When a human is asked, "Who was the president of France before Emmanuel Macron?" they don't just search for "president of France" and "Emmanuel Macron" as a single blob of text. They instinctively break the problem into two parts: identifying the current anchor (Macron) and then looking for the entity that occupied that same role in the preceding time interval.
Most artificial intelligence models, however, are "time-blind." They treat complex temporal queries as a "bag of words," often failing to distinguish between "A before B" and "B before A." To solve this, a growing body of research advocates for Temporal Query Decomposition: the process of disentangling a query's semantic core from its temporal constraints to allow specialized solvers to handle each dimension independently.
The Challenge of Temporal Blindness
Standard Retrieval-Augmented Generation (RAG) and Vision-Language Models (VLMs) frequently suffer from "temporal hallucinations." As Li Sun 0007 and colleagues observed, Multimodal LLMs often fail to pinpoint specific timestamps in videos because they cannot reason through the sequential logic of a description. Similarly, Siyue Zhang notes that standard RAG systems struggle with "time-sensitive" questions where the correct answer depends entirely on a specific date or order of events.
The root of the problem is computational and logical. When temporal math and propositional logic are fused into a single representation, the search space grows exponentially. By decoupling them, researchers can identify a "tractable core" that allows for efficient, polynomial-time reasoning, a concept pioneered by Eddie Schwalb, Kalev Kask, and Rina Dechter in their work on Conditional Temporal Networks (CTNs).
Disentangling What from When
The core mechanism of this research direction is the separation of the semantic core (the subject matter) from the temporal parameters (the timeframe or sequence).
In the MRAG framework developed by Siyue Zhang, a query is first stripped of its temporal markers. The system retrieves information based on the semantic intent and then uses symbolic logic to filter those results against the original time constraints. This "disentanglement" ensures that the retrieval engine isn't distracted by dates while searching for facts, and the reasoning engine isn't confused by facts while calculating dates.
This strategy is mirrored in Knowledge Base Question Answering (KB-QA). The TEQUILA framework by Abdalghani Abujabal and Gerhard Weikum acts as a "temporal enabler" for existing systems. It decomposes a query into a non-temporal sub-question to find candidate entities and a temporal sub-question to extract relevant time constraints, allowing standard engines to handle time-sensitive queries without internal architectural changes.
Decomposition as a Planning Strategy
In more complex scenarios, such as long-form video retrieval or robotics, decomposition takes the form of sequential planning. Instead of matching a whole paragraph to a video, models like CadenceRAG by Heng Liu break narrative queries into a sequence of "beats" or sub-queries.
This approach works because it transforms a high-level reasoning requirement into a structured set of atomic tasks:
- Identify the anchor: Find the primary event or object.
- Locate the sub-events: Use an LLM to "translate" the query into chronological steps, such as "start state" and "end state," as seen in the Point to Span (P2S) framework by Mingyu Jeon.
- Verify the sequence: Use specialized vision-language models to confirm each step in order.
By treating a complex query as a sequence of simpler semantic matches, systems can bridge the gap between descriptive natural language and raw data. Khanh-An C. Quan demonstrated that this sequential decomposition can double the accuracy of video retrieval models by preventing them from being distracted by later parts of a description while searching for the initial segments.
The Tractable Core: Why Modularity Wins
The intuition behind this research is that modularity reduces complexity. Historically, Michael Fisher 0001 introduced the Separated Normal Form (SNF) to isolate temporal operators from classical logic, allowing standard resolution techniques to handle the majority of reasoning tasks.
Modern researchers apply this same logic to Large Language Models. Alfredo Garrachón Ruiz and Daniel Borrajo found that LLMs perform significantly better when they act as "orchestrators" that call deterministic external APIs for tasks like chronological sorting or duration calculation, rather than trying to solve the math internally.
By delegating temporal logic to specialized modules—whether they are symbolic solvers, fuzzy constraint propagators as explored by Lluís Godo, or structured task planners like the Plan of Knowledge (PoK) by Xinying Qian—researchers are moving away from "black box" temporal reasoning toward a more robust, interpretable, and efficient architecture.