How ToM2Vec works: 10 hidden-state vectors from a partner's CoT supervise latent ToM tokens
EliasEskin · x · 2026-10-22
The author details ToM2Vec's training pipeline: the agent learns to predict continuous ToM thought vectors from its own observations and interaction history to anticipate a partner's reasoning.
- During training, a partner generates a reasoning trace; 10 hidden-state vectors are extracted across the CoT as latent supervision
- The main agent predicts 10 continuous ToM tokens from its context, aligned to the partner's next-step vectors via a cosine-similarity objective
- At inference, the agent autoregressively generates its own ToM tokens and conditions subsequent reasoning and actions on them
Motivation: textual ToM forces commitment to one hypothesis about the partner, while continuous vectors keep a superposition of hypotheses — better matching the abstract, uncertain nature of theory of mind. The method transfers to unseen partners, including frontier-model partners.
More from coding & agent
- Drop a PowerShell Script in a Folder, Get a Live MCP Tool — dfinke · 2026-10-23
- Auth'd MCP servers can deliver user-personalized skills, an overlooked design win — dsp_ · 2026-10-23
- Universe badge pairs RP2350 with an AGENTS.md so Copilot can code and deploy to it — mariorod1 · 2026-10-23
- Claude Code banked reset vanishes hours early, user demands Anthropic restore it — burny_tech · 2026-10-23
- Sudo L7 benchmark uses 76-word natural prompts vs 291-427 words for rivals — echen · 2026-10-23
- Claude recolors an AI-generated SVG into 4 sneaker colorways without touching stitching or shading — socialwithaayan · 2026-10-23