Propose-Rectify Framework Overview

• The Propose-Rectify Framework is a design pattern that separates the generation of candidate hypotheses from their subsequent corrective update using diagnostic evidence and optimization constraints.
• It is applied across domains such as autoformalization, multi-view clustering, and multi-agent systems through mechanisms including probabilistic modeling, chain-of-thought revisions, and targeted editing.
• Empirical results demonstrate notable accuracy gains and noise robustness, although challenges remain regarding initial proposal quality and context-specific limitations.

Across recent literature, the “Propose-Rectify Framework” denotes a recurrent design pattern in which a system first produces a candidate object or hypothesis and then modifies that candidate using diagnostic evidence, structural priors, or explicit optimization constraints. The candidate may be a noisy-sample assignment in multi-view clustering, a Lean 4 formalization, a rectifying vector for noisy-label learning, a chain-of-thought solution, a CLIP logit map, a trajectory rollout, a diffusion sampling direction, a parking-spot alignment, an inter-agent message, or a manipulation-localization proposal; rectification then acts by reweighting, editing, pruning, projecting, or re-optimizing that proposal rather than treating the initial output as final (Yang et al., 27 May 2025, Wang et al., 6 Jul 2026, Sun et al., 2021, Xue et al., 2023, Yadav et al., 24 Mar 2026, Zhang et al., 25 Aug 2025). In this sense, the framework is not a single algorithm but a family of architectures and optimization schemes that separate hypothesis formation from corrective update.

1. Core abstraction and recurrent loop structure

A common formal pattern is an iterative map of the form “propose, diagnose, rectify, repeat.” In autoformalization, an autoformalizer proposes a Lean 4 statement $T_0$, a diagnostic model computes a structured diagnosis $D(S,T)=\langle v,c,\ell,T'\rangle$, and a rectification operator returns $T_{k+1}=R(T_k,D(S,T_k))$, with stopping when the verdict is Aligned or a budget is exhausted (Wang et al., 6 Jul 2026). In reflective visual generation, the loop is written as

$$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$

where the proposal is an initial image, reflection produces a verdict, explanation, and action draft, and rectification is an editing update (Wang et al., 19 May 2026).

The same decomposition appears in non-language settings. In robust deep contrastive multi-view clustering, AIRMVC operationalizes a Propose-Rectify paradigm in which “Propose/Identify” detects anomalous samples with a Gaussian mixture model on learned embeddings and soft cluster predictions, while “Rectify” uses hybrid cross-view correction and a noise-robust contrastive mechanism (Yang et al., 27 May 2025). In systems with linear and complementarity constraints, ReCR first solves a relaxation and then rectifies the objective or linearization using residual information, explicitly casting feasibility search as a relaxation-rectification loop rather than a one-shot feasibility solve (AlAli et al., 26 Jun 2026).

A useful distinction in this literature is that “proposal” need not mean “final answer candidate.” In AIRMVC it is an anomaly-identification stage; in ReCLIP++ it is bias extraction via a learnable Reference prompt and positional projection; in trajectory prediction it is a multi-modal rollout at the current replanning step; in AgentDropoutV2 it is a message emitted by an agent inside a multi-agent system (Wang et al., 2024, Yadav et al., 24 Mar 2026, Wang et al., 26 Feb 2026). This broader usage indicates that Propose-Rectify is a control structure over intermediate representations as much as over end-task outputs.

2. Proposal mechanisms

Proposal stages differ sharply by domain, but they share a preference for inexpensive, structured, or weakly constrained candidate generation.

In AIRMVC, each view $v$ has an encoder $F^v(\cdot;\Theta^v)$ that produces latent embeddings $E_i^v$, and a classifier head $h(\cdot)$ that yields view-wise soft cluster assignments $y_i^v\in\Delta^{K-1}$. The proposal stage then reformulates noisy identification as anomaly identification using a $K$-component GMM over embeddings, with likelihood

$D(S,T)=\langle v,c,\ell,T'\rangle$0

anomaly score $D(S,T)=\langle v,c,\ell,T'\rangle$1, and a two-component GMM over agreement scores $D(S,T)=\langle v,c,\ell,T'\rangle$2 to obtain clean probabilities $D(S,T)=\langle v,c,\ell,T'\rangle$3 (Yang et al., 27 May 2025). The proposal is therefore probabilistic and view-specific.

