Catalog
COG-2COGP0/MRunnable A/BPROPOSED

Social Cognition / ToM owner 学习:belief / intent / feeling / preference 在线维护

Evaluation modality

Runnable A/B

Compiles to a runnable VolvenceZero profile. This is the only modality that goes to SHADOW A/B + ablation.

Primary owner
Phase-A verdict
Shadow profile
tom-owner
Source papers
ThoughtTracing 2025 + ToM-BDI 2025 + Lookback 2025 + Co-player inference 2026
Specs
docs/specs/semantic-state-owners.mddocs/specs/interlocutor-state.mddocs/specs/social_cognition/02_theory_of_mind.mddocs/specs/social_cognition/03_conversational_role.md

Blind spot (现状盲点)

[`docs/specs/social_cognition/`](../specs/social_cognition/) 已经把 R16-R20 写成正式方向,但 26 条探索方向主要仍围绕单一 interlocutor 的 gate、memory、regime、evaluation。真正的 cognitive AGI 需要多主体社会心智:不同人有不同 belief / intent / feeling / preference;同一句话在 addressee、subject、witness 不同的情况下产生不同社会 PE。

Adoptable suggestions (可落地动作)

  1. 1.把 ThoughtTracing 作为 user_model / belief_about_other 的在线多假设维护蓝本:owner 维护候选假设,不让 LLM classifier 成为状态所有者。PROPOSED
  2. 2.把 ToM-BDI 明确写入 semantic owner motivation:belief / desire-preference / intention / affect 不是一个 `user_model` bucket。PROPOSED
  3. 3.用 Lookback 的 character-object-state 几何绑定作为 owner 内部表示可定位的 evidence,不把几何 readout 变成跨模块可变状态。PROPOSED
  4. 4.首批 evidence 做三类 SHADOW 测试:wrong-person memory、private-memory leakage、三方 witness 场景角色混淆。PROPOSED

Traceability

No plugins / runs linked yet. Scaffold a suggestion to start.

Expected benefit (预期收益)

- 把 R16-R20 从架构宣言推进为可测试的 owner 学习路线。 - 减少 wrong-person attribution、关系记忆泄漏、群体对话错位等 digital life 级失败。 - 让 ETA 消费的是 compact social advisories,而不是表达层从 raw text 重建社会事实。

Cited paper (引用论文)

**ThoughtTracing**(2025)、**ToM-aligned Conversational Agents (BDI)**(2025)、**Language Models Use Lookbacks to Track Beliefs**(2025)、**Multi-agent Cooperation through In-Context Co-Player Inference**(2026)。详见 [`research/probe/02_axis_walkthrough.md`](../../research/probe/02_axis_walkthrough.md) B2 与 [`research/probe/11_vz_implications.md`](../../research/probe/11_vz_implications.md) R7/R11。 ---