The Archive

In this paper, we study Reinforcement Learning in Parametrized Action Markov Decision Processes (PAMDP), where each decision consists of a symbolic action and numerical parameters. In such settings Reinforcement Learning algorithms typically determine parameters with one-shot estimators, which makes their training sample inefficient. Though in most PAMDP environments explicit but incomplete knowledge (e.g., rules, safety constraints, or expert heuristics) is available, it is rarely directly used to increase the sample-efficiency of training Reinforcement Learning agents. We step into this gap...

Stochastic-process models are, as a rule, far easier to simulate than to condition. Non-linear observations, non-Gaussian likelihoods, black-box information, and global constraints all induce intractable conditional laws, requiring bespoke, model-specific constructions. We introduce LatentFlow, a single framework for conditioning stochastic processes, with no learned neural approximations and no training. Our starting point is to write the stochastic process as the deterministic image of a tractable latent innovation, $f_0 = T_{\vartheta}(ξ_0)$, with $ξ_0$ sampled from a simple reference dist...

CoCo loss function for learning normalized embeddings with intra-class collapse and inter-class contrast, with theoretical analysis vs. cross-entropy.

Physics-informed fall detection framework using dual-LTC architecture for CoM and BoS subsystems on edge platforms.

Theoretical analysis showing Randomized Hamiltonian Monte Carlo achieves accelerated mixing time for log-concave distributions.

MemOps benchmark for granular evaluation of LLM agent memory operations in long-horizon conversations beyond final QA correctness.

UR-VC method for unsupervised correction of time-derived progress labels in robot learning with contact-rich manipulation.

Energy-based learning framework for inverse problems in tensegrity structure form-finding and physical property prediction.

Pythia multi-agent system for autonomous clinical symptom extraction using open-weights LLMs without fine-tuning.

Study showing LLM judges exhibit bias and over-generosity in no-reference evaluation settings across multiple languages.

Benchmark evaluating LLM capability to identify and correct patient misconceptions across multi-turn medical conversations.

Deep4ge: public dataset of 14,227 DNN training runs with fault annotations for detection and diagnosis from training trajectories.

New York has become the first state to temporarily halt approval of large data centers, as Gov. Kathy Hochul argues the AI-driven building boom shouldn’t come at the expense of higher electricity costs, water supplies, or local control.

ROBIN method localizes and repairs bias in transformer attention heads via inference-time intervention without retraining.

EEC explainer reveals mechanisms of emergent collective behaviors in multi-agent RL for robot swarms despite black-box neural policies.

ChartGenEval introduces corruption-tested benchmark for rhythm-game chart generation evaluation beyond reference-note agreement.

Open-weight reasoning models fine-tuned via RLVR for thermal energy storage control, achieving building-scale load shifting with 30 prompts.

New York’s data center moratorium may become the blueprint for anti-AI movement.

Anthropic launches Claude for Teachers, an educational product tier for classroom use with curriculum resources and safety guardrails.

Superparamagnetic nanodot reservoir computing substrate demonstrates temperature robustness for unconventional ultra-low energy computing.

Meta-evaluation of LLM-generated rubrics for paper reproduction tasks; first systematic study on checklist-based evaluation reliability.

ANGLE framework applies generative learning for non-parametric distributional regression on circular/angular data in computer vision and biology.

KLLMs pretraining paradigm suppresses parametric recall via entity anonymization, forcing evidence-grounded reasoning over memorized knowledge.

Survey of efficient inference techniques for diffusion-based LLMs, covering algorithms and systems beyond theoretical parallel generation claims.

Philosophical analysis of Hempel's statistical ambiguity problem via maximally specific laws; applications to causal AI.

Reflection AI has signed a $1 billion deal to access Nebius's compute. Reflection was founded in 2024 and is developing open source AI technology.

Study examines how iterative AI agent interaction creates neuroplastic conditioning loops, reshaping human cognitive patterns through high-frequency feedback cycles.

Causal intervention analysis reveals Vision-Language Models' Chain-of-Thought reasoning relies on early visual token access, not continuous image grounding during generation.

Hugging Face CEO Clem Delangue says enterprises increasingly want open models, due to cost, accessibility, and ownership. Do frontier models still matter if most production AI ends up running on open models?

PixelLoop introduces pixel-space loop closures for topological navigation, extending SLAM-style corrections to dense relative geometry representations.