When a shopper refines a whole outfit through conversation ("the shoes are too casual", "make the top darker"), the system has to do two hard things at once. It has to keep the items compatible with each other, because an outfit's value comes from how the pieces combine and not from how good each piece is alone. It also has to cope with critiques that are often wrong: the user names the wrong item, overstates a change, or contradicts an earlier turn. Current systems handle at most one of these problems. Conversational bundle recommenders treat every answer as true. Noise-robust image retrieval methods tolerate imperfect text, but they retrieve one image at a time and do so in a single shot.
We propose to treat the system's state during outfit refinement as a probability distribution over whole candidate outfits, and not as a single query that each critique overwrites. Each critique updates this distribution through a learned likelihood, meaning an estimate of how probable that critique is if a given outfit were the user's target. The likelihood has two parts that are hidden from the system and must be inferred: which item the user was actually talking about, and how reliable this user is. Because no outfit is ever given zero probability, a bad critique lowers the chances of the right outfit but does not eliminate it, and later turns can recover it. The likelihood itself is trained on web-scraped relative captions, which are text descriptions of how one image differs from another. That data is known to be noisy. We clean it and estimate its per-example noise, and we use that noise estimate as a starting guess for how noisy live users will be.
Core Claim
The claim is specific and testable. When critiques are wrong 20–40% of the time, a set-level belief with two ingredients will reach the target outfit in fewer turns and recover more often after an erroneous critique than systems that rewrite the query deterministically. The first ingredient is coupling between items through a compatibility model. The second is latent attribution and reliability. We further claim that its advantage over a strong implicit alternative, a history-conditioned model trained with noise augmentation, appears mainly in recovery after contradictory feedback and in calibration. We do not expect a large advantage on clean dialogues. If that implicit alternative matches us on recovery, the explicit belief is not needed, and we will report that result.
What Prior Work Gives and Leaves Open
The motivation comes from two gaps. Bundle MCR frames multi-turn bundle recommendation as a Markov decision process solved by a transformer (Bundle BERT). The system alternates between asking attribute questions and showing bundles. Its feedback is limited to predefined attribute slots and assumed correct, so one wrong answer discards compatible candidates. Composed image retrieval (retrieving a target image from a reference image plus a text modification) has also become noise-aware. TME separates clean from mislabeled triplets by fitting a two-component mixture to per-sample losses, and repairs bad texts from visual differences. Multi-grained uncertainty regularization jitters target features to model one-to-many queries. Css-Net enforces consensus across several compositors. All three work on single, one-shot triplets. None models an outfit, and none models a person giving feedback over several turns.
The mechanism borrows from interactive learning that drops the assumption of a perfect human oracle. Visual Recognition with Humans in the Loop treats user answers as stochastic evidence in a Bayesian posterior, so wrong answers do not break the decision process. It also chooses questions by information gain. ABLUF learns the trainer's feedback noise online while optimizing the policy. We reuse both ideas. What is new is applying them to a combinatorial space of outfits and adding a latent variable for which item a critique refers to, a failure mode that does not exist in scalar control or single-item classification. HuGE constrains the design: it shows that human feedback used directly as reward makes agents fragile. We therefore use critiques only to update beliefs and never as a training reward for the retriever.
The set model comes from OutfitTransformer, a permutation-invariant transformer that scores whole-outfit compatibility, and from type-aware embeddings, which keep compatibility from being forced into a single similarity space. OutfitTransformer retrieves only one missing item at a time. Our system instead maintains a joint belief over all slots of the outfit at once. InfoCIR shows that interactive refinement greatly helps single-image retrieval. It also reports a cold-start weakness: the loop needs a relevant anchor among the first results. We address that weakness directly by keeping probability mass on outfits that were never shown.
Research Program
The belief is represented with particles. Each particle is one full candidate outfit, built from the top-$M$ catalog candidates per category slot and scored by OutfitTransformer compatibility. After a critique $c_t$ at turn $t$, each particle outfit $O$ is reweighted by
$$ p(c_t \mid O, S_t) = \rho \sum_{j} \pi_j(c_t, S_t)\, f(c_t \mid s_j, o_j) + (1-\rho)\, u(c_t), $$
where $S_t$ is the outfit shown at turn $t$ and $s_j, o_j$ are its $j$-th item and the particle's $j$-th item. The term $\pi_j$ is the probability that the critique refers to slot $j$, predicted from the text and the shown outfit. The term $f$ is a per-item relative-change likelihood, trained as a composed-retrieval scorer. The term $u$ is a broad background model for uninformative critiques, and $\rho$ is this user's reliability. $\rho$ starts at a prior taken from web data and is updated online by expectation-maximization, as in ABLUF. After each turn we resample particles and propose new ones by swapping items in slots with high $\pi_j$. We then either show the most probable outfit or ask a clarifying question chosen by expected information gain.
The web-noise component determines where $f$ and the prior on $\rho$ come from. We train $f$ on FashionIQ relative captions using TME-style loss-mixture separation. The fitted clean-probability of each training triplet gives an empirical noise rate, which we use as the initial $\rho$. We also use it to calibrate $f$ so that its confidence tracks how often similar captions were mislabeled. The hypothesis is that web annotators and live users make similar mistakes, so noise learned offline is informative about noise online. This is an empirical question and we test it directly.
