TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting Fan Zhang1 , Shiming Fan1 , Hua Wang2, * , 1
Shandong Technology and Business University, 2 Ludong University {zhangfan,2024410061}@sdtbu.edu.cn, [email protected]
arXiv:2604.12648v1 [cs.LG] 14 Apr 2026
Abstract Despite the recent success of large language models (LLMs) in time-series forecasting, most existing methods still adopt a Deep Synchronous Fusion strategy, where dense interactions between textual and temporal features are enforced at every layer of the network. This design overlooks the inherent granularity mismatch between modalities and leads to what we term semantic perceptual dissonance: highlevel abstract semantics provided by the LLM become inappropriately entangled with the lowlevel, fine-grained numerical dynamics of time series, making it difficult for semantic priors to effectively guide forecasting. To address this issue, we propose TimeSAF, a new framework based on hierarchical asynchronous fusion. Unlike synchronous approaches, TimeSAF explicitly decouples unimodal feature learning from cross-modal interaction. It introduces an independent cross-modal semantic fusion trunk, which uses learnable queries to aggregate global semantics from the temporal and prompt backbones in a bottom-up manner, and a stage-wise semantic refinement decoder that asynchronously injects these high-level signals back into the temporal backbone. This mechanism provides stable and efficient semantic guidance while avoiding interference with lowlevel temporal dynamics. Extensive experiments on standard long-term forecasting benchmarks show that TimeSAF significantly outperforms state-of-the-art baselines, and further exhibits strong generalization in both few-shot and zero-shot transfer settings.
1
Introduction
Long-term time series forecasting (LTSF) plays a crucial role in a wide range of real-world applications, including power load management (Fan et al., 2025a; Qiu et al., 2025c), traffic flow analysis (Kieu et al., 2024; Shen and Zhang, 2026; Qiu * Corresponding author.
et al., 2025a; Wang et al., 2026b), weather prediction (Liu et al., 2026), and financial markets (Ariyo et al., 2014; Zhang et al., 2026a). Traditional time series analysis methods typically rely on statistical models or deep learning architectures to capture temporal dependencies from historical observations (Zeng et al., 2023; Wang et al., 2024; Han et al., 2024; Ma et al., 2025; Wang et al., 2026a). However, as illustrated in Fig. 1(a), these models are often confined to the numerical modality and overlook the rich contextual information behind the series, such as metadata and event descriptions (Jin et al., 2024; Ge et al., 2025a). This separation between numerical dynamics and semantic context leads to a semantic perceptual deficit, which limits the model’s ability to generalize across domains and makes it difficult to adapt to data-scarce scenarios (Zhao et al., 2025; Ding et al., 2025) such as few-shot and zero-shot forecasting. In recent years, large language models (LLMs) have been introduced into time-series forecasting to compensate for the lack of semantic priors, leveraging their strong reasoning ability and rich parametric knowledge. Existing LLM-based methods typically adopt the deep synchronous fusion strategy shown in Fig. 1(b), i.e., layer-wise semantic coupling, where textual and temporal features are tightly aligned at every layer via dense crossattention or feature concatenation (Wang et al., 2025; Liu et al., 2024c). However, this design ignores the semantic perceptual dissonance between discrete text and continuous time series: high-level abstract semantics are compressed into the same representational scale as low-level numerical fluctuations, leading to heavily entangled features that are hard to interpret or control. We term this effect semantic perceptual dissonance, where LLM priors cannot effectively guide temporal forecasting and may even cause negative transfer when fusion is performed at inappropriate depths. To address these issues, we propose a hierarchi-
Forecasting
Forecasting Traditional Model
Forecasting
L Layer
L Layer
Trainable
Frozen
1 Layer
Semantic-Perceptual Dissonance
Inject
1 Layer
0.215
1 Layer
Fusion Memory
LLMs
LLMs
ETT m1
TimeLLM UniTime TimeCMA CALF Ours
Weather
TS Token
TS Token
(b) Deep Synchroncious Fusion Forward
…
(a) Time Series Foundation Methods
L Layer
Stage-wise Semantic Refinement
…
TS Token
…
… 1 Layer
Semantic Perceptual Deficit
L Layer Learnable Query
Layer-wise Semantic Coupling
(c) Hierarchical Asynchronous Fusion (Ours) Modal Fusion
Predicted value
Token
(d) MSE of LLM-based models in LTSF
Figure 1: Comparison of strategy and performance between TimeSAF and other methods.
