WeStar: Dual Frameworks in AI & Depth Estimation

• WeStar (conversational AI) is a lite-adaptive stylized question answering framework that combines context-grounded RAG with on-the-fly LoRA modules for style control.
• WeSTAR (depth estimation) is a weakly-supervised self-training adaptation framework that leverages dense pseudo-labels, semantic normalization, and ordinal depth cues.
• Both systems focus on parameter-efficient specialization using LoRA modules and tailored training pipelines to enhance scalability and robustness across diverse applications.

WeStar denotes two distinct research frameworks introduced in 2025 under closely related names. In large-scale conversational AI, WeStar is a lite-adaptive framework for stylized contextual question answering that combines context-grounded generation via RAG with style-aware generation using Parametric RAG (PRAG), where LoRA modules are dynamically activated per style cluster for millions of official accounts (Fan et al., 22 Sep 2025). In monocular depth estimation, WeSTAR denotes Weakly-Supervised Self-Training Adaptation with Regularization, a parameter-efficient framework that performs weakly supervised adaptation of depth estimation foundation models through dense self-training, semantically-aware hierarchical normalization, pairwise ordinal depth annotations, and weight regularization of LoRA updates (Huang et al., 18 Nov 2025).

1. Scope and nomenclature

The two frameworks share a near-identical name but address different technical problems. The conversational AI system is presented as “One Agent to Serve All: a Lite-Adaptive Stylized AI Assistant for Millions of Multi-Style Official Accounts” and focuses on stylized, context-grounded response generation for industrial-scale official account platforms. The monocular depth estimation system is presented as “Enhancing Generalization of Depth Estimation Foundation Model via Weakly-Supervised Adaptation with Regularization” and focuses on improving robustness and generalization of MDE foundation models in unseen and diverse domains (Fan et al., 22 Sep 2025, Huang et al., 18 Nov 2025).

In the supplied literature, the distinction is substantive rather than cosmetic. The former uses RAG, PRAG, style clustering, and SeDPO to control response style while preserving contextual relevance. The latter uses teacher-student self-training, semantics-informed normalization, sparse ordinal supervision, and regularized LoRA adaptation to improve depth prediction under distribution shift. A common source of confusion is therefore terminological: the label “WeStar” does not identify a single method family, but two separate systems with different objectives, datasets, and evaluation regimes.

2. WeStar for stylized contextual question answering

In the conversational setting, WeStar is defined as a lite-adaptive stylized question answering framework designed to efficiently serve millions of official accounts. The target problem is response generation that is simultaneously contextually grounded in account-specific content and stylistically aligned with the author’s interactive style. The paper positions existing alternatives as inadequate for this deployment regime: chain-of-thought prompting induces significant latency due to multi-turn reasoning, per-account fine-tuning is computationally prohibitive, and long prompt-based methods degrade the model’s ability to grasp injected context and style (Fan et al., 22 Sep 2025).

The architectural core is the combination of two knowledge-injection mechanisms. Retrieval-Augmented Generation (RAG) supplies contextual grounding by retrieving articles relevant to the user question and author, summarizing them, and injecting them into the model’s input prompt. Parametric RAG (PRAG) supplies style control through parameters rather than long prompts: each style cluster is associated with a corresponding set of LoRA modules, and the LoRA weights for the cluster matching the author’s style are loaded on-the-fly into the model. This design is explicitly intended to preserve contextual capacity while enabling efficient style adaptation.

The framework is also described as multi-dimensional, cluster-based parameter sharing. Rather than allocating a separate model for each account, WeStar groups authors with similar style profiles and allows all authors in a cluster to share a single parameter set. This reduces storage and training cost while preserving stylistic diversity. The paper’s stated motivation is that pure prompt injection of style is ineffective for scaling, causes context window overflow, and weakens style control; the cluster-based design is the proposed alternative.

