Test of Time: A Benchmark for Evaluating LLMs on Temporal Reasoning

Thanks for your review

Thank you for the thoughtful questions. Regarding question generation, for ToT Semantic, the questions are generated using multiple templates, one for each question type. We have added Appendix B to a revision (to be uploaded within 24 hours) which includes algoboxes detailing this process. For ToT Arithmetic, the questions were mainly written by the people who helped create questions, and in some cases modified by authors to make them into full-form questions. Regarding the date/time format for ToT Arithemetic, we included many variations of both date and time in the dataset. We consider this a robustness strategy; however, we did not drill down into the empirical performance differences between date/time format. Regarding baselines, we did not consider few-shot, self-consistency, or code generation. We consider these important avenues for future modeling innovation that can make use of our ToT benchmark framework. Regarding Section 4.1 and 4.1.1, all these analyses are about ToT-Semantic.

Response

Thank you for the detailed review! Please find our replies below: ### **Data Contamination** > Could you add some quantitative evidence showing the GPT-4 or Gemini-1.5 Pro baselines have pre-training data contaminations? This is difficult to ascertain for any proprietary LLM, however here we are referring to Temporal Reasoning tasks like TempTabQA[1] or MenatQA [2] which are based on real events (with well known entities) that are documented on resources like Wikipedia, social media, and/or the Internet at large. These resources are well known to be used for training LLMs [3]. Using entities and relationships a LLM has seen in pre-training detracts from the ability to measure its reasoning capability. [1] https://arxiv.org/abs/2311.08002 [2] https://arxiv.org/pdf/2310.05157 [3] https://platform.openai.com/docs/bots > “LLMs could even potentially guess the original entities due to their adjacent relations” We are happy to remove this statement, but believe it to be true based on current capabilities of LLMs: 1. TGQA uses a table to map entities from one name to another of the same type. This mapping is less probable than the real relationship (which the LLM has seen in its training data). 2. We also know that analogical reasoning is well studied in LLMs (e.g. [4]). 3. “De-masking” a TGQA example then, is asking a LLM to find a more likely assignment of entities to relationships than the one offered. For more information on the related task of text deanonymization, we refer to [5]. [4] https://arxiv.org/pdf/2404.12728v1 [5] https://arxiv.org/pdf/2303.12429 ### **Generation Question** Thanks for the question, in brief: - The generation uses templates for each question type. - First a graph is sampled - Then facts are sampled from the graph until the proper question template can be created. We will add this to the manuscript explicitly as an algorithm block in the Appendix. ### **Related Work** Thank you for bringing these three papers to our attention. Each of them addresses aspects of LLM temporal reasoning distinct from the purpose of ToT: * The questions in TRAM (Wang and Zhao, 2023) are all multiple-choice, and do not require LLMs to reason through a large list of temporal facts from a knowledge graph. * The questions in TimeBench (Chu et al. 2023) are collected from ten existing real-world datasets, one of which requires reasoning through temporal facts provided in the context. ToT goes beyond such datasets by providing controllable, comprehensive temporal relationship collectinos via synthetic graph generation. * The questions in Timo (Su et al. 2024) are grouped into two categories: math-time and pure-time. The math-time category contains questions that are similar in nature to our ToT-Arithmetic dataset; however, we cover several more categories that do not seem to be included in the math-time subset of Timo. This includes questions about calendar schedules, timezones, trick questions that tend to mislead models, questions about conversions between BC and AD times, and questions that require applying multiple different operations to find the final answer. The pure-time category is in the same vein as our ToT-Semantic subset; however, they mostly focus on commonsense temporal reasoning with a focus on real-world facts and temporal knowledge, whereas our focus is mainly on multi-hop temporal reasoning on generic facts which require can be only answered through reasoning and not based on the parametric knowledge of the models. In that regard, we believe the two datasets are complementary. We will add these details to our related work section.

Response

Thank you for the comments and questions! ### **Long distance temporal dependencies** Regarding your first comment, yes, we have explicitly disclaimed this limitation in Appendix F, noting it as an area for future work. We regret that due to space limitations, we could not explore this area further. ### **Template Generation** Regarding your second comment, we agree that our reliance on templates makes the examples in our benchmark potentially less-realistic than real-world examples in terms of natural language. However, this is a trade-off we make to enable us to generate highly-complex temporal and relational structures. The purpose of ToT is to isolate LLM capacity for reasoning over structured (temporal) knowledge as opposed to natural language comprehension. ### **Static and Temporal Facts** We do not believe that the results would change much. Our randomly generated examples for ToT-Semantic already contain facts that vary in their duration (ie. it has both “short” and “long” facts). A static fact would then be one with maximum length -- maintaining its truth throughout the example. Nonetheless, we believe that the combination of static and temporal facts is an interesting area for future work.

Response

Thanks for your time and thoughtful review! Let us briefly address the questions you raised: ### **Graph Generation** Great question! We agree the description of our algorithm for ensuring temporal and relational consistency was limited (explained partially at L195): > "Then, for each edge (u, v) labeled with a relation r, we assign a valid time interval [t1, t2] that respects the relation types." Our approach for ensuring overall graph correctness is similar. At each generation step -- graph generation, relation generation, time interval generation -- we check to make sure the generated nodes, edges, relations, and time intervals are consistent with the existing objects in the temporal knowledge graph. If not, we re-generate until they are, or until a pre-specified step count has been reached. If the latter, we throw out the example and start again. We will include these details in our revision. Regarding your second point about the "distinction between generating static graphs and those with temporal dynamics", we note that every temporal graph has a static *backbone* -- the structure remaining if all timestamps are removed. As discussed in S3.1, we use static graph generators to generate the backbones, and then we impute timestamps consistent with the relation types and logical rules. ### **Are Graphs Necessary?** To your second question, we note that every collection of temporal facts of the form (object, relation, subject, time_interval) can be expressed as a temporal graph. ToT-SEMANTIC specifically targets LLM reasoning ability over such collections. We do not claim that graph generators are the only way to construct such a benchmark. However, because all temporal fact collections contain an underlying graph, we are proposing a graph-generator-based framework to produce benchmark examples. We argue that a framework that exposes generation of the static graph backbone is more controllable and allows for a final dataset that is more comprehensive w.r.t. the variety and complexity of temporal relationships between generated entities (L49). We will extend the discussion to further motivate graph generation as a useful tool for generating different temporal dependency structures.