以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

Qwen3-8B

6 atoms · 跨 3 天 · 首见 2026-05-23 · 最近 2026-07-01

三色: 🟦 fact 5 · 🟥 take 1 · stance ▲1/▼0/◆5

来源: X 6

时态: fresh:6

标签: 好数字:3 · 好思考:3 · 好信源:1

2026-07-01 · 自我解释跟踪行为漂移能力
self-explanations track a model's own behavior as that behavior changes, and it shows promise in making introspection training a part of scalable post-training pipelines
tags: 好思考 · → daily
value: qty=true · date=post-SFT

2026-05-23 · 数学基准测试得分提升
DelTA lifts Qwen3-8B by 3.26 points on 7 math benchmarks.
tags: 好数字·好信源 · → daily
value: qty=3.26 points · direction=up
2026-07-01 · 训练后行为自我解释能力
SFT-ed model explains its own current behaviors better than the base model's behaviors
tags: 好思考 · → daily
value: qty=better · date=post-SFT
2026-07-01 · 跨模型标签训练自我解释效果
We train Qwen3-8B on labels from Llama-3.1-8B, and the model still explains itself better than its training data source
tags: 好思考 · → daily
value: qty=better · date=post-training
2026-06-07 · BFCL raw recovery at 36.2% substrate
On Qwen3-8B BFCL, raw recovery at the 36.2 percent substrate was 19.1 percent.
tags: 好数字 · → daily
value: qty=19.1%
2026-06-07 · BFCL recovery with post-attribution objective
With the right post-attribution objective, recovery rose to 84.6 percent at the same channel budget.
tags: 好数字 · → daily
value: qty=84.6%