EReLiFM: Reliability-Aware Residual Flow Meta-Learning

• The paper presents an innovative filter‐expand‐recycle pipeline using evidential loss clustering, residual flow matching, and meta-learning to tackle open-set challenges under noisy labels.
• The method uses a dedicated UTS-ELC stage for reliable clean/noisy partitioning, followed by domain- and category-conditioned residual flow matching to augment data robustly.
• Empirical results demonstrate significant gains in Acc, H-score, and OSCR metrics on benchmark datasets, validating the efficacy of the integrated approach.

Evidential Reliability-Aware Residual Flow Meta-Learning (EReLiFM) is a training framework for Open-Set Domain Generalization under Noisy Labels (OSDG-NL) that combines evidential reliability estimation, residual flow matching, and meta-learning to improve known-class recognition and unseen-class rejection in unseen target domains [2510.12687]. The method addresses the joint difficulty of domain shift, open-set recognition, and corrupted source supervision by organizing training into a filter (\rightarrow) expand (\rightarrow) recycle pipeline: it first identifies likely clean and noisy samples using Unsupervised Two-Stage Evidential Loss Clustering (UTS-ELC), then trains Domain- and Category-Conditioned Residual Flow Matching (DC-CRFM) on the clean subset, and finally reuses noisy samples through pseudo-labeling and evidential supervision [2510.12687]. Within the broader literature, EReLiFM can be situated as a synthesis of evidential uncertainty modeling [2510.08938], residual or flow-based conditional transport [2503.19371], and residual reliability modulation [2604.21349], while remaining specifically tailored to OSDG-NL [2510.12687].

1. Task setting and conceptual scope

EReLiFM is defined for Open-Set Domain Generalization under Noisy Labels, where training uses multiple source domains and evaluation is performed on a held-out unseen target domain containing both known and unseen classes [2510.12687]. The label space is partitioned as
[
\mathcal{Y}=\mathcal{Y}_k \cup \mathcal{Y}_u,
]
with (\mathcal{Y}_k) denoting known classes available during training and (\mathcal{Y}_u) denoting unseen classes that appear only at test time [2510.12687]. Source supervision is itself unreliable because source labels are corrupted.

The paper presents three explicit difficulties. First, domain generalization requires learning from multiple source domains while generalizing to a new target domain. Second, open-set recognition requires the model both to classify known classes and to reject unknown classes. Third, noisy labels corrupt source-domain knowledge, which degrades both closed-set classification and open-set confidence estimates [2510.12687]. The method is motivated partly by the limitations of HyProMeta, which relies on hyperbolic prototypes for clean/noisy separation and label correction but is described as sensitive to noisy early features, limited in clean/noisy partitioning quality, and constrained by interpolation-style augmentation such as MixUp [2510.12687].

At a high level, EReLiFM replaces prototype-centric noise diagnosis with evidential training-dynamics-based reliability estimation, and replaces linear interpolation augmentation with structured residual flow modeling [2510.12687]. This makes the method distinct from work that uses meta-learning mainly as dynamic hyperparameter control for evidential loss terms, such as the Meta-Policy Controller (MPC), which adapts the KL divergence coefficient and class-specific Dirichlet prior strengths but does not include residual flows, normalizing-flow density modeling, or an episodic task-distribution formulation [2510.08938].

2. Core architecture and training pipeline

The paper summarizes EReLiFM with the sequence filter (\rightarrow) expand (\rightarrow) recycle [2510.12687]. Each stage corresponds to a concrete subsystem.

Component	Role	Mechanism
UTS-ELC	Reliability-aware filtering	FINCH + 2-component GMM on evidential loss trajectories
DC-CRFM	Structured augmentation	Domain- and category-conditioned residual flow matching
Meta-learning stage	Joint optimization on clean and noisy data	Meta-train on clean/augmented clean, meta-test on noisy samples

In the filter stage, a backbone (\mathbf{M}\alpha) is trained on the full noisy training set while recording evidential losses over (N_e=10) epochs [2510.12687]. These trajectories are used to partition the dataset into
[
\mathcal{T}{clean},\quad \mathcal{T}_{noisy}.
]
The method treats these trajectories as reliability signals rather than relying on feature prototypes [2510.12687].

In the expand stage, only the clean subset is used to train DC-CRFM, a conditional residual flow model that learns residual distributions for two transition types: domain residuals for same-category cross-domain variation, and category residuals for same-domain cross-category variation [2510.12687]. These residuals are added to clean samples to form augmented batches.