In WarPI, the proposal is a latent rectifier rather than a label. A meta-network $D(S,T)=\langle v,c,\ell,T'\rangle$4 takes logits $D(S,T)=\langle v,c,\ell,T'\rangle$5 and a possibly noisy label $D(S,T)=\langle v,c,\ell,T'\rangle$6, and proposes a factorized Gaussian $D(S,T)=\langle v,c,\ell,T'\rangle$7 over a per-sample, per-class rectifying vector $D(S,T)=\langle v,c,\ell,T'\rangle$8. The paper’s formulation is explicitly hierarchical and probabilistic:

$D(S,T)=\langle v,c,\ell,T'\rangle$9

with $T_{k+1}=R(T_k,D(S,T_k))$0, $T_{k+1}=R(T_k,D(S,T_k))$1 (Sun et al., 2021). Here the proposal stage is designed to expose ambiguity rather than suppress it.

In RCoT, the proposal is a chain-of-thought solution $T_{k+1}=R(T_k,D(S,T_k))$2 to a problem $T_{k+1}=R(T_k,D(S,T_k))$3. Detection then reverses this solution to reconstruct a problem $T_{k+1}=R(T_k,D(S,T_k))$4, making factual inconsistency measurable in terms of overlooked conditions, hallucinated conditions, and question misinterpretation (Xue et al., 2023). In FormalRx, the proposal is $T_{k+1}=R(T_k,D(S,T_k))$5, a candidate Lean 4 translation of an informal mathematical statement $T_{k+1}=R(T_k,D(S,T_k))$6; the framework deliberately separates that step from later semantic diagnosis (Wang et al., 6 Jul 2026).

Proposal can also be architectural. ReCLIP++ proposes a bias logit map $T_{k+1}=R(T_k,D(S,T_k))$7 by independently encoding class-preference bias with a learnable “Reference” prompt and space-preference bias with a projection of CLIP positional embeddings, then combining them by matrix multiplication (Wang et al., 2024). TODSynth proposes mask-conditioned remote-sensing images with a Multimodal Diffusion Transformer using unified triple attention over text, image latent, and mask tokens (Yang et al., 18 Dec 2025). In autonomous driving, LMFormer-D proposes a multi-modal set of trajectory hypotheses $T_{k+1}=R(T_k,D(S,T_k))$8 at each closed-loop step (Yadav et al., 24 Mar 2026).

3. Rectification operators

Rectification stages are considerably more heterogeneous than proposal stages. They can be grouped into four broad types: confidence-weighted correction, edit or patch application, optimization-state correction, and reject-or-prune mechanisms.

Confidence-weighted correction is exemplified by AIRMVC. For $T_{k+1}=R(T_k,D(S,T_k))$9, it constructs mixed soft targets

$$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$0

where the first view is assumed relatively clean and serves as an anchor. The corresponding rectification loss is a cross-entropy over these rectified targets, and a $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$1-gated contrastive loss excludes low-confidence cross-view pairs (Yang et al., 27 May 2025). WarPI performs a different confidence-sensitive correction: the sampled rectifying vector $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$2 multiplicatively warps logits, so the classifier is trained with $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$3, thereby reshaping gradients away from noisy labels (Sun et al., 2021).

Edit or patch application is central in FormalRx, RCoT, and reflective visual generation. FormalRx’s diagnostics return not just a verdict and error category but also a minimal code span $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$4 and a corrected statement $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$5, so rectification can be either direct replacement by $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$6 or a patch derived from $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$7 (Wang et al., 6 Jul 2026). RCoT converts inconsistencies between $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$8 and $$R_t=\pi_\theta(P,\mathbf{I}^{(t)})=\langle v_t,e_t,a_t\rangle,\qquad \mathbf{I}^{(t+1)}=\Phi(\mathbf{I}^{(t)},\langle e_t,a_t\rangle),$$9 into fine-grained feedback—such as “You have ignored some real conditions” or “You use some wrong candidate conditions”—and prompts a revision of the original solution (Xue et al., 2023). In R$v$0-Refiner, rectification is an executable text edit instruction $v$1 applied by an external editor $v$2 to produce $v$3 (Wang et al., 19 May 2026).

Optimization-state correction appears in trajectory prediction, controllable synthesis, and mathematical programming. “Rectify, Don’t Regret” replaces fully differentiable simulation with a detached receding horizon rollout, applying stop-gradient at every step boundary so that the model must recover from drifted states instead of exploiting non-causal gradients flowing through induced simulator states (Yadav et al., 24 Mar 2026). TODSynth’s Control-Rectify Flow Matching modifies the predicted diffusion velocity field during early high-plasticity steps:

$v$4

with the correction derived from downstream semantic segmentation loss (Yang et al., 18 Dec 2025). ReCR updates relaxation objectives via loss-like rectification terms such as $v$5, or linearizes the concave part of a quadratic complementarity reformulation and resolves a convex subproblem (AlAli et al., 26 Jun 2026).

