The Archive

ViToS: dual-stream RL framework for visual token pruning in medical multimodal reasoning on sparse visual evidence.

Evaluation of ML-based intrusion detection systems on Gotham2025 IoT network security dataset.

Optimizer choice drives 7x variance in emergent misalignment severity in fine-tuned Qwen models.

Requirements engineering framework for ML systems to ensure alignment with stakeholder needs and trustworthiness.

ZEBRA framework addresses base-to-novel generalization gap in audio-language models via entropy-regularized prompt learning.

DPPE rethinks camera-based positional encoding for scaling multi-view transformers in 3D vision and novel view synthesis.

Proposes robustness measure bounding neural network MSE under input perturbations via black-box analysis.

Localized conformal prediction improves uncertainty quantification for image classification with vision-language models.

Tutorial on differential equations, SDEs, and Fokker-Planck framework for generative modeling with ELBO derivations.

Relative Surprisal Index method for adaptive token selection in RLVR to improve LLM reasoning and training efficiency.

Physics-aware Neural Operator Transformer for real-time temperature field reconstruction in EAST fusion divertor.

MPL-MAE addresses positional leakage in 3D masked autoencoders for point cloud representation learning.

FLARE-AI audits AI flaw reporting systems, identifies fragmentation across 12 platforms, proposes standardized triage framework.

ACE module enables LLM agents to dynamically manage context windows by elastically retrieving discarded information, addressing trajectory length bottlenecks.

Knowledge graph linking CVE vulnerabilities to MITRE ATT&CK tactics/techniques via classification and relation extraction for threat response.

Compares room-acoustic simulation fidelity (geometrical vs. wave-based) impact on SpatialNet multichannel speech enhancement training.

AutoTrainess enables LLM agents to autonomously conduct post-training via planning, data construction, job scheduling, and checkpoint evaluation.

Artificial intelligence is transforming what is possible in agriculture, but industry leaders should be wary of investing in AI without first laying the groundwork. The use cases are promising, especially for an industry navigating volatile fertilizer costs, unpredictable weather, and margins that leave little room for error. Research shows AI-enabled predictive models can improve crop…

“After college, my plan was to come to New York and get a record deal.” | Cath Virginia / The Verge, Getty Images Watching Elon Musk fulminate at Bill Savitt during Musk v. Altman - the case in which Musk sued Sam Altman and OpenAI instead of seeing a therapist about his AI failures - was a bit like watching a toddler have a temper tantrum at his nursery school teacher. Savitt's questions were "designed to trick me," Musk said. He also told Savitt at one point, "You mostly do unfair questions." Savitt, who has the approximate demeanor of a handsome Droopy Dog, gently told Musk, "I am trying t...

Solver for ARC-AGI-2 visual reasoning benchmark using modality-driven search across text/image/code channels with holistic trace judging.

Dynamical Lie algebra analysis showing quantum circuits suffer underfitting via expressivity-trainability paradox; proposes structured architectures.

Time-series classification framework predicting individual absenteeism in high-demand sectors using sequential behavioral structure under class imbalance.

Theoretical analysis of Self-Improving Alignment (SAIL) convergence; proposes SAIL-RevKL with reverse-KL regularization to ensure strong concavity.

FinPersona-Bench benchmark measures Mandate Salience Decay in autonomous financial agents, quantifying behavioral drift over long market horizons.

RaBitQCache uses rotated binary quantization and binary-INT4 arithmetic for efficient sparse attention on long-context LLM KV cache.

Classification framework for LLM-agent orchestration balancing autonomy, traceability, and correctness in business process management.

Preregistered study showing self-repair feedback in frozen code models depends on external falsification, not re-exposure to errors.

Open-source ASR system and benchmark for Bambara child reading assessment with 55 hours of field-collected speech data.

Prediction-error surprise signal for plasticity and metacognition in continual learning; 17.7-point retention gain on frozen encoders.

Systematic framework defining and characterizing robustness in robotic manipulation across subfields.