cal asynchronous fusion strategy, as illustrated in Fig. 1(c). Unlike designs that enforce synchronous interaction at every layer, the proposed scheme employs stage-wise semantic refinement to restrict cross-modal interaction to a few discrete stages: it first aggregates semantic representations from the time-series backbone and the prompt backbone in a bottom-up manner, and then injects these highlevel semantics into deeper layers of the temporal backbone in a top-down manner. This asymmetric, cross-layer interaction prevents numerical modeling and semantic interaction from being entangled at all layers. Building on this strategy, we develop TimeSAF, a multimodal time-series forecasting framework. Architecturally, TimeSAF constructs a compact semantic memory bank between temporal patterns and textual prompts via a fusion trunk parameterized by learnable queries, while embedding gated asynchronous refinement blocks into both unimodal backbones so that the temporal branch can selectively read from the fusion memory and update its representations. This design deliberately decouples feature extraction and multimodal fusion along the temporal depth, mitigating the interference caused by layer-wise synchronous fusion, while the fine-grained top-down refinement continuously aligns and injects task-relevant semantics from the fusion trunk into the numeric backbone. Extensive experiments show that TimeSAF achieves superior performance to existing methods across multiple LTSF benchmarks and multimodal settings. In summary, the contributions of this paper are as follows: • Fusion strategy. We propose a hierarchical asynchronous fusion strategy that decouples
unimodal encoding from cross-modal interaction, effectively alleviating the entanglement between numerical features and textual semantics. • Model architecture. Building on this strategy, we introduce TimeSAF, which incorporates an independent cross-modal semantic fusion trunk and stage-wise semantic refinement decoder. The architecture first aggregates global semantics in a bottom-up manner and then injects the fused semantics back into the temporal backbone in a top-down fashion. • Empirical validation. Extensive experiments on seven public benchmarks demonstrate that the proposed method consistently achieves state-of-the-art performance compared with both LLM-based and non-LLM baselines.
2
Related Work
With the rapid advances of deep learning (Li et al., 2025; Xiao et al., 2026a) in domains such as computer vision (Li et al., 2026b; Chen et al., 2026), video understanding (Chen et al., 2025b), and multimodal representation learning (Xiao et al., 2026b; Ge et al., 2025b), data-driven neural models have also become increasingly prevalent in time series forecasting. Early time series forecasting predominantly relied on classical statistical models such as ARIMA, VAR, and STL with trend–seasonal decomposition (Siami-Namini et al., 2018; Schorfheide, 2005; Cleveland et al., 1990). These approaches are robust in short-term and single-task settings, but are typically built on stationarity assumptions and are sensitive to high-dimensional nonlinear de-
pendencies and complex noise. With the rise of deep learning, methods based on RNNs, CNNs, Transformers, and MLPs have become mainstream (Sherstinsky, 2020; Wu et al., 2023; Han et al., 2024): RNN/LSTM/GRU model temporal dependencies via recurrent hidden states, convolutional architectures capture local temporal patterns and inter-variable relations, and Transformer-style models leverage global self-attention to better handle long-range dependencies (Li et al., 2026a; Chen et al., 2025a; Hu et al., 2026; Zhang et al., 2026b; Fan et al., 2025b). Building on this, PatchTST (Nie et al., 2023) enhances long-sequence modeling through temporal patching and channelindependent design, while iTransformer (Liu et al., 2024d) and MoE/subspace-based methods (Qiu et al., 2025b) mitigate multivariate heterogeneity and non-stationarity via channel reordering and pattern grouping. Despite this growing architectural