TIME: Temporally Intelligent Meta-reasoning Engine for Context-Triggered Explicit Reasoning

` Blocks:** Optional, short reasoning bursts that can appear anywhere in a response (start, middle, or end), allowing for re-triggering of reasoning within a single turn.
3. Tick Events: User turns containing only a timestamp, representing silent time passage without new text.

Training Curriculum

The framework aligns Qwen3 dense models (4B, 8B, 14B, and 32B parameters) using a four-phase supervised fine-tuning (SFT) curriculum:

• Phase 1 (Structural Seeding): Introduces primitives and output formats using synthetic single-turn data to teach compact, well-delimited reasoning.
• Phase 2 (Temporal Exposure): Uses two-turn dialogues with time gaps and tick events to teach the model to condition on temporal metadata, revise assumptions after silence, and suppress verbosity when unnecessary.
• Phase 3 (Contextual Modulation): Extends to multi-turn settings, training the model to suppress or re-trigger reasoning blocks under changing context and increased tick frequency.
• Phase 4 (Gradient-Aligned Convergence): A decisive alignment step using a small (128 samples), maximally diverse full-batch. This set isolates the target invariant: reasoning bursts triggered only by temporal or discourse cues. This phase uses deterministic full-batch updates to concentrate gradients on the policy while avoiding sampling variance.

Evaluation: TIMEBench

The authors introduce TIMEBench, a diagnostic benchmark consisting of 77 scenarios across seven categories. Unlike benchmarks focused on temporal fact recall, TIMEBench tests reasoning under temporal discontinuity, anomaly, and contextual re-anchoring. Categories include:

• Chronological Retrospection
• Invalid Time Detection
• Temporal Adaptivity (e.g., approaching deadlines)
• Temporal Contextual Awareness (e.g., holidays, late-night contexts)
• Temporal Flow Anomaly Detection (e.g., impossible time jumps)
• Time Gap Awareness
• Timezone Sensitivity

Evaluation uses an LLM-as-a-Judge (blind to prompts and timestamps) with binary scoring and bootstrapped confidence intervals.

Key Results

Across all model scales (4B to 32B), TIME models significantly outperform base Qwen3 models in both "no-thinking" and "thinking" modes on TIMEBench:

• Performance Gains: TIME-32B achieved a score of 64.81, compared to 37.40 for the base Qwen3-32B in thinking mode and 31.82 in no-thinking mode. Similar improvements were observed across 4B, 8B, and 14B models.
• Statistical Significance: Improvements were statistically significant ($p < 0.001$) across all scales via Wilcoxon signed-rank tests.
• Efficiency and Structure:
  • Token Reduction: TIME models reduced explicit reasoning tokens by roughly an order of magnitude (e.g., ~84 reasoning tokens per run for TIME-32B vs. ~911 for base thinking mode).
  • Placement Shift: The location of reasoning shifted from front-loaded (100% at the start for base models) to predominantly mid-turn (75.6% for TIME-32B).
  • Degeneracy: TIME models showed significantly lower rates of output degeneracy (e.g., 3.64% for TIME-32B vs. 18.18% for base thinking mode).

Significance and Claims

The paper claims that TIME demonstrates that explicit reasoning can be aligned as a selective, context-triggered policy rather than a fixed decoding style. The significance of the work lies in:

1. Behavioral Shift: Moving from "always-on" or "never-on" reasoning to a policy where reasoning is deployed selectively in response to latent state changes signaled by temporal cues.
2. Improved Auditability: By using short, localized reasoning bursts, the mapping between specific claims and their justifications becomes more direct.
3. Temporal Cognition: The framework shows that models can learn to use temporality not just as content to be recalled, but as a structured probe for latent context changes (e.g., stale assumptions, urgency) that require re-anchoring.

Limitations and Scope

The authors explicitly note the following limitations:

• Model Scope: Experiments are limited to dense hybrid reasoners from the Qwen3 family; generalizability to pure instruction models or mixture-of-experts architectures is not claimed.
• Benchmark Scope: TIMEBench is a preliminary diagnostic tool with 77 scenarios; it does not evaluate performance on broader task-based benchmarks like general math or coding.
• Evaluation Method: Scores rely on LLM-as-a-Judge rather than human annotation, introducing potential evaluator noise.
• Mechanistic Interpretability: The work addresses when and how reasoning is surfaced but does not provide mechanistic interpretability of the underlying model states.
• Language: Experiments are conducted only in English.

The paper concludes that TIME and TIMEBench provide a foundation for teaching models to deploy explicit reasoning selectively, though broader transfer and more robust benchmarking remain open areas for future work.