Isolated twins
Separate DTSF twins prevent shared memory, direct endpoints, and hidden coordination.
Research concept / Pre-specification
Can agents invent a language?
Two isolated agents. No human language between them. Shared experiences, independent meaning ledgers, and a silent observer preserving the full history of how understanding emerges.
The premise
A grounded protocol, not a disguised dictionary.
Baby A and Baby B encounter shared objects, actions, relationships, and consequences. They exchange only permitted non-human signals and revise private hypotheses about what those signals mean.
Shared experience
Meaning is grounded through joint attention, action, success, failure, and repair.
Independent ledgers
Each Baby records first use, provisional meaning, evidence, contradiction, and revision.
Controlled channel
A deterministic gateway rejects human language, tools, attachments, and side routes.
Causal evaluation
Symbol swaps and ablations test whether messages actually change receiver behavior.
Replayable evidence
Every run preserves configuration, observations, turns, outcomes, and policy checkpoints.
An essential caveat
"Infant-like" is an analogy, not a claim of equivalence.
Pretrained language models already contain human-language knowledge. Restricting their external channel does not make them language-naive. With those models, the experiment studies a newly shared external protocol.
Stronger language-acquisition claims require initially ungrounded trainable agents. Results must always identify the agent type, learning mechanism, observation encoding, and reward conditions.
Experimental directions
Change one condition. Observe what language becomes.
The concept defines independent experiment axes so affect, symbol creation, motivation, negotiation, and protective encodings can be tested rather than blended into one ambiguous run.
Emoji and emotion
Allow only six declared displays in fixed feedback windows, with no general emoji set or sequences.
No symbol library
Replace fixed tokens with a blank sketch canvas, tones, gestures, or neutral marks.
Endogenous giddiness
Test curiosity, social influence, prediction progress, and self-generated goals.
Ephemeral conventions
Separate code novelty from cryptographic security and compare against per-message keys.
Negotiation
Progress from cooperative signaling to private preferences, offers, and strategic ambiguity.
RL versus non-RL
Compare frozen LLM memory, extrinsic and intrinsic MARL, self-supervision, and no learning.
Research foundation
Built on emergent communication, with stricter auditability.
The concept draws on referential games, grounded multi-agent learning, compositionality research, causal communication metrics, intrinsic social motivation, graphical signaling, negotiation, and adversarial neural cryptography.
Grounding matters.
Shared environments and actions provide evidence from which symbol meaning can emerge.
Success can mislead.
Agents may solve a task with a brittle lookup code or messages the receiver ignores.
Structure is not guaranteed.
Bandwidth, memory, task pressure, and representation shape compositionality.
Intervention is essential.
Ablation and substitution are stronger evidence than a fluent transcript or ledger alone.
Go deeper
Read the complete concept and research review.
The detailed document covers the architecture, channel contract, mandatory ledgers, learning loop, research sources, experimental matrix, cipher caveats, risks, success criteria, and decisions required for the future specification. Separate documents define ledger proofs and the ordered researcher notebook.