Paper: AdapTime: Enabling Adaptive Temporal Reasoning in Large Language Models (arXiv 2604.24175). Accepted to ACL 2026 Findings. Authors from Xi’an Jiaotong University, City University of Hong Kong, and Tencent Jarvis Lab. Code: https://github.com/Applied-Machine-Learning-Lab/ACL2026-AdapTime

Problem

LLMs reason well over general knowledge but are weak at temporal questions — anything hinging on time: “Who was CEO of X before Y?”, multi-hop date arithmetic, event ordering, “what was true as of 2019?”. Prior approaches had two flaws: they lean on external tools or manual verification (poor generalizability), and they apply one fixed pipeline to every question. A fixed pipeline wastes compute on easy questions and under-reasons on hard ones.

Core idea: adaptivity

Instead of a static chain, an LLM planner inspects each question and decides which reasoning actions to invoke and in what order, based on the question’s temporal complexity. Simple lookups get a short path; complex multi-hop temporal questions get the full treatment. No external tools required — it works on top of an off-the-shelf LLM.

The three actions the planner orchestrates

  • Reformulate — decomposes a complex temporal question into simpler, focused sub-questions. The heavy lifter for multi-hop queries and intricate time expressions.
  • Rewrite — restates/normalizes the question or intermediate form so temporal constraints are explicit and unambiguous (resolving relative time references, pinning the reference point).
  • Review — a verification/self-check step that catches temporal inconsistencies before committing to an answer.

Results

Consistently beats strong baselines, with the biggest gains on temporally complex tasks, and robust generalizability across scenarios (the point of avoiding scenario-specific tooling).

Why it’s relevant to AI application engineering

Two transferable ideas, even outside temporal QA:

  1. Planner-decides-the-pipeline is a clean, model-agnostic way to do adaptive-depth reasoning — spend tokens proportional to difficulty instead of a one-size chain. Directly applicable to cost/latency control in production agents.
  2. Reformulate -> rewrite -> review is a reusable scaffold for any task where ambiguity and multi-hop decomposition cause errors — essentially a structured self-correction harness you can prompt-engineer onto an existing model without fine-tuning or tools.

Sources