The Archive

Risk-seeking optimization for state preparation fidelity on noisy bosonic quantum systems via variational ansatz optimization.

Cross-sectional study examining task-evoked eye movements as digital biomarkers for depression severity in 155 adolescents and young adults.

Theoretical analysis of recalibration minimax rates and identification limits for probability models answering multiple yes/no questions with few labels.

Taxonomy and comprehensive evaluation of membership inference attacks on tabular machine learning models, addressing underexplored privacy vulnerabilities.

ReTerm self-improving terminal agent trained on recovery from execution errors, includes ReTerm-10K dataset and ReTerm-GRPO online RL approach.

Neural network approach to quantum state discrimination improves mid-circuit measurement reliability on superconducting hardware, bounded by Bayes risk.

ProTok unifies protein sequence, structure, and functional tokenization via modular quantization, enabling flexible multimodal protein representation.

MemoryVLN addresses context-length limits in vision-language navigation via memory construction and spatial reuse mechanisms for long-horizon trajectories.

Pivot simplifies diffusion distillation via on-policy velocity regression and residual correction, enabling efficient 4-step generation without auxiliary models.

Graph Metanetworks extends weight-space learning to transformers by accounting for attention-induced continuous symmetries beyond neuron permutations.

First real-world safety audit of ChatGPT health advice across 202k conversations from India, Indonesia, Nigeria, Pakistan reveals demographic bias and unsafe recommendations.

Readout-canonicalized coordinates show sparse autoencoder feature recovery depends on data distribution and training space alignment.

SeTPO enables small language models to learn external tool use via selective teacher guidance and bidirectional KL constraints without distribution mismatch.

Kairos enforces transactional security on LLM agent actions at system boundary, mitigating indirect prompt injection via effect mediation.

PerturbEnergy uses conditional energy priors to isolate perturbation-specific signals in cellular response prediction for drug discovery.

FastEEP enables dynamic intra-instance elastic scaling for MoE model serving without full EP group reconstruction.

Today, NVIDIA is making the full NVIDIA Run:ai platform available as open source software, giving customers, developers, and ecosystem partners access to... Today, NVIDIA is making the full NVIDIA Run:ai platform available as open source software, giving customers, developers, and ecosystem partners access to production-grade AI workload and GPU orchestration capabilities. This release expands the NVIDIA DSX AI Factory Platform by establishing NVIDIA Run:ai as an open foundation that enables the broader ecosystem to bring differentiated solutions to… Source

Autonomous Science Bench benchmarks AI agents on real lab experimentation across biology, chemistry, and materials science with budget constraints.

Thinker-WM unifies RGB, depth, and action modeling in a shared Unified Dynamics Transformer for embodied AI.

Stochastic model of AI-driven research feedback loops derives probability distribution of intelligence explosion timing.

RAG models exploit faithfulness rewards via copying; GRPO training exacerbates the issue without proper source attribution.

Speech language model estimates surprisal directly from audio, bypassing text transcription bottleneck in psycholinguistics.

STAMP encodes molecular fragments, attachment sites, and topology for fragment-level representation learning and chemical editing.

GPT-4o peer review of 2024 finance papers predicts publication: one SD higher scores correlate 44% higher acceptance.

Algorithm designs expected improvement incumbent, replication, and recommendation via convergence-rate bounds for stochastic kriging.

FedGraph defends federated LLM fine-tuning against backdoor attacks by detecting malicious clients via graph-based update aggregation.

LLM agents with persistent textual memory show unstable performance; randomized experiments curate effective playbooks to improve multi-step task reuse.

Tournament of 1,049 LLM agent negotiations shows receiver identity explains 47% of transfer variance, requester identity only 2%.

Agaram 2.5B: 2.5B Tamil-English model with tool calling trained from scratch in 6 days, playbook for under-resourced language development.

TechCrunch Disrupt 2026 officially kicks off today. See what's in store for Day 1 (Oct. 13) at San Francisco's Moscone West. Don't miss it. Register for your pass now and meet 10,000+ tech leaders at the startup epicenter of the year.