AI Research

AI Research

How Test Set Contamination Skews Generative AI Evaluations

New research quantifies how training data contamination affects generative model benchmarks, revealing critical implications for evaluating deepfake detectors and synthetic media generators.

LLM Training

SIGMA: Spectral Analysis Reveals Early Warning Signs of LLM Colla

New research introduces SIGMA, a scalable spectral method using eigenvalue analysis to detect model collapse during LLM training before performance degrades catastrophically.

Explainable AI

New Framework Explains Why AI Generates What It Does

Researchers introduce prompt-counterfactual explanations, a new method for understanding generative AI behavior by identifying minimal prompt changes that alter outputs.

LLM infrastructure

Joint KV-Cache Encoding: A New Approach to Scalable LLM Serving

New research proposes joint encoding of KV-cache blocks to improve memory efficiency in large language model inference, addressing a key bottleneck in scalable AI deployment.

digital twins

Digital Twins Meet World Models: AI's Path to Physical Reality

New survey explores how Digital Twin AI evolves from LLMs to world models, enabling AI systems to simulate and predict physical reality with unprecedented accuracy.

LLM Agents

New Benchmark Tests LLM Agents Against Messy Real-World APIs

Researchers challenge the assumption that LLM agents work reliably with perfect APIs, revealing how real-world complexity degrades AI performance.

LLM

Entropy-Aware Speculative Decoding Boosts LLM Reasoning

New research introduces entropy-based adaptive speculation that detects reasoning phases in LLMs, dynamically adjusting decoding strategies to improve both speed and output quality.

LLM

STED Framework: New Method for Evaluating LLM Output Reliability

New research introduces STED and Consistency Scoring, a systematic framework for measuring how reliably large language models produce structured outputs—critical for production AI systems.

LLM Inference

Yggdrasil: New Tree-Based Decoding Cuts LLM Latency

New research introduces Yggdrasil, a tree-based speculative decoding architecture that bridges dynamic speculation with static runtime for faster LLM inference.

LLM Agents

LLM Agents Roleplay as VCs to Predict Startup Success

New research uses multi-agent LLM systems simulating venture capitalists to evaluate startups, achieving notable predictive accuracy through collective roleplay-based reasoning.

AI Governance

New Framework Maps Risks and Governance for Agentic AI Systems

Researchers propose comprehensive framework for governing agentic AI systems, mapping capabilities to risks and establishing safety protocols as autonomous agents become more prevalent.

reinforcement learning

Liquid AI's Pure RL Approach Redefines Small Model Training

Liquid AI's LFM2-2.6B-Exp uses pure reinforcement learning without supervised fine-tuning, achieving dynamic hybrid reasoning that outperforms larger models on key benchmarks.