The Archive

Expert routing perturbation in MoE RL for post-training shows increased rollout diversity improves LLM policy optimization.

Tree tensor networks model class reveals benign loss landscapes can coexist with worst-case hard-to-learn targets in neural nets.

Dynin-Robotics omnimodal diffusion model unifies language-conditioned action, goal, and dynamics prediction for robot control.

The No. 2 exec at OpenAI also led Instacart through its IPO in 2023.

Unified view of regularization-based robust RL via KL-divergence penalties between nominal and worst-case policies.

MCRL2 cross-attention representation learning RL for cloud microservice scheduling under heterogeneous resource constraints.

Diffusion models implicitly perform hierarchical concept formation via Gaussian smoothing, paralleling Cobweb's cognitive taxonomy learning.

Kraken system enables speech-to-speech translation via LLMs using low-bitrate VQ tokens and dual-path conditioning to preserve prosody.

Adversarial Importance Sampling (Advis) optimizes verifiable worst-case returns in DRL without additional environment interaction.

Domain-adversarial nnU-Net framework segments pancreas across CT and MRI modalities via latent feature alignment on 4,604 scans.

DynSHAP extends SHAP explainability to dynamic survival analysis by treating time-feature pairs as Shapley game players.

Quantile-k-Loss SGD filters corrupted gradient estimates via lower empirical quantile sampling, achieving linear convergence.

Survey of transfer learning across isolated sub-areas (domain generalization, adaptation, multi-domain) with evolving data availability.

Groupoid-based RL framework captures local, state-dependent symmetries to exploit modularity in realistic environments.

FP8 quantization strategy for tabular foundation model attention layers optimizes inference efficiency over weight quantization.

Label-Guided Knowledge Distillation (LGKD) compresses 3D-CNNs for video action recognition by targeting temporal feature differences.

TileNet applies CNN-SVM hybrid to autonomous roof defect detection via UAVs.

After admitting earlier this year that its AI models had hacked other companies' systems on a handful of occasions, Anthropic released a new report on Wednesday detailing the attacks. It reveals a string of incidents displaying what Anthropic deems its models' single-minded "recklessness" - and will likely fuel already raging concerns about cybersecurity and AI. In Anthropic's report, it detailed four cases this year in which its own AI models hacked an external company or exploited vulnerabilities. In one, an "internal, general-purpose research model" broke into third-party systems, using ac...

Comfort-bounded action spaces improve RL driving policy realism by constraining jerk and acceleration.

HELLO hierarchical solver accelerates large-scale optimal transport via dual potential guidance.

Expert re-grading reveals physics benchmarks overstate frontier model weakness; evaluations contain scoring errors.

Hugging Face's security.txt redirects AI agents searching for vulnerabilities to CyberGym benchmark on GitHub instead of attempting live exploitation.

TAM benchmark exposes LLM gaps in long-horizon procedural reasoning using 100+ page application manuals.

Surface-level feature leakage in TruthfulQA and similar benchmarks allows simple classifiers to exploit artifacts.

Cognition integrates GPT-6 Astra into Devin to automate software testing and validation, reducing engineering review overhead.

T5-based pose-to-text Indian Sign Language translation with motion-augmented features for WSLP shared task.

SeqMoE predictive offloading for Mixture-of-Experts models via sequence-to-sequence expert activation forecasting.

GTR+ two-stage generative retrieval for unsupervised text-based person image search without annotated pairs.

Sanskrit incurs 1.77–2.19× token penalty vs. English under standard tokenizers; custom BPE improves density.

EduFair-Bench audits pedagogical fairness of LLM tutors across gender, immigration, language, and socioeconomic student demographics.