Reject-or-prune rectification is explicit in AgentDropoutV2. Each agent output is intercepted, checked against retrieved failure indicators, iteratively regenerated if violations are detected, and pruned if no correction passes within a preset budget $v$6. A fallback reset is triggered when the count of valid messages falls below a safety threshold $v$7 (Wang et al., 26 Feb 2026). This is a stricter form of rectification than editing: irreparable candidates are removed from the information flow entirely.

4. Formal semantics, optimization, and guarantees

Theoretical treatment of Propose-Rectify systems is domain-specific, but several papers make the structure mathematically explicit.

FormalRx defines semantic alignment using interpretation functions $v$8 and $v$9, with correctness given by $F^v(\cdot;\Theta^v)$0 iff $F^v(\cdot;\Theta^v)$1. It then operationalizes diagnosis through four outputs—verdict, category, location, correction—and resolves ambiguous error labels through the SCI taxonomy’s strict priority ordering $F^v(\cdot;\Theta^v)$2-type-specific $F^v(\cdot;\Theta^v)$3-location-based (Wang et al., 6 Jul 2026). The framework therefore formalizes rectification not as heuristic post-processing but as a structured approximation to semantic proof obligations.

AIRMVC provides a mutual-information theorem for robust representation learning. Under assumptions involving clean soft predictions $F^v(\cdot;\Theta^v)$4, noisy predictions $F^v(\cdot;\Theta^v)$5, and representations $F^v(\cdot;\Theta^v)$6, it proves

$F^v(\cdot;\Theta^v)$7

which the paper interprets as retaining clean information while discarding noisy information (Yang et al., 27 May 2025). This places the proposal-rectification loop inside an information-theoretic robustness argument.

ReCR supplies the most explicit optimization-theoretic foundation. Its Universal Relaxation Theory states that feasibility of systems with linear and complementarity constraints can be reformulated as a bilinear optimization problem in a lifted space, and that a feasible solution exists iff there is a cost vector for which a primal-dual pair of optimal solutions satisfies a componentwise dominance relation. The resulting Relaxation-Rectification algorithms are then analyzed in terms of boundedness, stationarity, and infeasibility certification (AlAli et al., 26 Jun 2026). In that setting, rectification is not metaphorical; it is a rigorously defined update rule on objective parameters or convexified subproblems.

In trajectory prediction, the theoretical contribution is causal rather than variational. By inserting stop-gradient between simulation steps, the method enforces

$F^v(\cdot;\Theta^v)$8

in the backward graph, eliminating the gradient leakage that otherwise allows future ground-truth signals to revise earlier predictions non-causally (Yadav et al., 24 Mar 2026). This establishes a rectification principle based on causal training signals.

5. Representative instantiations across domains

The framework has been instantiated across clustering, formal reasoning, segmentation, diffusion sampling, multi-agent systems, optimization, and digital forensics.

Paper	Domain	Propose / Rectify instantiation
AIRMVC (Yang et al., 27 May 2025)	Noisy multi-view clustering	GMM-based anomaly identification / anchor-view interpolation and gated contrastive learning
FormalRx (Wang et al., 6 Jul 2026)	Autoformalization	Lean 4 proposal / verdict, category, location, correction loop
WarPI (Sun et al., 2021)	Noisy-label learning	Probabilistic rectifier proposal / logit warping in meta-learning
RCoT (Xue et al., 2023)	Arithmetic reasoning	CoT solution / reverse reconstruction and targeted feedback
ReCLIP++ (Wang et al., 2024)	Unsupervised semantic segmentation	Bias extraction / logit subtraction and decoder refinement
TODSynth (Yang et al., 18 Dec 2025)	Remote-sensing data synthesis	MM-DiT generation / CRFM velocity correction
AgentDropoutV2 (Wang et al., 26 Feb 2026)	Multi-agent systems	Agent message proposal / rectify-or-reject pruning
ReCR (AlAli et al., 26 Jun 2026)	Complementarity systems	Relaxation proposal / residual-driven rectification

Additional variants extend the pattern further. LLM-TPC uses a Think-Program-reCtify loop for 3D situated reasoning: the LLM decomposes a question, grounds it into executable Python over 3D perception APIs, and revises the plan and code when execution fails (He et al., 2024). “Rectify and Align GPS Points to Parking Spots via Rank-1 Constraint” first proposes candidate parking spot coordinates from road geometry and then jointly rectifies and aligns corrupted GPS points under a rank-1 low-rank prior using ADMM (Deng et al., 15 Oct 2025). In image manipulation localization, a forensic-adapted LLaVA produces initial analysis and coarse suspicious regions, which are then rectified by a Forensics Rectification Module using SRM, Bayar, Sobel, and Noiseprint++ features, followed by forensic-aware SAM embedding enhancement (Zhang et al., 25 Aug 2025). In open-vocabulary remote sensing segmentation, DR-Seg decouples CLIP channels into semantics-dominated and structure-dominated subspaces, rectifies only the structure-dominated branch with a prior-driven graph under DINO guidance, and fuses it back using uncertainty-guided adaptive fusion (Feng et al., 2 Apr 2026).

