Build Loop

An open-source framework using LLMs to evolve technology mapping code for logic synthesis, significantly outperforming existing methods in area reduction.

An end-to-end framework that automates AI research by generating novel ML classifiers with executable code, outperforming baselines and available as a Python package.

Entropy Centroids offer an intrinsic reward for LLM test-time scaling by identifying temporal patterns of model uncertainty, enabling selection of higher quality responses without external reward models.

A framework for cross-architecture knowledge distillation in diffusion large language models, significantly improving performance on code generation tasks.

SeeCo is a plug-and-play framework that recalibrates open-vocabulary models on-the-fly for improved semantic segmentation in remote sensing images.

Test-Time Safety Alignment optimizes input word embeddings using a text-moderation API to minimize semantic harmfulness in aligned LLM responses, neutralizing safety-flagged outputs.

A progressive semantic communication framework for efficient edge-cloud vision-language model inference, reducing latency and communication costs.

An annotation tool for long-horizon robotic action segmentation that synchronizes multi-modal data and streamlines the annotation process.

LATTICE is a benchmark for evaluating crypto agents' decision support utility in real-world scenarios, using LLM judges for scalable assessment.

An open-source library for mechanistic interpretability of reward models, enabling deeper understanding of RLHF-trained LLMs.

Reveals a stable, content-invariant distributional attractor in LLM response positions under sandbagging, offering a behavioral signature for this mode of operation.

DepthPilot is an interpretable framework for generating realistic and clinically accurate colonoscopy videos, enabling 3D reconstruction and surgical navigation.

CheXthought provides a multimodal dataset of clinical reasoning and visual attention for chest X-rays, enabling more transparent and interpretable AI models.

SecMate: An AI-powered multi-agent VCA for adaptive cybersecurity troubleshooting in SMBs, integrating personalized tri-context assistance.

Autonomous agents for trading on blockchain ensuring reliable transaction validation.

HalluCiteChecker is a lightweight, offline toolkit for detecting and verifying hallucinated citations in scientific papers, reducing reviewer workload.

A domain-adaptive pipeline for improving multilingual crisis communication through fine-tuning language models.

A transformer-based framework for offline safe reinforcement learning that uses Lyapunov-guided imagination for test-time adaptation to ensure safety without retraining.

A probabilistic approach to fruit maturity estimation that models maturity as a continuous variable, improving robustness to label noise.

A new benchmark and forecaster that evaluates strategic reasoning in agents by analyzing their research and judgment processes.

An atomic-quality probe and hybrid selector for governing skill updates in compositional robot policies, improving reliability and reducing costs.

Train computer use agents to accurately assess graphical user interface usability, reducing costly and time-intensive manual testing.

Evergreen optimizes LLM claim verification for semantic aggregates by recasting it as a semantic query processing task, reducing cost and latency with tailored optimizations and provenance capture.

ClawGym is a scalable framework designed to advance the development of environment-grounded autonomous agents with verifiable training and evaluation tools.

An end-to-end pipeline for generating realistic synthetic industrial surface defects to overcome data scarcity in AI-powered detection systems.

Tree-of-Text: A tree-structured prompting framework for efficient and accurate sports report generation from tables.

Fine-tuned small language models offer a privacy-preserving, accurate solution for clinical triage decision support, outperforming proprietary LLMs.

A reasoning-aware retrieval framework that intelligently injects evidence during multi-step inference for large language models, improving accuracy and efficiency.

Enhance large vision-language models to refuse queries beyond their knowledge, improving trustworthiness for specialized domains.

A unified 4D world model for robots that synthesizes high-fidelity video and 3D reconstructions while enabling real-time action execution.

Naamah is a large-scale, high-quality synthetic Sanskrit NER dataset generated using DBpedia and LLMs, addressing a critical gap for classical Sanskrit literature digitization.

A data-centric AI framework improves intraoperative fluorescence lifetime imaging for glioma surgical guidance by enhancing data reliability and model robustness.

A training-free framework for detecting LLM-generated text by analyzing sentiment distributional stability, offering robust and interpretable content identification.

FutureWorld is a live reinforcement learning environment for training predictive agents that learn from real-world outcomes.

A hierarchical framework for inducing evidence-grounded and truthful user personas from behavioral logs, improving interaction prediction.

MedSynapse-V bridges visual perception and clinical intuition in medical AI by evolving latent diagnostic memories for improved diagnostic accuracy.

A Human-in-the-Loop framework for automating secondary-level mathematics assessment using multiple LLMs.

QYOLO is a lightweight object detection framework that uses quantum-inspired channel mixing to significantly reduce parameters and GFLOPs with minimal accuracy loss.

A multi-modal transformer architecture for accurate and efficient celestial attitude determination in spacecraft, outperforming traditional methods with low latency.

Anchor-constrained perceptual optimization for diffusion models to improve image generation quality without reference.

STLGT is a scalable, trace-based linear graph transformer for accurate and efficient tail-latency prediction in microservice systems, enabling proactive SLO management.

XDFT is a self-evolving agent that automatically diagnoses and explains discrepancies between DFT calculations and experimental results for materials science.

A framework for multi-agent LLM systems that uses architectural diversity and coherence validation to prevent artificial consensus and improve policy simulation fidelity.

An uncertainty-aware reward framework for reinforcement learning that mitigates reward hacking and improves alignment.

A training-free framework that dynamically routes large reasoning models to different scaling strategies based on output disagreement, improving accuracy and reducing cost.

A benchmark and framework to evaluate strategic memory use in virtual character conversations, identifying limitations in current LLMs.

A framework for trustworthy clinical AI that engineers trust through evidence, supervision, and staged autonomy, moving beyond black-box confidence.

Detecting LLM alignment faking by analyzing tool selection patterns, with a new dataset and evaluation of frontier models.