In the recycle stage, noisy samples are reintroduced in a meta-test objective. Pseudo-labels are obtained from the most confident prediction,
[
\mathbf{y}{pseudo} = \arg\max(\mathbf{M}\alpha(\mathbf{B}_{noisy})),
]
and combined with evidential learning and a cross-entropy term on the original noisy label [2510.12687]. The paper describes this as optimizing the model such that the update direction on the clean set maximizes the loss decrease on the noisy set, although the body instantiates this procedurally rather than through an explicit MAML-style bilevel gradient equation [2510.12687].

This design places EReLiFM within a broader class of systems that use learned residual transformations and reliability-aware supervision, but its combination is specific: reliability estimation precedes flow-based augmentation, and augmentation precedes cautious noisy-data reuse [2510.12687].

3. Evidential reliability awareness and UTS-ELC

The evidential component appears in two places: reliability diagnosis and meta-test supervision [2510.12687]. The evidential learning loss is written as
[
\mathcal{L}{EL} = \sum{i=1}^\mathcal{C} \left[\mathbf{y}{i} \left(\log S{EL} - \log( \mathbf{M}{\alpha}(\mathbf{x})_i + 1) \right)\right], \tag{1}
]
where the intended semantics are that (\mathbf{M}\alpha(\mathbf{x})i) is the evidence for class (i), the Dirichlet parameter is (\alpha_i = \mathbf{M}\alpha(\mathbf{x})i + 1), and (S{EL}) is the Dirichlet strength [2510.12687]. The notation is described in the source as somewhat nonstandard and slightly ambiguous, but the evidential interpretation is explicit.

For each sample, the evidential loss is recorded across epochs as
[
\mathbf{l} = [l_1,l_2,\dots,l_{N_e}],
]
and these trajectory vectors form the feature set for partitioning [2510.12687]. Clustering is performed separately within each source-domain and known-class subset:
[
\Omega_d^y = {\omega_1,\omega_2,\dots,\omega_{N_p}} \leftarrow \text{FINCH}(\mathcal{X}d^y), \tag{2}
]
[
\Omega = {\Omega_d^y \mid d\in\mathcal{D}_S,\; y\in\mathcal{Y}_k}. \tag{3}
]
Partition-level averages are then computed,
[
\hat{\mathcal{X}{d}^{y} = {\mu(\omega_1(\mathcal{X}{d}^{y})), \mu(\omega_2(\mathcal{X}{d}^{y})),\dots,\mu(\omega_{N_p}(\mathcal{X}_{d}^{y}))}, \tag{4}
]
and a 2-component GMM assigns lower-average-loss components to the clean group and higher-average-loss components to the noisy group [2510.12687].

The paper’s rationale is that embedding geometry can be unreliable under noisy labels, whereas training dynamics provide a stronger signal: clean samples generally fit earlier and show lower or more stable evidential losses, while noisy samples remain high-loss or erratic longer [2510.12687]. The per-domain, per-class grouping is intended to avoid confounding due to heterogeneous difficulty across domains and categories. A related but distinct theme appears in the dynamic evidential calibration literature, where reliability is also treated as a first-class objective, but there the emphasis is on meta-controlling the KL coefficient and class-specific Dirichlet prior during training rather than on clustering per-sample loss trajectories [2510.08938].

A plausible implication is that EReLiFM’s evidential reliability notion is more diagnostic than purely probabilistic: it uses evidence not only to quantify uncertainty in prediction, but also to identify which training samples are trustworthy enough to support later residual-flow learning.

4. Residual flow matching and structured augmentation

The paper defines Residual Flow Matching (RFM) as modeling residual distributions rather than direct sample generation [2510.12687]. Given a source sample (\mathbf{I}s), a target sample (\mathbf{I}_t), and a condition (\mathbf{q}), the target residual is
[
\mathbf{r}_1 = \mathbf{I}_t - \mathbf{I}_s \sim p_r^{(\mathbf{q})}.
]
Training uses the objective
[
\mathcal{L}{\mathrm{RFM} = \mathbb{E}{(\mathbf{q},\,\mathbf{r}_0,\,\mathbf{r}_1,\,t)} \big[ | f\theta(\mathbf{r}t, t, \psi(\mathbf{q})) - (\mathbf{r}_1 - \mathbf{r}_0)|_2^2 \big], \quad \mathbf{r}_t = (1-t)\mathbf{r}_0 + t\mathbf{r}_1, \tag{6}
]
with (\mathbf{r}_0 \sim \mathcal{N}(0,1)) and (t\sim \mathcal{U}(0,1)) [2510.12687]. At generation time, the model integrates
[
\frac{d\mathbf{r}}{dt}=f\theta(\mathbf{r},t,\psi(\mathbf{q})),
]
starting from Gaussian noise to produce a residual sample (\mathbf{r}), and then constructs the augmentation
[
\mathbf{I}_{aug}=\mathbf{I}_s+\mathbf{r}.
]
The paper interprets this as enabling “diverse and uncertainty-aware transfer paths” beyond interpolation-based augmentation [2510.12687].