A recurring misconception is that Propose-Rectify is inherently tied to language-model self-correction. The cited work shows otherwise: in ReCLIP++ it is explicit bias modeling and subtraction; in AIRMVC it is confidence-weighted target interpolation; in ReCR it is relaxation plus objective rectification; in trajectory prediction it is detached receding-horizon replanning (Wang et al., 2024, Yang et al., 27 May 2025, AlAli et al., 26 Jun 2026, Yadav et al., 24 Mar 2026).

6. Empirical behavior, limitations, and broader significance

Empirical evidence across these papers is consistently favorable, though the gains are mechanism-dependent. AIRMVC outperforms state-of-the-art multi-view clustering baselines on six benchmark datasets under noise ratios from $F^v(\cdot;\Theta^v)$9 to $E_i^v$0, with examples including $E_i^v$1 ACC and $E_i^v$2 NMI on UCI-digit at $E_i^v$3 noise (Yang et al., 27 May 2025). FormalRx-8B achieves F1-scores of $E_i^v$4 for verdict and $E_i^v$5 for categorization, with localization accuracy $E_i^v$6 and correction accuracy $E_i^v$7, and its structured feedback improves Accumulated Pass@8 from $E_i^v$8 to $E_i^v$9 in a 3-round test-time scaling study (Wang et al., 6 Jul 2026). AgentDropoutV2 reports an average accuracy gain of $h(\cdot)$0 percentage points on math benchmarks (Wang et al., 26 Feb 2026). R$h(\cdot)$1-Refiner improves Reflective Verdict Score by $h(\cdot)$2 and Rectification Score by $h(\cdot)$3 on R$h(\cdot)$4-Bench (Wang et al., 19 May 2026). In autonomous driving, detached receding horizon rollout reduces target collisions by up to $h(\cdot)$5 relative to fully differentiable closed-loop training at high replanning frequencies (Yadav et al., 24 Mar 2026).

The reported limitations are equally consistent. AIRMVC can suffer from GMM misspecification and from failure of its clean-anchor assumption when the first view is corrupted (Yang et al., 27 May 2025). FormalRx notes that SCI is Lean 4-specific in its implementation dimension and that ambiguous natural language can still confound categorization and correction (Wang et al., 6 Jul 2026). RCoT is less effective for purely computational errors because its reverse-chain mechanism targets factual inconsistency rather than arithmetic execution (Xue et al., 2023). TODSynth depends on the quality of the external segmentation model used for guidance, and over-applying CRFM can trigger adversarial-like artifacts or mode collapse (Yang et al., 18 Dec 2025). AgentDropoutV2 can over-correct when too many indicators are retrieved or too many rectification rounds are allowed (Wang et al., 26 Feb 2026). ReCR’s infeasibility certification may become expensive because its branching procedure can be exponential in the size of the active violation set (AlAli et al., 26 Jun 2026).

A broader implication, suggested repeatedly in the literature, is that proposal quality and rectification quality are separable design axes. FormalRx explicitly argues that evaluation becomes actionable only when verdicts are accompanied by category, location, and correction (Wang et al., 6 Jul 2026). AIRMVC shows that direct fusion of noisy views can underperform even a single clean view, whereas hybrid rectification preserves complementary information while filtering noise (Yang et al., 27 May 2025). ReCLIP++ finds that prompt engineering alone is insufficient; bias must be explicitly modeled and subtracted (Wang et al., 2024). The recurrent lesson is that correction is most effective when it is grounded in domain-specific evidence rather than generic self-revision.

Taken together, these works present the Propose-Rectify Framework as a reusable systems principle: first externalize a tentative structure, then constrain, filter, or revise it using evidence that the proposal stage alone cannot reliably encode. In some domains the evidence is probabilistic, in others symbolic, geometric, causal, or forensic; the framework’s unifying feature is the explicit separation between hypothesis formation and corrective enforcement (Yang et al., 27 May 2025, Wang et al., 6 Jul 2026, AlAli et al., 26 Jun 2026, Zhang et al., 25 Aug 2025).