T3LLM Framework: Time Series QA

• T3LLM is a framework using three LLM agents (worker, reviewer, and student) to generate, verify, and fine-tune answers for time series questions.
• It employs structured prompting and explicit chain-of-thought generation with reviewers detecting and correcting reasoning errors step-by-step.
• Empirical results on CTQA and TMQA benchmarks demonstrate superior accuracy, reduced forecasting errors, and improved efficiency.

The T3LLM (Time-series-Task-Tailored LLM) framework is an architecture for robust multi-step reasoning and self-correction in time series question answering (TSQA). It is motivated by the observation that, unlike text-only QA, TSQA allows each reasoning step to be verified against numerical input, facilitating explicit consistency checking and systematic correction. T3LLM employs three LLM agents—a worker, a reviewer, and a student—which respectively generate chain-of-thought (CoT) solutions, identify and correct reasoning errors, and internalize corrected reasoning traces for direct inference. Empirical results confirm that T3LLM achieves state-of-the-art performance on established TSQA benchmarks (Su et al., 27 Dec 2025).

1. Architectural Design

T3LLM employs a three-model division to modularize generation, review, and learning:

• Worker LLM ($f_{\mathrm w}$): Receives a TSQA input comprising question $Q_n$, time series $\mathcal T$, and a working prompt $\mathcal P_W$. It generates a step-wise CoT $\mathcal R^{(0)} = [R^{(0)}_1, …, R^{(0)}_L]$ and a provisional answer $\hat A^{(0)}_n$.
• Reviewer LLM ($f_{\mathrm r}$): Consumes $$(Q_n, \mathcal T, A_n, \mathcal R^{(t)}, \hat A^{(t)}_n)$$ along with a reviewing prompt $\mathcal P_R$, inspects each reasoning step, locates the first inconsistency, truncates the CoT after the erroneous step, and inserts an explicit [Reflection] correction.
• Student LLM ($f_{\mathrm s}$): Fine-tuned on corrected CoT-answer pairs to directly answer TSQA prompts in inference, eliminating the need for external review.

This design leverages strong LLMs for reasoning and error localization, while the student model, often made more compact, delivers inference-time efficiency (Su et al., 27 Dec 2025).

2. Structured Prompting and CoT Generation

Explicit prompt structuring underpins the reliability of T3LLM's reasoning:

• Working Prompt ($Q_n$0): Mandates step-wise logic within > …</think>, enforces numerical grounding, and restricts final outputs to <answer>…</answer>. Each reasoning step is explicitly enumerated. > > Example format: > ``` > <think> > [Step 1] Compute differences: ... > [Step 2] Apply volatility measure: ... > [Step 3] Map result to answer choices: ... <answer> C </answer> $Q_n$1python for each (Q_n, T, A_n): R^0, Â^0 = f_w(Q_n, T; prompt=P_W) for t in range(MCR): tune = f_r(Q_n, T, A_n, R^t, Â^t; prompt=P_R) if tune == NO_CHANGE: R* = R^t; Â* = Â^t; break else: R_tilde = tune R^t+1, Â^t+1 = f_w(Q_n, T, R_tilde; prompt=P_C) collect (Q_n, T, R*, A_n)

output = f_s(Q, T; prompt=P_W) ``` At inference, only the student model is deployed; worker and reviewer are used exclusively during training (Su et al., 27 Dec 2025).

6. Empirical Results and Ablations

T3LLM achieves prominent empirical advances on two major TSQA benchmarks:

Dataset	Task Type	Best Baseline	T3LLM Performance
CTQA	MCQ accuracy	0.467	0.665
TMQA	T/F accuracy	0.637	0.766
TMQA	Forecasting RMSE	11,926	7,586
TMQA	Imputation MAE	64.91	50.20

• Datasets: CTQA (trend/volatility MCQ), TMQA (classification, anomaly detection, forecasting, imputation).
• Metrics: Accuracy, F1/mF1 for discrete tasks; RMSE/MAE for sequence prediction.
• Ablation studies reveal performance gains saturate after 3–5 correction rounds, and quality improves with stronger worker/reviewer backbone models (e.g., DeepSeek-R1) (Su et al., 27 Dec 2025).

7. Insights, Limitations, and Future Directions

Key insights:

• Verifiable chaining of numeric reasoning using CoT induces a strong inductive bias for TSQA, enabling systematic error localization and correction at each reasoning step.
• Stopping review at the first detected error prevents error propagation and promotes concise corrective feedback.
• Training a student model on self-corrected CoTs yields a lightweight, standalone QA agent suitable for deployment.

Limitations include:

• Runtime overhead due to multiple worker–reviewer interleavings.
• Reliance on reviewer strength and prompt structure; insufficient reviewers may under-correct or over-prune.
• Prompt engineering may require adaptation for new TSQA task classes or formats.

Potential avenues for extension:

• Joint reinforcement-based training of worker and reviewer models.
• Application to domains with verifiable multi-step reasoning, such as program induction or table QA.
• Incorporation of uncertainty estimation within reflection steps.
• Development of assemblies involving multiple reviewers to increase correction reliability.

The T3LLM framework demonstrates that explicit, verifiable, and correctable CoT reasoning is highly effective for time-series question answering, establishing new standards for both methodology and empirical performance in the field (Su et al., 27 Dec 2025).