The Archive

LASER uses low-rank SVD of activations to understand and optimize recursive model architectures like Tiny Recursive Models.

Region-Affinity Attention applies Deep Ultraviolet imaging for whole-slide breast cancer classification, a medical imaging application.

Proposes bilinear input modulation for Mamba SSMs to improve memory retention and computational expressiveness via Koopman forms.

Uses LLM-based multi-agent simulation to study cognitive biases and coordination in supply chain dynamics at scale.

Derives non-asymptotic PAC-Bayes generalization bounds for Gibbs posteriors using singular learning theory.

Analyzes text shortcut learning in Vision-Language Models via adversarial evaluation framework measuring visual-textual trade-offs.

Proposes gradient-based sample selection to preserve safety alignment during fine-tuning by identifying high-gradient harmful samples.

Evaluates LLM performance on fine-grained medical entity recognition in clinical narratives beyond standard benchmarks.

Introduces safety token regularization to preserve alignment properties during domain-specific fine-tuning of LLMs.

Proposes DREAM framework for medical report generation from retinal images using adaptive multi-modal fusion with limited data.

Proposes CDSA-Net for coronary digital subtraction angiography by decoupling vascular structure from background noise.

Reconciles theory-practice gap in online alignment methods by analyzing temperature-zero regret vs. KL-regularized regret criteria.

Framework for calibrating model-based summarization metrics without reference summaries or human annotations, addressing reliability in automatic evaluation.

Loss function approach for controlling summary generation across multiple quality dimensions while managing trade-offs between completeness and conciseness.

Odds-only probabilistic models for sports forecasting and market efficiency analysis via betting data conversion methods.

Study of using LLM-generated graph priors to improve agent coordination in multi-agent reinforcement learning without hand-specified topologies.

Reward-based optimization framework combining reasoning traces with preference alignment for faithful multi-role dialogue summarization.

Requirements engineering methodology using personas to design explainable multi-agent educational systems for clinical training scenarios.

Hybrid neural-symbolic framework combining LLM-based code synthesis with compiler feedback for automated vulnerability repair via adaptive routing.

Structural economic model of Bitcoin transaction fee formation using mempool queueing data and mechanism design theory.

Analysis of Mixture-of-Experts routing locality and KV cache sharing overlap across layers in multi-candidate code generation from shared prefixes.

Survey of decentralized trust mechanisms for IoT edge networks including federated learning, Zero Trust, and lightweight blockchain.

New CogBiasESC dataset for training LLMs to detect and address cognitive distortions in emotional support conversations.

Empirical study measuring representational change depth profiles across 240 fine-tuning runs on 15 transformer and state-space models.

Framework for autonomous spacecraft guidance using reasoning models to interpret mission intent and generate safe trajectories.

RosettaSearch: LLM-based inference-time optimization for protein sequence design using RosettaFold3 structure prediction rewards.

CognitiveBench: new benchmark measuring LLM performance on four cognitive dimensions (emotion, stance, thinking style, intention) jointly.

CCCL: GPU-native compression-coupled collective communication library for LLM training reducing overhead in tensor and expert parallelism.

Comparison of system prompt changes between Claude Opus 4.6 and 4.7, analyzed via git history visualization.

Asset pricing study arguing capacity sparsity and factor sparsity are complementary in high-dimensional financial feature discovery.