Technology
Regime-Adaptive Continual Learning for Portfolio Management
ReCAP is a regime-adaptive continual learning framework for portfolio management that dynamically synthesizes policy vectors to adapt to market shifts. It outperforms baselines by safeguarding prior knowledge while ensuring rapid responsiveness to new regimes.
BudgetDraft: Acceptance-Aware Multi-View Training for Sparse-KV Speculative Decoding
BudgetDraft aligns sparse and full KV caches via multi-view training, restoring acceptance rates for mid-to-long context inference. It achieves up to 6.55x speedup over autoregressive decoding while maintaining memory efficiency.
Multi-Contrast MRI Motion Correction via Parameter-Informed Disentanglement and Adaptive Experts
This framework uses parameter-guided disentanglement and adaptive experts to correct multi-contrast MRI motion. It outperforms SOTA methods, showing strong generalization to unseen clinical data.
RAFT: Data Refinement and Adaptive Distillation for Domain Fine-Tuning with Alleviated Forgetting
RAFT mitigates catastrophic forgetting in domain fine-tuning via data refinement and adaptive distillation. It boosts domain accuracy by 23.2% while preserving general capabilities.
Persona Attack: Incremental Memory Injection Jailbreak Attack against Large Language Models
Persona Attack is a novel jailbreak technique that incrementally injects instructions into LLMs' memory, bypassing safety alignments. It achieves up to 95% success by exploiting conversational context retention.
StemBind: When MLLMs Get Lost Between Rules and Instances in Abstract Visual Reasoning
StemBind isolates MLLM failures in abstract visual reasoning, revealing a "binding gap" where models correctly perceive and identify rules but fail to apply them. This diagnostic benchmark pinpoints rule-to-instance mapping as the primary bottleneck.
Emergence of Exploration in Policy Gradient Reinforcement Learning via Retrying
ReMax introduces a retry-based objective that naturally fosters exploration in policy gradient RL without explicit bonuses. RePPO, its PPO variant, optimizes this via a continuous retry parameter, outperforming baselines on MinAtar and Craftax.
PrivacyPeek: Auditing What LLM-Based Agents Acquire, Not Just What They Say
PrivacyPeek audits LLM agents’ data acquisition, revealing pervasive unnecessary collection of sensitive information. Current defenses fail to mitigate these vulnerabilities, highlighting critical privacy risks beyond output leakage.
Benchmarking Multimodal LLMs on Code Generation for Complex Interactive Webpages
WebIGBench evaluates MLLMs on generating code for complex interactive webpages, addressing gaps in existing static benchmarks. It features 103 live sites and an automated pipeline to assess interaction consistency.
DiffCrossGait: Trajectory-Level Alignment for 2D-3D Cross-Modal Gait Recognition via Latent Diffusion
DiffCrossGait aligns 2D-3D gait recognition via trajectory-level latent diffusion, decoupling training from inference. It achieves state-of-the-art performance on SUSTech1K and FreeGait benchmarks.
Interpreting FCDNNs via RG on Exponential Family
This study extends RG-based DNN interpretability to continuous exponential family data, showing trained networks match RG fixed points. This reveals DNN training as feature-distilling RG calculations, explaining real-world performance.
A physics-informed foundation model for quantitative diffusion MRI
PIGMENT is a physics-informed foundation model that enables robust, quantitative diffusion MRI from sparse data. It enhances microstructure mapping across diverse clinical settings and accelerates acquisition by tenfold.
A Protocol-Language Model for Network Intrusion (Without Deep Packet Inspection)
PLM-NIDS detects intrusions using encrypted traffic metadata via a RWKV-4 model, achieving 0.94 PR-AUC without deep packet inspection. It outperforms LSTMs by learning traffic "grammar" from Layer 3/4 features.
Improving IoT Intrusion Detection Through SMOTE-Based Oversampling and Extended Multi-Model Evaluation on Side-Channel Power Data
This study improves IoT intrusion detection by applying SMOTE to balance side-channel power data. Random Forest and Extra Trees models achieved superior accuracy and speed compared to prior research.
DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning
DataShield filters safety-degrading samples in benign LLM fine-tuning using a compliance-aware score. It efficiently identifies risky data without high computational costs, preserving model safety.
ChurnNet: A Optimized Modern AI for Churn Prediction
ChurnNet benchmarks traditional ML against a modern time-series model, finding that Random Forests, XGBoost, and SVMs offer superior accuracy and efficiency for churn prediction.
MyoSem: Aligning Electromyography to Natural-Language Action Semantics for Hand Action Understanding
MyoSem aligns EMG signals with natural-language semantics for bidirectional retrieval, surpassing baselines. It enables robust, queryable hand action understanding across diverse users and unseen classes.
UF-AMA: A unified framework for cross-domain emotion recognition via adaptive multimodal alignment
UF-AMA addresses cross-domain emotion recognition by adaptively aligning EEG and eye-tracking data. It uses confidence-aware screening and multi-level domain adaptation to mitigate distribution shifts.
Digital-to-Physical Transfer of Adversarial Patches for Aerial Vehicle Detection
This study transfers digital adversarial patches to physical space, revealing that the "ON" configuration outperforms digital predictions in aerial vehicle detection. Results highlight significant security vulnerabilities in DNN detectors despite weather augmentation failures.
Beyond Augmentation: Score-Guided Pathological Prior for EEG-based Depression Detection
This study introduces Score-Guided Classification, using anomaly scores as pathological priors to detect depression via EEG without data augmentation. It also features a spatial adaptation module for handling inconsistent channel configurations across datasets.