RAD-DPO: Dual Domain Methods

• RAD-DPO is an acronym representing two distinct technical constructs in planetary radiation science and e-commerce retrieval.
• The RAD Dose Prediction Operator uses empirical fits from Martian RAD data to predict surface dose rates based on atmospheric pressure and heliospheric modulation.
• Robust Adaptive Denoising Direct Preference Optimization enhances generative retrieval by mitigating gradient conflicts, pseudo-negative issues, and multi-label probability squeezing.

Searching arXiv for both uses of “RAD-DPO” to ground the article in the cited papers. RAD-DPO is an acronym with two unrelated technical uses in arXiv literature. In planetary radiation science, it denotes the RAD Dose Prediction Operator, an empirically calibrated operator derived from Mars Science Laboratory Radiation Assessment Detector measurements to predict galactic-cosmic-ray-induced surface dose rate and dose equivalent on Mars as functions of atmospheric pressure and heliospheric modulation (Guo et al., 2015). In recommender and retrieval modeling, it denotes Robust Adaptive Denoising Direct Preference Optimization, a preference-optimization objective for generative retrieval in e-commerce that modifies DPO to accommodate hierarchical Semantic IDs, noisy implicit feedback, and multi-label relevance (Chen et al., 27 Feb 2026). The shared acronym masks a complete separation of domain, objective, mathematical structure, and operational interpretation.

1. Terminological scope and disambiguation

The ambiguity of RAD-DPO is substantive rather than notational. In the Martian-radiation setting, “RAD” refers to the Radiation Assessment Detector on board Curiosity, and “DPO” refers to a Dose Prediction Operator fitted from one-sol-averaged surface data. In the e-commerce retrieval setting, “RAD-DPO” expands to Robust Adaptive Denoising Direct Preference Optimization, where “RAD” characterizes a set of denoising and robustness mechanisms layered onto DPO for alignment of generative retrieval models (Guo et al., 2015, Chen et al., 27 Feb 2026).

A common misconception is that RAD-DPO names a single optimization framework with cross-domain reuse. The literature in fact uses the acronym for two independent constructs. One is a radiation-environment estimator operating on physical variables such as pressure $P$ and modulation potential $\Phi$; the other is a training objective operating on queries $x$, positives $\mathcal{Y}_{\rm pos}$, negatives $\mathcal{Y}_{\rm neg}$, and sequence probabilities under $\pi_\theta$. The only commonality is acronymic coincidence.

2. The RAD Dose Prediction Operator for the Martian surface

In Guo et al., the RAD Dose Prediction Operator is introduced as an empirical representation of how Martian-surface radiation dose varies with atmospheric and heliospheric conditions during the first MSL Martian year. The underlying physical definition of the surface dose rate is written as

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$

where $j$ labels particle species, $E$ is particle kinetic energy, $\epsilon$ the energy deposited in the detector, $\Phi$0 the surface spectral fluence at atmospheric pressure $\Phi$1 and heliospheric modulation $\Phi$2, $\Phi$3 the detector yield, and $\Phi$4 the detector mass (Guo et al., 2015).

Within the range of the first MSL Martian year, the two dominant, independent drivers of $\Phi$5 were found to be the surface pressure $\Phi$6 and the modulation potential $\Phi$7. Two empirical forms were fitted from one-sol-averaged RAD data:

$\Phi$8

for Model I, and

$\Phi$9

for Model II. Here $x$0 is in Pa, $x$1 in MV, and $x$2 in $x$3; the fitted parameters are detector-specific. The paper attributes short-term diurnal variations to daily thermal tides, long-term seasonal changes to Martian atmospheric pressure changes, and long-term solar-cycle and solar-rotation effects to heliospheric modulation of the primary GCR flux (Guo et al., 2015).

This operator is explicitly empirical. It does not replace transport modeling; rather, it quantitatively demonstrates how long-term influences of pressure and solar modulation are related to measured dose rates, enabling predictions of Martian surface radiation environment under different solar modulations and atmospheric conditions.

3. Calibration, fitted coefficients, corrections, and validity limits

All fits for the Martian RAD Dose Prediction Operator were carried out for two RAD dose channels—silicon detector B and plastic, tissue-equivalent detector E—after removing SEP excursions and diurnal-pressure oscillations. A bootstrap Monte Carlo was used to propagate measurement and binning uncertainties (Guo et al., 2015).