diversity, these models still rely solely on historical numerical sequences to make deterministic predictions, which limits cross-domain generalization and leads to suboptimal performance in zero-shot and few-shot regimes. More recently, large language models (LLMs) have been incorporated into time series forecasting (Gruver et al., 2023; Chang et al., 2025). TimeLLM (Jin et al., 2024) adopts a time-series encoder with LLM interaction, jointly feeding encoded temporal features and textual prompts into the LLM so that it can assist trend understanding and pattern reasoning. GPT4TS (Zhou et al., 2023) converts raw time series into token sequences via encoding, quantization, or descriptive prompts, and lets a pretrained LLM directly generate future trajectories. CALF (Liu et al., 2025b) adopts a dual-stream architecture with dedicated loss functions to achieve deep cross-modal alignment; TimeCMA (Liu et al., 2025a) focuses on cross-channel alignment to alleviate channel entanglement; DualSG (Ding et al., 2025) employs a decoupled dual-stream design, where a numeric stream models fine-grained temporal dynamics and a semantic stream, driven by an LLM, performs trend-level semantic correction and channel-wise reasoning. Our Work. Unlike prior LLM-based forecasters, TimeSAF first lets the temporal and semantic branches learn stable single-modal representations, and only then injects LLM-derived semantics back into the temporal backbone at a few staged levels, achieving a better balance between semantic guidance and robust numerical modeling.
3
Problem Formulation and Preliminary
3.1
Problem Formulation
Given a multivariate time series X = {x1 , ..., xL } ∈ RL×N where L denotes the length of the historical window and N the number of variables, we further construct for each variable an offline LLM-derived prompt embedding E ∈ RDllm ×N where Dllm is the dimensionality of the textual feature space. Given an observation window Xt−L+1,t and its associated prompts E, the objective under a forecasting horizon of length H is to learn a mapping: fθ (Xt−L+1,t , E) 7→ Ŷt+1:t+H ∈ RH×N 3.2
(1)
Time Series Encoding Branch
In practical applications, time series often exhibit strong non-stationarity (Liu et al., 2024a). To alleviate this, we first apply reversible instance normalization (RevIN) (Kim et al., 2021) to the input sequence. We then perform segmentation and embedding along the temporal dimension. Given a sequence of length L, with window length P and stride S, the time axis is partitioned into Np = L−P + 1 temporal patches, each containS ing P consecutive time steps, forming the initial ime temporal patch tokens PTt−L+1,t ∈ RNp ×P : ime PTt−L+1,t = Patching(Xt−L+1,t ).
(2)
A projection layer g(·) is then applied along the temporal dimension to map each patch to a Ddimensional representation. With an added learnable positional embedding epos , we obtain the final temporal tokens, which are used as input to the subsequent encoder: pos Xtime . i,t−L+1,t = g(Pi,t−L+1,t ) + e
3.3
(3)
LLM-based Prompt Encoding Branch
To inject external priors and semantic structure into the model, we introduce an LLM-based prompt encoding branch. This branch leverages a pretrained and frozen GPT-2 backbone to process natural language descriptions associated with segments of the input sequence. Following (Liu et al., 2025a), we automatically generate, for each variable in X, a prompt describing its statistics over the observation window (see Appendix D for details). Each prompt is tokenized by the GPT-2 tokenizer and fed into the frozen GPT-2 encoder, yielding prompt representations E ∈ RDllm ×N , where N is the number of
Space Block
S5
Input History 𝐿 timesteps: 𝐗 ∈ ℝ𝐋×𝐍 Var 1
Patching & Embedding
Unimodal Temporal Encoding Block
Instance Norm
Feed-Forward Network
Self-Attention
…
Reversible Instance Normalization (RevIN)
Self-Attention Cross-Attention
D P0
P1 P2
P3
P4 P5
Feed-Forward Network
Self-Attention
Cross-Attention Learnable Query Slots
Time
Prompt
Temporal backbone: progressivelymodeling patch-level dynamics andnon-stationarity.
