KairosAgent: Agentic Time Series Forecasting with Fused Semantic Reasoning¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://foundation-model-research.github.io/KairosAgent
Area: LLM Reasoning
Keywords: Multimodal Time Series, Time Series Foundation Models, Agentic Reasoning, Morphology Fusion, Process-Level Credit Assignment
TL;DR¶
KAIROSAGENT introduces a reason-then-forecast agentic framework that dynamically invokes 23 statistical tools to deduce qualitative future morphology descriptions, fuses these semantic priors into a lightweight time series foundation model via gated cross-attention, and trains the reasoner using turn-level reinforcement learning on the 40k-trajectory T-STAR corpus for superior zero-shot forecasting.
Background & Motivation¶
Cross-domain multimodal time series forecasting represents a fundamental frontier in modern temporal intelligence, requiring models to simultaneously master fine-grained continuous numerical dynamics and comprehend complex semantic background knowledge across heterogeneous domains. Existing literature has sharply polarized into two distinct paradigms. Time Series Reasoning Models (TSRMs) leverage large language models to process time series serialized into text, exploiting strong linguistic comprehension; however, standard tokenizers fragment continuous numerical tokens, leading to severe numerical hallucinations and degraded quantitative predictions. Conversely, Time Series Foundation Models (TSFMs) and their multimodal variants preserve the native numerical modality to ensure robust statistical modeling, but operating primarily as black-box pattern fitters, they lack transparent interpretability and the ability to extrapolate future trends through qualitative reasoning.
Recent explorations have attempted to bridge this gap by deploying LLMs within agentic workflows, yet they continue to struggle with three structural and optimization bottlenecks. First, inappropriate serialization persists, as raw continuous sequences are naively fed into LLMs as plain text strings, preventing the extraction of reliable statistical features. Second, a severe modality disconnect isolates the reasoning output from the downstream forecaster: existing frameworks use the LLM either to predict raw numbers directly or as an external controller while the numerical forecaster remains confined to unimodal time series inputs, missing genuine representation-level fusion. Third, current training paradigms either rely on supervised fine-tuning (SFT) that merely memorizes annotated trajectories without generalization, or resort to outcome-only reinforcement learning with sparse terminal rewards, making it impossible to assign credit to intermediate tool invocations across multi-turn trajectories.
To resolve these tensions, this paper decouples qualitative semantic deduction from quantitative continuous generation into a harmonious, closed-loop agentic framework. Core idea: propose KAIROSAGENT, a modular reason-then-forecast framework that empowers an LLM reasoner to deduce qualitative future morphology descriptions using 23 statistical tools, deeply injects these semantic priors into a lightweight TSFM decoder via gated cross-modal attention, and optimizes multi-turn exploration using turn-level advantage credit assignment on the curated 40k T-STAR trajectory corpus.
Method¶
Overall Architecture¶
The inference pipeline of KAIROSAGENT decomposes end-to-end multimodal forecasting into two coordinated stages: 1. Tool-Augmented Morphology Reasoning (Stage I): Given historical observations \(X \in \mathbb{R}^L\) and optional textual context \(c\), the LLM reasoner \(\pi_{\text{reason}}\) interacts with an analytical toolset \(\mathcal{T}\) over multiple turns. It inspects both global trajectories and local regimes to produce a compact, qualitative morphology description \(r\) (formulated as two distinct paragraphs addressing short-term and long-term horizons, strictly avoiding exact numbers or dataset identifiers to prevent hallucination). 2. Morphology-Conditioned Forecasting (Stage II): The numerical forecaster \(f_{\text{forecast}}\) (built upon the pretrained Kairos backbone) encodes \(r\) using a GIST text encoder and projection MLP into semantic priors \(E_r\), which are fused layer by layer into the Kairos decoder via gated cross-attention to generate final predictions \(\hat{Y} \in \mathbb{R}^H\).