For the diurnal pressure coefficient $x$4, obtained by fitting one-Mars-sol binned perturbations $x$5 versus $x$6 through

$x$7

the reported values are as follows:

Parameter	Detector E	Detector B
$x$8	$x$9	$\mathcal{Y}_{\rm pos}$0

In the long-term DPO formulae, $\mathcal{Y}_{\rm pos}$1 was used for both Models I and II. After normalizing to the reference pressure $\mathcal{Y}_{\rm pos}$2, the fitted Model I coefficients are $\mathcal{Y}_{\rm pos}$3 and $\mathcal{Y}_{\rm pos}$4 for detector E, and $\mathcal{Y}_{\rm pos}$5 and $\mathcal{Y}_{\rm pos}$6 for detector B. For Model II, the corresponding normalized coefficients are $\mathcal{Y}_{\rm pos}$7, $\mathcal{Y}_{\rm pos}$8, and $\mathcal{Y}_{\rm pos}$9 for detector E, and $\mathcal{Y}_{\rm neg}$0, $\mathcal{Y}_{\rm neg}$1, and $\mathcal{Y}_{\rm neg}$2 for detector B (Guo et al., 2015).

The correction conventions are equally explicit. Detector E is inherently tissue-equivalent and includes both charged and neutral contributions, whereas detector B sees mostly charged. To convert silicon-detector dose $\mathcal{Y}_{\rm neg}$3 into water-equivalent dose, the prescribed factor is $\mathcal{Y}_{\rm neg}$4. Dose equivalent is then obtained from

$\mathcal{Y}_{\rm neg}$5

with mean quality factor $\mathcal{Y}_{\rm neg}$6, adopted from Hassler et al. (2014). Operational use proceeds by selecting $\mathcal{Y}_{\rm neg}$7 and $\mathcal{Y}_{\rm neg}$8, choosing Model I or Model II, selecting detector channel, computing

$\mathcal{Y}_{\rm neg}$9

and optionally converting to dose equivalent with the stated $\pi_\theta$0 (Guo et al., 2015).

The stated validity range is limited to fits based on data with $\pi_\theta$1 and $\pi_\theta$2. Extrapolation well outside those windows carries large model uncertainty, and divergence of Model I versus II at $\pi_\theta$3 is specifically noted. Both models also assume linear independence of $\pi_\theta$4 and $\pi_\theta$5 effects; simulations are said to hint that at very low $\pi_\theta$6 or very high $\pi_\theta$7, second-order coupling may appear. SEP events are excluded from the operator and must be added separately, while dust opacity and atmospheric composition changes are described as negligible at RAD’s sensitivity (Guo et al., 2015).

4. Robust Adaptive Denoising Direct Preference Optimization in generative retrieval

In the 2026 e-commerce literature, RAD-DPO addresses alignment of Generative Retrieval (GR) models that retrieve items by autoregressive decoding of structured Semantic IDs (SIDs). Given a query or context $\pi_\theta$8, the model $\pi_\theta$9 generates an SID

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$0

where each token corresponds to one level of a hierarchical taxonomy, such as category, subcategory, and item code. The generated SID is then mapped back to one or more SKUs for ranking or display (Chen et al., 27 Feb 2026).

The work situates itself as a modification of Direct Preference Optimization. The classic pairwise DPO loss, ignoring SFT, is given by

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$1

Three failure modes are identified when this is applied directly to structured SIDs. First, standard DPO induces gradient conflicts on shared prefixes because positives and negatives often share the first $$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$2 tokens. Second, it is vulnerable to pseudo-negatives arising from noisy implicit feedback. Third, in queries with multiple relevant items, it creates probability squeezing among valid candidates, degrading recall (Chen et al., 27 Feb 2026).

The proposed RAD-DPO therefore combines three mechanisms: token-level gradient detachment to protect shared hierarchical prefixes, similarity-based dynamic reward weighting to mitigate label noise, and a multi-label global contrastive objective integrated with global SFT loss to expand positive coverage. This suggests that the method is less a minor loss reweighting than a composite redesign of the preference-learning pipeline around the structural peculiarities of SIDs.

5. Objective construction and optimization mechanics

The token-level detachment mechanism begins from the longest common prefix between a positive-negative pair $$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$3:

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$4

For the negative sequence, standard log-likelihood is

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$5

RAD-DPO defines

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$6

and applies stop-gradient to shared-prefix terms:

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$7

The resulting gradient is

$$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$8

so the first $$D(\Phi,P) = \sum_{j} \iint_{E,\epsilon} \lambda_{j}(E,\epsilon)\,F_{j}(\Phi,P,E)\,dE\,d\epsilon / m,$$9 tokens incur zero gradient from the negative sample, while the positive sequence remains fully differentiable (Chen et al., 27 Feb 2026).

