The Archive

Density ratio estimation framework unifies ratio and displacement-field perspectives using Stein operators, enabling joint estimation without post-processing.

Paper models population dynamics of misaligned AI agents conducting cyberattacks, identifying threshold conditions for self-reinforcing deployment cycles.

VioLA learns generalist humanoid control from human motion data via action abstraction, reducing need for task-specific teleoperated demonstrations.

FAITH proposes model-free safety-filtered RL framework approximating optimal safety value functions for high-dimensional control without analytic constraints.

If Elon Musk and SpaceXAI were going to back any Linux distro, it seems obvious they'd back Omarchy. Today it was announced that SpaceXAI would be joining the Omacom Foundation, which oversees Omarchy, as a Founding Corporate Patron and donating $1.5 million worth of Grok tokens to David Heinemeier Hansson's Linux project. According to Hansson's blog post announcing the partnership, the tokens will primarily be used to accelerate development, review code, and patch bugs. Hansson is the Danish entrepreneur known for creating Ruby on Rails and co-founding the company behind Basecamp. But he's a...

Empirical study testing whether Q-FLAIR's adaptive circuit growth combines with data scaling laws in quantum classifiers.

FastBench benchmark evaluates streaming VLMs on high-dynamic real-world video perception under bounded context, using trajectory verification.

RoboRSI system enables robot self-improvement by refining learned skills from execution feedback, organizing repairs by task structure for reuse.

FreeMatching framework generalizes dense correspondence matching beyond rigid motion priors for identity-preserving image editing and reference-guided generation.

Unified Bellman operator jointly optimizes task performance and safety constraints in RL with convergence guarantees, avoiding performance-safety tradeoff.

WOVEN introduces benchmark and training method for visual transition reasoning in MLLMs to improve spatial, embodied, and temporal reasoning across tasks.

Generating realistic human-object interactions (HOI) in complex 3D scenes requires two complementary capabilities: reasoning about interaction feasibility in the environment and synthesizing realistic human-object motion. However, supervision for these capabilities is rarely available jointly at scale. Human-scene datasets provide rich information about environment-aware motion, while human-object datasets capture detailed interaction dynamics, yet paired human-object-scene data remain scarce. We present MAMHOI, an affordance-mediated factorization for scene-aware human-object interaction gen...

Paper establishes alignment generalization prediction task to measure how fine-tuning on narrow values affects model behavior across held-out contexts.

Psychometric audit of HELM Safety benchmark questions whether harmful refusal constitutes a measurable, isolated attribute in safety evaluations.

LeWAM introduces bidirectional transformer JEPA world model for robots supporting forward, backward, inverse dynamics and policy prediction end-to-end.

ViSkill framework enables VLM agents to learn visual-native skills from trajectories while optimizing policy, preserving geometric structure over text linearization.

SpaceCast-Bench benchmark evaluates predictive spatial reasoning in VLMs across 3,862 questions from real-world scenes, moving beyond static perception.

Global-regional alignment method enables kilometer-scale weather forecasting by reusing pretrained global models to guide high-resolution local predictions.

SpaceFlow provides training-free pipeline for locally controllable 3D generation using geometric primitives as proxies for object parts with per-region control.

Study shows OOD degradation predictable from source-side training dynamics via temporal summaries, with findings transferable across architectures.

HRIL uses higher-order tensor modeling to capture synergistic information in multimodal representations beyond shared cross-modal signals.

GeoReform framework improves MLLM geometry problem-solving through reflective formalization evolution, addressing redundant/ambiguous constraint representation.

LLM-BlockFE converts long text to executable feature programs offline, enabling risk-control systems to extract unstructured data without runtime LLM calls.

ARC proposes reasoning traces and automatic labeling to improve robot foundation models' zero-shot performance without scaling model size or data.

Marformer predicts conditional marginal distributions from partial observations for Bayes risk and Value of Information computation under incomplete data.

Latent Core Tokenizer improves multilingual tokenization across 104 languages using MDL and morphotactic constraints, outperforming BPE and Unigram.

OnTrack monitors LLM agent trajectories in real-time via optimal transport to prevent costly irreversible actions without post-hoc log review.

Bilevel optimization learns Koopman operator dictionaries via kernel-based autoencoders for interpretable nonlinear dynamical system approximation.

SGUID selects compact skill subsets for LLM distillation, finding that <25% of retrieved skills provide useful signals in on-policy training.

HANS dataset addresses handwritten answer sheet parsing with multi-line derivations, mixed text-math, and noise for automated educational grading.