acceptodds
Under review as a conference paper at ICLR 2027

IntervalRAG: Hierarchical Interval Graphs for Time-Grounded Retrieval-Augmented Generation

Abstract

IntervalRAG is a retrieval-augmented generation framework for questions whose answers depend on when a fact remains valid. Such questions often concern a persistent state, while the supporting evidence is distributed across appointments, departures, and other transitions in a narrative. Retrieving an isolated observation can therefore leave the generator to reconstruct which state applies to the queried period and how its boundaries are established. IntervalRAG organizes this evidence in a two-level temporal graph: macro units identify broad states or phases, and micro units preserve individual facts and transitions together with the temporal information needed to place them in context. Hierarchical edges retain the association between each micro-event and its source context, while local temporal edges expose neighboring evidence for assembling a short chronology. The query first locates a relevant macro context and then guides the selection of nearby micro-events and source passages around the requested period. The resulting context presents the relevant state together with the transitions that explain its temporal boundaries. We evaluate IntervalRAG on TimeQA, TempReason, and ComplexTR, where it obtains accuracies of 60.93%, 91.79%, and 68.39%, respectively. Code and data will be released upon publication.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Retrieval-Augmented Generation, Question Answering, Temporal Graphs, Information Retrieval, Time-Aware Retrieval, Interval Algebra, Hierarchical Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ff88OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents IntervalRAG, a temporal retrieval-augmented generation framework that represents persistent states or phases as Macro-Interval Units (MIUs) and events or transitions as Micro-Event Units (MEUs). It builds hierarchical links between the two unit types and local temporal edges among events. At inference time, the system retrieves and reranks macro-level candidates, expands them through bounded temporal graph traversal, and supplies the resulting event evidence together with selected source passages to LLMs. The authors also conduct some experiments and ablation studies to show the effectiveness of the porposed algorithm.

Critical strengths
  1. The paper targets an important temporal QA limitation: point-event retrieval can miss persistent states and the transitions that delimit them.
  2. The MIU–MEU representation, interval-aware retrieval, and local graph traversal form a coherent end-to-end pipeline, supported by useful ablations and stage-wise retrieval diagnostics.
  3. The experiments report high answer-inclusion accuracy on TempReason and ComplexTR and include both an extraction-quality audit and an efficiency analysis.
Critical weaknesses
  1. The paper seems only have incremental novelty. MIUs provide a useful state-oriented formulation, but interval-valued event/state representations and hierarchical temporal retrieval are already present in related work (IA-RAG models interval events and hierarchical thematic units with interval-algebra-guided retrieval, while DyG-RAG already concatenates semantic and Fourier time features in its retrieval embeddings.)
  2. It is unclear whether the comparison with the baseline methods are reasonable. More specifically, it may be hard to tell whether the better performance of the propose method is becuase of the framework or merely just because more context is return to LLM generation stage. Also, although table 3 shows the comparison to graph-free baselines, the IntervalRAG seems also brings some addition event information so making it unfair in terms of the retrieval constraints.
  3. It may need to futher explanation for the experiment results. The Acc_full metric does not penalize extra incorrect or contradictory claims. The supplementary short-answer results also show metric sensitivity: on TimeQA, IntervalRAG’s EM is 27.67%, compared with 28.17% for DyG-RAG, despite the higher answer-inclusion score. The paper should foreground a stricter answer-level metric or human correctness audit and explain how Acc_full relates to established inclusion-based accuracy. A more reasonable evaluation should include validation of the entity, time and relations.
  4. The offline indexing cost is unclear. It seems the paper describes separate LLM extraction passes for macro intervals and micro events, yet does not report actual per-chunk or per-query request counts, input/output tokens, monetary cost, or the extra query cost of anchor-event recovery.
  5. It is unclear what will be the cost of updating the information in the proposed framework.
Questions for authors
  1. How sensitive are the results to key search hyperparameters—particularly the 2,000-vector candidate pool, 20 macro seeds, traversal limits, and two-chunk passage budget—and were these values selected using validation data?
  2. How does macro–micro linking handle events that span multiple chunks, conflicting intervals, or gaps in source evidence, and has this behavior been audited on manually labeled examples?
Reviewer_XYGiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

