Build Loop

TriEx is a game-based framework for explaining multi-agent LLM reasoning through structured self-reasoning, belief states, and oracle audits.

MMCORE is a unified framework for multimodal image generation and editing that leverages pre-trained VLMs to reduce computational overhead and improve synthesis quality.

iTARFlow advances Normalizing Flows for image generation with an end-to-end likelihood-based objective and iterative denoising for competitive performance.

A novel framework for AI adjudication that explicitly identifies missing evidence, significantly improving accuracy and reliability in legal decision-making.

A frontier 4B deep research agent for edge-scale deployment trained on limited open data, outperforming larger models.

EmbodiedMidtrain bridges Vision-Language Models and Vision-Language-Action Models by curating data during mid-training, improving robot manipulation performance with existing backbones.

A multi-agent system for iterative HTML5 game generation that uses programmatic signals and mechanic-guided planning to enable progressive evolution.

A defense framework for embodied AI agents to distinguish between legitimate environmental signals and malicious visual injections.

A framework for interactive spatial narrative refinement that aligns human and LLM intent for sensemaking.

A framework that fine-tunes vision-language models with core skills for improved environmental understanding and reliable task execution in embodied agents.

A diagnostic benchmark for vision-language models to identify image distortions, revealing significant weaknesses in low-level perception.

Forage V2 is an agent organization architecture that enables knowledge evolution and transfer across autonomous agents, improving reliability and performance in open-world tasks.

Leveraging LLMs for scalable depression risk assessment from social media, providing cost-effective psychological monitoring.

ChipCraftBrain uses multi-agent orchestration and symbolic-neural reasoning for high-accuracy RTL generation, validated on industrial benchmarks and FPGA.

Infection-Reasoner is a compact AI tool for classifying wound infections with evidence-based clinical reasoning.

This research fine-tunes Vision Transformers using human saliency maps to align their attention with human cognition, improving interpretability without sacrificing classification performance.

AI framework to optimize NOx emissions in cement factories using multi-plant data for significant cost and efficiency improvements.

This research reveals how name-conditioned evaluative framing in LLM-generated resume summaries can introduce subtle biases, potentially destabilizing hiring processes and evading conventional fairness audits.

A transformer-encoded normalizing flow on the sphere for rapid and accurate neutrino direction reconstruction in IceCube, achieving state-of-the-art angular resolution.

AI agents in prediction markets show potential for information aggregation but struggle with complex information structures and reasoning about others' knowledge.

Uncovering and diagnosing two distinct failure modes in LLM quantization to improve model efficiency.

Develop a foundational model for single-cell proteomics to enhance biological insights.

This framework separates LLMs from probabilistic reasoning for medical diagnosis, offering privacy, modularity, and adjustable accuracy-coverage tradeoffs.

An AI framework that formalizes hospital quality improvement processes by co-optimizing natural language specifications and AI pipelines to discover modifiable factors for better healthcare delivery.

This paper introduces a compositional architecture for AI agents that enables hypothesis-space restructuring for improved causal reasoning, outperforming existing agents in experimental trials.

A method to expand Mixture-of-Experts models during continued pre-training to improve capacity and reduce training costs.

Investigates the impact of generalization techniques on membership inference attacks, showing significant reduction in attack performance.

This paper explores automated reasoning and dual bounds for Condorcet dimensions of elections, providing a framework for searching ranked voting data and a path toward proving smaller committee sufficiency.

This paper introduces a diagnostic framework for evaluating AI-generated reviews at the concern level, moving beyond simple verdict agreement to assess how AI systems identify and prioritize review concerns.

A computational model based on active inference to simulate how road users resolve space-sharing conflicts through uncertainty reduction.

This research analyzes user interactions with AI companions to understand how different AI roles influence user behavior and potential signs of addiction, highlighting ethical design concerns.

This paper formalizes self-modification in artificial superintelligence using operator algebra, exploring its collapse and connection to paradoxes and philosophical concepts.