The method distinguishes two conditional residual families. Domain residuals keep the category fixed while changing domain:
[
(\mathbf{y}c,\mathbf{y}_c),\; (\mathbf{y}_d,\hat{\mathbf{y}_d}),
]
whereas category residuals keep the domain fixed while changing category:
[
(\mathbf{y}_c,\hat{\mathbf{y}_c}),\; (\mathbf{y}_d,\mathbf{y}_d).
]
Algorithmically, domain residuals form (\mathbf{B}{dr}) and category residuals form (\mathbf{B}_{cr}) by adding sampled residuals to clean batches [2510.12687]. Category-residual samples are assigned an additional class (\mathbf{y}_a), intended to improve awareness of unknown or open classes [2510.12687].

This use of flows differs from FLoWN, where conditional flow matching is applied in latent space to generate task-specific neural network parameters for few-shot adaptation [2503.19371]. FLoWN learns a conditional ODE
[
\frac{d z_t}{dt} = v_\theta(z_t,t;c)
]
over latent encodings of weights, conditioned on support data, and is best understood as a latent conditional weight generator [2503.19371]. By contrast, EReLiFM’s equations are written in terms of image-like variables (\mathbf{I}_s,\mathbf{I}_t), so the residuals are modeled in sample space rather than weight space [2510.12687]. The connection is therefore conceptual rather than architectural: both use conditional transport, but only EReLiFM uses it to synthesize structured cross-domain and cross-category residuals for OSDG-NL.

5. Meta-learning formulation and optimization behavior

The meta-learning stage is organized around a clean/noisy split [2510.12687]. On clean and flow-augmented clean data, the meta-train loss is
[
\mathcal{L}{m\text{-}train} = \mathcal{L}{CE}(\mathbf{B}{clean}, \mathbf{y}_c) + \mathcal{L}{CE}(\mathbf{B}{dr}, \mathbf{y}{dr}) + \mathcal{L}{CE}(\mathbf{B}{cr}, \mathbf{y}_{cr}).
]
Here the original clean sample retains its known-class label, domain-residual augmentations keep the same category label, and category-residual augmentations are assigned to the additional class (\mathbf{y}_a) [2510.12687].

For noisy samples, the meta-test objective uses both pseudo-labels and original noisy labels:
[
\mathbf{y}{pseudo} = \arg\max(\mathbf{M}{\alpha}(\mathbf{B}{noisy})),
]
[
\mathcal{L}{m\text{-}test} = \mathcal{L}{EL}(\mathbf{B}{noisy}, \mathbf{y}{pseudo}) + \mathcal{L}{CE}(\mathbf{B}{noisy}, \mathbf{y}{nc}).
]
The final objective is
[
\mathcal{L}{total} = \mathcal{L}{m\text{-}train} + \mathcal{L}_{m\text{-}test}.
]
The paper does not provide explicit weighting coefficients between these losses; the faithful reading is that they are summed directly [2510.12687].

The evidential pseudo-label loss is justified as a correction mechanism, while the cross-entropy term on the original noisy labels is retained because some samples in the noisy split may in fact be clean but hard examples incorrectly partitioned by the filter stage [2510.12687]. The method therefore does not simply discard noisy data or replace them wholesale with pseudo-labels. Instead, it uses what the paper calls a “competition” between pseudo-label and original label to make training more robust to imperfect clean/noisy separation [2510.12687].

The abstract states that “the model is optimized such that the update direction on the clean set maximizes the loss decrease on the noisy set,” but the body does not provide a full second-order MAML-style derivation [2510.12687]. This has interpretive importance. EReLiFM is meta-learning in the sense of decoupled clean/noisy task roles inside a joint optimization framework, not in the strict sense of a formally written inner/outer-loop higher-order objective. This distinguishes it from both classical episodic meta-learning and from meta-controllers such as MPC, where a policy network explicitly learns how to configure evidential loss parameters over time through REINFORCE-style policy gradients [2510.08938].

A useful contrast also arises with Trust-SSL, where reliability enters learning through a stop-gradient additive residual term rather than multiplicative gating [2604.21349]. EReLiFM does not adopt that specific mechanism, but the comparison clarifies that “residual” can refer to different loci of intervention: EReLiFM uses residuals as learned transport objects for augmentation, whereas Trust-SSL uses residual reliability modulation in the loss [2604.21349].

