All Articles
17670 articles total
One Reflection Is Not Enough: Self-Correcting Autonomous Research via Multi-Hypothesis Failure Attribution
Introduction of SAGE, an autonomous research agent that uses Multi-Hypothesis Failure Attribution (MHFA) to better recover from experimental failures.
Surprise as a Signal for Plasticity and Metacognition
Research on using prediction-error signals as a mechanism for plasticity and metacognition in AI, enhancing memory and VLM behavior.
Design and Implementation of Agentic Orchestrations and Orchestration of Agents
A classification framework and metrics for the design and implementation of agentic orchestrations in business process management.
A time-series classification framework for individual-level absenteeism prediction under severe class imbalance
A time-series classification framework for predicting individual-level absenteeism in high-demand work environments using deep learning.
Modality-Driven Search with Holistic Trace Judging for ARC-AGI-2
A high-scoring solver for ARC-AGI-2 that treats reasoning modalities as search operators and employs holistic trace judging.
ACE: Pluggable Adaptive Context Elasticizer across Agents
Introduction of ACE, a plug-and-play module for adaptive context management in LLM-based agents to prevent information loss during truncation.
Which Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal Index
Proposal of the Relative Surprisal Index (RSI) and RSI Selection (RSI-S) to improve token selection during Reinforcement Learning with Verifiable Rewards (RLVR).
Scientific Explanations in Health Sciences: Causality, Trust, and Epistemic Adequacy
A critical review examining the philosophical foundations of explainability in medical AI (XAI) and its alignment with clinical needs.
Think in English, Answer in Korean: Efficient Adaptation of Multilingual Tool-Using Agents
Development of LuckyStar 111B, a hybrid reasoning model for Korean-English enterprise agents optimized for memory-constrained deployment.
World-Model Collapse as a Phase Transition
Researchers identify a 'phase transition' in long-horizon language agents where small changes in state load or horizon cause sudden world-model collapse.
Wisdom Of The (AI) Crowd: Investigating Artificial Swarm Intelligence In Large Language Models
A study explores artificial swarm intelligence in LLMs, finding that aggregating responses from multiple models can significantly reduce estimation errors.
Xiaomi-GUI-0 Technical Report
Xiaomi introduces Xiaomi-GUI-0, a multimodal GUI agent trained in a real-device closed loop to improve stability and usability in real mobile environments.
Learning to Select, Not Relearn: Hard-Routed Mixtures of Reasoning LoRAs
The Hard-Routed MoR-LoRA framework allows for the composition of frozen reasoning LoRA experts using hard top-1 routing, preserving expert behavior with fewer parameters.
BP-TTA: Balanced and Prototype-Guided Test-Time Adaptation in Dynamic Scenarios
BP-TTA is a new test-time adaptation method that handles class imbalance and continual domain shifts using balanced sampling and prototype-guided updates.
Ask the World Before Acting: Budgeted Environment Probing for World-Model Calibration
The study introduces a budgeted environment probing mechanism to calibrate an agent's internal world model, reducing terminal errors in long-horizon tasks.
CDR-Bench: Evaluating Faithful Execution of Compositional, Order-Sensitive Data Refinement Recipes
CDR-Bench is introduced to evaluate LLMs on compositional and order-sensitive data refinement recipes, revealing a lack of procedural faithfulness in current models.
Who Determines the Meaning of an Emotion? Affective Sovereignty as an Epistemic Consequence of Measurement Limits
This paper argues for 'affective sovereignty,' suggesting that the final authority over the meaning of an emotion should remain with the human subject rather than AI.
CSTrader: A Testbed for Language-Grounded Trading in a Community-Driven Virtual Asset Market
CSTrader is a multi-agent framework for language-grounded trading in the CS2 skin market, demonstrating that LLMs can turn unstructured text into stable profits.
CLOUDADV: Decision-Aligned Instance Sizing with Zero-Shot Foundation Models under Drift
CLOUDADV uses zero-shot foundation models to provide decision-aligned cloud instance sizing recommendations to reduce overprovisioning and costs.
How information theory saved my word game
An exploration of how information theory principles can be applied to optimize the mechanics and design of a word game.