The training framework follows a three-stage progressive pipeline: (I) Reasoner SFT warms up tool-calling and trajectory synthesis on T-STAR; (II) Multimodal alignment fine-tunes the forecaster and text adapter under pinball quantile loss; and (III) RL with turn-level credit assignment refines the reasoner using the frozen forecaster as an objective reward engine under GRPO.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Historical Series X & Text Context c"] --> B["Tool-Augmented Statistical Feature Extraction<br/>23 multi-scale time series analysis tools"]
B --> C["Qualitative Morphology Prior Synthesis<br/>Short-term and long-term future description r"]
C --> D["Semantic Prior Projection<br/>GIST text encoder + MLP projection into Er"]
D --> E["Gated Cross-Modal Feature Fusion<br/>Adaptive layer-wise injection into Kairos decoder"]
E --> F["Horizon-Decoupled Parallel Prediction<br/>Independent horizon query heads output Y_k"]
Key Designs¶
1. Tool-Augmented Statistical Feature Extraction: Overcoming Serialization Hallucination
Feeding continuous floating-point sequences directly as serialized strings degrades LLM numerical reliability. KAIROSAGENT provides the reasoner with 23 specialized statistical tools adapted from tsfresh across 7 categories: Trend (linear_trend), Volatility (standard_deviation, mean_abs_change), Distribution (quantile), Periodicity (autocorrelation), Extrema (number_peaks, mean_n_absolute_max), Frequency (fft_coefficient, spkt_welch_density), and Dynamics (augmented_dickey_fuller, cid_ce). Operating over arbitrary user-specified windows \([left, right)\), the agent executes a structured diagnostic policy: first evaluating full-window properties for baseline trends, then targeting local tail regimes for recent volatility shifts, and finally confirming periodic recurrence or turning points.
2. Gated Cross-Modal Feature Fusion: Bridging the Structural Modality Disconnect
Rather than keeping textual reasoning external or naively prepending text embeddings, KAIROSAGENT embeds the morphology description \(r\) into the latent space of the Kairos foundation model: \(E_r = \text{Proj}(\text{TextEnc}(r)) \in \mathbb{R}^{M \times D_h}\). At each decoder layer \(l\), the hidden state \(H^{(l)}\) is updated via gated cross-attention:
where \(g^{(l)} \in \mathbb{R}^{D_h}\) is a learnable gating parameter initialized to \(-2.197\), ensuring \(\sigma(g^{(l)}) \approx 0.1\) at the beginning of training. This progressive gating preserves the well-calibrated numerical representations of the pretrained TSFM while allowing semantic conditioning to activate smoothly without disruptive catastrophic interference.
3. Horizon-Decoupled Parallel Prediction: Eliminating Autoregressive Rollout Drift
Standard patch-indexed decoders struggle when rolled out autoregressively across arbitrary horizons due to cascading error accumulation. KAIROSAGENT decouples decoder queries into \(K\) horizon-specific learnable embeddings \(\{q_k\}_{k=1}^K\), corresponding to target horizons \(\{H_1, \dots, H_K\}\). Each query attends to the shared encoded history \(H_{\text{en}}\) and morphology prior \(E_r\), generating horizon-dedicated embeddings that feed directly into distinct readout heads:
This formulation supports multi-horizon inference in a single forward pass, granting each head temporal specialization while enforcing global morphological consistency.
Loss & Training¶
The training procedure unfolds across three systematic stages:
- Stage I: Reasoner SFT: Fine-tunes Qwen3.5-4B using standard cross-entropy over all assistant tokens on 80% of the curated T-STAR corpus (41,864 validated multi-turn trajectories across 29 datasets from 9 domains), establishing tool-selection and morphology-summarization foundations.
- Stage II: Multimodal Alignment: Fine-tunes the Kairos forecaster backbone, text adapter, and prediction heads using the full T-STAR dataset under a temporally weighted quantile pinball objective:
- Stage III: Turn-Level Advantage RL (GRPO): Refines the reasoner on the held-out 20% T-STAR subset using the frozen Stage II forecaster as a reward provider. At each interaction turn \(i \in [1, N]\), the reasoner produces an intermediate morphology hypothesis \(r_i\) (with \(r_0\) as the tool-free baseline), scored by \(S_i = -L_q(X, r_i; Y)\). The marginal step improvement \(\Delta_i = S_i - S_{i-1}\) generates discounted returns \(R_i = \sum_{j=i}^N \gamma^{j-i} \Delta_j\). Advantages \(\hat{A}_i^g\) are computed via group normalization across \(G\) sampled rollouts, attributing fine-grained credit directly to turn \(i\) tokens under the GRPO objective.
