Iterative Contrastive Refinement (ICR)

• Iterative Contrastive Refinement (ICR) is a strategy that repeatedly updates contrastive signals—such as similarity distributions, labels, or embeddings—to enhance model performance.
• ICR leverages techniques like teacher-student distillation, prototype updating, and interactive feedback to mitigate noisy supervision and improve learning reliability.
• Applications of ICR span self-supervised learning, cross-modal alignment, and community detection, addressing the limitations of fixed one-shot contrastive methods.

Iterative Contrastive Refinement (ICR) denotes a class of procedures in which contrastive supervision is repeatedly updated rather than fixed as a one-shot positive/negative assignment. Across the cited literature, the refined object may be a similarity distribution, a soft label, a sequence embedding, a cross-modal alignment, a prototype, a community assignment, or an explicit intent state. What remains consistent is the recursive dependency: a current representation, label, or memory is used to define a better contrastive signal, and that improved signal is then used to update the next representation or decision state (Tejankar et al., 2020, Tsai et al., 29 Sep 2025).

1. Conceptual basis

Standard contrastive learning, as described in self-supervised vision work, treats similarity as a binary distinction between positive pairs and negative pairs. ISD argues that not all random images are equal and replaces the binary treatment of negatives with soft similarity over anchors, while the self-labeling refinement literature shows theoretically that inaccurate label assignment heavily impairs generalization for semantic instance discrimination and that more accurate labels benefit generalization (Tejankar et al., 2020, Zhou et al., 2021).

Representative ICR formulations differ mainly in what is iteratively refined.

Refined object	Refinement mechanism	Representative work
Similarity distribution	Teacher-student distillation with $KL(p_t \| p_s)$	ISD (Tejankar et al., 2020)
Stepwise sentence embedding	Stepwise contrastive loss plus refinement regularization	GIRCSE (Tsai et al., 29 Sep 2025)
Self-labels	Self-Labeling Refinery and Momentum Mixup	(Zhou et al., 2021)
Cross-modal alignment	Recurrent Alignment with layerwise contrastive losses	ICAF (Zhang et al., 2021)
Class prototypes	Moving-average prototype updates and pseudo-labeling	IPR (Sun et al., 2024)
Intent memory	Positive/negative memory with contrastive alignment	IntRec (Shamsolmoali et al., 19 Feb 2026)

This suggests that ICR is better understood as a design pattern than as a single canonical loss. In some papers it is an explicit objective, while in others it is the operative principle behind iterative distillation, pseudo-label correction, recurrent alignment, or interactive disambiguation.

2. Refining similarity and labels in self-supervised representation learning

ISD, introduced as "Self-Supervised Learning by Iterative Similarity Distillation," departs from standard contrastive learning by replacing hard positive/negative supervision with a teacher-defined soft similarity distribution over random anchors. The teacher evolves as an exponential moving average of the student,

$\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$

and the student is trained to match the teacher’s anchor-similarity distribution with

$L = KL(p_t \,\|\, p_s).$

The anchor embeddings are always from the teacher, a MoCo-like memory bank provides a large set of anchors, and no direct positive pair loss is used because the query image is not compared to its other augmentation as an anchor. The method is presented as less constrained than recent contrastive learning methods and is reported to perform better when the unlabeled data is unbalanced, precisely because semantically similar items in the negative set are assigned high similarity rather than being forced into the negative class (Tejankar et al., 2020).

The self-labeling refinement method of (Zhou et al., 2021) formalizes a closely related idea at the level of label quality. Its Self-Labeling Refinery (SLR) estimates semantic similarity between a query and all keys, combines that estimate with the vanilla one-hot label, and iteratively generates more accurate and informative soft labels:

$$\ell_i^{(t)} = (1-\alpha_1^{(t)}-\alpha_2^{(t)})y_i + \alpha_1^{(t)}p_i^{(t)} + \alpha_2^{(t)}q_i^{(t)}.$$

A second module, Momentum Mixup (MM), randomly combines queries and positives to increase semantic similarity between the generated virtual queries and their positives. The paper states that generalization error in contrastive learning linearly depends on discrepancy between assigned labels and true semantic labels, and further claims that SLR can exactly recover true semantic labels of label-corrupted data and supervise networks to achieve zero prediction error on classification tasks under its assumptions (Zhou et al., 2021).

Taken together, these methods instantiate ICR as a replacement for static one-hot supervision. The iterative step is not merely repeated optimization; it is repeated relabeling of similarity itself.

