TIME
Temporally Intelligent Meta-reasoning Engine for Context-Triggered Explicit Reasoning
I wanted a model that could decide to reason explicitly when the conversation actually called for it halfway through an answer, instead of committing to one global thinking mode up front.
What bugged me
“Thinking mode” is oddly all-or-nothing.
Many reasoning models make the choice once at the beginning: produce a long visible reasoning block, or don’t. That felt unnecessarily rigid.
Conversations keep changing.
Time passes, assumptions go stale, users return later, and new information can suddenly make an earlier line of reasoning worth revisiting. Temporal cues gave me a clean way to probe that.
What I tried
TIME treats explicit reasoning as a context-triggered policy. I added lightweight dialogue primitives for elapsed time and silent time passage, then trained Qwen3 models from 4B through 32B with a four-stage supervised curriculum. The goal was not to make a clock-reading model; temporality is a controlled source of contextual change.
<time>Optional timestamps expose elapsed time and delayed returns.
<think>Reasoning bursts can appear where needed, not only in one front-loaded block.
tickSilent passage of time without inventing extra user content.
How I tested it
I built TIMEBench: 77 scenarios across seven categories, with ten trials per scenario per model. It is meant to diagnose reasoning under temporal and latent contextual change rather than temporal fact recall.