The Agent Architect #42: Multi-Criteria Weighted Scoring

今週のパターン

Multi-Criteria Weighted Scoring

概要:

複数の重み付き基準で候補を採点し、合計スコアで順位づけしてトレードオフを可視化し、意思決定を裏づけます。

使いどころ:

関係者ごとに重視する要素が異なり、透明性が求められる状況で、競合する選択肢から選ぶとき。

注意点:

重みや採点尺度は主観的です。わずかな変更で順位が逆転し、本当の対立が価値観の違いであることを覆い隠します。

今週のエージェントAI

• DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Since Day 0vLLM

  vLLM achieved 1.9x speedup and 5x agentic throughput improvement for DeepSeek-V4.1-Flash within three weeks.

• vLLM Support for NVIDIA Vera Rubin NVL72: 7.8x Throughput over GB200 NVL72vLLM

  vLLM achieves 7.8x higher per-GPU throughput on NVIDIA Vera Rubin NVL72 versus GB200 NVL72 on AgentX.

• Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech ModelarXiv cs.AI

  Distilled text-to-speech model reduces parameters from 82M to 8M with 15x less compute.

• Ai2 at COLM 2026: Open research, from models to agents for scienceAi2

  Ai2 presented advances in open language models, training infrastructure, and AI for science at COLM 2026.

• Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real HarnessesMicrosoft Research

  Agent Lightning is a 3,500-line framework for training RL agents with existing tools without rebuilding.