LLaTiSA: VLM for TSR & LATIS Survey

• LLaTiSA is a dual-purpose term representing a hierarchical vision-language model for difficulty-stratified time series reasoning and a major Lyα tomography survey in extragalactic astronomy.
• The vision-language model utilizes a structured curriculum and chain-of-thought supervision to improve numerical grounding, pattern recognition, and semantic reasoning in time series data.
• The LATIS survey delivers extensive spectroscopic data and 3D tomographic maps of the cosmic web at z≈2–3, advancing studies of protoclusters and IGM-galaxy correlations.

LLaTiSA encompasses two distinct but prominent meanings in contemporary academic literature: (1) a hierarchical Vision-LLM architecture and methodology for difficulty-stratified time series reasoning, and (2) an informal abbreviation, occasionally used playfully, for the Lyα Tomography IMACS Survey (LATIS), a major spectroscopic and tomographic data release for extragalactic astronomy at $z\sim2$–3. Both uses are documented in technical literature and involve rigorous pipeline development, benchmarking, and public data releases (Ding et al., 19 Apr 2026, Newman et al., 9 Oct 2025).

1. LLaTiSA for Time Series Reasoning: Overview and Motivation

LLaTiSA, as introduced by RainingNovember et al., refers to a Vision-LLM (VLM) specifically optimized for time series reasoning tasks. The model addresses key deficiencies in existing LLMs and VLMs, particularly their inability to robustly parse, analyze, and semantically reason about structured temporal data from both visual and numerical sources. The core challenges motivating LLaTiSA include:

• Fragmented Task Definitions: Contemporary benchmarks intermix classification, forecasting, and descriptive tasks under ambiguous definitions of "reasoning," making performance comparison and development of unified models difficult.
• Ambiguous Labels: Datasets often support multiple valid answers and lack sufficient context, inhibiting reliable evaluation.
• Deficient Numerical Grounding: Vision-only models recognize visual patterns but hallucinate precise values, while text-only models are inefficient at targeting specific locations in long sequences.

To systematically address these, LLaTiSA employs four design principles:

1. Difficulty-Stratified Curriculum: Structured training progression from basic value extraction (L1) through pattern recognition (L2) to semantic integration (L3).
2. Dual-View Multimodality: Each instance provides both a time-series plot and a precision-calibrated numerical grid, exposing the model to complementary qualitative and quantitative evidence.
3. Chain-of-Thought Supervision: Each training sample is paired with a verified stepwise reasoning trajectory, enforcing the logic flow from numerical grounding to pattern discovery to semantic judgment.
4. Plug-and-Play Backbone: LLaTiSA builds atop the Qwen3-VL-8B-Instruct VLM, only requiring minimal augmentation for an additional vision input (Ding et al., 19 Apr 2026).

2. Four-Level Taxonomy for Time Series Reasoning

The LLaTiSA framework formalizes Time Series Reasoning (TSR) as a function $\text{TSR}:(x, C) \to y$, where $x$ is the time-series input, $C$ contextual metadata, and $y$ the answer. The taxonomy is stratified by reasoning complexity:

Level	Name	Example Task Type
L1	Numerical Read-out	"At which timestamp is the max?"
L2	Pattern Perception	"How many turning points?", "Match the overall shape"
L3	Semantic Reasoning	"Diagnose turbine shutdown from the series"
L4	Predictive Inference	"Predict the next patch"

The operator $D(\cdot)$ encodes increasing task difficulty: $D(L1) < D(L2) < D(L3) < D(L4)$ (Ding et al., 19 Apr 2026).

3. HiTSR Dataset and Chain-of-Thought Annotation

LLaTiSA is enabled by HiTSR, a hierarchical, taxonomy-aligned dataset consisting of approximately 83,000 labeled tasks distributed across L1 to L3:

• HiTSR-L1 (~30,000): Synthetic, subtasked into Min/Max grounding, multi-series comparison, start/end, and subseries localization.
• HiTSR-L2 (~50,700): Synthetic, including local/global pattern differentiation and numerical perception.
• HiTSR-L3 (~3,100): Real-world series from energy, finance, health, requiring semantic understanding.

L2 and L3 items are annotated with explicit Chain-of-Thought (CoT) reasoning paths, generated and cross-validated by GPT-5 then human-audited to Cohen’s $\kappa \geq 0.85$. The annotation strictly enforces a three-stage logic: numerical grounding $\to$ pattern identification $\to$ semantic judgment. The test set is difficulty-stratified and contains only unambiguous queries with a single correct answer (Ding et al., 19 Apr 2026).

