#评测

共 121 篇带此标签

121 articles Newest 2026-11-21
Archive · #评测 11 articles
Tools

dettato’s 21M Audio-to-MIDI Model Beats 1.3B

dettato claims strong audio-to-MIDI results with roughly 21 million parameters and a model size about 60 times smaller than muscriptor 1.3b. Its open weights make the project easy to test, though independent benchmarks, training details, and licensing remain necessary to verify.

2026-11-21 Source 2026-11-17 7 min 33 views
LLMs

LLM Weight Rewriting Is Not Self-Evolution

A close look at arXiv paper 2611.21885, which explores how language models can modify their own weights—and why evaluation, rollback, and reproducibility matter more than a single benchmark gain.

2026-11-21 Source 2026-11-20 8 min 33 views
LLMs

Gemini 4 Neon Leads MEI: Engineering, Not Benchmarks

Gemini 4 Neon has taken the top spot on the Models Engineering Index, but the ranking measures performance on practical engineering tasks rather than overall model capability. A closer look at MEI’s methodology, limitations, and model-selection process.

2026-11-20 Source 2026-11-19 7 min 33 views
LLMs

Gemini 4 Neon: Look Beyond Benchmark Scores

Choosing an AI Agent requires more than comparing model benchmarks. A practical evaluation should account for tool use, context caching, retries, human intervention, and the cost of migrating production workloads.

2026-11-19 Source 2026-11-18 7 min 33 views
Tools

64% of Agent Compute Is Spent on Failed Tasks

An analysis of 27,280 coding agent runs from FrontierHarness Eval v1.1 shows that failed tasks consumed 64% of total inference spending. The findings point to the need to measure retries, tool usage, and failed execution paths alongside success rates.

2026-11-19 Source 2026-11-17 6 min 33 views
LLMs

GPT-6.2 Astra vs. Sol: An Agent Selection Guide

OpenAI has released GPT-6.2 Astra and Sol on the same day. This guide compares task success rates, per-task costs, and retry behavior, then outlines a practical workflow for testing single-model and dynamically routed Agent setups.

2026-11-18 Source 2026-11-17 7 min 33 views
Opinion

Can Astra Really Rewrite Qwen’s Weights?

A close look at Astra’s claim that it can “read and write weights to improve Qwen,” with a practical checklist for verifying parameter changes, baselines, controls, and evaluation contamination.

2026-11-17 Source 2026-11-03 8 min 33 views
Opinion

Filling the Post-Training Data Gap for AI Agents

OpenAthena argues that a shared commons for post-training data does not exist. For AI Agent teams, the priority is not collecting more samples, but building high-quality task trajectories that capture tool state, failure causes, and acceptance criteria.

2026-11-16 Source 2026-11-03 8 min 33 views
LLMs

Qwen4-Max: Agent Selection Goes Beyond Benchmarks

Qwen4-Max reached the Hacker News front page with 483 points and 197 comments, but Agent teams should evaluate more than model rankings. A reproducible baseline of 30 tasks, each run five times, reveals the real cost, latency, and recoverability of each option.

2026-11-16 Source 2026-11-16 8 min 33 views
LLMs

Can Tencent’s Rebuilt Hy4 Architecture Be Reproduced?

Hy4 reached the Hacker News front page, but its official summary omits model parameters, training details, code, weights, and benchmark data. This analysis outlines the reproducibility checklist a model release should meet and what Tencent still needs to disclose.

2026-11-16 Source 2026-11-04 7 min 33 views
Case Studies

The Hidden Cost of Non-English LLM Requests

Research from the University of Michigan finds that generative AI generally consumes more energy on non-English text. This article examines translation overhead, tokenization differences, and Agent retries, then outlines practical ways to monitor and route costs by language.

2026-11-15 Source 2026-11-10 8 min 33 views