Study Compares Single and Multi-Agent LLMs on Financial Review
A controlled study tests the common guidance that dividing a document review among multiple LLM agents improves accuracy.
Research Digest
Preprints and papers, rewritten in plain English.
TrajEvolve: Credit Assignment for Long-Horizon RL in LLMs
A new arXiv paper proposes TrajEvolve, a trajectory-evolution approach to improve credit assignment in reinforcement learning for large language models on long-horizon tasks.
RACE: RoPE-Aligned Attention Compression for KV Cache
A new head-dimension compression method for key-value caches aims to avoid the decoding FLOPs overhead of existing approaches.
Multi-Objective RL for Pareto-Optimal RTL Code Generation
A new approach uses multi-objective reinforcement learning to generate RTL circuits that balance power, performance, and area.
InsightRAG aggregates scholarly judgments for survey generation
The system aims to improve automatic survey writing by incorporating distributed scholarly judgments, addressing the open-ended search problem in existing methods.
Autonomous Construction via LLMs and Reinforcement Learning
A new arXiv paper proposes pairing large language models with reinforcement learning to automate construction tasks and track progress.
Edge-Deployable Agentic AI Aims to Ease Radiology Report Structuring
A new arXiv paper proposes an edge-deployable, modality-aware agentic system to address the GPU bottleneck in LLM-based radiology report structuring.
MAMath-Bench evaluates math reasoning processes in multi-agent LLMs
A new benchmark prioritizes coherent and verifiable reasoning steps over final answers for multimodal multi-agent mathematical problem solving.
Weakly-Supervised ECG Foundation Model Targets Interpretability Gap
A new foundation model for electrocardiograms aims to make cardiac machine learning interpretable enough for clinical adoption.
LLM Agents Take on Autonomous Planning in Smart Manufacturing
A new arXiv paper describes how large language models are pushing manufacturing agents beyond predefined workflows toward autonomous planning and execution, while flagging the difficulty of dynamic physical environments.
Point2Acoustic predicts room acoustics from 3D point clouds in milliseconds
A new learning-based framework estimates reverberation time and reverberant-to-direct ratio directly from scene geometry and materials, enabling fast acoustic updates for XR.
Geographic side data sharply raises voice-privacy risk
A new semi-synthetic dataset shows that pairing anonymized speech with location metadata can dramatically increase the chance a speaker is singled out.
EMMA Uses Generative Video to Boost Robot Control Robustness
A new framework augments robot training with synthetic multi-view videos, improving zero-shot visual performance by over 92%.
Filter Bank Keeps AR Glasses Noise Cancellation Robust
A new approach makes active noise control for augmented reality glasses resilient to changing noise directions without continuous recalibration.
Kamino: GPU-based rigid-body simulator handles kinematic loops natively
A new GPU-based physics solver lets robots with closed-loop mechanisms be simulated in massive parallel batches without simplifying their topology.
PAKE Framework Improves Whole-Body Loco-Manipulation for Quadruped Robots
A new control framework decomposes loco-manipulation into reference generation and imitation control, improving precision and workspace coverage on a quadruped with a six-DoF arm.
Transformers plus refinement estimate plate-reverb parameters from IRs
Two Transformer-based pipelines recover six physical parameters of a synthetic plate-reverb model from impulse responses, with refinement cutting errors by orders of magnitude.
MARE: Open-Source Lunar Rover Simulator with Shadow Ground Truth
A new Unreal Engine 5-based simulator provides pixel-accurate shadow ground truth to help develop lunar rover perception systems for extreme polar lighting.