Temporal Reasoning

The foundational ability to sequence events along a timeline. Models must recognize explicit temporal markers (before, after, during) and implicit ordering cues to arrange facts correctly.

• Resolves relative positioning of events without absolute dates
• Handles anachronistic references where narrative order differs from chronological order
• Critical for answering 'What happened first?' or 'What was the sequence of events?' queries

Example: Correctly ordering 'The company announced layoffs after the stock dropped' requires understanding that the stock drop preceded the announcement.

The capacity to perform mathematical operations over time intervals. Models must compute differences between dates, add or subtract durations, and handle granularity mismatches (days vs. months vs. years).

• Calculates elapsed time between two timestamps
• Aggregates durations across multiple events
• Handles calendar irregularities like leap years and varying month lengths

Example: Computing 'How many days between March 15, 2023 and July 8, 2024?' requires precise date arithmetic spanning a year boundary.

The ability to identify and categorize the specific logical relationship between two temporal entities. Goes beyond simple ordering to recognize nuanced Allen's interval algebra relations.

• Distinguishes overlap from containment (event A happens entirely within event B)
• Identifies meets relations where one event ends exactly as another begins
• Resolves simultaneity and sequentiality in parallel timelines

Example: Understanding that 'The CEO served during the recession' implies the CEO's tenure contains or overlaps the recession period, not merely precedes or follows it.

The process of converting natural language time expressions into standardized, machine-readable formats. Resolves relative references ('last quarter', 'next Tuesday') against a reference date.

• Maps vague expressions ('recently', 'earlier', 'a few years ago') to approximate ranges
• Resolves deictic temporal anchors relative to document publication date or current time
• Standardizes diverse formats ('Q2 2024', 'April-June 2024', 'second quarter of 2024')

Example: Normalizing 'the previous fiscal year' in a document published March 2024 to the range 2023-01-01 to 2023-12-31 for a calendar-aligned fiscal year.

The reasoning process of satisfying multiple temporal constraints simultaneously to infer a consistent timeline. Models must detect and resolve temporal contradictions across evidence sources.

• Propagates constraints through transitive chains (A before B, B before C → A before C)
• Detects inconsistencies when constraints conflict
• Infers implicit temporal relationships not explicitly stated

Example: If Document A states 'The merger closed in June' and Document B states 'The CEO resigned before the merger,' the system infers the resignation occurred no later than May.

The ability to recognize when multiple textual mentions refer to the same real-world event, enabling accurate temporal aggregation. Prevents double-counting and timeline fragmentation.

• Links 'the 2020 acquisition' with 'the deal announced in January 2020'
• Merges partial temporal information from co-referring mentions
• Resolves event identity across documents with conflicting or incomplete timestamps

Example: Recognizing that 'the product launch' in one paragraph and 'the Q3 release' in another refer to the same event, consolidating temporal metadata into a single timeline entry.

The foundational preprocessing step that converts relative and absolute natural language time expressions into a standardized, machine-readable format. This involves resolving phrases like 'next quarter' or 'the following Friday' into precise ISO 8601 date ranges using a reference timestamp. Without accurate normalization, a model cannot anchor a query to a specific temporal coordinate, making subsequent arithmetic or comparison operations impossible. Modern systems use rule-based SUTime libraries combined with neural sequence tagging to handle the vast diversity of human temporal expression.

The task of identifying the chronological sequence of events described in text, often using Allen's Interval Algebra to classify relationships. A model must determine if one event precedes, overlaps, contains, or is simultaneous with another. This goes beyond simple date extraction to build a causal or narrative timeline from unstructured documents, enabling reasoning over event sequences even when explicit dates are absent.

An extension of standard knowledge graph traversal that incorporates a temporal dimension to facts. Instead of querying a static triple like (Elon Musk, founded, SpaceX), a time-aware system queries a temporal knowledge graph with a quadruple: (Elon Musk, founded, SpaceX, [2002]). This allows the system to correctly answer questions about past states, such as 'Who was the CEO of Twitter in 2021?', by traversing relationships valid only within a specific time interval, preventing the retrieval of stale or anachronistic facts.

The computational ability to perform mathematical operations on time spans, such as calculating the difference between two dates or adding a duration to a timestamp. This is critical for answering questions like 'How many days passed between the signing and the ratification?' or 'What is the date 90 business days from next Tuesday?' . Robust systems must handle calendar irregularities, including leap years, time zones, and business day conventions, often by invoking a dedicated calculator tool via Tool-Augmented Reasoning.

A reasoning process used to resolve a set of implicit temporal constraints to find a valid timeline. When a query states 'The meeting is after lunch but before the 3 PM call,' the system must construct a constraint graph where nodes are events and edges are temporal relations. It then solves this graph to infer that the meeting must occur in the interval between the end of lunch and 3 PM, a critical capability for scheduling and logistical reasoning.

The synthesis of temporal reasoning with multi-hop retrieval to answer questions requiring the connection of time-stamped facts across multiple documents. For example, answering 'Which company acquired the startup that launched its first product in 2019?' requires a first hop to find a startup with a 2019 product launch, and a second hop to find the acquiring company of that specific entity. This interleaves Bridge Entity resolution with temporal filtering.