Temporal Reasoning in Language Models

Type 2 diabetes case reports describe complex clinical courses, but their timelines are often expressed in language that is difficult to reuse in longitudinal modeling. To address this gap, we developed a textual time-series corpus of 136 PubMed Open Access single-patient case reports involving glucagon-like peptide 1 receptor agonists, with clinical events associated with their most probable reference times. We evaluated automated LLM timeline extraction against gold-standard timelines annotated by clinical domain experts, assessing how well systems recovered clinical events and their timings. The best-performing LLM produced high event coverage (GPT5; 0.871) and reliable temporal sequencing across symptoms (GPT5; 0.843), diagnoses, treatments, laboratory tests, and outcomes. As a downstream demonstration, time-to-event analyses in diabetes suggested lower risk of respiratory sequelae among GLP-1 users versus non-users (HR=0.259, p<0.05), consistent with prior reports of improved respiratory outcomes. Temporal annotations and code will be released upon acceptance.

Large language models cannot estimate how long their own tasks take. We investigate this limitation through four experiments across 68 tasks and four model families. Pre-task estimates overshoot actual duration by 4--7$\times$ ($p < 0.001$), with models predicting human-scale minutes for tasks completing in seconds. Relative ordering fares no better: on task pairs designed to expose heuristic reliance, models score at or below chance (GPT-5: 18% on counter-intuitive pairs, $p = 0.033$), systematically failing when complexity labels mislead. Post-hoc recall is disconnected from reality -- estimates diverge from actuals by an order of magnitude in either direction. These failures persist in multi-step agentic settings, with errors of 5--10$\times$. The models possess propositional knowledge about duration from training but lack experiential grounding in their own inference time, with practical implications for agent scheduling, planning and time-critical scenarios.

Real-world dialogue usually unfolds as an infinite stream. It thus requires bounded-state memory mechanisms to operate within an infinite horizon. However, existing read-then-think memory is fundamentally misaligned with this setting, as it cannot support ad-hoc memory recall while streams unfold. To explore this challenge, we introduce STEM-Bench, the first benchmark for STreaming Evaluation of Memory. It comprises over 14K QA pairs in dialogue streams that assess perception fidelity, temporal reasoning, and global awareness under infinite-horizon constraints. The preliminary analysis on STEM-Bench indicates a critical textit{fidelity-efficiency dilemma}: retrieval-based methods use fragment context, while full-context models incur unbounded latency. To resolve this, we propose ProStream, a proactive memory framework for streaming dialogues built on a hierarchical structure. It enables ad-hoc memory recall on demand by reasoning over continuous streams with multi-granular distillation. Moreover, it employs Adaptive Spatiotemporal Optimization to dynamically optimize retention based on expected utility. It enables a bounded knowledge state for lower inference latency without sacrificing reasoning fidelity. Experiments show ProStream delivers higher reasoning fidelity than prior baselines while maintaining substantially lower latency than full-context alternatives.

Large Language Models (LLMs) have demonstrated remarkable success in general-purpose reasoning. However, they still struggle to understand and reason about time series data, which limits their effectiveness in decision-making scenarios that depend on temporal dynamics. In this paper, we propose Thoth, the first family of mid-trained LLMs with general-purpose time series understanding capabilities. As a pivotal intermediate stage, mid-training achieves task- and domain-agnostic alignment between time series and natural language, for which we construct Book-of-Thoth, a high-quality, time-series-centric mid-training corpus. Book-of-Thoth enables both time-series-to-text and text-to-time-series generation, equipping LLMs with a foundational grasp of temporal patterns. To better evaluate advanced reasoning capabilities, we further present KnoTS, a novel benchmark of knowledge-intensive time series understanding, designed for joint reasoning over temporal patterns and domain knowledge. Extensive experiments demonstrate that mid-training with Book-of-Thoth enables Thoth to significantly outperform its base model and advanced LLMs across a range of time series question answering benchmarks. Moreover, Thoth exhibits superior capabilities when fine-tuned under data scarcity, underscoring the effectiveness of mid-training for time series understanding. Code is available at: https://github.com/thuml/Thoth.

