The Archive

Regularized Q-functions close horizon gap in adversarial MDP policy optimization, matching occupancy-measure algorithms without loss factor penalty.

Counterfactual audit reveals open-weight LLM legal reasoning cites correct authorities (67-100%) but verdicts remain unchanged when substituting false ones.

Framework for evaluating LLM agents' epistemic humility when faced with knowledge conflicts via Identify-Solve-Escalate behavioral dimensions.

Theoretical analysis of Gaussian regression sample complexity on Boolean cube subspaces with sparse undersampling effects.

GPD distillation method improves vision-language model spatial reasoning by leveraging 3D geometry during training without inference overhead.

SplitJEPA decomposes latent world representations into invariant and variant factors for robotics without requiring full reconstruction.

Prior barrier concept quantifies pretrained model support disparity across concepts; proposes SFT technique for long-tail concept learning.

AAArena benchmark evaluates Adversarial Heuristic Learning paradigm where agents refine game policies via executable code without weight updates.

RefineMix framework trains discrete diffusion models with out-of-distribution data at selective noise levels to improve data-scarce generalization.

VFold symmetry-aware KV cache compression reduces LLM memory at long contexts via inter-layer value merging without architectural changes.

ContiLNN augments 2D medical image restoration with continuous-time liquid neural modules for cross-slice anatomical continuity modeling.

asdex JAX library automates sparse differentiation by detecting sparsity patterns and graph coloring to reduce AD computational cost.

RiCo models rigid-body contact via local surface neighborhoods rather than end-to-end scene prediction, improving physics simulation accuracy.

Statistical method for fusing RCT and observational data to estimate treatment effects with fewer confounding assumptions.

Theoretical analysis of oracle complexity for nonsmooth nonconvex composite optimization including neural network training.

SparseDecoding addresses distribution shift in Hessian-guided LLM pruning by using self-generated tokens, reducing decoding latency.

LLM agents selecting PDE neural operator fine-tuning configurations outperform Bayesian optimization, with causal analysis of feedback vs. prior usage.

Multi-agent Q-learning in spatial public goods shows cooperators self-organize into clusters; learning rates affect welfare dynamics.

Anthropic updates Usage Policy for 2026; details restrictions and permitted use cases for Claude API users.

Anthropic is making changes to its usage policy for the first time in over a year to reflect new and high-risk cases of misuse - including election interference, weapons development, surveillance, and health and financial uses. But one of the most significant changes prohibits "sustained and needless abusive or cruel behavior" toward Claude. Last August, the company announced it would allow Claude to end conversations with "persistently harmful or abusive" users as part of its research into "model welfare." The new update says that terminating conversations is still the "primary enforcement m...

LLM compliance systems show verdict invariance to rule perturbations; internal representations largely ignore regulatory rule changes.

Structural theory proposing explicit representational primitives for context, invariance, and abstraction in learning systems.

Small-world connectivity in LLM attention networks correlates with reasoning performance, linking internal organization to fluid reasoning ability.

GRASP combines LLM hypothesis generation with functional gradient optimization to induce interpretable probabilistic logic programs from relational data.

Periodic Labs' Fedus and Cubuk discuss synthesis superintelligence across semiconductors and superconductors on Latent Space podcast.

Natura’s $99 Interface smart ring lets you summon AI agents with the press of a finger to complete tasks, capture thoughts, and control devices — while doubling as a health tracker.

Oracle deploys ChatGPT and Codex across recruiting, engineering, and operations to automate specialist workflows.

Goodfire just launched what it says is a cheaper way to keep AI agents in check: Instead of paying a second AI to read everything an agent does, its monitors peek inside the model while it works and only call in backup when something looks fishy.

Google has released an experimental note-taking app that can transcribe meetings and audio files entirely offline, as reported earlier by TechCrunch. The app, called Google AI Edge Foresight, is free to use and runs on macOS using the company's on-device EmbeddingGemma 2 model. Similar to AI note-taking apps like Granola and Wispr Flow, Foresight summarizes your meetings while providing a space where you can take your own notes. Google says you can jot down shorthand bullet points during your meeting, and the app will transform them into "polished notes" based on the information in the transc...

Ambrosia Energy CEO Ben Longmier and Bloom Energy SVP Bill Thayer join the Smart Systems Stage at TechCrunch Disrupt. Register now to save up to $100. Grab a second of the same pass to save 50%.