IntervalRAG studies questions whose answers depend on when a fact holds, such as a person's employer in a particular year. It separately extracts macro-interval units and micro-event units from source passages. Macro units describe a state or phase, while micro units record individual facts and changes. The graph links micro units to macro contexts from the same passage and connects nearby events in time. A query first retrieves macro contexts using semantic, temporal, and entity information. The system then collects related micro-events and source passages, orders the events by time, and passes the evidence to a language model. Using Qwen2.5-14B-Instruct, the authors report scores of 60.93%, 91.79%, and 68.39% on the RAG versions of TimeQA, TempReason, and ComplexTR. These scores measure whether the complete response contains all required answers. The paper also reports component ablations, results with different model sizes, short-answer F1/EM, retrieval coverage, checks of extracted events and dates, and running times.

Critical strengths
  1. The method addresses a clear difficulty in temporal question answering. A nearby event may not be enough to determine a person's affiliation or role at a requested date. Finding the relevant state together with events that mark its beginning or end is a reasonable approach. Including the original passages also allows the extracted facts and dates to be checked. The prompt tells the model to trust the original text when it conflicts with an extracted event.

  2. The short-answer results and ablations support parts of the method. Table 9 applies the same rules to remove explanatory text from the answers, without consulting the reference answers. Relative to DyG-RAG, IntervalRAG improves short-answer token F1 by 4.99, 3.16, and 6.27 percentage points on the three datasets. This provides evidence of improvement beyond the main score. Table 2 keeps the event graph and replaces local traversal with top-K event retrieval. This supports the use of traversal in the current method. The authors also acknowledge that retrieving neighboring events does not by itself prove that a state continued between them, and that finding the answer names in the text does not establish the correct subject, relation, or time.

Critical weaknesses
  1. The main metric can count a wrong final answer as correct. Equation (8) in Appendix H checks whether all required answer strings appear anywhere in the complete response. Prompt C in Appendix N asks the model to list candidates for the same subject and relation and state whether each should be included in the final answer. If all required names appear in this list, the response can receive credit even when the model rejects a correct candidate and gives a wrong final answer. Extra incorrect candidates or answers do not lower this score. The model can follow the prompt and still receive credit for a wrong final answer. The prompt also asks the model to include AMBIGUOUS-IN candidates when the context does not rule them out. This makes it especially important to check whether the final answer contains the right entities. Mentioning more correct names does not show that the model has chosen the facts that hold at the queried time. The authors acknowledge what the main metric measures and provide short-answer F1/EM in Table 9. These results are useful, but IA-RAG is missing from the comparison, and TimeQA EM is 27.67%, slightly below DyG-RAG's 28.17%. Token overlap and exact string matching also differ from checking whether the model selected the right set of entities, including alternative names for the same entity. The paper needs to show how much of the reported advantage remains when the final answers are scored this way.

  2. The comparisons do not show what the two-level structure adds. IA-RAG already uses interval events, a hierarchy, and Allen relations to guide retrieval. IntervalRAG changes how the hierarchy is built: it extracts macro and micro units separately from the source text and links them using their common source passage. IA-RAG instead builds its thematic forest by recursively grouping and summarizing events. This is a specific difference, but the current results do not convincingly show why the proposed design is better. The w/o Traversal comparison in Table 2 supports local traversal, but does not test the links between the macro and micro levels separately. A comparison with flat interval-aware retrieval followed by time ordering would help. Both methods should have access to the same extracted records, use the same generation settings, and have the same limit on input length. The NoGraph comparisons in Table 3 also change the information given to the model. Even when both methods receive two source chunks, IntervalRAG receives additional event descriptions. The authors state this difference. These results support the complete system, but do not show how much the graph itself helps. Given how much the method shares with IA-RAG, I find the evidence for the added contribution insufficient.

Questions for authors
  1. How well do the final answers perform? Please use the same requested answer format for IntervalRAG, DyG-RAG, and IA-RAG, extract their final answers in the same way, and account for alternative names for the same entity. Report entity-set precision, F1, and exact set match, and explain the filtering rules used in Table 9. Among responses counted as correct by Equation (8), separately report the fraction whose final answers omit required entities and the fraction containing extra incorrect entities. Please state the denominators and how often both errors occur in the same answer.

