The Archive

Multi-task vehicle routing solvers using locally augmented preferences and representation disentanglement to overcome RL reward-scale disparities and preference-optimization stagnation.

Isotonic Bellman calibration for offline RL detects and reduces residual occupancy-balance violations in importance-weighted policy evaluation.

BrowserForge: parallel browser sandbox framework scaling web agent training data from thousands to millions of diverse interaction trajectories across websites.

FedV-KGQA: federated framework for multi-hop question answering over vertically partitioned knowledge graphs preserving data privacy across silos.

LAION-BVD: 10-million-hour open video dataset (1.3B URLs, 80M videos) with synthetic captions for multimodal pre-training across video, audio, image.

LLM financial analysts show retrieval-integration gap: risk disclosures lose influence on investment judgments with added context despite maintained retrieval accuracy.

LLM framework detects redundant content in educational assessments using dual semantic/structural similarity analysis.

Constrained entity selection method grounds LLM-generated knowledge graph QA answers without full SPARQL parsing.

BioKERN learns multimodal spatial representations for histology-transcriptomics by incorporating biological kernel regularization.

Anthropic launches $5M grant program for independent research on AI's impact on user wellbeing.

Geometric framework analyzes robustness of neighborhood-based fairness audits under adversarial perturbations.

Effective learning rate (ELR) ratio governs loss dynamics in LLM pretraining across scales, optimizers, and architectures.

MDTE uses diffusion to balance minority-class node representation in temporal graph classification.

Strictly causal video anomaly detector using state-space models with O(1) per-frame computation for edge deployment.

Crase bounded agentic graph search outperforms open-ended agents on scholarly paper ranking with explicit pruning and stopping.

MCGenie generates standardized model cards from repository metadata; MCTidy reorganizes existing cards for consistency.

StarHarness evolves agent harnesses (prompts, tools, MCP providers, subagent structure) without retraining model weights.

LION applies Clifford algebra to multimodal-attributed graph neural networks, improving representation and downstream task performance.

CAFE co-evolves search agent and feedback critic to improve intermediate trajectory corrections without waiting for terminal rewards.

Perturbation audit reveals medical chain-of-thought rationales often don't drive model outputs; proposes clinically-motivated failure mode classification.

Linear probes achieve robust, sample-efficient detection of machine-generated text with better OOD generalization than supervised baselines.

StepGuard audits agent tool invocations step-by-step; introduces StepGen to auto-generate safe/unsafe trajectories for guardrail training.

Ensemble CNN system predicts stroke risk from 11 features using seven supervised ML algorithms on medical imaging data.

Approximates Bayesian Ideal Observer performance bounds for detection tasks via denoising score matching, avoiding MCMC resampling per image.

AtlasNav identifies Evidence Blindness in LLM agents using direct corpus interaction; proposes navigation strategy to surface and expose critical documents.

MoTE uses task-level sparse mixture-of-experts for multi-task video understanding, improving over dense decoders for action recognition and forecasting.

Google Search tips for home decor shopping and DIY projects; consumer lifestyle content unrelated to frontier AI.

FactoMap models disentangled representations with factor-space structure (domains, identifications, anisotropy) beyond Euclidean product assumptions.

Framework for generating CAD programs from natural-language descriptions by inferring missing spatial constraints using design experience.

RACE: Statistical framework for evaluating functional consistency of Transformer neurons at scale via forward-pass perturbation analysis.