TIME: Temporally Intelligent Meta-reasoning Engine for Context Triggered Explicit Reasoning
Susmit Das
TL;DR
The paper addresses inefficiencies and auditability challenges of front-loaded chain-of-thought reasoning in dialogue systems. It introduces TIME, a Temporally Intelligent Meta-reasoning Engine, which treats explicit thinking as a context-sensitive resource activated by discourse and temporal cues using three primitives: optional <time> tags, transient <think> blocks, and tick turns. TIME is trained via a four-phase curriculum on Qwen3 models (4B–32B) with adapter-based fine-tuning and a small, maximally diverse full-batch Phase 4 alignment, aiming to produce brief in-place reasoning bursts that can be re-triggered mid-response. Evaluation with TimeBench, a 77-scenario temporally grounded dialogue benchmark, shows consistent improvements over base Qwen3 in temporal reasoning across model sizes while substantially reducing reasoning tokens, demonstrating the effectiveness of a policy-based, context-triggered approach. Overall, TIME converts explicit reasoning from a fixed template into a reactive, concise practice aligned with conversational state changes, enabling more efficient and auditable dialogue systems, albeit with limitations such as language scope and the absence of reinforcement learning or safety considerations.
Abstract
Reasoning oriented large language models often expose explicit "thinking" as long, turn-global traces at the start of every response, either always on or toggled externally at inference time. While useful for arithmetic, programming, and problem solving, this design is costly, blurs claim level auditability, and cannot re-trigger explicit reasoning once the model begins presenting. Dialogue models are also largely blind to temporal structure, treating replies after seconds and replies after weeks as equivalent unless time is stated in text. We introduce TIME, the Temporally Intelligent Meta-reasoning Engine, a behavioral alignment framework that treats explicit reasoning as a context sensitive resource driven by discourse and temporal cues. TIME augments dialogue with optional ISO 8601 <time> tags, tick turns that represent silent gaps, and short <think> blocks that can appear anywhere in a reply. A four-phase curriculum including a small, maximally diverse full-batch alignment step trains Qwen3 dense models to invoke brief, in-place reasoning bursts and keep user facing text compact. We evaluate with TIMEBench, a temporally grounded dialogue benchmark probing chronology, commonsense under gaps and offsets, anomaly detection, and continuity. Across 4B to 32B scales, TIME improves TIMEBench scores over base Qwen3 in both thinking and no-thinking modes while reducing reasoning tokens by about an order of magnitude. Our training data and code are available at https://github.com/The-Coherence-Initiative/TIME and TIMEBench is available at https://github.com/The-Coherence-Initiative/TIMEBench