Phantom Preference Drift in Exploratory Shopping Agents

When people shop or search without a clear target, they often decide what they want only after the system shows them something. Suppose an assistant mentions that a jacket is "packable into its own pocket," and the user replies "oh, I want that." The user's preference has just changed. If the jacket really has that feature, the system has helped. If the assistant made the feature up, the user is now chasing something no real product may satisfy. We call this phantom preference drift: a change in what the user wants that was caused by a false claim from the system. We hypothesize that phantom drift is not random. It should concentrate where the system's search engine is biased toward popular products. When the item that fits a user's niche, newly formed need sits in the catalog's long tail, a popularity-skewed search cannot surface it. The language model agent then tends to fill the gap by attributing the wanted feature to a popular item it did find.

This note proposes to measure that coupling and then break it. We first build an attribute-level audit for interactive shopping benchmarks. The audit checks every feature the agent asserts against the product's catalog record, and it labels each user preference change as genuine or phantom. We then run a dose-response experiment that turns the popularity skew of the search engine up and down, to test whether hallucinated features are downstream of retrieval failure. The proposed fix is a small control rule called verification-triggered tail retrieval. When a feature claim fails catalog verification, the agent does not assert it and does not simply abstain. It reissues the search with popularity correction switched on. Only if that also fails does it tell the user the need cannot be met and offer the nearest real alternative.

Core Claim

Our claim is that in multi-turn exploratory search, popularity bias in retrieval causes a measurable share of feature hallucinations by language model agents. These hallucinations then corrupt the user's evolving preferences in ways that current success metrics may not register. Applying popularity correction only when verification fails should recover tail items without hurting users who genuinely want mainstream products. Applying it to every query should cost head-item accuracy. Abstaining on every failed claim should cost task completion.

The pieces of this claim come from separate literatures that have not been joined. AcCoRD models shopping and travel collaborations as partially observed decision problems with "triggered" preferences that appear only after the agent shows an option. It reports that triggered preferences drive some models' scores as low as 0.09, but it does not track whether the triggering content was true. ToolRec stops a language model recommender from inventing nonexistent items by restricting it to retrieval tools and checking results against an item directory. It assumes preferences are fixed, it leaves the retrieval tools' popularity bias uncorrected, and it underperforms on niche local-business data (Yelp2018). KGHaluBench shows on encyclopedic knowledge, not catalogs, that larger models mostly stop failing to recognize entities while still getting fine-grained facts about less prominent entities wrong. Its fact-level hallucination falls only about 24% across model scale. Lin et al. show that conversational recommenders default to popular items because popular item embeddings grow long. They study non-generative backbones, so the bias they measure cannot produce invented features.

Together these motivate the hypothesis, but none tests it. AcCoRD supplies the evaluation setting, ToolRec supplies the grounding technique we reuse, and Lin et al. and KGHaluBench supply the reason to expect the coupling.

Research Program

The first component is the audit. We will log every assertion the agent makes about a product's attributes or options on AcCoRD-Shop, which is built on WebShop's catalog of about 1.2 million products. We will also log the AcCoRD-Travel scenarios, which average 11.25 preferences each. Each assertion is matched against the structured catalog record by a rule-based matcher, with an LLM judge for paraphrases. We will check the judge against human labels on a sample, since KGHaluBench's automated verifier reached only up to 87.74% agreement with humans in its own setting.

Every preference change in the simulated user is then traced to the agent turn that triggered it. It is labeled phantom if the triggering claim was false. We will also rescore episodes against catalog ground truth, in case the LLM user simulator accepts a hallucinated feature as satisfied. The audit, the phantom labels, and the rescored success rate are the first deliverable, and they are useful even if the later hypotheses fail.

The second component is a controlled manipulation of popularity. WebShop products do not come with a clean popularity signal in the material we have. We will therefore inject a synthetic popularity prior into the search engine's ranking score with a tunable weight. Popularity is assigned so that it is independent of whether an item satisfies a given user's hidden preferences. We will also run a natural variant that uses review counts where the catalog provides them.

Sweeping the prior's weight gives a dose-response curve. On one axis is retrieval skew. On the other are claim-level hallucination rate and phantom drift rate, stratified by whether the satisfying item is head or tail. The causal story predicts that hallucinations rise with skew mainly in tail-target episodes. A flat curve would mean hallucination is driven by the language model alone, not by retrieval.

The third component is the intervention. A verifier checks each candidate feature claim against the catalog record before it reaches the user, following ToolRec's directory-checking idea. On failure, the agent reruns its search with the popularity prior removed counterfactually. This follows the subtraction logic of MACR, which separates a direct popularity effect from user-item matching and subtracts it at inference. If a real item with the feature exists, the agent presents it. If none exists, the agent states that the request cannot be met and proposes the closest verified item. This is the negotiation behavior AcCoRD finds missing for unachievable preferences.

