PulseAugur
中
实时 16:16:54
English(EN) Self-distillation trains only the student. Its teacher can use the extra information to exploit shortcuts unavailable to the student.

Perplexity 的 JOLT+ 方法以更少的训练 token 优于 GRPO

Perplexity 推出了 JOLT+,一种新颖的训练后方法,其性能显著优于 GRPO 等现有的强化学习技术。JOLT+ 采用一种自蒸馏方法,其中一个模型同时充当教师和学生,教师向学生提供反馈。该方法在包括数学和编码任务在内的各种基准测试中都表现出卓越的性能,与 GRPO 相比,在训练 token 数量大大减少的情况下实现了更高的准确性。 AI

影响 这种新的训练方法可能导致更高效的语言模型微调,降低计算成本并提高复杂任务的性能。

排序理由 该集群描述了一种新的研究方法及其基准性能,而非产品发布或前沿模型。

在 X — Perplexity 阅读 →

AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →

Perplexity 的 JOLT+ 方法以更少的训练 token 优于 GRPO

本文如何被排名

Signal score
0 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Research
该集群描述了一种新的研究方法及其基准性能,而非产品发布或前沿模型。
Source corroboration
4 independent sources
Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics
model release, product
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
31 days old
Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.

完整方法见我们的编辑标准。

报道来源 [4]

  1. X — Perplexity TIER_1 English(EN) · perplexity_ai ·

    在数学、编码、工具使用和终端任务方面,JOLT+ 在包括 AppWorld 的 9.9 分在内的所有五项基准测试中均优于 GRPO。

    Across math, coding, tool use, and terminal tasks, JOLT+ beat GRPO on all five benchmarks, including 9.9 points on AppWorld. https://t.co/KFHdZ0tuFn https://t.co/rKojjZn1sd

  2. X — Perplexity TIER_1 English(EN) · perplexity_ai ·

    自蒸馏只训练学生模型。其教师模型可利用额外信息来利用学生模型无法利用的捷径。

    Self-distillation trains only the student. Its teacher can use the extra information to exploit shortcuts unavailable to the student. JOLT also trains the teacher on outcome rewards from its own attempts, with a penalty when its predictions diverge from the student's.

  3. X — Perplexity TIER_1 English(EN) · perplexity_ai ·

    像GRPO这样的强化学习方法每次尝试只给一个奖励,这并不能显示哪些步骤有帮助。

    Reinforcement learning methods like GRPO give one reward per attempt, which doesn't show which steps helped. Self-distillation adds per-token feedback from a teacher. The same model plays the teacher, with extra information like a verified solution, and the student, without it. …

  4. X — Perplexity TIER_1 English(EN) · perplexity_ai ·

    新研究:我们开发了JOLT,一种训练后方法,其中一个模型既充当教师又充当学生,教师经过训练以提供学生反馈

    New research: We developed JOLT, a post-training method where one model acts as both teacher and student, with the teacher trained to give feedback the student can use. On MATH, it matched GRPO's accuracy with 6.5x fewer training tokens. https://t.co/KFHdZ0tuFn https://t.co/Y4S…