The Archive

BOTTLED benchmark tests LLM agents' ability to autonomously create cost-effective task-specific solutions under compute budgets.

AdvSim2Real co-evolves web agent defenses and adaptive prompt injection attacks in a world model to improve robustness.

Fredholm stabilization control design for Kuramoto-Sivashinsky PDE with spatially-varying anti-diffusion coefficients.

VeriFine co-evolves policy, curriculum, and verification judge for self-improving embodied reasoning agents.

WorldSonus adds real-time spatial audio synthesis to world models via video-to-audio framework with interactive control.

Neural Petri flows preserves chemical reaction semantics (valence rules, bond conservation) in learned neural architectures.

Dual minimal pair framework detects inconsistent stereotype preferences in LLM bias evaluations via data augmentation.

Linear bandits under sliding-window constraints with transition diameter-based rare-switching algorithm.

RLCP adapts recommendation action sets via critic scores and conformal thresholds for dynamic slate sizing.

Polynomial-time algorithm for robust bandits with optimal O(√T) regret when learner's model misspecified.

H-CDLMs combine hierarchical representations with continuous diffusion for improved parallel language generation.

Polynomial-time deterministic algorithm achieves optimal O(√d) regret in online inverse linear optimization.

datasette-atom 0.11a0 fixes compatibility with Datasette 1.0a41 alphas.

Systematic study of semantic DocID space design in generative information retrieval with efficient evaluation metrics.

EgoLAP pre-trains vision-language-action models on human+robot trajectories via language-based motion intent chains.

Study shows agent history weakly predicts when context compaction harms performance on TRACE corpus of 590 AppWorld boundaries.

LLM agents generate physics solvers via code to simulate complex dynamics for embodied AI, games, and scientific problem-solving.

Mathematical analysis of unbiased evaluation for Best-of-N selection using holdout scoring matrices across K independent judge scores.

Finetuning-induced knowledge loss in LLMs is often recoverable; spurious forgetting mechanics explained via associative memory model.

Joint training of interpolation paths and velocity fields in flow matching identifies path overfitting as failure mode.

Ablation study isolates effects of decoder paradigm, document identifiers, and decoding strategy in generative retrieval on NQ320K and MS300K.

Statistical method for spatiotemporal forecasting with confidence intervals using labeled recent data and long-term unlabeled covariates.

Cross-model LLM consensus on student math error diagnosis shows agreement does not validate accuracy in K-12 tutoring dialogue analysis.

nanoMuse: open-source personal agent framework supporting multi-device accounts, persistent memory, and proactive conversation.

GeneICL: transcriptomics-aware tabular foundation model using in-context learning for gene expression-based clinical outcome prediction.

ScienceClaw benchmark evaluates AI agents' ability to self-improve code across 23 scientific disciplines via iterative task solving and verification.

Framework for counterfactual inference in Gaussian-process causal models with mixed discrete/continuous variables using noise-abduction procedures.

Study of turn-taking timing between two full-duplex speech models in self-play dialogue reveals coupled floor-transfer patterns distinct from human-model interaction.

Systematic evaluation of small LMs on ARC-TGI abstract reasoning benchmark probes transferable rule learning vs. distribution fitting across 1,000+ runs.

First security analysis of speculative decoding for LLMs reveals vulnerabilities when draft models generate unverified candidate tokens accepted by target models.