FUTURE WORK / INTERFACE
Conversation Flow tags
A proposed control vocabulary for MacroCF.
At each decision step, the interface would return one control tag, optionally preceded by a brief <think> block. The tags describe the next conversational decision; the application determines how that decision maps to playback, speech generation, or listening.
| Tag | Meaning | Intended interpretation |
|---|---|---|
<EOT/> | End of turn | The target speaker has finished a turn; a response may begin. |
<INT/> | Interruption | A participant is taking the floor; the application can yield or stop its own speech. |
<LTN/> | Listen | Continue listening and waiting; there is not yet enough evidence to act. |
<BC/> | Backchannel | A brief acknowledgement is appropriate while the other participant retains the floor. |
<REJ/> | Reject / no response | Choose not to respond to the current input, for example when the speech is not directed at the agent. |
Listening and non-response are different
LTN means waiting for further context before deciding. REJ means deciding that the current input should not receive a response. Neither tag prevents reconsideration when new context arrives.
Optional <think> block
Reasoning-enabled example:
<think>The speaker is still developing the request; continue listening.</think> <LTN/>
Direct-tag example:
<LTN/>
These are illustrative interface examples, not recorded model outputs. The <think> block is optional; one control tag remains required for each output step.
Decision cadence and actions
The design targets a decision every 200 ms using accumulated context. A repeated tag need not retrigger an application action: continuous listening can remain active across multiple steps, while discrete actions need application-level event handling.
Evaluation scope
The five-tag interface is a research direction for MacroCF. TurnBench EOT and INT event lists are a separate evaluation format. The existing MicroCF results do not measure LTN, BC, or REJ, nor do they validate this proposed interface.