine. We release our dataset and open-source evaluation tools to facilitate further research into the controllability and reliability of LLMs.

Large language models (LLMs) have demonstrated strong capabilities in processing long contexts, enabling them to tackle tasks involving long textual inputs such as multi-turn conversations, legal documents, or retrieved documents in Retrieval-Augmented Generation (RAG) systems. However, despite their ability to handle long sequences, the resulting decoding latency and memory overhead remain substantial, posing challenges for real-world deployment. Recent advances in KV Cache reuse have shown potential to mitigate these costs, but still suffer from notable performance degradation. To address this issue, we conduct an in-depth investigation of recomputation-based reuse methods and observe that the recomputed tokens often fail to align with the context segments most relevant to the question. This misalignment hinders proper updates to the critical contextual representations. Therefore, we propose the $\textbf{A}$ttention-$\textbf{A}$ware $\textbf{A}$ccurate KV Cache Fusion algorithm ($A^3$), which precomputes and selectively fuses the KV Cache of text chunks based on their relevance to the question, achieving accurate integration with minimal computational overhead. Extensive experiments on various benchmarks and LLMs demonstrate that $A^3$ achieves the best task performance compared to four baselines while reducing the time-to-first-token (TTFT) by 2$\times$.

Recent advancements in Text-to-Speech (TTS) technology have been remarkable, enabling current models to clone arbitrary unseen speakers and synthesize high-quality, natural-sounding speech. However, corresponding evaluation techniques appear to be lagging: Existing Mean Opinion Score (MOS) estimation models typically perform regression-based scoring on entire speech segments-while a failed synthesized speech usually contains problematic elements in only a few isolated words rather than throughout the entire utterance. In this context, we presents an intriguing finding: encoder-decoder ASR models, such as Whisper, leverage their extensive pre-training to precisely capture word-level mismatches between speech and text within their cross-attention mechanisms, thereby providing a fine-grained reward signal. Building upon this insight, we propose a novel TTS optimization method, which we term Word-level TTS Alignment by ASR-driven Attentive Reward (W3AR). Instead of relying on any explicit reward annotations, W3AR leverages the attention information within a pre-trained ASR model, enabling finer-grained alignment and optimization of the sequences predicted by the TTS model. Experimental results demonstrate that W3AR not only effectively improves the TTS generation quality of existing models but also further enhances zero-shot robustness based on both in-domain and out-of-domain prompt speakers. Additionally, our findings and proposed methodology offer a new insight for generative tasks: understanding models can potentially serve as evaluators, providing highly fine-grained and valuable feedback for generation.

This paper presents a gate-level Boolean evolutionary geometric attention neural network that models images as Boolean fields governed by logic gates. Each pixel is a Boolean variable (0 or 1) embedded on a two-dimensional geometric manifold (for example, a discrete toroidal lattice), which defines adjacency and information propagation among pixels. The network updates image states through a Boolean reaction-diffusion mechanism: pixels receive Boolean diffusion from neighboring pixels (diffusion process) and perform local logic updates via trainable gate-level logic kernels (reaction process), forming a reaction-diffusion logic network. A Boolean self-attention mechanism is introduced, using XNOR-based Boolean Query-Key (Q-K) attention to modulate neighborhood diffusion pathways and realize logic attention. We also propose Boolean Rotary Position Embedding (RoPE), which encodes relative distances by parity-bit flipping to simulate Boolean ``phase'' offsets. The overall structure resembles a Transformer but operates entirely in the Boolean domain. Trainable parameters include Q-K pattern bits and gate-level kernel configurations. Because outputs are discrete, continuous relaxation methods (such as sigmoid approximation or soft-logic operators) ensure differentiable training. Theoretical analysis shows that the network achieves universal expressivity, interpretability, and hardware efficiency, capable of reproducing convolutional and attention mechanisms. Applications include high-speed image processing, interpretable artificial intelligence, and digital hardware acceleration, offering promising future research directions.

Recent progress in diffusion-based generative models has enabled high-quality image synthesis conditioned on diverse modalities. Extending such models to brain signals could deepen our understanding of human perception and mental representations. However,electroencephalography (EEG) presents major challenges for image generation due to high noise, low spatial resolution, and strong inter-subject variability. Existing approaches,such as DreamDiffusion, BrainVis, and GWIT, primarily adapt EEG features to pre-trained Stable Diffusion models using complex alignment or classification pipelines, often resulting in large parameter counts and limited interpretability. We introduce SYNAPSE, a two-stage framework that bridges EEG signal representation learning and high-fidelity image synthesis. In Stage1, a CLIP-aligned EEG autoencoder learns a semantically structured latent representation by combining signal reconstruction and cross-modal alignment objectives. In Stage2, the pretrained encoder is frozen and integrated with a lightweight adaptation of Stable Diffusion, enabling efficient conditioning on EEG features with minimal trainable parameters. Our method achieves a semantically coherent latent space and state-of-the-art perceptual fidelity on the CVPR40 dataset, outperforming prior EEG-to-image models in both reconstruction efficiency and image quality. Quantitative and qualitative analyses demonstrate that SYNAPSE generalizes effectively across subjects, preserving visual semantics even when class-level agreement is reduced. These results suggest that reconstructing what the brain perceives, rather than what it classifies, is key to faithful EEG-based image generation.

These proceedings include the Late-Breaking Abstracts accepted for the Evo* 2025 Conference, hosted in Trieste (Italy), from April 23th to 25th. These extended abstracts were presented through short talks at the conference, providing an overview of ongoing research and initial results on the application of diverse Evolutionary Computation strategies and other Nature-Inspired methodologies to practical problem domains. Collectively, these contributions point to encouraging directions for future work, underscoring the potential of nature-inspired approaches-- especially Evolutionary Algorithms -- for advancing research and enabling new applications.