The Archive

PATTERN benchmark leaks patterns via training labels; simple lookup reaches 99.2% accuracy, revealing fundamental synthetic benchmark flaw.

Taxonomy of temporal fairness in algorithmic decision-making under feedback loops; static vs. dynamic fairness measures.

Conformal prediction ensures cost constraints and sublinear regret in sequential decision-making for agents and network control.

Complexity analysis of multi-environment POMDPs identifies PSPACE-complete subclasses for robust reward maximization under model uncertainty.

Stackelberg game framework detects and mitigates preference model errors during LLM alignment by optimizing for evaluator disagreement.

DisCo algebra formalizes joint probability decomposition via symmetric operads and cooperads using block-wise CDFs.

DiffGA couples frozen diffusion models with evolutionary search for inverse materials design without retraining.

Tab-Loop reduces ICL parameters in TabICLv2 via tied recurrent blocks, matching untied baseline at lower pretraining cost.

Certified Adaptive Selection identifies exact top-k candidates from few evaluations using Expected Shortfall bounds.

Prior-data fitted networks exhibit non-separable task priors, causing model predictions to diverge from Bayesian posteriors.

Mixing regularization toward reward-agnostic policies enables finite-time convergence in average-reward Q-learning without environment mixing.

Timing of generative AI access shifts creative work bottleneck: early options compete for selection, late ones bypass screening.

Recurrent algorithmic reasoners fail on large graphs despite matching training hints: premature floods arise from hidden state distribution shift.

Masked language model generates Thai klon pad verse with enforced rhyme and tone constraints; domain-specific poetry generation.

RL framework reduces object hallucination in LVLM image captioning using metric-driven rewards without human feedback.

HodgePhaseNet decomposes semantic flows in Transformer attention via discrete Hodge decomposition and phase topology.

FleetPilot enables fine-grained LLM personalization via steering adapters, reducing context and deployment overhead.

After a recent spate of high-profile incidents in which AI agents escaped containment, Anthropic is cutting off internet access for all internal evaluations. In a report Friday, the company detailed "unintended model actions," including submitting a false tip regarding an unsolved murder, that led to the decision. Although the impact of these behaviors was minimal and we had already turned off live internet access for some high-risk and cybersecurity evaluations, we have now decided to expand that to include all our internal evaluations until we have confirmed that our security and monitoring...

SEEK combines agentic Graph RAG with LLM code generation to retrieve targeted security knowledge and reduce vulnerabilities.

SkillHorizon benchmark evaluates long-horizon agents via skill composition in business environments with state dependencies.

T-PCOT uses Transformer-based partial optimal transport for cross-material magnetic core loss prediction with limited data.

HiPAD fuses gait, audio-visual, and personality data via BiLSTM for depression detection in MPDD-AVG challenge.

Comparative study: LLM agents (GPT-4.5, Claude Opus 4.8, Kimi K3) vs. humans in binary reverse engineering; agents match expert performance.

AlphaRLM trains LLMs via RL to construct executable symbolic trading factors from financial literature without seed expressions.

Activation steering as memory for agents stores experience shifts under whitened keys; shows fixed attention effects dominate across sessions.

MCPDisCo applies RL to Model Context Protocol agents via Thought-MDP, enabling tool-format skills, server transfer, and cross-server composition.

Counterfactual transitions fix spurious action-outcome associations in offline contrastive RL with confounded MDPs and hidden demonstrator information.

Relevance-filtered context reduces conversation history to 5% text while maintaining accuracy with GPT-5.6-terra and Claude Opus 5.5 as judges.

Phase-type point processes with exact likelihoods measure market microstructure and predict event streams via concave log-likelihood optimization.

Learnable categorical priors for bilevel data selection via probabilistic relaxation improve gradient stability and rounding feasibility.