Prompt Embedding
Variable-wise LLM prompts
1
2
3
Unimodal LLM Encoding Block
𝜅𝑠
Flatten Head
Self-Attention
Cross-Attention
Fusion memory: ℱ 𝑠
Learnable Positional Embeddings
𝜅𝑺𝒇𝒊𝒏𝒂𝒍
𝒃=𝟏
Var N
Patch Token Future 𝐻 timesteps:
𝑻𝒆𝒙𝒕
Self-Attention Cross-Attention
Self-Attention
Feed-Forward Network
Semantic backbone: Adapt LLM prompts to forecasting feature space.
Cross-Attention
Fusion 𝒇𝒊𝒏𝒂𝒍 memory: 𝓕 𝑺
Fusion Trunk : Aggregation of high-level temporal patterns and LLM semantics.
+ 𝜶𝓡(𝜣) 𝟐
Var 1 Var 2
𝑻𝒊𝒎𝒆
𝓗𝜿𝒔 Self-Attention Self-Attention Self-Attention Self-Attention Cross-Attention Cross-Attention Cross-Attention Cross-Attention Adapter Adapter Adapter Adapter Gate Gate Gate Gate
𝓗𝜿𝑺
𝟐
𝑩
𝒃 −𝒀 𝒃 𝓛𝐭𝐚𝐬𝐤 = 𝒀
𝑻𝒆𝒙𝒕
Learnable Query Slots
LLM-Derived Prompt Embedding
Projection RevIN Inverse
Feed-Forward Network
Self-Attention
Frozen GPT-2 Prompt Encoder
Self-Attention Self-Attention Self-Attention Cross-Attention Cross-Attention Cross-Attention Cross-Attention Adapter Adapter Adapter Adapter Gate Gate Gate Gate
𝓗𝜿𝒔
4
…
Var1: From [T1] to [Tn], the values were x1 ,..., xn Var5: From [T1] to [Tn], every f. The total trend x ,..., x valuethe wasvalues T. were Var4:1 From n[T1] to [Tn], every f. The total trend the values were x1 ,..., xn value was T. every f. The total trend value was T.
𝑻𝒊𝒎𝒆
𝓗𝑳
𝑻𝒆𝒙𝒕 𝓗𝜿𝟏
…
value was T.
Fusion memory: ℱ 1
Cross-Attention
Var1: From [T1] to [Tn], the values were x1 ,..., xn Var2: From [T1] to [Tn], every f. The total trend ,..., x[T valuethe wasvalues T. were n ] to [T ], Var1:x1From 1 n every f. The trend the total values were x1 ,..., xn value wasevery T. f. The total trend
…
Output
Semantic Refinement Decoder
Learnable Query Slots 𝜅1
𝑻𝒊𝒎𝒆 𝓗𝜿𝟏 Self-Attention
Var 2 Var N
Cross-Modal Semantic Fusion Trunk
𝑻𝒊𝒎𝒆
𝓗𝜿𝑺 Self-Attention Self-Attention Self-Attention Self-Attention Cross-Attention Cross-Attention Cross-Attention Cross-Attention Adapter Adapter Adapter Adapter Gate Gate Gate Gate
Time
Y ∈ ℝH×N Trainable Frozen Learnable token Forward Patch Token
Figure 2: Overall architecture of TimeSAF.
variables and Dllm is the LLM embedding dimension (768 for GPT-2). To ensure consistency within a mini-batch, all prompt sequences are padded to a unified length. We then introduce a learnable semantic adaptation module l(·) that maps E from the original LLM embedding space to the model semantic space RD , producing node-wise semantic features. A learnable positional embedding e is further added to each variable, resulting in the textual modality embedding XT ext ∈ RD×N , which serves as the input to the subsequent semantic encoding branch.
