Technology
Argument Collapse: LLMs Flatten Long-Form Public Debate
LLMs flatten public debate by converging on a limited repertoire of arguments, producing significantly less unique content than humans. This "argument collapse" risks homogenizing long-form discourse.
THRD: A Training-Free Multi-Turn Defense Framework for Jailbreak Attacks on Large Language Models
THRD is a training-free framework defending LLMs against multi-turn jailbreaks by modeling temporal risk accumulation. It reduces attack success rates to 0.2–4.0% while preserving model utility.
SECUREVENT: Hybrid AI/ML Security Monitoring for Distributed Event-Based Systems
SECUREVENT integrates AI/ML with traditional security to monitor distributed event-based systems. It outperforms static rules in detecting dynamic attacks while maintaining low false positives.
Construction of Historical Knowledge Graphs Based on BERT and Graph Neural Networks
This study integrates BERT and GNNs to construct historical knowledge graphs, outperforming traditional methods in entity and relationship extraction. It also introduces a novel image retrieval system using FastRQNet and Vilt-qaformer+RoBInet.
Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams
Adaptive Auto-Harness enables LLM agents to sustain self-improvement in dynamic, open-ended task streams. It outperforms baselines by adapting harnesses to shifting problem distributions via multi-agent evolution and routing.
FLARE: Diffusion for Hybrid Language Model
FLARE converts hybrid LLMs to support both autoregressive and diffusion decoding via a unified checkpoint. This approach preserves model quality while boosting throughput for low-latency applications.
Breaking the Information Silo: Semantic Personas for Cross-Domain Recommendation
SPHERE enables cross-domain recommendations for disjoint platforms by using LLMs to create semantic personas, bypassing the need for shared users or items.
Multilinguality of Large Language Models From a Structural Perspective
This study analyzes LLM multilinguality structurally, finding low-resource languages diverge more from English. Post-training modifies these structures without disrupting inter-language relationships.
MOSS-Audio Technical Report
MOSS-Audio is an audio-language model using DeepStack injection and time markers for precise temporal grounding. It supports captioning, transcription, and reasoning via 4B/8B variants, excelling in speech and general audio understanding.
STaR-KV: Spatio-Temporal Adaptive Re-weighting for KV Cache Compression in GUI Vision-Language Models
STaR-KV compresses GUI VLM KV caches via spatio-temporal re-weighting, cutting memory by 40% with negligible overhead. It outperforms existing methods in accuracy while maintaining high efficiency.
ProbeScale: Probing Analysis to Optimize Neural Scaling Laws for Efficient Small Language Model Inference
ProbScale optimizes small language models by using probing to identify efficient subnetworks, reducing parameters 5-10x while retaining 95-98% of original accuracy.
"I've Seen How This Goes": Characterizing Diversity via Progressive Conditional Surprise
The authors introduce Decan, a novel diversity metric using in-context learning to score creative outputs without training. It effectively identifies diversity loss in AI models and correlates with human judgments.
Dynamic Trust-Aware Sparse Communication Topology for LLM-Based Multi-Agent Consensus
DySCo introduces dynamic, trust-aware sparse communication for LLM multi-agents, reducing overhead by selecting high-value links. It maintains consensus quality while lowering costs across reasoning tasks.
Learning Implicit Bias in Generative Spaces for Accelerating Protein Dynamics Emulation
This method accelerates protein dynamics emulation by applying a history-dependent bias to generative models, boosting diversity and coverage. It achieves up to 37x faster sampling and identifies more low-energy states than unbiased approaches.
LayerRoute: Input-Conditioned Adaptive Layer Skipping via LoRA Fine-Tuning for Agentic Language Models
LayerRoute uses LoRA fine-tuning to enable agentic LLMs to skip transformer layers based on input type, reducing FLOPs by 15.25% for tool calls while maintaining high quality with minimal trainable parameters.
Unveiling the Limits of Large Language Models in Inferring Pragmatic Meaning from Non-Verbal Responses
This study reveals LLMs struggle significantly with non-verbal pragmatics, showing 60% lower accuracy than verbal tasks. While in-context learning helps, models still face major challenges in interpreting indirect non-verbal intent.
Suppressing Forgery-Specific Shortcuts for Generalizable Deepfake Detection
The S³ framework suppresses forgery-specific shortcuts via subspace modeling, enhancing deepfake detection’s generalization. It uses training-time suppression and inference-time attenuation to improve cross-method performance without sacrificing in-domain accuracy.
Boosting Multimodal Federated Learning via Chained Modality Optimization
FedMChain optimizes multimodal federated learning via sequential modality phases and sparse sign-guided aggregation, reducing modality competition and communication costs while boosting accuracy.
RadioMaster: Multi-Agent System for Autonomous Radio Signal Generation
RadioMaster is a multi-agent system that autonomously converts user commands into physical wireless signals using domain knowledge and hardware-aware verification. It outperforms existing baselines in signal fidelity and configuration viability.
Physics-Guided Attention in a Lightweight TCN for Efficient WiFi CSI-Based Human Activity Recognition
This study introduces a lightweight TCN for WiFi HAR using physics-guided attention. It outperforms deeper models while significantly reducing computational overhead.