Key Experimental Results¶
Main Results¶
On the large-scale multi-domain Time-MMD benchmark, KAIROSAGENT was evaluated across 8 real-world domains against leading zero-shot foundation models and fully supervised baselines:
| Domain | Metric | KAIROSAGENT (Ours, Zero-Shot) | Aurora (Zero-Shot) | Sundial (Zero-Shot) | Moirai (Zero-Shot) | ChronosBolt (Zero-Shot) | T3Time (Full-Shot Multimodal) |
|---|---|---|---|---|---|---|---|
| Agriculture | MSE / MAE | 0.194 / 0.282 | 0.282 / 0.356 | 0.327 / 0.366 | 0.239 / 0.306 | 0.218 / 0.302 | 0.229 / 0.303 |
| Climate | MSE / MAE | 0.863 / 0.739 | 0.863 / 0.747 | 0.920 / 0.765 | 0.982 / 0.792 | 0.948 / 0.788 | 1.206 / 0.894 |
| Economy | MSE / MAE | 0.186 / 0.335 | 0.275 / 0.412 | 0.216 / 0.348 | 0.198 / 0.345 | 0.192 / 0.342 | 0.239 / 0.384 |
| Energy | MSE / MAE | 0.217 / 0.330 | 0.251 / 0.370 | 0.234 / 0.337 | 0.261 / 0.347 | 0.263 / 0.355 | 0.266 / 0.378 |
| Environment | MSE / MAE | 0.378 / 0.435 | 0.276 / 0.379 | 0.379 / 0.443 | 0.412 / 0.446 | 0.427 / 0.462 | 0.489 / 0.507 |
| Security | MSE / MAE | 76.658 / 4.340 | 72.763 / 4.085 | 83.403 / 4.836 | 74.249 / 4.129 | 73.977 / 4.117 | 72.113 / 4.070 |
| Social Good | MSE / MAE | 0.769 / 0.376 | 0.828 / 0.506 | 0.819 / 0.377 | 0.868 / 0.391 | 0.951 / 0.388 | 0.998 / 0.432 |
| Traffic | MSE / MAE | 0.151 / 0.231 | 0.162 / 0.289 | 0.228 / 0.292 | 0.186 / 0.263 | 0.222 / 0.249 | 0.289 / 0.368 |
| Top Rank Count | Count | 6 (MSE) / 6 (MAE) | 2 / 1 | 0 / 0 | 0 / 0 | 0 / 0 | 0 / 0 |
On the irregular multimodal benchmark Time-IMM, KAIROSAGENT was assessed zero-shot across diverse temporal irregularity patterns:
| Dataset | Irregularity Type | KAIROSAGENT (Zero-Shot) | TiMi (Full-Shot SOTA) | IMM-TSF (Full-Shot) | Time-MMD (Full-Shot) | PatchTST (Full-Shot Unimodal) |
|---|---|---|---|---|---|---|
| GDELT | Event-based logging | 1.075 / 0.642 | 1.029 / 0.689 | 1.051 / 0.694 | 1.235 / 0.773 | 1.067 / 0.705 |
| RepoHealth | Adaptive/reactive logging | 0.429 / 0.307 | 0.532 / 0.416 | 0.552 / 0.418 | 0.579 / 0.453 | 0.573 / 0.419 |
| FNSPID | Resource-aware collection | 0.111 / 0.180 | 0.127 / 0.222 | 0.128 / 0.221 | 0.141 / 0.245 | 0.130 / 0.224 |
| ClusterTrace | Operational-window sampling | 0.887 / 0.667 | 0.689 / 0.676 | 1.037 / 0.839 | 0.830 / 0.745 | 2.037 / 0.990 |
| CESNET | Scheduling jitter/delay | 1.040 / 0.769 | 0.896 / 0.727 | 1.124 / 0.819 | 0.971 / 0.761 | 1.318 / 0.880 |
| EPA-Air | Multi-source asynchrony | 0.583 / 0.553 | 0.599 / 0.579 | 0.620 / 0.599 | 0.651 / 0.598 | 0.620 / 0.595 |
Ablation Study¶
Ablations isolating the reinforcement learning optimization strategy and semantic modality fusion:
| Configuration / Variant | Reasoning Acc. (Climate / Energy / Traffic %) | Time-IMM MAE (ClusterTrace / EPA-Air / RepoHealth) | Time-IMM MASE (ClusterTrace / EPA-Air / FNSPID / RepoHealth) |
|---|---|---|---|
| Pretrained Forecaster (w/o text / Kairos-Base) | — | — | 13.10 / 3.02 / 6.76 / 4.28 |
| Fine-Tuned Numerical Forecaster (T-STAR values only) | — | — | 13.56 / 3.20 / 7.58 / 4.82 |
| KAIROSAGENT (SFT-Only Reasoner) | 97.80 / 45.21 / 40.56 | 0.692 / 0.570 / 0.313 | — |
| KAIROSAGENT (+ Outcome-Level Reward RL) | 96.70 / 43.33 / 38.27 | 0.679 / 0.561 / 0.309 | — |
| KAIROSAGENT (+ Turn-Level Credit Assignment RL) | 98.08 / 50.47 / 43.88 | 0.667 / 0.553 / 0.307 | 11.08 / 2.56 / 4.90 / 3.67 |
Key Findings¶
- Turn-Level Credit Assignment Transcends Sparse Terminal Rewards: Outcome-level RL causes slight performance degradation compared to SFT alone (Traffic accuracy drops from 40.56% to 38.27%), because assigning a single scalar return across long trajectories fails to distinguish informative tool actions from redundant exploration. Turn-level RL isolates step-wise marginal improvement \(\Delta_i\), unlocking superior reasoning accuracy and downstream forecasting gains.