3. Stepwise objectives in generative and multimodal encoders

GIRCSE makes ICR an explicit training objective for generative sentence embeddings. Instead of supervising only a final embedding, it autoregressively generates $K$ soft tokens and pools the first $k$ generated token representations to obtain

$$\mathbf{z}_k = \frac{1}{k}\sum_{i=1}^k \mathbf{g}_i.$$

A contrastive loss is applied at every generation step, and a refinement regularizer penalizes steps where the loss worsens:

$$\mathcal{L}_{\text{reg}} = \frac{1}{K-1}\sum_{k=1}^{K-1}\max\left(\log \mathcal{L}_{k+1}-\log \mathcal{L}_k,\,0\right),$$

with total objective

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{contrast}} + \lambda \mathcal{L}_{\text{reg}}.$$

The paper reports that GIRCSE outperforms strong LLM-based embedding baselines on MTEB and instruction-following tasks, achieves top-5 on MTEB, top-2 on instruction following, and overall best rankings, and exhibits an emergent test-time scaling property: generating more tokens at inference steadily improves embedding quality. Figure evidence is described as showing monotonic improvements for $K=1,3,5,10,15,20$ (Tsai et al., 29 Sep 2025).

ICAF applies the same logic to multimodal abstractive summarization. Its Recurrent Alignment layer uses Cross-modal Attention and a gated Renovation Addition Module to gradually investigate fine-grained semantical relationships between image patches and text tokens across $\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$0 layers. At each step during encoding, cross-modal contrastive losses are applied directly in both directions, $\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$1 and $\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$2, and the overall objective combines generation and contrastive terms:

$\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$3

On MSMO, the reported scores for ICAF are ROUGE-1 56.11, ROUGE-2 36.97, ROUGE-L 49.71, relevance 0.452, and human evaluation 3.52 / 3.61, outperforming the stated baselines (Zhang et al., 2021).

In these models, iteration occurs inside the representation-building process itself. Refinement is stepwise in generation length or encoder depth, not only across training epochs.

4. Prototypes, graphs, and community structure

Iterative Prototype Refinement (IPR) for ambiguous speech emotion recognition uses a small precise-labeled set to initialize class prototypes,

$\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$4

then introduces ambiguous and unlabeled samples after a warm-up phase. For each ambiguous embedding $\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$5, similarity to each prototype is computed as

$\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$6

a pseudo-label is assigned by

$\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$7

and the winning prototype is updated by a moving average,

$\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$8

provided that the pseudo-labels for the sample and its augmentation agree. Contrastive learning then treats pairs with matching pseudo-labels as positives. On IEMOCAP, IPR achieves an Acc of 70.75% on D3, compared with 64.25% for the baseline, 66.85% for the baseline with pseudo-labels, 67.0% for MCIL, and 72.3% for the fully supervised upper bound (Sun et al., 2024).

RoGRAD extends the refinement loop to deficient text-attributed graphs. Its Semantic-Guided Retrieval-Augmented Generation and Graph Enrichment module iteratively produces drafts, retrieves same-class exemplars, diagnoses redundancy and off-category drift, and revises the generated text in a retrieval–diagnosis–revision loop. Its Retrieval-Refined Iterative Contrastive Learning module then constructs two graph views and applies a supervised contrastive loss over same-class positives. The process is rerun at fixed intervals, updating the embedding store and regenerating refined graph views. The paper presents RoGRAD as the first iterative paradigm for LLM-on-graph augmentation and reports up to 82.43% average improvement over conventional GNN- and LLM-enhanced baselines (Wang et al., 2 Oct 2025).

ReCon is a model-agnostic post-processing framework for signed community detection that repeats four steps: structural refinement, boundary refinement, contrastive learning, and clustering. Structural refinement combines a neighborhood score and a community score; boundary refinement uses unbalanced triangles or purge likelihood; contrastive learning applies node-level and community-level losses across augmented views,

$\theta_t \leftarrow m \theta_t + (1-m)\theta_s,$9

Ablation results reported in the paper show that using all three refinement steps gives the best detection accuracy, while contrastive learning alone can degrade performance because it may reinforce incorrect clusters from noisy input. Across eighteen synthetic and four real-world networks using four community detection methods, ReCon is reported to consistently improve ARI and modularity (Shin et al., 23 Jan 2026).

These variants make the contrastive signal depend on intermediate discrete structure—prototypes, graph augmentations, or community assignments—rather than on a fixed minibatch definition alone.