No benchmark exists for multi-turn, set-level composed retrieval, so we build one from Polyvore Outfits. Each episode pairs a target outfit with a starting outfit that shares some slots and differs in others. A critique simulator produces relative captions for the differing items. It uses a captioner trained on FashionIQ, whose dress, shirt, and top categories map onto outfit slots. We inject controlled noise of three kinds: critiques attributed to the wrong item, reversed direction ("lighter" instead of "darker"), and contradictions of earlier turns. Noise rates range from 0 to 40%, the range where HuGE reports its own tolerance. A small human study then collects real critiques on 100–200 episodes. These give a check on calibration and success rate that does not depend on the simulator.
Baselines
The naive reference point is one-shot retrieval with no refinement. We run OutfitTransformer complementary retrieval slot by slot from the starting outfit and report success at turn zero. This measures how much any refinement adds.
The deterministic-refinement baseline treats each critique as a hard edit. An LLM rewrites the critique into per-slot modification texts. A per-slot composed retriever (the same $f$ backbone) replaces the named item, and OutfitTransformer then reranks for compatibility. This is the "trust the user" loop that InfoCIR and CoV-RAG-style systems implicitly assume.
The strongest competitor is a history-conditioned set retriever. It is an OutfitTransformer-style encoder that takes the current outfit and all past critiques as tokens and scores whole candidate outfits. It is trained on the same simulator with the same noise distribution. It could learn to discount unreliable critiques without any explicit belief, which is why it is the real threat. It gets the same backbone, the same simulator episodes, the same hyperparameter search budget, and the same candidate pools per slot.
We also adapt Bundle MCR's Bundle BERT by mapping critiques onto its attribute slots, so the comparison includes the existing conversational bundle method. For reference points, the file reports OutfitTransformer at Recall@10 of 9.58 for complementary item retrieval on Polyvore Outfits and 67.10% FITB accuracy. TME holds 74.58% average recall on CIRR at 80% synthetic noise. Bundle MCR reports 61.56% to 196.72% relative accuracy gains over non-conversational baselines. We found no reported result on Lacuna for multi-turn set-level composed retrieval under noisy critiques. Our benchmark therefore defines the comparison, and all systems run on identical episodes.
Beating the history-conditioned model means three things. First, a clear gain in success after a contradictory critique, meaning the target outfit is recovered within three turns after an injected error, at noise rates of 20% or higher. Second, better calibration of the predicted probability of success. Third, no loss at zero noise.
Three ablations isolate the ingredients. Removing compatibility coupling, so each slot has an independent belief, tests whether the joint set belief matters. Fixing $\rho$ instead of learning it tests whether online reliability estimation matters. Training $f$ on raw rather than denoised FashionIQ, and dropping the web-derived $\rho$ prior, tests whether the web-noise component earns its place.
Why not just…?
Why not just train the history-conditioned transformer with noisy simulated users and skip the explicit belief? This is the most serious alternative, and it may be enough on the simulator it was trained on. Our concern is transfer. Implicit robustness is learned for one noise distribution. The explicit model instead estimates each user's reliability at test time, as ABLUF does. We settle this with two experiments. We train both systems at one noise rate and test them at others, and we compare both on real human critiques. If the implicit model matches us in both, we will recommend it.
Why not just retrieve each slot independently with a noise-robust composed retriever such as TME? Per-slot retrieval ignores two things. It ignores that "more formal" means different changes for shoes and for tops, and it ignores whether the replacements still go together. The type-aware and OutfitTransformer results show that set-level scoring beats pairwise and independent scoring. Our decoupled-slot ablation measures directly how much of that advantage survives under interactive noise.
Why not just append the critique to an LLM prompt and regenerate the search? This treats every critique as a hard constraint. The Visual 20 Questions result shows the cost: probabilistic answers raised accuracy from about 19% to about 66% over the deterministic approach. The deterministic-refinement baseline implements this option. The recovery-after-error metric shows whether it drifts after a misattributed critique.
Why not just use Bundle MCR with attribute questions, which avoids free-text noise? Attribute answers are also noisy, and categorical slots cannot express relative visual changes. We still run BUNT on the same episodes. If it matches us when feedback is restricted to attributes, the advantage of free-text critiques is limited to cases where attributes are too coarse.
Decisive Checks
We run three checks, each of which could falsify part of the proposal. The first is whether the web-derived noise estimate predicts real user noise. We compare the calibrated $\rho$ prior against the rate of erroneous critiques we label in the human study. A large mismatch would mean the web-noise component is not useful. The second is the recovery curve: probability of recovering the target outfit as a function of turns after an injected error, compared against the history-conditioned baseline. The third is whether success at a fixed turn budget holds up on real human critiques, beyond the simulator. If the first check fails but the second and third succeed, the contribution narrows to set-level belief refinement, and we will report it that way.
References
- Bundle MCR: Towards Conversational Bundle Recommendation
- Learning with Noisy Triplet Correspondence for Composed Image Retrieval
- Composed Image Retrieval with Text Feedback via Multi-grained Uncertainty Regularization
- Collaborative Group: Composed Image Retrieval Via Consensus Learning from Noisy Annotations
- Visual Recognition with Humans in the Loop
- Learning Behaviors with Uncertain Human Feedback
- Breadcrumbs to the Goal: Supervised Goal Selection from Human-in-the-Loop Feedback
- OutfitTransformer: Learning Outfit Representations for Fashion Recommendation
- Learning Type-Aware Embeddings for Fashion Compatibility
- InfoCIR: Multimedia Analysis for Composed Image Retrieval