# 马东锡 NLP：SSI 或正构建可反复改进"工作区"的模型

- 来源：Dongxi 东锡 NLP (@dongxi_nlp)
- 发布时间：2026-08-14 05:18
- AIHOT 分数：30
- AIHOT 链接：https://aihot.virxact.com/items/cmss2dsoz04ncroh6xqnnqqi5
- 原文链接：https://x.com/dongxi_nlp/status/2088012274602557685

## AI 摘要

马东锡 NLP 认为“工作区”是模型作答前内部持有、评估和修正想法的区域，SSI 或正构建能反复改进该工作区、判断思考方向并保留有用发现的模型。其推荐 BIGAI 研究团队的博客《Reasoning as Value-Guided Latent-Space Optimization》，阐释模型如何利用额外测试时计算改进隐藏状态。

## 正文

I began to like the word "Workspace".

It is the internal area where a model holds, evaluates, and revises ideas before producing an answer.

Based on Ilya Sutskever's recent emphasis on generalization, value functions, and continual learning, it seems that SSI may be building a model that can repeatedly improve this workspace, judge whether its thinking is moving in the right direction, and retain useful discoveries.

Anthropic's Jacobian Lens offers a related clue: it helps researchers understand how changes in middle-layer representations shape a model's eventual output.

The Jacobian Lens reads the workspace.
Latent reasoning updates the workspace.
A value function tells it which direction to move. Continual learning preserves what worked.

Highly recommended a blog from @ZilongZheng 's BIGAI Research Team.

This excellent blog explains the latent-reasoning part of that picture:

How a model can use extra test-time compute to improve its hidden state before answering.

Blog:

Reasoning as Value-Guided Latent-Space Optimization

https://latentreasoning.github.io/test-time-latent-reasoning/