All Articles
16523 articles total
Lifting Embodied World Models for Planning and Control
A method for 'lifting' embodied world models by mapping high-level actions to low-level joint sequences, reducing search complexity for humanoid robot control.
Bridging the Last Mile of Circuit Design: PostEDA-Bench, a Hierarchical Benchmark for PPA Convergence and DRC Fixing
Presentation of PostEDA-Bench, a hierarchical benchmark for evaluating LLM agents in the final stages of circuit design, focusing on DRC fixing and PPA convergence.
GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding
Introduction of Group-Query Latent Attention (GQLA), a modification of MLA that enables hardware-adaptive decoding paths for different GPU architectures without retraining.
Global Automation Atlas
A large-scale study using LLMs to map automation exposure across 124 economies, analyzing how task content and country-level conditions affect labor displacement.
Tunable MAGMAX: Preference-Aware Model Merging for Continual Learning
Tunable MAGMAX is proposed as a preference-aware model merging framework to mitigate catastrophic forgetting in continual learning by adjusting task-specific performance.
The Cognitive Kardashev Scale: Quantifying the Material Envelope of Civilisational Computation
A theoretical exploration of the 'Cognitive Kardashev Scale', quantifying the maximum possible machine cognition based on energy supply and hardware efficiency.
Codeberg Bans Cryptocurrency Projects
Codeberg, an open-source git hosting service, has implemented a ban on projects related to cryptocurrency.
Run large language models at home, BitTorrent‑style
A discussion on running large language models (LLMs) locally using a BitTorrent-style peer-to-peer distribution method.
NixOS is more complicated than you think but OpenCode fixed it
Discussion regarding the complexity of NixOS and how a project called OpenCode aims to simplify its usage.
When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM Pipelines
Research identifying a 'selection bottleneck' in multi-agent LLM pipelines, suggesting that selector quality is more critical than model diversity for output quality.
Doctorina MedBench-ICD10: A Dialogue-Based Benchmark and Evaluation Framework for Agent-Based Medical AI
Introduction of Doctorina MedBench, a dialogue-based evaluation framework for agent-based medical AI using simulated physician-patient interactions.
M-RAG: Semantic Key-Value Indexing for Retrieval-Augmented Generation
Proposed M-RAG, a semantic key-value indexing layer for RAG that separates retrieval keys from generation payloads to optimize token budgets.
FVRuleLearner: Operator-Level Reasoning Tree (Op-Tree)-Based Rules Learning for Formal Verification
FVRuleLearner introduces an Operator Reasoning Tree (Op-Tree) to automate formal verification and improve the generation of SystemVerilog Assertions.
Robust Reasoning Benchmark
The Robust Reasoning Benchmark (RRB) reveals that open-weights LLMs are highly susceptible to textual perturbations and 'Intra-Query Attention Dilution'.
CPGRec+: A Balance-oriented Framework for Personalized Video Game Recommendations
CPGRec+ is a balance-oriented framework for personalized video game recommendations utilizing LLMs and GNNs to improve accuracy and diversity.
Why Do Vision Language Models Struggle To Recognize Human Emotions?
Study analyzes why Vision Language Models struggle with human emotion recognition, citing head-class bias and inability to handle dense temporal sequences.
SKETCH: Semantic Key-Point Conditioning for Long-Horizon Vessel Trajectory Prediction
A new framework called SKETCH improves long-horizon vessel trajectory prediction by using semantic key-points to maintain global directional consistency.
Toward Learning POMDPs Beyond Full-Rank Actions and State Observability
Researchers propose a method to learn Partially Observable Markov Decision Processes (POMDPs) using spectral approaches and tensor decomposition to reason about systems with hidden states.
Training and Simulation of Quadrupedal Robot in Adaptive Stair Climbing and Descending for Indoor Firefighting: An End-to-End Reinforcement Learning Approach
A two-stage reinforcement learning approach allows quadruped robots (Unitree Go2) to adaptively climb and descend diverse indoor staircases for firefighting missions.
LinguistAgent Technical Report: A Reflective Multi-Model Platform for Automated Linguistic Annotation
LinguistAgent is a reflective multi-model platform designed to automate complex linguistic annotation tasks like metaphor identification.