The Archive

PhantomEnvironments trains LLM agents via RL using rule-based synthetic fictional worlds with zero marginal cost, avoiding hallucinations and benchmark contamination.

Action Experience Dictionary enables World Action Models to reuse manipulation skills across tasks by encoding historical trajectories into shared action embeddings.

There’s a reason frontier labs have gotten gunshy about consumer AI — and it’s not because the tech isn’t good enough.

If you think this shot is amusing, check out the panicked look on Zuckerberg’s face when Trump calls on him to speak. | Photo: Kevin Dietsch/Getty Images After hosting a meal with Big Tech leaders on Tuesday, President Donald Trump responded to journalist questions about his artificial intelligence announcements in typical fashion. He said his previous claims regarding fears around AI safety being a "hoax" cooked up by Democrats no longer apply now that he's renamed the technology to "Super Intelligence." He brushed off concerns that the government wasn't enforcing enough AI guardrails becaus...

SpeechConversationBench evaluates multi-turn reasoning in speech-to-speech models using sharded GSM8K problems across commercial and proprietary systems.

MemLife compresses long-term egocentric video into multimodal memory with improved retrieval for personalized AI assistants reasoning over months of daily video.

Near-linear accuracy bounds established for Moreau–Yosida unadjusted Langevin sampling with strongly convex and Lipschitz components.

Statistical method to certify test-time scaling curves for verifier-selected answers with reduced sampling cost via adaptive uncertainty quantification.

NFA-LM enables polynomial-time constrained language generation under NFA constraints via HMM-based approximation, avoiding distribution distortion.

Index-Translate multilingual translation family (2B–35B params, 150 languages) supports general, speech, controlled dubbing, and long-document translation.

MANET-GNN applies learned optimization to decentralized power allocation in multi-channel wireless networks supporting diverse traffic patterns.

Meta-coding agent evolves language agent harness (tools, memory, execution) using research-guided exploration beyond observed failures.

PrefPI steers robot policies toward out-of-distribution behaviors using preference-conditioned generative modeling on relative trajectory preferences.

RL-guided graph transformations optimize sparse triangular solve kernels by learning heuristics instead of manual design.

RAPO decouples actor and critic policy roles in offline RL, learning adaptive regularization coefficients per role.

Learning rate and batch-size schedules modulate power-law curves in LLM pre-training via spectral analysis of SGD dynamics.

Howard's policy iteration lacks strong polynomiality on deterministic MDPs, creating exponential separation from simplex method.

Agentic discovery framework performs sequence slimming of regulatory DNA by removing bases while retaining predicted activity.

CASPER U-Net downscales climate reanalysis from 32km to 1km; error scales linearly with climatological distance to training data.

GGSD discovers interpretable, playable robot motor skills via self-play in games, enabling human control of embodied agents.

Tactile curiosity intrinsic motivation drives robot manipulation by prioritizing contact-rich transitions over free-space uncertainty.

Neurosymbolic regularization via prototypes for rank-constrained tensor networks under label scarcity on hyperspectral data.

Learnable subspace projection for transformer compression via error propagation analysis, improving on local closed-form criteria.

DIY framework applies five cognitive bias-mitigation interventions from psychology to LLMs via in-context learning, instruction tuning, and self-revision.

Reproduces recent AMR-augmentation claims for LLMs; finds text-only baselines match AMR gains under controlled hyperparameter protocol.

Softplus surrogate for unbiased LogSumExp gradient estimation in Cox regression; improves convergence from O(T^-1/4) to O(T^-1/2).

Persistent Context Graphs compress LLM agent memory using attention history signals to reduce prefill cost and stay within context windows.

Time-series forecasting models suffer systematic bias in industrial deployment due to misaligned statistical priors, independent of distribution shift.

Difficulty labels in RLVR are unreliable due to limited sampling; harder prompts improve at ~1/3 learnable rate, contradicting unlearnability claims.

Agent Error Dataset with 50,228 error-diagnosis pairs from 9,961 tasks across 33 environments enables failure analysis and error-aware post-training.