For noise mitigation, the method uses cosine similarity between final hidden-state vectors at EOS:

$j$0

The first $j$1 similarity scores are collected into $j$2, quartiles $j$3, $j$4, and $j$5 are computed, and anchors are updated periodically. The negative-sample weight is then defined piecewise as

$j$6

This gives softer penalties to highly ambiguous negatives (Chen et al., 27 Feb 2026).

For multi-label settings, the positive set is

$j$7

The global SFT term is

$j$8

while the global contrastive preference term selects the highest-score positive $j$9 under the current model and contrasts it against the negative pool. The unweighted form is

$E$0

and the weighted form incorporates both detachment and $E$1. The total RAD-DPO objective combines $E$2 and $E$3; the training algorithm updates $E$4 using the gradient of $E$5 (Chen et al., 27 Feb 2026).

6. Training procedure, inference profile, and reported results

The reported training algorithm takes a pretrained SFT model $E$6 and a preference dataset $E$7. The listed hyperparameters are learning rate $E$8, batch size $E$9, temperature $\epsilon$0 (tuned), smoothness $\epsilon$1, and warm-up on the first $\epsilon$2 pairs. Each minibatch requires forward passes for all $\epsilon$3, extraction of hidden states $\epsilon$4, quartile updates beyond warm-up, computation of $\epsilon$5 and the detached, weighted $\epsilon$6, and gradient descent on their sum (Chen et al., 27 Feb 2026).

At inference, the method uses beam-search decoding with beam size $\epsilon$7, generates top-$\epsilon$8 SIDs, maps each to candidate SKUs, and merges them with other retrieval branches. The stated serving profile is latency below $\epsilon$9 at $\Phi$00 on a single RTX 4090 (Chen et al., 27 Feb 2026).

The reported offline evaluation is on JD.com’s 700 M-log dataset, test day $\Phi$01, for a 1.7 B model. The SFT baseline reports Halluc. $\Phi$02, $\Phi$03 $\Phi$04, $\Phi$05 $\Phi$06, $\Phi$07 $\Phi$08, and MRR $\Phi$09. Standard DPO reports Halluc. $\Phi$10, $\Phi$11 $\Phi$12, $\Phi$13 $\Phi$14, $\Phi$15 $\Phi$16, and MRR $\Phi$17. RAD-DPO reports Halluc. $\Phi$18, $\Phi$19 $\Phi$20, $\Phi$21 $\Phi$22, $\Phi$23 $\Phi$24, and MRR $\Phi$25. The summary statement is that RAD-DPO reduces hallucination by approximately $\Phi$26 relative to DPO and boosts $\Phi$27 by $\Phi$28 absolute; it is also reported to scale across model sizes from $\Phi$29 B to $\Phi$30 B and to remain robust when preference data is reduced from $\Phi$31 M to $\Phi$32 M (Chen et al., 27 Feb 2026).

Online A/B testing places the 1.7 B RAD-DPO model as a parallel generative branch in JD.com’s live search engine. The paper reports a $\Phi$33 absolute lift in User Conversion Rate over one week of testing on hundreds of millions of users, with the same sub-$\Phi$34 latency at $\Phi$35 (Chen et al., 27 Feb 2026).

7. Comparative interpretation and domain-specific significance

The two RAD-DPO formulations exemplify how identical acronyms can encode entirely different epistemic roles. The Martian RAD Dose Prediction Operator is a measurement-driven empirical predictor: its parameters are fitted coefficients, its inputs are environmental variables, and its output is a dose or dose-equivalent rate under quiet-Sun conditions. Robust Adaptive Denoising Direct Preference Optimization is a training-time alignment objective: its parameters are neural-network weights, its inputs are preference-annotated retrieval instances, and its output is an updated generative retrieval model (Guo et al., 2015, Chen et al., 27 Feb 2026).

Their methodological contrast is similarly sharp. The Martian operator assumes linear independence of pressure and solar-modulation effects over a bounded validity window and explicitly excludes SEP events. The retrieval objective, by contrast, is built around sequence-level and token-level differentiable optimization, representation-space similarity weighting, and multi-label contrastive learning. A plausible implication is that the acronym overlap is most usefully handled as a disambiguation problem in bibliographic and indexing systems, because the surrounding terminology alone—Mars surface dose versus Semantic IDs and SKUs—determines the intended referent.

In both cases, however, RAD-DPO denotes an attempt to make a complex observational or behavioral system operational. In planetary science, that operationalization supports empirical prediction of the Martian surface radiation environment under varying pressure and heliospheric conditions. In e-commerce retrieval, it supports alignment of generative retrieval models with structured preference data while reducing gradient conflicts, pseudo-negative sensitivity, and multi-label probability squeezing (Guo et al., 2015, Chen et al., 27 Feb 2026).