برچسب: مدل‌های زبانی

Enterprise AI Assistant Compliance Under Pressure

PACT examines whether enterprise AI assistants continue following rules under urgency, user pressure, and tempting shortcuts—and shows that correct answers in calm conditions are not enough for operational…

SIEVE: Selective Unlearning in Vision-Language Models via Attention-Value Suppression

SIEVE suppresses attention values for forget examples and matches them to a reference model for retain examples, reducing recovery of sensitive image–question–answer associations. Reported MLLMU-Bench results also show…

A Rubric Landscape for Evaluating Clinical Reasoning in Large Language Models

This preprint, whose independent peer review is unconfirmed, examines how far existing tools can evaluate clinical reasoning by large language models over longitudinal, multi-document records.

In Video Safety, “Unsafe” Is Not a Sufficient Answer

A close reading of a preprint on multi-label video safety detection and controlling false-positive versus false-negative errors during training.

When Translations Need to Reach the Same Router

This research proposes aligning expert-selection distributions, rather than hidden states directly, to bring language processing closer together in MoE models—an encouraging idea with clear boundaries.

Cross-Lingual Alignment in MoE Models Through Router Outputs

This paper introduces a method for aligning processing across languages in decoder-only MoE models by bringing the mean expert-routing distributions of translated sequences closer together. Experiments on four…

Global Coherence in Multi-Agent Collaboration

This paper studies cases where each agent acts correctly from its local view, yet the team’s combined result violates shared constraints. Its central proposal is an external harness…

Mem++: Non-Destructive Long-Term Memory for Organizational Agents

Mem++ preserves complete organizational documents with dates and authors, deferring the choice among conflicting versions until question answering.

When the Same Reward Teaches Different Visual Skills

This preprint argues that improved visual-benchmark scores need not come from learning from images; task design must make visual evidence necessary for reward.

VISTA: Lossless Visual Memory for Multimodal Agents

VISTA is a visual harness that lets multimodal models retrieve and re-examine past frames at their original detail. The authors report that this design substantially improves the underlying…