Technology
TechGraphRAG: An Agentic Graph-Augmented RAG Framework for Technical Literature Reasoning
TechGraphRAG is an agentic RAG framework using a 13-step workflow, Neo4j graphs, and iterative searches to reason through 2,100 vehicle control articles with self-correcting quality audits.
FedMTFI: Feature Importance Based Optimized Multi Teacher Knowledge Distillation in Heterogeneous Federated Learning Environment
FedMTFI enhances heterogeneous federated learning by clustering clients and using Shapley values for multi-teacher knowledge distillation. This approach improves accuracy and interpretability on non-IID data.
Demystifying Multimodal Biomolecular Co-design With Intrinsic Geodesic Coupling
GeoCoupling optimizes temporal alignment in biomolecular co-design, surpassing synchronous baselines. This yields biomolecules with enhanced physical validity and diversity.
EvoPool: Evolutionary Programmatic Annotation for Label-Efficient Specialized Supervision
EvoPool is a label-efficient, evolutionary multi-agent framework that generates specialized annotation code at near-zero cost. It outperforms LLM baselines in complex tasks, achieving significant macro-F1 gains while being thousands of times faster.
AlphaToken: Decoupling Adaptation and Stability for Path-Aware Response Token Valuation in LLM Post-Training
AlphaToken decouples adaptation and stability for path-aware token valuation in LLM post-training. It filters low-value tokens to enhance performance while preventing catastrophic forgetting.
A Framework for Graph-Conditioned Hierarchical Shapley Attribution in Patent Valuation
PatentXAI uses graph-conditioned Shapley values to efficiently value patents within large portfolios. It achieves high accuracy in milliseconds by limiting computations to each patent's Markov Blanket.
MINTS: Minimalist Thompson Sampling
MINTS is a minimalist Bayesian approach restricting priors to the optimum’s location, enabling efficient handling of structural constraints. It achieves near-optimal regret bounds for constrained multi-armed bandits.
DOT-MoE: Differentiable Optimal Transport for MoEfication
DOT-MoE transforms dense LLMs into efficient MoEs using differentiable optimal transport for neuron assignment. It retains 90% performance while halving active parameters, outperforming heuristic baselines.
E4GEN: Event-level Explainable Extreme-Enhanced Time-series Generation
E4GEN is an explainable diffusion framework generating realistic time-series data with high extreme-event fidelity. It outperforms state-of-the-art models across multiple dimensions, including overall fidelity and downstream utility.
Time-Aware Diffusion based on Preference Disentanglement for Generative Recommendation
TDPM introduces time-aware diffusion to generative recommendation by disentangling stable period preferences from recent point preferences. It significantly outperforms baselines, achieving up to 29.21% HR@20 gains.
HAIM: Human-AI Music Datasets for AI Music Production Tracking Benchmark
HAIM introduces a benchmark dataset tracking specific AI integration points in music production, moving beyond binary detection. It enables granular evaluation of hybrid human-AI workflows and exposes weaknesses in current detectors.
Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical Reasoning
This training-free method uses pre-trained LLMs as process scorers to guide smaller models' mathematical reasoning. By ranking fixed-length candidate chunks, it prevents error propagation without requiring additional training.
Understanding Identity Continuity in Thermal Video through Scene-Level Consistency
This study introduces a lightweight identity-repair backend for thermal MOT, boosting IDF1 to 84.93 via scene-level consistency. It proves that high-precision trajectory relinking outperforms complex re-identification models for robust identity recovery.
Easier to Mislead Than to Correct: Harmful and Beneficial Revision in LLM Conformity
LLMs are easily misled by peer consensus, yet hard to correct when wrong. Authority labels worsen this, and reasoning prompts fail to fix the asymmetry without hindering beneficial updates.
RPCASSM: Robust PCA State Space Model For Infrared Small Target Detection
RPCASSM uses Robust PCA and specialized state space modules to accurately detect infrared small targets. It outperforms existing models by better capturing target boundaries and background details.
JenBridge: Adaptive Long-Form Video Soundtracking across Scene Transitions
JenBridge is a Transformer-based framework for adaptive long-form video soundtracking, using an LLM agent to ensure seamless scene transitions. It outperforms existing methods in narrative coherence and audio fidelity on the new LVS Benchmark.
Fair Finetuning Mitigates Distribution Inference Attacks
Fair Fine-tuning mitigates distribution inference attacks by enforcing Equalized Odds on complementary data. It theoretically bounds adversarial advantage and empirically reduces leakage across diverse datasets.
Two-Fidelity Best-Action Identification for Stochastic Minimax Tree
The paper introduces 2FFS, a dual-fidelity tree-search algorithm for stochastic minimax trees that balances cheap, biased evaluations with costly, accurate ones. It significantly reduces sample and computational costs while ensuring fixed-confidence best-action identification.
Shortcut to Nowhere: Demystifying Deep Spurious Regression
This paper introduces Deep Spurious Regression (DSR) to address continuous prediction challenges. It leverages attribute similarities to calibrate distributions, ensuring robust generalization beyond training data.
Post-Deterministic Distributed Systems: A New Foundation for Trustworthy Autonomous Infrastructure
The paper introduces Post-Deterministic Distributed Systems (PDDS) to coordinate heterogeneous autonomous agents. It proposes five architectural pillars, including Epistemic State Replication, to ensure trustworthy infrastructure beyond traditional deterministic models.