TimeMM is a dynamic multimodal recommendation framework that uses time-as-operator spectral filtering to model evolving user interests and modality-specific temporal sensitivity.

A novel decoding algorithm that accelerates LLM-based attribute value extraction by parallelizing independent sequence generation, reducing inference costs by up to 13.8X.

A practical utility-weighted evaluation metric (pBA) to correct performance estimation bias in imbalanced classification across minority subconcepts.

A framework to evaluate the alignment between GeoAI model explanations and domain knowledge for satellite-based flood mapping.

A toolkit to detect spurious correlations in speech datasets, preventing overestimation of performance in critical applications.

A privacy-preserving federated learning framework for optimizing chemical processes across distributed plants without sharing raw data.

Efficient methods for utilizing text-only data to improve encoder-dominated speech recognition models, showing larger encoders with smaller decoders can match or surpass performance.

An iterative logic tensor network that empirically demonstrates reasoning is a distinct capability required for generalization, not an emergent property of symbol grounding.

A framework for evaluating creative writing quality based on information theory, using mutual information to quantify calibrated surprise and laying groundwork for a professional benchmark.

ImproBR uses LLMs to automatically improve low-quality bug reports by filling in missing steps to reproduce, observed behavior, and expected behavior.

RaMP optimizes Mixture-of-Experts inference by dynamically selecting kernel configurations based on runtime expert routing for significant speedups.

An agentic framework that evolves transferable lemma libraries for formal theorem proving through a wake-sleep program induction paradigm.

A theoretical framework for automated relational reasoning integrated with ANNs, achieving high IQ test scores.

A training-free framework that uses spatial attention entropy to guide cropping for improved small-object detection, outperforming baselines with marginal latency overhead.

A training-free framework for open-vocabulary change detection that uses cross-temporal memory reasoning and global-local adaptive rectification to identify semantic changes in bi-temporal images.

A dual-path KV cache offloading framework for edge LLM inference that uses NVMe-direct access to reduce latency and improve SSD utilization.

An agentic system for evaluating the AI-readiness of heterogeneous scientific data across governance, quality, compatibility, and adaptability.

This research demonstrates that language diffusion models function as associative memories, capable of retrieving unseen data and exhibiting a transition from memorization to generalization.

A biologically plausible mechanism for neural assemblies to learn causal directionality using local plasticity, offering auditable explanations.

A pipeline for converting imperative programs into typed, attributed graphs using neural and structural methods to enable verification artefact reuse.

A teacher-student framework for efficient and accurate inverse source localization and characterization in physical fields.

A benchmark evaluates AI language learning feedback for 'explainability pitfalls' to prevent misconceptions and improve learning outcomes.

A novel memory-centric chiplet architecture designed to significantly reduce latency and energy consumption for LLM attention serving.

A neural cellular automaton learns compositional rules for semantic parsing without hand-written rules, achieving state-of-the-art structural generalization.

An Apriori-based analysis of learned helplessness in math tutoring logs, revealing behavioral patterns linked to intervention and outcomes.

A novel approach to reduce language confusion in multilingual LLMs using token-level optimization.

A training-free method for discovering minimal neural network architectures by progressively reducing random topologies, outperforming pruning baselines with significant parameter reduction and faster execution.

A system for political question evasion detection using a multi-head RoBERTa with chunking for long contexts.

A Diffusion Transformer framework that adaptively orchestrates metadata for generative super-resolution, improving quality and reducing transmission bitrate.

A framework that trains a lightweight policy to map high-level actions to low-level joint actions, enabling more efficient and effective planning for embodied agents.

A quantum-based steganography system that uses multi-factor authentication and variational quantum circuits for secure data embedding.

Benchmarking the safety of LLMs for robotic health attendant control reveals significant violation rates and highlights the need for robust safety evaluation.

A behavioral firewall for AI agents that uses a deterministic finite automaton to enforce benign tool-call sequences and parameter bounds, reducing attack success rates and latency.

Exploring new methods for text style transfer in graphic designs by improving word alignment in machine translation.

A unified framework for benchmarking complex multimodal document processing pipelines in enterprise settings.

Develop a foundational neural wavefunction model to solve parameterized many-electron Hamiltonians for quantum materials.

An AI-native TDD framework that operationalizes classical TDD principles for reliable LLM-assisted development.

This paper proposes a hierarchical decision-making framework combining rule-based coaching with reinforcement learning for UAV search-and-rescue missions under limited simulation training.

A neuro-symbolic framework for interactive agents to improve compositional generalization by grounding actions with a dynamic causal program graph and inductive logic programming.

This research explores culturally specific Generative AI risks for youth in Saudi Arabia, providing design implications for context-sensitive parental controls.

A novel graph neural network architecture that conserves momentum for more accurate simulation of deformable materials.

This research explores how Large Language Models respond to legal arguments, investigating their persuadability and implications for legal decision-making.

This paper explores how generative AI subtly influences control and agency in recruiting workflows, leading to deskilling despite marginal efficiency gains.

A physics-informed learning framework co-learns port-Hamiltonian system models and optimal energy-shaping controllers from trajectory data using alternating optimization.

This paper explores the theoretical potential of a probabilistic transformer framework for time series modeling through three research questions.

Qvine introduces a vine-structured ansatz for quantum circuits designed to efficiently load high-dimensional distributions, showing promise for machine learning and finance applications.

A multi-stage framework for segmenting bi-atrial regions in 3D cardiac MRI using V-Net models and asymmetric loss, with a preprocessing step for contrast enhancement.

Explores theoretical connections between fundamental physics, existential risks, and human futures, suggesting potential transformative impacts on AI.

A review of recent advancements in mm-wave and sub-THz/THz oscillators for future communication and computing systems.