The Archive

As the debate rages over whether the recent spate of rogue AI agents is a step toward AGI or a more conventional engineering problem, Nvidia is offering its own answer to problem. Nvidia CEO Jensen Huang on Monday introduced a toolkit of software and hardware products that add independent security layers around AI agents to […]

In March, Janice Malone began getting calls about suspicious activity from her nonprofit organization, Vivian's Door. Vivian's Door, headquartered in Alabama, typically provided training, resources, and community to underserved and minority-owned businesses. The work sometimes put it in close contact with these companies' financial data, which was stored on its systems. But suddenly, concerned callers from all over the world warned they'd been getting emails "begging for money" - which she hadn't sent. The organization's third-party IT team pulled its systems offline for three days while they...

Anthropic has released the newest version of its mid-range model, boasting faster response times and less token burn.

FurE reconstructs animal fur geometry from multi-view images using latent fields and PCA decoding, addressing fine-scale detail without fur datasets.

Telescopic Language Models train nested-capacity Transformers via stochastic prefix supervision, enabling single model to serve multiple compute budgets.

PDMD improves video diffusion model distillation by projecting critic errors, reducing instability and artifacts during training.

UMM-Ref trains multimodal models to self-repair images via interleaved reflection and rendering using joint RL over full loop.

Benchmark of 189 biblical intertextual references in Karen Blixen's texts, evaluating TF-IDF and sentence encoders on retrieval task.

MGFlow unifies distributional training for one-step visual generation via Wasserstein gradient flow, connecting global objectives to pointwise updates.

NstAgent scales LLM story generation to full-length novels via training-free agentic framework tracking structured narrative state.

TokenCast forecasts token consumption during LLM agent execution by learning composable cost representations per execution segment.

Hard-EM framework identifies contractive dynamical representations for adaptive control under coupled disturbances via Kalman smoothing.

NHMO solves elliptic PDEs on variable domains using transformer-based boundary kernels supervised by Walk-on-Spheres sampling.

Foil method enables efficient looped mixture-of-experts by flattening experts and decoupling attention, improving parameter utilization in sparse models.

KV-streams enables efficient context compaction for long-horizon agentic LLM training by avoiding repeated prefilling, increasing throughput without degradation.

Teacher-Assisted Communication Training (TACT) reduces communication cost in multi-agent negotiation while maintaining social goal attainment.

Study finds looped transformers achieve steeper accuracy-compute scaling slopes than non-looped baselines during test-time decoding, though with suboptimal absolute performance.

Linear Vision Transformers initialized from Softmax ViT weights via attention transfer outperform from-scratch training, reducing efficiency-accuracy trade-off.

FinAutoRubric generates task-specific evaluation rubrics for financial research agents using expert guidance rules and reusable criteria across horizons.

Retrospection-Only Fine-Tuning (ROFT) improves agentic LLM behavior via explanation-based training on agent experience without reinforcement learning.

Harness learning trains proposer models to revise agent harnesses (tool-calling programs) via execution feedback, enabling test-time task adaptation.

Failure-Transparent Agents (FTA) benchmark isolates post-failure reporting errors in tool-using LLMs via controlled failure states and evidence auditing.

X-Reset resolves exploration in object manipulation RL via cross-embodiment resets using human hand-object demonstrations without per-task reward shaping.

ScAn-Bench introduces benchmarks for systematically evaluating scaling law methodology across LLM and VLM architectures using 12k+ checkpoints.

AI Night-Scientist uses reinforcement learning to teach LLMs to balance structured reasoning with creative divergence for scientific discovery.

Doubly-spectral stochastic expansion enables unified uncertainty representation for GNNs across calibration, OOD detection, and robustness.

MeqMuon improves the Muon optimizer for LLM pretraining by balancing both row and column dimensions in update matrices.

Distillation defenses against LLM replication attacks are defeated after adversarial reinforcement learning, revealing realistic attack vectors.

Dually Regularized AIL provides finite-sample complexity bounds for adversarial imitation learning with reward and policy regularization.

Test-time alignment via factorized ranking models enables personalized LLM generation, outperforming monolithic reward models at Best-of-N scaling.