4. LLaTiSA Architecture and Curriculum Fine-Tuning

The model leverages a dual-view architecture:

• Inputs: $\text{TSR}:(x, C) \to y$0 (plot image), $\text{TSR}:(x, C) \to y$1 (numerical grid image), $\text{TSR}:(x, C) \to y$2 (textual context).
• Encoding: Shared vision encoder $\text{TSR}:(x, C) \to y$3 for both images ($\text{TSR}:(x, C) \to y$4, $\text{TSR}:(x, C) \to y$5), and frozen LLM text encoder $\text{TSR}:(x, C) \to y$6 ($\text{TSR}:(x, C) \to y$7).
• Fusion: Concatenated embeddings $\text{TSR}:(x, C) \to y$8 pass through cross-modal Transformer layers with standard self-attention, resulting in final decoding and answer generation.

Training follows a three-stage supervised fine-tuning (SFT) curriculum, minimizing cross-entropy:

• Stage 1 (HiTSR-L1): 1 epoch, LR$\text{TSR}:(x, C) \to y$9
• Stage 2 (HiTSR-L2): 2 epochs, LR$x$0
• Stage 3 (HiTSR-L3): 2 epochs, LR$x$1

Ablation experiments demonstrate that curriculum learning enhances high-level generalization (e.g., OOD L3 accuracy drops from 67.0% to 57.0% with joint training). Removing CoT annotations further reduces OOD L3 accuracy from 67.0% to 55.0% (Ding et al., 19 Apr 2026).

5. Empirical Results and Comparative Evaluation

LLaTiSA demonstrates robust out-of-distribution performance:

Task	LLaTiSA (%)	Closest Baseline (%)	Δ(pp)
L1 Min/Max grounding	86.8	54.2 (GPT-4o, multimodal)	+32.6
L2 Local pattern	75.6	72.2 (GPT-4o, plot-only)	+3.4
L2 Global pattern	97.5	96.7	+0.8
L3 Semantic	67.0	59.0 (ChatTS)	+8.0

Encoding strategy ablation shows dual-view inputs outperform plot-only or text-only pipelines. On real-world ECG grounding, LLaTiSA achieves 62.8% in-domain diagnosis accuracy and excellent lead-wise coverage (84.0%), using 2.5% of the training samples required by specialized models such as GEM-LLaVA. Failure analysis reveals that GPT-4o misreads step-like features—an error corrected by LLaTiSA’s dual-modality and CoT supervision.

6. Limitations and Future Work

LLaTiSA currently focuses on interpretive (L1–L3) rather than generative (L4) tasks, with predictive inference earmarked as a priority for extension. The reliance on supervised SFT constrains adaptation to unseen domains; integration with reinforcement learning fine-tuning (RFT) represents an open research avenue. Planned investigations include curriculum policy learning, prompt-based adaptation, and the development of multi-patch L4 evaluations (Ding et al., 19 Apr 2026).

7. "LLaTiSA" as the LATIS Survey: Data Release and Tomographic Mapping

In a separate context, “LLaTiSA” is also used as a colloquial abbreviation for the Lyα Tomography IMACS Survey (LATIS), a major optical spectroscopic and IGM tomography project (Newman et al., 9 Oct 2025). Key aspects include:

• Spectroscopic Scope: 7,408 optical spectra (3890–5830 Å, $x$2), via IMACS on Magellan-Baade, encompassing 5,575 high-confidence redshifts at $x$3.
• Tomographic Mapping: $x$4 Lyα forest pixels processed into 1.55 deg$x$5 3D maps at $x$6–2.8, with $x$7 cMpc spatial resolution through Wiener filtering and Gaussian smoothing.
• Scientific Applications: Enables identification of cosmic web structures, protoclusters (e.g., via peaks in $x$8), and direct IGM-galaxy cross-correlation studies.
• Caveats: Effective resolution and interpretability limited by smoothing scale, continuum placement, sky subtraction, and redshift uncertainties.

This dual usage underscores the importance of context when referencing “LLaTiSA”; in most vision-language contexts, the term denotes the hierarchical TSRM framework, while in extragalactic astronomy, it often refers to the LATIS project and its tomographic data products (Ding et al., 19 Apr 2026, Newman et al., 9 Oct 2025).