3. Style representation, data construction, and SeDPO

WeStar’s data pipeline begins with the construction of CQA triplets consisting of context, question, and answer. The paper states that these are built through both forward-thinking and bottom-up approaches: one route uses LLM-generated question-answer pairs from articles, and the other uses LLM-simulated user queries followed by model-generated answers. The purpose is to balance diversity and realism.

Style labeling is carried out on 12 carefully-selected stylistic features spanning four dimensions: semantic, grammatical, syntactic, and lexical. Authors are then hierarchically clustered by these style labels via recursive splitting, producing a style tree in which each leaf node is a style cluster of authors with similar multi-dimensional style profiles. The paper gives the clustering schema in terms of style standards $S = \{s_1, \ldots, s_D\}$ with $D = 12$, and recursive splitting proceeds when a candidate split yields subsets of size $> k$.

From these ingredients, the framework constructs CQSA triplets by rewriting CQA answers into stylized answers using the style labels and in-context exemplars from the cluster itself. Training data are then filtered by automatic evaluation using four metrics: Contextual Alignment, Question Relevance, Stylistic Strength, and Fluency. The paper states that the top 10,000 CQSA samples are used for training each cluster.

The optimization procedure is Style-enhanced Direct Preference Optimization (SeDPO). For each CQSA sample, a rejected sample is selected from sibling clusters with high contextual similarity but a critical style label difference. This creates preference pairs that emphasize fine-grained stylistic distinctions while controlling for context. The paper defines the DPO-style objective as

$$L_{\text{DPO}} = -\log \frac{\exp[\beta \cdot (\log p_\theta(y^+|x) - \log p_\theta(y^-|x))]}{1 + \exp[\beta \cdot (\log p_\theta(y^+|x) - \log p_\theta(y^-|x))]}$$

and states that LoRA with rank 16 is used for each style cluster (Fan et al., 22 Sep 2025).

4. Online serving and empirical characteristics of WeStar

At inference time, WeStar follows a four-stage serving pipeline. Relevant article segments are first retrieved and inserted into the prompt. The author’s style cluster is then identified from the style tree. The LoRA weights for that cluster are dynamically loaded into the base LLM, specified in the paper as Qwen3-32B, and the model then generates a response intended to be both contextually accurate and stylistically faithful (Fan et al., 22 Sep 2025).

The reported motivation for this design is industrial-scale deployment efficiency. The paper states that LoRA modules are lightweight (<1% parameters of the full model) and can be loaded quickly for each request. This suggests that the system is engineered for low-overhead personalization rather than per-account retraining.

Empirically, the paper evaluates WeStar on a realistic, large-scale dataset from an official accounts platform and reports scores on four dimensions. WeStar achieves 4.43 on Question Relevance, 4.55 on Contextual Alignment, 4.25 on Stylistic Strength, and 4.77 on Fluency. The corresponding figures reported for SFT-Prompt are 4.26, 4.30, 3.73, and 4.70; for R1-Prompt, 4.38, 4.45, 4.25, and 4.75; and for LoRA-SFT, 4.35, 4.43, 3.92, and 4.73. The paper also reports an inference-time comparison in which WeStar achieves a ~1.2x speedup relative to SFT-Prompt, with 2.08s vs. 2.47s per sample. Qualitative analysis in the paper further states that WeStar more reliably reflects subtle target style traits, including grammatical quirks and persona idiosyncrasies.

5. WeSTAR for adaptation of monocular depth estimation foundation models

In monocular depth estimation, WeSTAR is introduced against the backdrop of foundation models such as the Depth Anything series, with the observation that these systems have substantially advanced zero-shot generalization in MDE but still remain suboptimal on novel domains, especially under distribution shifts or image corruptions. The central question posed is whether such models can be further improved when some downstream data are available. The proposed answer is WeSTAR, a parameter-efficient framework for Weakly supervised Self-Training Adaptation with Regularization (Huang et al., 18 Nov 2025).