Time series reasoning demands both the perception of complex dynamics and logical depth. However, existing LLM-based approaches exhibit two limitations: they often treat time series merely as text or images, failing to capture the patterns like trends and seasonalities needed to answer specific questions; and when trained on a mix of simple and complex tasks, simpler objectives often dominate the learning process, hindering the development of deep reasoning capabilities. To address these limitations, we propose the Pattern-Aware Alignment and Balanced Reasoning model (PATRA), introducing a pattern-aware mechanism that extracts trend and seasonality patterns from time series to achieve deep alignment. Furthermore, we design a task-aware balanced reward to harmonize learning across tasks of varying difficulty, incentivizing the generation of coherent Chains of Thought. Extensive experiments show that PATRA outperforms strong baselines across diverse Time Series Question Answering (TSQA) tasks, demonstrating superior cross-modal understanding and reasoning capability.

Evaluating language models in streaming environments is critical, yet underexplored. Existing benchmarks either focus on single complex events or provide curated inputs for each query, and do not evaluate models under the conflicts that arise when multiple concurrent events are mixed within the same document stream. We introduce StreamBench, a benchmark built from major news stories in 2016 and 2025, comprising 605 events and 15,354 documents across three tasks: Topic Clustering, Temporal Question Answering, and Summarization. To diagnose how models fail, we compare performance with and without structural cues, which organize key facts by event. We find that structural cues improve performance on clustering (up to +4.37%) and temporal QA (up to +9.63%), helping models locate relevant information and separate distinct events. While temporal reasoning remains an open challenge inherent to current LLMs, consistent gains across tasks show that structural cues are a promising direction for future work in massive document streams.

Multimodal Large Language Models (MLLMs), particularly smaller, deployable variants, exhibit a critical deficiency in understanding temporal and procedural visual data, a bottleneck hindering their application in real-world embodied AI. This gap is largely caused by a systemic failure in training paradigms, which lack large-scale, procedurally coherent data. To address this problem, we introduce TPRU, a large-scale dataset sourced from diverse embodied scenarios such as robotic manipulation and GUI navigation. TPRU is systematically designed to cultivate temporal reasoning through three complementary tasks: Temporal Reordering, Next-Frame Prediction, and Previous-Frame Review. A key feature is the inclusion of challenging negative samples, compelling models to transition from passive observation to active, cross-modal validation. We leverage TPRU with a reinforcement learning (RL) fine-tuning methodology, specifically targeting the enhancement of resource-efficient models. Experiments show our approach yields dramatic gains: on our manually curated TPRU-Test, the accuracy of TPRU-7B soars from 50.33% to 75.70%, a state-of-the-art result that significantly outperforms vastly larger baselines, including GPT-4o. Crucially, these capabilities generalize effectively, demonstrating substantial improvements on established benchmarks. The codebase is available at https://github.com/Stephen-gzk/TPRU/ .

Time Series Language Models (TSLMs) promise reasoning over real-world temporal data, but their ability to retrieve and reason over long time-series remains largely untested. We introduce TS-Haystack, a multi-domain retrieval benchmark with ten event-grounded question-answering tasks over contexts from 100 seconds to 24 hours, spanning direct retrieval, temporal reasoning, multi-step reasoning, and contextual anomaly detection. Existing TSLMs exhibit severe long-context degradation: accuracy declines with context length, direct-tokenization models run out of memory beyond 100 seconds on high-rate signals, and time-interval-grounded tasks collapse toward near-zero accuracy when increasing the time-series lengths, aligning with existing literature on text and multi-modal long context retrieval. An agentic retrieval framework using specialized time-series classifier tools matches or outperforms SoTA TSLMs on 9 of 10 tasks, highlighting agentic retrieval as a promising approach for long-context TSLMs.

Tool use, such as web search, has become a standard capability even in freely available large language models (LLMs). However, existing benchmarks evaluate temporal reasoning mainly in static, non-tool-using settings, which poorly reflect how LLMs perform temporal reasoning in practice. We introduce Time Puzzles, a constraint-based date inference task for evaluating iterative temporal reasoning with tools. Each puzzle combines factual temporal anchors with (cross-cultural) calendar relations and may admit one or multiple valid dates. The puzzles are algorithmically generated, enabling controlled and continual evaluation. Across 13 LLMs, even the best model (GPT-5) achieves only 55.3% accuracy without tools, despite using easily searchable facts. While web search improves performance, models perform substantially better when constraints are rewritten with explicit dates, removing the need for factual lookup. These results reveal a gap in reliable tool use for iterative temporal reasoning.