5. Interactive and human-guided refinement

IntRec formulates ICR as interactive object retrieval. Its Intent State at turn $L = KL(p_t \,\|\, p_s).$0 is

$L = KL(p_t \,\|\, p_s).$1

where $L = KL(p_t \,\|\, p_s).$2 stores positive anchors and $L = KL(p_t \,\|\, p_s).$3 stores negative constraints. Candidate regions are ranked by a contrastive alignment function,

$L = KL(p_t \,\|\, p_s).$4

This replaces one-shot retrieval with a loop of retrieval, user confirmation or rejection, Intent State update, and reranking. On LVIS, IntRec achieves 35.4 AP, outperforming OVMR, CoDet, and CAKE by +2.3, +3.7, and +0.5, respectively. On LVIS-Ambiguous, it improves performance by +7.9 AP over its one-shot baseline after a single corrective feedback, with less than 30 ms of added latency per interaction (Shamsolmoali et al., 19 Feb 2026).

Schemex moves the refinement loop into interactive schema induction. Its workflow has three AI-assisted stages—clustering, abstraction, and contrastive refinement—and the third stage explicitly compares schema-guided outputs against gold examples. The contrastive feedback interface proposes revisions with tags such as [ADD], [DEEPEN], [REFINE], and [RESTRUCTURE], after which the schema is updated and reapplied. Expert ratings on a 7-point scale report improvements from o1-pro to Schemex Full on cluster quality (5.1 to 6.4), dimension/attribute quality (5.0 to 6.8), example fit (5.6 to 6.6), usefulness (5.3 to 6.7), output alignment (5.1 to 6.4), and effectiveness of generations (5.5 to 6.3). The paper also notes a trade-off: contrastive refinement increases precision and depth, sometimes at the cost of generalizability (Wang et al., 16 Apr 2025).

A related human-in-the-loop formulation appears in small-sample adaptive text classification through iterative and contrastive description refinement. Starting from as few as 20 labeled samples per class, the method generates natural-language category descriptions, validates them with additional samples, and triggers refinement when accuracy drops below a threshold such as <80%. Iterations typically stop after target accuracy is exceeded or after 4 rounds, beyond which improvements plateau. Contrastive prompting explicitly teaches the model to differentiate closely related classes, and new classes can be introduced by supplying representative samples and a natural-language description, without retraining. Reported results are 91.1% on AGNews (seen), 83% on AGNews (unseen), 84% on DBPedia (seen), and 87% on DBPedia (unseen) (Rajeev et al., 1 Aug 2025).

In these systems, ICR is not limited to representation learning in the narrow sense. It also serves as a mechanism for interactive disambiguation, schema revision, and dynamic category definition.

6. Empirical regularities, limitations, and terminological overlap

A recurring empirical theme is that iterative refinement is most valuable where the original contrastive supervision is unreliable: false negatives in self-supervised vision, ambiguous labels in speech emotion recognition, missing or noisy structure in graphs and signed networks, and user ambiguity in open-vocabulary retrieval. ISD emphasizes unbalanced unlabeled data, IPR emphasizes blended emotional states, ReCon emphasizes noisy or conflicting edge signs, and IntRec emphasizes complex scenes with multiple similar objects (Tejankar et al., 2020, Sun et al., 2024, Shin et al., 23 Jan 2026, Shamsolmoali et al., 19 Feb 2026).

The same literature also records important failure modes. ISD avoids direct positive pair loss because including the query itself as an anchor would create degenerate distributions; GIRCSE states that naively increasing the number of refinement steps can yield redundant or stagnant embeddings; ReCon reports that contrastive learning alone can degrade performance; and Schemex observes that additional refinement may narrow the schema toward the training examples (Tejankar et al., 2020, Tsai et al., 29 Sep 2025, Shin et al., 23 Jan 2026, Wang et al., 16 Apr 2025). This suggests that ICR is beneficial when refinement is coupled to a stabilizing mechanism such as a momentum teacher, prototype warm-up, structural correction, retrieval grounding, or user feedback.

The acronym “ICR” is not unique to iterative contrastive refinement. "Towards Generalizable Implicit In-Context Learning with Attention Routing" introduces In-Context Routing, also abbreviated ICR, an unrelated method that biases attention logits using Principal ICL Directions and an input-conditioned router (Li et al., 26 Sep 2025). Likewise, "Reducing Predictive Feature Suppression in Resource-Constrained Contrastive Image-Caption Retrieval" uses “ICR methods” to mean image-caption retrieval, and its main contribution is Latent Target Decoding as a constraint-based auxiliary objective for contrastive retrieval rather than an iterative refinement loop (Bleeker et al., 2022).

Across the cited work, the most defensible generalization is therefore limited but clear: ICR refers not to a single universally adopted algorithm, but to a recurrent strategy in which contrastive relations are revised in light of improved representations, auxiliary structure, or interaction. Under that interpretation, the literature spans self-distillation, self-labeling, generative embedding, recurrent cross-modal alignment, prototype updating, retrieval-augmented graph learning, community post-processing, and interactive retrieval.