The framework uses an EMA teacher model to generate pseudo-labels for unlabelled target samples within a teacher-student architecture. Each input receives a weak augmentation $T_w$ for the teacher and a strong augmentation $T_s$ for the student. The main self-supervision signal is a dense self-training objective. Because self-training in MDE is challenged by scale and shift ambiguity, WeSTAR normalizes both pseudo-labels and predictions using robust local statistics: $$\Phi(d_p, \mathcal{C}_p) = \frac{d_p - t(\mathcal{C}_p)}{s(\mathcal{C}_p) + \epsilon},$$ where $t(\mathcal{C}_p)$ and $s(\mathcal{C}_p)$ are the median and median absolute deviation of depths in context $\mathcal{C}_p$. The per-pixel self-training loss is given as

$D = 12$0

A distinctive component is Semantically-Aware Hierarchical Normalization (SA-HDN). Traditional normalization based on fixed grids or global statistics is described as potentially fragmenting semantic objects and yielding unstable normalization. WeSTAR instead uses instance-level segmentation maps automatically generated by SAM2. For each pixel, the normalization contexts include both the global scene and the corresponding instance mask. The stated effect is to enforce both global consistency and object-level fine structure.

6. Weak supervision, LoRA regularization, and robustness results in WeSTAR

Beyond dense pseudo-label supervision, WeSTAR introduces pairwise ordinal depth annotations as a cost-efficient weak supervision signal. Each weak label is a tuple $D = 12$1, where $D = 12$2 indicates “farther,” “equal,” or “nearer.” The ranking loss is defined piecewise in terms of the predicted depth difference $D = 12$3, with a margin-based term when $D = 12$4 and an absolute-difference term when $D = 12$5. The paper states that these ordinal constraints mitigate local topological errors that dense pseudo-labels cannot resolve (Huang et al., 18 Nov 2025).

Parameter efficiency is achieved through LoRA rather than full fine-tuning. Each attention layer’s weights are adapted as $D = 12$6, where $D = 12$7, $D = 12$8, and $D = 12$9. A weight regularization loss is then used to anchor the LoRA updates, keep the parameters near their initialized values, ensure training stability, and preserve the model’s generalizable knowledge. The total training objective is

$> k$0

The empirical evaluation spans NYU-V2, KITTI, Sintel, DIODE, corrupted variants such as NYU-C, KITTI-C, DIODE-C, Sintel-C, and real-world scenarios including NuScenes-Night and DrivingStereo-Weather. The reported metrics are $> k$1, defined as the fraction of pixels with relative depth error below $> k$2, and AbsRel, defined as

$> k$3

Baselines include zero-shot source models, source-free DA methods such as TTT++, FR, TTAC, and SSA, self-supervised iBOT, and weakly-supervised SGRL.

The paper reports that WeSTAR consistently improves generalization and achieves state-of-the-art performance across a wide range of benchmarks. On Sintel, the cited example moves from 74.8 / 20.3 for the source model to 82.2 / 16.9 for WeSTAR in $> k$4/AbsRel. On Sintel-C, the source model is reported at 60.3 / 30.6, the best prior weakly-supervised or DA figures are 66.5 / 29.9 and 63.7 / 28.5, and WeSTAR reaches 71.8 / 24.1, corresponding to +11.5 $> k$5 and -6.5 AbsRel relative to the source. Ablation studies are reported to show that dense self-training, weakly-supervised ranking, weight regularization, and SA-HDN contribute additively, that LoRA adaptation is crucial for robust but non-destructive adaptation, and that SA-HDN outperforms both global normalization and classic HDN using grid contexts.

These two systems therefore occupy different positions in contemporary machine learning practice. WeStar for official accounts addresses scalable style-conditioned generation through retrieval and parameter-efficient adaptation, whereas WeSTAR for depth estimation addresses weakly supervised domain adaptation of foundation models under corruption and out-of-distribution shift. The shared name masks fundamentally different problem formulations, but in both cases the central design pattern is parameter-efficient specialization rather than full-model retraining.