Abstract
Temporal Knowledge Graph Question Answering (TKGQA) demands multi-hop reasoning under temporal constraints. Prior approaches based on large language models (LLMs) typically rely on rigid, hand-crafted retrieval workflows or costly supervised fine-tuning. We show that simply granting an off-the-shelf LLM autonomy—letting it decide what to do next—already yields substantial gains even in a strict zero-shot setting. Building on this insight, we propose AT2QA, an autonomous, training-free agent for temporal question answering that iteratively interacts with the temporal knowledge graph via a general search tool for dynamic retrieval. Experiments on MultiTQ demonstrate large improvements: AT2QA achieves 88.7% Hits@1 (+10.7% over prior SOTA), including a +20.1% gain on challenging multi-target queries, showing that agentic autonomy can decisively outperform fine-tuning for temporal question answering.
The Problem: The Fragility of Rigid Workflows
Temporal Knowledge Graph Question Answering involves time-sensitive facts, such as "Who was the French foreign minister before the 2012 election?" This requires models to perform multi-hop reasoning while respecting strict temporal constraints. Current LLM-based systems suffer from several critical flaws:
- Rigid Workflows: Most systems use static pipelines. If an initial retrieval step fails or returns noisy data, the error propagates through the entire process with no mechanism for recovery.
- High Costs: State-of-the-art (SOTA) performance often depends on expensive supervised fine-tuning (SFT) or Reinforcement Learning (RL).
- Lack of Autonomy: Existing methods force LLMs to follow human-designed paths rather than allowing the model to dynamically determine what information it needs next based on intermediate findings.
AT2QA: An Autonomous Agentic Approach
AT2QA (Autonomous and Training-free Agent for Temporal Question Answering) reimagines the LLM as an agent interacting with a Temporal Knowledge Graph (TKG) environment. Instead of a single-pass workflow, the agent follows an iterative cycle: Thought → Action (Search) → Observation → Self-Correction.
Core Components
- The Agentic Loop: If the agent retrieves insufficient data, it can autonomously rewrite its search query, adjust time windows, or roll back its reasoning state. This "meta-cognitive" behavior allows it to avoid the "greedy acceptance" trap of static RAG systems.
- Structured Temporal Search Tool: The agent uses a tool supporting symbolic filtering (by time or entity), semantic ranking (via dense embeddings), and hybrid sorting (chronological or relevance-based).
- Training-Free Experience Mining: To boost performance without parameter updates, the authors use a selection scheme to identify "advantage experiences"—successful reasoning paths from training questions—and use them as few-shot demonstrations to elicit the model's latent capabilities.
Key Results and Insights
Evaluated on MultiTQ, a benchmark with ~500k question-answer pairs, AT2QA established a new state-of-the-art. The most striking finding is that an off-the-shelf LLM with a search tool can outperform models specifically fine-tuned for the task.
- Superior Accuracy: AT2QA achieved 88.7% Hits@1, a +10.7% absolute gain over the previous best model (Temp-R1). Even in a zero-shot setting, it achieves 84.4%, surpassing fine-tuned competitors.
- Multi-Target Mastery: On complex queries requiring multiple answers, the agent saw a +20.1% leap in accuracy, demonstrating the power of iterative exploration.
- Cost Efficiency: The framework is model-agnostic and cost-effective; testing the entire 56,000-question set cost only ~$150 in API fees, far less than the compute required for large-scale fine-tuning.
Conclusion
AT2QA demonstrates that the primary bottleneck in TKGQA is often not the model's intelligence, but the rigid interfaces we impose upon it. By granting LLMs autonomy to search and self-correct, we can unlock significant latent reasoning capabilities. Future work will focus on reducing the latency inherent in multi-turn interactions and improving early-stopping policies to balance thoroughness with efficiency.
Summary updated