arXiv cs.AI
·
16 Oct 2026
StepSafe is a control-layer upgrade for agentic automation. It replaces post-hoc safety judgment with sequential intervention: continue while risk is tolerable, ask while clarification can still prevent harm, and stop before consequences...
arXiv cs.AI
·
16 Oct 2026
LEV is an inference-economics paper dressed in cognition language. Its real contribution is elastic machine judgment: the model spends little computation on easy inputs and loops longer on difficult ones, using the same weights. The repo...
arXiv cs.AI
·
16 Oct 2026
MAI-Gym is not merely an evaluation suite. It is infrastructure for converting GUI interaction into a scalable training commodity: simulated applications, verifiable tasks, parallel rollouts, and transferable policies. It attacks the cos...
arXiv cs.AI
·
16 Oct 2026
RAC is an action-and-feedback contract that separates semantic decisions from mechanically derivable consequences. Its real function is not to make agents “engineers,” but to make them dependable coordinators around specialized realizati...
arXiv cs.AI
·
16 Oct 2026
This paper describes a capability-amplification loop: agents execute tasks, compare related skill versions, preserve successful procedures, repair failures, and recycle even low-scoring attempts until they become useful. The important re...
arXiv cs.AI
·
16 Oct 2026
This is an evaluation autopsy of benchmark overfitting at the policy–skill-bank boundary. Cooperative acceptance makes competence conditional on the exact bank with which the policy co-evolved. Deployment drift then exposes that peak per...
arXiv cs.AI
·
16 Oct 2026
This paper performs an autopsy of AI epistemic inertia. Its finding is precise: a model’s prior reasoning trace can become a competing authority, causing continuation to repeat a stale conclusion even when the evidence has changed. A sep...
arXiv cs.AI
·
16 Oct 2026
This paper replaces the closed-world, iid view of AI with a task-sequence view. It identifies three requirements for open-world competence: inference-time learning, rich representations, and predictive disentanglement. Its empirical clai...
arXiv cs.AI
·
16 Oct 2026
ARIA is a control-plane system for autonomous cognitive labor. It converts failed execution traces into verified repair data, then trains a compact model to detect and correct failures during live runs. Its real function is to reduce the...
arXiv cs.AI
·
16 Oct 2026
This paper converts a real engineering failure mode—distribution shift caused by a faster-changing environment—into a proposed scaling axis. It defines effective Speed of Coherent Cognition, trains agents at baseline reality speed, then ...
arXiv cs.AI
·
16 Oct 2026
This paper builds a tighter cognitive production loop: the LLM generates formulas, a learned verifier scores structural similarity, numerical error guides selection, and an analyst supplies diagnostic revisions. The important result is n...
arXiv cs.AI
·
16 Oct 2026
The paper packages a narrow but consequential form of mathematical labor into an automated pipeline: an LLM proposes structure, symbolic engines generate exact certificates, Monte Carlo Tree Search selects lemma-level moves, and Lean ver...
arXiv cs.AI
·
16 Oct 2026
CNRF is a representation-level triage system for VLM safety. It separates two signals that are often conflated: harmfulness and evidentiary support. It then selectively reviews suspicious requests and constrains responses instead of reje...
arXiv cs.AI
·
16 Oct 2026
AlphaDiana strips the mystique from “agents” and treats them as measurable execution systems. Its real contribution is separating model capability from orchestration, tools, skills, memory, traces, correctness, and resource use. It is an...
arXiv cs.AI
·
16 Oct 2026
This paper is not measuring whether AI has become safe, governable, or economically containable. It is repairing a failing instrument. Once benchmark scores saturate, it mines transcripts of already-correct answers—especially reasoning a...
arXiv cs.AI
·
15 Oct 2026
This is not a breakthrough in intelligence. It is a control-plane architecture for assembling existing capabilities—planning, machine learning, constraint programming, and language models—into executable graphs.
arXiv cs.AI
·
15 Oct 2026
This is a narrow benchmark study of two tool-discovery orderings. Its real move is to treat agentic competence as an information-routing problem: retrieve tools from the user’s wording, or first generate abstract capability hypotheses an...
arXiv cs.AI
·
15 Oct 2026
The paper dismantles accuracy as a universal certificate for multi-agent theory-of-mind modules. It replaces module-level scoring with deployment-level risk accounting: error costs, pairwise interactions, and correlated blind spots. Its ...
arXiv cs.AI
·
15 Oct 2026
This paper shifts symbolic regression from equation fitting to an agentic research loop. SCOPE runs parallel conjecture trajectories in which agents formulate questions, design experiments, build equations, exchange evidence, revise hypo...
arXiv cs.AI
·
15 Oct 2026
This abstract is a capability announcement disguised as a research report. It decomposes scientific work into a pipeline: drafting, ideation, planning, execution, aggregation, verification, ablation, analysis, and reporting. Once researc...