  2. Does the two-level structure still help when the available records and input length are kept the same? Please compare the full method with flat interval-aware retrieval followed by time ordering, and separately test the links between macro and micro units. Use the same pool of extracted macro and micro records, reranker, and generation prompt, with the same limit on the total number of input tokens. Each retrieval method should still select its own evidence. This would help show whether the hierarchy improves the results beyond the information already in the extracted records.

  3. Why do the DyG-RAG results differ between Tables 1 and 4? Were the baselines evaluated in the same way? Table 1 reports 58.78/84.75/55.62 for DyG-RAG, while the 14B row of Table 4 reports 58.94/82.38/59.57. Which scores were taken from prior work, and which were measured in this paper? Please explain whether the methods used the same questions, generation prompts, decoding settings, and rules for counting failures. The paper states that questions with empty context and other failures are kept in the total; please confirm that this also applies to the baselines. Please also explain how IA-RAG was run on these datasets.

  4. What are the exact rules for selecting events? Please define temporal similarity in Equation (4), including the time units. How is the first event selected for an open-ended interval? How does each temporal relation determine the search window? Please also explain when the search stops, how the limits on depth, visited nodes, and branches work together, and what evidence is returned when the search falls back to the macro context. Pseudocode or anonymous code would help make these steps clear.

Reviewer_sytdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets temporal question answering and proposes a framework for retrieval-augmented generation (RAG). For a given question, the framework first retrieves relevant passages, and then establishes a hierarchical graph with retrieved sentences as nodes and temporal relations as edges. A local graph traversal then explores temporal neighborhoods in the graph, selecting a set of events that are provided to the LLM as context for answering.

Experiments on three benchmarks for temporal QA (TimeQA, TempReason, ComplexTR) demonstrate performance improvements of the proposed approach compared to the State of the Art.

Critical strengths
  1. The empirical evaluation is extensive, and the approach achieves state-of-the-art performance across three distinct temporal QA benchmarks (TimeQA, TempReason, ComplexTR). Those results provide strong evidence that the proposed temporal structuring effectively addresses the limitations of standard RAG in temporal reasoning.

  2. The introduction of a hierarchical graph-based approach for temporal evidence organization and retrieval offers a compelling architectural pattern for handling temporal dependencies.

Critical weaknesses
  1. The research novelty is unclear. Hierarchical evidence representation and multi-level retrieval have already been explored in prior work, including RAPTOR (Sarthi et al., 2024). Recent temporal RAG approaches, such as TG-RAG (Han et al., 2025) and IA-RAG (Wang et al., 2026), also explicitly organize temporal information at multiple granularities. Those are the two main contributions listed by the authors, but their novelty relative to prior work remains unclear. Without a clearer distinction between IntervalRAG and these recent works, the paper's primary contribution remains difficult to evaluate.

  2. The description of the extraction of Micro and Macro units lacks sufficient technical detail. Based on the examples provided in the Appendix (Tables 15-17), the Macro units often appear sufficient for answering the question, with Micro units providing additional detail or even redundancy. The utility of both Micro and Macro units should be verified in dedicated experiments.

Questions for authors
  • Experiments in Table 3 provide an incomplete picture. How would IntervalRAG perform with 1 and 5 chunks? Is the hard interval filter applied before or after semantic retrieval? And how would IntervalRAG compare against semantic retrieval followed by reranking, with the same cross-encoder used within IntervalRAG?

  • The last experiment on query-evidence temporal distance should be clarified. How is the query-evidence temporal distance estimated, and how are these distances separated into near, middle and far (fixed thresholds or percentiles)?

  • The main paper lacks a proper description and example of how macro and micro units are derived. Without looking in the Appendix, it is unclear how each of these would look like. Since this is the main contribution of this work, a proper description in the main paper would be required. How are the time intervals extracted (e.g., "[2009-07-01, 2010-06-30] Sidwell played for Aston Villa during the 2009/10 season." in Table 17), and what is the contribution of this for the entire framework? Retrieval of macro intervals appears to be sufficient for answering most of the questions.

  • Please clarify the need for Micro units. Based on Figure 1 and other examples in the paper, it is unclear why finer-grained units are required. The question is already answerable from the Macro unit. Would Micro units recover failures in Macro unit extraction?

  • How is the graph traversal / local search implemented? There is no description in the main paper, and even the description in the Appendix is vague ("query relation defines the search window" => how?, "neighboring intervals and the span between adjacent events provide the relevance test" => which relevance test?).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.