All Articles
16670 articles total
VTubing: How a Japanese Phenomenon Is Going Worldwide
An exploration of the global expansion of VTubing, focusing on its cultural impact and growth from Japan.
Is Your Model Thinking or Just Stagnating? PUMA: Diagnosing Reasoning Pathology via Phase-Momentum Alignment
Introduces PUMA, a training-free framework designed to diagnose and mitigate 'overthinking' in Large Reasoning Models by aligning phase-momentum.
Toward Anthropomorphic Dialogue: A Closed-Loop Framework for Human-Like Chat Generation, Evaluation, and Preference Alignment
Presents AnthroDial, a closed-loop framework for generating and evaluating human-like, anthropomorphic dialogue in LLMs.
A Systematic Evaluation of Trajectory Data Curation for LoRA Fine-Tuning of Code Agents
A systematic study on how the quality and quantity of trajectory data affect the LoRA fine-tuning of code agents like Qwen2.5-Coder.
Constrained Path Reasoning: Measuring When Committed Stages Earn Their Cost
Discusses Constrained Path Reasoning (CPR), a method to measure the cost-effectiveness of committed stages in LLM reasoning pipelines.
LenGuard-GPC: Length Guarding with Guided-Prompt Consistency for Spatial Reasoning Reinforce Learning
Introduces LenGuard-GPC, a dense reward framework for Reinforcement Learning to optimize spatial reasoning and control response length in VLMs.
Coordinated Disentanglement with Iterative Mode Discovery Under Hidden Correlations
Proposes CoDID, an end-to-end framework for disentangled representation learning that discovers modes under hidden correlations.
An Explicit World Model Based on Data-First Ontology: DaoQL Multimodal Storage Validation and Counterfactual Reasoning Evaluation
Introduces DaoQL, a multimodal storage system that creates an explicit world model to reduce hallucinations and improve counterfactual reasoning in LLMs.
Lossless but Not Free: An Empirical Anatomy of Speculative Decoding on Consumer Hardware
An empirical study on the performance and pitfalls of speculative decoding on consumer Apple-silicon hardware.
Linux kernel will support $ORIGIN, sort of
The Linux kernel is adding limited support for $ORIGIN to improve how shared libraries are located.
Expected Free Energy as Belief-Dependent Utility for rho-POMDPs
Researchers demonstrate that minimizing Expected Free Energy (EFE) in active inference is equivalent to solving a rho-POMDP, eliminating the need for manual tuning of exploration weights.
PriorProof: A Point-in-Time Measure of Technique Novelty for Formal Proofs
PriorProof is introduced as a tool to measure the novelty of formal mathematical proofs in Lean by comparing them against historical snapshots of Mathlib.
Reward-Driven LLM Agent Workflows: Synthesizing POMDP Routing and Self-Correction for Autonomous Decision-Making
A new LLM agent workflow combines POMDP routing and self-correction to improve long-horizon planning and reduce hallucinations in autonomous decision-making.
When LLMs Over-Answer: Measuring and Mitigating Quality Issues in LLM-Based Hardware Description Language Question Answering
A study on LLMs used for Hardware Description Languages (HDL) finds they often 'over-answer' with redundant info, leading to a proposed multi-agent framework to improve quality.
Bridging the Information Gap: Semantic Densification and Hindsight Distillation for Cold-Start Prediction
SemRaD is a new framework for cold-start prediction in e-commerce that uses structured semantic reasoning and hindsight distillation to improve LTV and CVR predictions.
Otap:Structure-Aware Optimal Transport for Evaluating Planning and Execution in Agent Trajectories
Otap (Optimal Transport for Agentic Planning) is proposed as a new metric to evaluate LLM agent trajectories by measuring distance between execution and solution graphs.
Fourier Geometric Wind Power Forecasting with Numerical Weather Prediction
A new multimodal framework using Fourier Neural Operators (FNO) improves wind power forecasting by integrating SCADA data with numerical weather predictions.
Evidence Interfaces Shape How Retrieval-Augmented Readers Use Support
Research on RAG systems shows that how retrieved evidence is presented (the 'evidence interface') significantly impacts the reader's ability to utilize support for multi-hop QA.
A Diagnostic Framework for AI Agent Behavior
A new diagnostic framework for AI agents proposes 'layer attribution' to distinguish whether behaviors stem from computational architecture or behavioral modulation.
Tensor Notation in Mathup
Discussion regarding the implementation of tensor notation within the Mathup typesetting system.