4.1
Unimodal Encoding Backbones
After obtaining the encoded time-series tokens and LLM-based prompt embeddings, we construct two structurally symmetric unimodal backbones for the numerical and semantic modalities, respectively. Both backbones are composed of several stacked Unimodal Encoding blocks, each consisting of a self-attention layer and a feed-forward network, which progressively extract higher-level representations within each modality without introducing any cross-modal interaction. Let Hl denote the input to the l-th layer of a unimodal backbone: Ul = Hl + Self Attn(Hl )
4
(4)
Overall Architecture of TimeSAF
In this section, we outline the overall architecture of TimeSAF. As illustrated in Fig. 2, given a multivariate historical time series and its associated LLM-based prompts, TimeSAF first encodes the numeric input with a patching-based temporal encoder, while a frozen GPT-2 backbone produces prompt-derived semantic representations. A semantic fusion trunk is inserted at several predefined layers, where a set of learnable queries aggregates information from both the temporal and semantic branches. Subsequently, asynchronous refinement modules inject the fused semantics back into the temporal backbone, and a lightweight prediction head maps the refined temporal tokens to future forecasts.
Then, a position-wise feed-forward network further transforms the intermediate representation to produce the output of the (l+1)-th layer. Hl+1 = Ul + F F N (Ul )
(5)
Here, Self Attn(·) denotes a multi-head selfattention module with an internal layer normalization, and F F N (·) denotes a position-wise feedforward network. Stacking multiple Unimodal Encoding blocks on the temporal modality allows the model to progressively enrich the latent representation of the series at the patch level, while on the textual modality, the LLM-derived semantic vectors are gradually adapted to the feature space of the backbone network. It is worth noting that, at this stage, the two backbones evolve strictly within
their own modalities, providing stable and semantically well-formed unimodal representations for subsequent cross-modal fusion. 4.2
Cross-Modal Semantic Fusion Trunk
After the layer-wise encoding of the Unimodal Encoding Backbones, we obtain high-level representations for both the temporal and textual modalities. We then introduce an independent Cross-Modal Semantic Fusion Trunk, which performs explicit cross-modal aggregation at a set of designated fusion layers, and compresses information from both branches into a fixed-length fusion memory. Formally, each unimodal backbone contains dp blocks, and we predefine S fusion stages, yielding LS = dp/S depth intervals. Let κs be the layer index where the s-th fusion stage is triggered. At the beginning of this stage, we instantiate a set of learnable fusion queries QFs ∈ RPf ×Df , which are broadcast across samples and variables during the forward pass to form the initial fusion representaF ∈ R(BN )×Pf ×Df . When the temporal tion Hs,0 and textual backbones reach layer κs , their hidden states HκTsime and HκTsext are used as key–value inputs, and a bottom–up semantic aggregation is performed. We first apply self-attention over the fusion queries: F F H̃sF = Hs,0 + Self Attn(Hs,0 ).
(6)
Then, taking H̃sF as queries, we retrieve information from the temporal and textual branches via two cross-attention operations: H̃sF ← H̃sF + CrossAttnT ime (H̃sF , HκTsime ), H̃sF ← H̃sF + CrossAttnT ext (H̃sF , HκTsext ). (7) Finally, a position-wise feed-forward network integrates these signals and yields the fusion memory for the s-th stage: F (s) = H̃sF + F F N (H̃sF ).