6. Empirical performance, related methods, and interpretive issues

EReLiFM is evaluated on PACS, DigitsDG, and TerraInc, under symmetric label noise rates of (20\%), (50\%), and (80\%), and asymmetric noise at (50\%) generated according to semantic similarity using BERT textual features and cosine similarity [2510.12687]. The principal metrics are Acc, H-score, and OSCR, with
[
H_{score} = \frac{2 \times Acc_u \times Acc_k}{Acc_u + Acc_k}.
]
The paper emphasizes OSCR as the main threshold-free metric for open-set classification and rejection quality [2510.12687].

On PACS with ResNet18, compared with HyProMeta, EReLiFM improves average metrics by (+14.76\%) Acc, (+5.63\%) H-score, and (+13.70\%) OSCR at (20\%) symmetric noise, by (+11.56\%), (+4.68\%), and (+9.01\%) at (50\%) symmetric noise, by (+5.14\%), (+9.14\%), and (+4.93\%) at (80\%) symmetric noise, and by (+11.99\%), (+3.63\%), and (+8.20\%) at (50\%) asymmetric noise [2510.12687]. On PACS with ViT-Base, average OSCR increases from (54.97) to (61.24) at (20\%) symmetric noise, from (52.91) to (63.21) at (50\%), from (25.26) to (34.96) at (80\%), and from (43.44) to (50.60) at (50\%) asymmetric noise [2510.12687]. On DigitsDG with ConvNet, average OSCR rises from (55.34) to (61.34), (44.10) to (47.19), (10.63) to (11.57), and (48.79) to (50.97) across the same four noise settings [2510.12687].

Ablations attribute these gains to each major component. Removing DC-CRFM causes large OSCR drops on DigitsDG under (50\%) symmetric noise, for example from (69.88) to (58.91) on mnist and from (39.64) to (33.19) on syn [2510.12687]. Replacing DC-CRFM with MixUp lowers OSCR to (67.46) on mnist and (22.06) on syn, supporting the claim that residual flow augmentation is not equivalent to interpolation [2510.12687]. Removing UTS-ELC or replacing evidential loss clustering with non-evidential loss clustering also degrades performance, indicating that filtering before residual learning is important and that evidential trajectories are specifically useful [2510.12687]. In the meta-test stage, removing (\mathcal{L}{EL}) or removing (\mathcal{L}{CE}) both harms OSCR, showing that both pseudo-label correction and original-label retention contribute [2510.12687].

Several interpretive cautions follow from the source. Performance gains remain but are smaller under extreme noise, especially (80\%) symmetric noise [2510.12687]. The method is image-specific in its published form, and the mathematical treatment of meta-learning is more procedural than formally bilevel [2510.12687]. The phrase “residual flow” should also be understood in the paper’s own sense—learning conditional distributions over residuals (\mathbf{r}=\mathbf{I}_t-\mathbf{I}_s)—rather than as a claim about invertible residual-flow architectures or likelihood-based normalizing flows [2510.12687]. By comparison, FLoWN supplies a clearer flow-matching formalism in latent parameter space [2503.19371], and Trust-SSL supplies a clear design principle for reliability-aware residual modulation, but neither paper addresses OSDG-NL directly [2503.19371][2604.21349].

Within the evidential literature, EReLiFM’s reliability notion is closer to sample trustworthiness under training dynamics than to the dynamic calibration objective of MPC [2510.08938], and closer to clean/noisy partition quality plus pseudo-label correction than to the weakly supervised bag-instance uncertainty estimation of MIREL, which derives a residual operator
[
R(\mathbf{x})=T(\mathbf{x})+r_\pi(f_\psi(\mathbf{x}))
]
for Multi-Instance Uncertainty Estimation under weak supervision [2405.04405]. A plausible implication is that these related works illuminate different parts of the EReLiFM name: MPC sharpens the “reliability-aware” and dynamic evidential aspects, FLoWN sharpens the flow-matching viewpoint, Trust-SSL clarifies residual reliability modulation, and MIREL shows how residual evidential correction can be structurally meaningful even without flow or meta-learning [2510.08938][2503.19371][2604.21349][2405.04405].

EReLiFM’s central contribution is therefore not a single new module but a coordinated sequence: evidential reliability estimation to select trustworthy supervision, conditional residual flow matching to enlarge clean supervision with structured domain and category variation, and meta-learning-style reuse of noisy data through pseudo-labels and original-label regularization [2510.12687]. Empirically, the paper reports that this combination yields state-of-the-art performance on OSDG-NL benchmarks and especially strong gains in OSCR [2510.12687].