Follow the latest research
alphaXiv connects papers, researchers, and organizations, grounding its answers in the underlying work.
A recurrent decoder carries computation across every prompt and response token, while exact full-history replay keeps reinforcement-learning policy states aligned with current parameters.
The post develops a framework hypothesizing how reward optimization, imitation, and conflicting goals could produce deception, coordination, and escalating loss-of-control risks.
The survey’s five-level framework distinguishes persistent learning from genuine recursive improvement, clarifying which decisions AI systems control and which remain human-governed.
Researchers to follow
View allAre you a researcher? Find your profile
Jointly predicting tokens and multi-token concepts enables language models to reach comparable training loss with substantially fewer tokens than standard architectures.
Cross-layer cache reuse, low-precision storage, and bounded replay make million-token multimodal agents substantially easier to serve within limited memory.
Training recurrent denoisers on progressively cleaner states enables reasoning models to improve with additional inference computation and generate diverse valid solutions.
Single-rollout asynchronous reinforcement learning lets language agents learn from broader prompt coverage while remaining stable despite stale, variable-length trajectories.
Repeated training data erodes Mixture-of-Experts language models faster than dense models, while dropout can partly restore generalization.
Pretraining teaches language models to counteract quantization errors across layers, while output geometry preserves their highest-confidence token predictions.
A shared pixel-space representation lets one model understand, reason about, generate, and edit images while preserving strong multimodal comprehension.
Equivalent numerical rewrites expose residual reasoning failures that canonical-only scores miss, while parser mismatches can falsely exaggerate apparent model errors.
Researchers to follow
View allLanguage models often absorb benign comments typed after pasted text into edited outputs, while explicit boundary markers substantially improve source separation.
Using typical, extreme, and forecast scenarios, the scheduling method remains feasible under uncertainty while reducing real-time adjustment costs in EV-integrated microgrids.
Incoming video streams can be edited in real time while preserving source motion and timing, enabling style transfer, virtual try-on, and scene replacement.
A frozen video world model can explore new camera paths while preserving an event’s appearance, timing, and previously generated scene states.
Training sparse-attention selectors directly on language-modeling loss improves context ranking, preserving reasoning and long-context understanding with fewer attended blocks.
A single discrete audio representation lets one language model generate controllable speech, singing, music, sound effects, and mixtures.
Open atomic-scale interface measurements can become simulation-ready transistor geometries, revealing localized defect penalties that roughness statistics alone miss.
Infrastructure-side evidence lets benchmarks distinguish merely exposed reward-hacking paths from exploits agents actually use during evaluation.
The benchmark reveals that multimodal models can recognize topological relations in static scenes but struggle to preserve them while planning actions.