(8)
By repeating this procedure over all S stages, the model obtains a set of fusion memories {F (1) , . . . , F (S) } that compactly encode highlevel semantic information from both backbones. Cross-modal interactions occur only at the designated fusion layers, while the temporal and semantic branches evolve independently in the remaining layers. This stage-wise design allows sufficient multi-modal integration, while avoiding the heavy coupling and optimization instability that
arise when cross-attention is inserted at every layer, and provides a clean contextual interface for the subsequent asynchronous refinement modules. A simplified theoretical analysis under linearized assumptions is provided in Appendix E to further support the intuition behind this design. 4.3
Semantic Refinement Decoder
After obtaining the stage-wise fusion memory F (s) from the semantic fusion trunk, the Semantic Refinement Decoder feeds this high-level semantic context back into both unimodal backbones. Let m ∈ {T ime, T ext} index the temporal and textual (m) modalities, and denote by Hl the input to the l-th layer of modality m. Before cross-modal refinement, we first perform an intra-modal self-attention update: (m)
Ul
(m)
= Hl
(m)
+ SelfAttn(m) Hl
,
(9)
where SelfAttn(m) is a multi-head self-attention block with layer normalization. (m) The intermediate representation Ul is then (s) refined using the fusion memory F via crossattention: (m) (m) Zl = CrossAttn(m) Ul , F (s) , (10) where CrossAttn(m) (·, ·) shares the same form as the cross-attention used in the Fusion Block and treats F (s) as shared key–value context. To obtain modality-specific refinement while keeping the injection strength controllable, each modality is equipped with an independent linear (m) adapter Wad , and a scalar gate g ∈ R shared across modalities. The gated refinement residual is (m)
Rl
(m)
(m)
= σ(g) Wad Zl
,
(11) (m)
where σ(·) is the sigmoid function, and Rl represents the refined signal contributed by the fusion memory. Finally, we add this residual back to the intra-modal representation and apply a positionwise feed-forward network to obtain the layer output: (m)
(m)
Ĥl = Ul (m) + Rl (m) (m) (m) Hl+1 = Ĥl + F F N (m) (Ĥl )
(12)
Architecturally, the temporal and textual branches share the same fusion memory F (s) inside the Semantic Refinement Decoder, but project it
Table 1: Average MSE and MAE over four prediction lengths. All experiments fix the lookback length T = 96. The prediction length set is H ∈ {96, 192, 336, 720}. The best result is red, the second best result is underlined. Our full results are in Appendix A. Models
TimeSAF
Metrics
MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE
MAE
MSE MAE MSE MAE MSE MAE
ETTm1
0.377 0.390 0.395 0.390 0.392 0.405 0.399 0.402 0.385 0.399 0.410 0.409 0.396 0.400 0.407
0.410
0.387 0.400 0.502 0.502 0.448 0.452
ETTm2
0.272 0.320 0.282 0.321 0.289 0.332 0.286 0.332 0.293 0.334 0.296 0.340 0.293 0.338 0.288
0.332
0.281 0.326 1.216 0.707 0.304 0.349
ETTh1
0.426 0.428 0.443 0.435 0.451 0.449 0.442 0.434 0.442 0.447 0.460 0.449 0.456 0.450 0.454
0.447
0.469 0.454 0.620 0.572 0.440 0.459
ETTh2
0.367 0.394 0.371 0.394 0.407 0.419 0.382 0.406 0.377 0.402 0.389 0.408 0.456 0.413 0.383
0.407
0.387 0.407 0.942 0.683 0.436 0.449
Weather
0.240 0.268 0.249 0.273 0.253 0.283 0.270 0.288 0.253 0.276 0.274 0.290 0.278 0.297 0.257
0.278
0.259 0.281 0.258 0.315 0.308 0.360
Exchange 0.340 0.385 0.356 0.402 0.432 0.446 0.338 0.391 0.363 0.404 0.372 0.416 0.370 0.409 0.360
0.403
0.367 0.404 0.470 0.477 0.518 0.428
Electricity 0.174 0.264 0.175 0.265 0.194 0.287 0.216 0.299 0.215 0.304 0.223 0.309 0.217 0.307 0.178
0.270
0.205 0.290 0.244 0.334 0.213 0.326
CALF
TimeCMA
Time-FFM
UniTime
Time-LLM
back to their own representation spaces through (T ime) (T ext) separate adapters Wad and Wad . This shared but modality-specific refinement drives the two branches toward a compatible latent subspace. At the beginning layer κs of each fusion stage, a new fusion memory F (s) is computed by the CrossModal Semantic Fusion Trunk and then reused within that stage to update HlT ime and HlT ext via RTl ime and RTl ext , respectively, until the next stage produces a new fusion memory that replaces it as the contextual signal. 4.4
Output Projection and Optimization Objective
Finally, the last-layer temporal representation Hd pT ime is fed into a linear output head, which first flattens the patch-wise features and then maps them to the prediction horizon: Y = F latten(Hd pT ime )Wout + bout .