- Semantic Morphology Text Provides Real Inductive Bias: Solely fine-tuning the numerical backbone on T-STAR values without morphology text worsens forecasting error (MASE degrades to 13.56 on ClusterTrace and 4.82 on RepoHealth). Incorporating the morphology prior yields dramatic MASE reductions between 14.1% and 27.5%, proving that textual morphology acts as an indispensable global prior.
- Dynamic Diagnostic Tool Invocation: Tool distribution shifts from global properties (Trend and Volatility account for >70% in turns 1–2) to local refinements (Extrema, Quantiles, and Periodicity in later turns), dynamically adapting to domain characteristics (e.g., US Births prioritizes periodicity, while Alibaba traces prioritize volatility).
Highlights & Insights¶
- Morphology Deduction Mitigates Numerical Hallucination: Rather than forcing LLMs to output volatile numeric predictions, restricting the reasoner to qualitative morphology representations decouples semantic pattern synthesis from continuous regression, achieving optimal synergy across both modalities.
- Smooth Initialization via Gated Activation: Initializing the cross-modal fusion gate to \(-2.197\) (\(\sigma \approx 0.1\)) allows the pretrained numerical model to maintain baseline statistical stability before gradually absorbing semantic signals during training.
- High Inference Efficiency: The modified forecaster adds only 40.6M parameters and 0.161 GB of GPU memory over the unimodal backbone, achieving a forward-pass latency of only 0.075s on a TITAN RTX GPU.
Limitations & Future Work¶
- Single TSFM Backbone Validation: The current architecture has been implemented exclusively upon Kairos; empirical evaluations across alternative backbones (e.g., Chronos, PatchTST, Moirai) remain to be explored.
- Task Scope Restricted to Forecasting: The framework focuses on time series forecasting and morphology synthesis, leaving extensions to anomaly detection, classification, and event causality analysis for future investigation.
- Multi-Agent Expansion: The reasoner-forecaster pair operates as a single conversational agent pipeline; incorporating multi-agent peer reviews or iterative critic loops could further enhance reasoning robustness under severe distribution shifts.
Related Work & Insights¶
- vs TSRMs (Time-MMD, Timeomni-1): Direct text serialization suffers from tokenization fragmentation and severe hallucination; KAIROSAGENT retains native numerical generation in the TSFM while using LLMs strictly for qualitative pattern deduction.
- vs TSFMs (Chronos-2, Moirai-2, TiRex): Unimodal TSFMs lack contextual reasoning and interpretability; KAIROSAGENT introduces structured tool-grounded semantic priors that guide zero-shot forecasting in non-stationary and irregular environments.
- vs Tool-Using Agents (TimeART, Cast-R1): Previous agents either isolate text reasoning from forecasting models or rely on sparse outcome RL; KAIROSAGENT establishes deep latent fusion and turn-level process credit assignment.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Decouples qualitative morphology reasoning from quantitative continuous prediction, coupled with turn-level credit assignment.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive zero-shot evaluations across 8 regular domains and 6 irregular temporal dynamics with in-depth tool usage analysis.
- Writing Quality: ⭐⭐⭐⭐⭐ Cohesive narrative, clear mathematical formulation, and well-designed architectural and experimental visualizations.
- Value: ⭐⭐⭐⭐⭐ Establishes a highly efficient, interpretable, and reproducible benchmark and agent framework for multimodal time series intelligence.