(13)
The resulting forecast Y is then de-normalized via the inverse RevIN operation to obtain Ŷ ∈ RH×N on the original value scale. TimeSAF is trained with a mean squared error loss plus standard ℓ2 weight decay. Let Θ denote all trainable parameters; the overall objective is L = Lpred + αR(Θ),
(14)
where α ≥ 0 is a balancing coefficient, and B
Lpred =
2 1 X Ŷ(b) − Y(b) , B 2 b=1
R(Θ) =
X
θ2
θ∈Θ
(15) The objective L is minimized by back-propagation.
5
GPT4TS
iTransformer
PatchTST
Crossformer
FEDformer
Experiments
Datasets and Metrics. We evaluate TimeSAF on seven widely used multivariate time series benchmarks: the four subsets of the Electricity Transformer Temperature (ETT) dataset (ETTh1, ETTh2, ETTm1, and ETTm2), together with Electricity, Weather, and Exchange. In line with standard practice in forecasting studies, we adopt Mean Absolute Error (MAE) and Mean Squared Error (MSE) as our primary evaluation metrics. The detailed statistics of these datasets are summarized in Appendix B. Baselines. We compare TimeSAF against a diverse set of recent and representative forecasting models. (1) LLM-based models: CALF (Liu et al., 2025b), TimeCMA (Liu et al., 2025a),TimeFFM (Liu et al., 2024b) , UniTime (Liu et al., 2024c), Time-LLM (Jin et al., 2024), and GPT4TS (Zhou et al., 2023). (2) Transformer-based models: iTransformer(Liu et al., 2024d), PatchTST (Nie et al., 2023), Crossformer (Zhang and Yan, 2023), and FEDformer (Zhou et al., 2022). (3) CNN-based models: TimesNet (Wu et al., 2023) and MICN (Wang et al., 2023). (4) MLP-based models: DLinear (Zeng et al., 2023). Implementation Details. We conduct all experiments under a unified evaluation pipeline and adopt the same configuration as (Wu et al., 2023) to ensure a fair comparison with strong baselines. We use a pretrained GPT-2 model (the first six layers) (Wu et al., 2023) as the default LLM backbone. TimeSAF is optimized using the Adam optimizer, trained for up to 50 epochs with early stopping. All experiments are run on 8 NVIDIA GeForce RTX 3090 GPUs (24 GB each). See Appendix C for
Table 2: Few-shot forecasting performance on ETT datasets using only 10% of the training data. All experiments fix the lookback length T = 96. The prediction horizon is set to H ∈ {96, 192, 336, 720}. Our full results are in Appendix A. Models
TimeSAF
CALF
TimeCMA
Time-LLM
GPT4TS
PatchTST
Crossformer
FEDformer
TimesNet
DLinear
MICN
Metrics MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE ETTm1 0.483 0.449 0.504 0.462 0.594 0.504 0.636 0.512 0.608 0.500 0.557 0.483 1.340 0.848 0.696 0.572 0.673 0.534 0.567 0.499 0.970 0.674 ETTm2 0.294 0.330 0.302 0.330 0.309 0.344 0.308 0.343 0.303 0.336 0.295 0.334 1.985 1.048 0.356 0.392 0.321 0.354 0.329 0.382 1.073 0.716 ETTh1
0.622 0.527 0.644 0.541 0.721 0.575 0.765 0.584 0.689 0.555 0.683 0.645 1.744 0.914 0.750 0.607 0.865 0.625 0.647 0.552 1.405 0.814
ETTh2
0.422 0.423 0.419 0.427 0.451 0.448 0.589 0.498 0.579 0.497 0.550 0.487 3.139 1.378 0.553 0.525 0.476 0.463 0.441 0.458 2.533 1.158
Table 3: Zero-shot forecasting performance on the ETT datasets, where prediction lengths H ∈ {96, 192, 336, 720}. “h1→m1” indicates that models trained on ETTh1 are evaluated on ETTm1, and similarly for the other transfer settings. Our full results are in Appendix A. Models
TimeSAF
Metrics
MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE
CALF
TimeCMA
Time-LLM
GPT4TS
PatchTST
Crossformer
FEDformer
TimesNet
DLinear
MICN
h1 → m1 0.749 0.563 0.755 0.574 0.820 0.590 0.847 0.600 0.798 0.574 0.894 0.610 0.999 0.736 0.765 0.588 0.794 0.575 1.439 0.870 0.760 0.577 h1 → m2 0.313 0.355 0.316 0.355 0.329 0.365 0.315 0.357 0.317 0.359 0.318 0.362 1.120 0.789 0.357 0.403 0.339 0.370 2.428 1.236 0.399 0.439 h2 → m1 0.873 0.604 0.836 0.586 1.087 0.673 0.868 0.595 0.920 0.610 0.871 0.596 1.195 0.711 0.741 0.588 1.286 0.705 0.764 0.601 0.778 0.594 h2 → m2 0.317 0.360 0.319 0.360 0.349 0.386 0.322 0.363 0.331 0.371 0.420 0.433 2.043 1.124 0.365 0.405 0.361 0.390 0.527 0.519 0.496 0.496
more details. 5.1
Long-term Forecasting
Setups. For a fair comparison, we fix the input sequence length to L = 96 and consider four forecasting horizons H ∈ {96, 192, 336, 720}. Consistent with the TFB-based setup (Xu et al., 2023), we maintain the same configuration but do not use the “Drop-Last” trick during training to ensure fair comparison. Results. The overall results are summarized in Table 1. On the aggregated comparison across all datasets and forecasting horizons, TimeSAF consistently outperforms all baselines and achieves the best performance on all 14 aggregated metrics. Compared with state-of-the-art LLM-based methods (CALF, TimeCMA, Time-FFM, UniTime, and Time-LLM), TimeSAF reduces MSE by 3.10%, 8.83%, 6.58%, and 10.31%, respectively. It also clearly surpasses Transformer-, CNN-, and MLPbased models, with improvements over these baselines typically exceeding 5%. These experimental results demonstrate that our TimeSAF can fully utilize the temporal patterns and semantic information in a limited input sequence, thus enabling accurate predictions.
5.2
Few/zero-shot Learning
Setups. Given that large language models (LLMs) have demonstrated strong generalization in fewshot and zero-shot learning settings (Zhou et al., 2023; Jin et al., 2024), this property is particularly relevant for real-world time series forecasting under data-scarce scenarios. Therefore, we also evaluate the performance of TimeSAF in few-shot and zero-shot regimes. In the few-shot setting, we use the ETT datasets and restrict the training data to only 10% of the original training split. In the zeroshot setting, the model trained on one dataset is directly deployed to a new, unseen dataset without any additional training, while keeping the training hyperparameters identical to those used in the long-term forecasting experiments. Few-shot Learning. Table 2 reports the results on the challenging few-shot forecasting task under different prediction horizons. With only a small fraction of training samples, TimeSAF achieves the best performance on 7 out of 8 overall metrics, indicating that the model can effectively extract useful patterns from limited historical data. Compared with LLM-based baselines (CALF, TimeCMA, Time-LLM, and GPT4TS), TimeSAF achieves relative MSE improvements of 2.38%, 10.93%, 18.91%, and 15.09%, respectively. In addi-
tion, TimeSAF outperforms the strong Transformerbased baseline PatchTST by 11.46%, further confirming its advantage in the few-shot setting. Zero-shot Learning. To further assess the crossdataset generalization ability of TimeSAF, we conduct zero-shot transfer experiments, where the model is trained on one dataset and then directly evaluated on a different dataset without any additional fine-tuning. As reported in Table 3, TimeSAF achieves the lowest average MSE on three out of four transfer directions. Overall, its zeroshot performance is comparable to the strongest LLM-based baseline, CALF, while consistently outperforming TimeCMA, Time-LLM, and GPT4TS, with average MSE reductions of 10.60%, 3.29%, and 4.19%, respectively. These results indicate that the proposed asynchronous fusion framework offers highly competitive zero-shot transfer capability compared with existing LLM-enhanced forecasting models.