Gating the correction on verification failure matters for a specific reason. Always-on debiasing runs into the accuracy cost for users who really want popular items, and CSBR identifies all-head sessions as a hard case for calibrated debiasing.

Temporal drift enters this program as the regime, not as a training objective. Offline time-discounted alignment such as NS-DPO reweights old preference pairs and does not address within-session drift triggered by system output. We therefore do not build on it. We do, however, report whether phantom drift accumulates over turns, since once a user adopts a phantom feature, later turns condition on it.

Baselines

The naive reference is a ReAct agent, which interleaves reasoning steps with actions, run on AcCoRD-Shop with the default search engine and no verification. This is the setting where AcCoRD reports a perfect outcome rate below 30% even for its top model, Claude-4.5. For the underlying WebShop environment, the best trained agent reaches a 29.1% success rate and a 62.4 average score, against 59.6% and 82.1 for expert humans (WebShop). We found no reported claim-level hallucination rate or phantom-drift measure on AcCoRD in Lacuna, so these numbers are produced by our audit.

The strongest prompting competitor is UncReAct from AcCoRD, which makes the agent mark uncertainty before acting. The strongest grounding competitor is a ToolRec-style agent whose claims are verified against the catalog and dropped or abstained on when they fail. The strongest debiasing competitor applies MACR-style popularity subtraction to every query.

The ablations that isolate the key ingredient are two. The first combines always-on debiasing with verification, which tests whether gating matters. The second uses verification with plain re-search on failure, without popularity correction, which tests whether reformulation alone explains any gain. WebShop already identifies query reformulation as a bottleneck.

All agents share the same backbone model, interaction budget, catalog, and user simulator seeds. Each method gets an equal budget for tuning prompts and thresholds on a held-out scenario split.

Beating the field means three things on tail-target episodes. The full method must lower phantom drift and claim-level hallucination relative to the ToolRec-style verifier, and it must also raise perfect outcome rate relative to that verifier. On head-target episodes, it must stay within noise of the undebiased agent. If always-on debiasing plus verification matches it on both strata, the gating idea is unnecessary, and we will report that.

Why not just…?

Why not just verify every claim and abstain on failure? This is the most serious threat to the hallucination half of the claim. A verifier alone may drive hallucination near zero. Our hypothesis is that it does so by turning tail-target episodes into dead ends: the agent cannot find the right item, cannot invent it, and stops. KGHaluBench flags this kind of over-abstention as making agents unhelpful. The verifier-only baseline will show this directly if its perfect outcome rate on tail-target episodes stays low while its hallucination rate falls. If its completion rate holds up, the retrieval component adds nothing, and we will say so.

Why not just debias the search engine everywhere? Popularity correction such as MACR's subtraction or Adap-τ's embedding-length normalization targets the retrieval failure itself. It may remove most phantom drift without any verifier. Two things argue against it as a complete answer. First, debiasing cannot stop the model from asserting features no retrieved item has, so unachievable preferences still invite fabrication. Second, always-on correction penalizes users who want mainstream items, the pure-head conflict CSBR documents. The always-on ablation, stratified by head and tail target, settles both questions empirically.

Why not just prompt the agent to admit unavailability and consider niche items? AcCoRD already tested the closest version of this, UncReAct. It did not significantly improve outcomes, and weaker models over-asked and exhausted their budgets. We still include it as a baseline because AcCoRD did not measure hallucination, and prompting could plausibly reduce false claims even without raising success.

Why not just use a stronger model? KGHaluBench's scale results suggest fine-grained facts about non-prominent entities remain error-prone even at the frontier. Catalog attributes of tail products are exactly that kind of fact. We will run the dose-response sweep on two backbones of different strength to test whether the coupling weakens with scale.

Proposed Experiments

The first decisive check is the audit on unmodified AcCoRD-Shop and AcCoRD-Travel runs. If phantom drift accounts for a non-trivial fraction of triggered preference changes, and rescoring against catalog ground truth changes reported success, the evaluation gap is real. That holds regardless of what follows.

The second check is the popularity dose-response on AcCoRD-Shop. The hypothesis is supported if hallucination and phantom drift rise with injected skew specifically in tail-target episodes, with the effect stronger than in head-target episodes. It is refuted if the curves are flat or the strata behave identically.

The third check is the head-to-head comparison against the baselines above, at the strongest skew level and at the natural review-count popularity. We report perfect outcome rate, average task score, claim-level hallucination rate, and phantom drift rate. We also report tail and head success rates in the spirit of the tail success rate defined by Lin et al., plus turns used.

A secondary transfer test repeats the dose-response on conversational movie recommendation over ReDial, using item mention frequency as popularity. PopNudge shows that only about 35% of TG-ReDial items ever appear in training and raises base ReDial Hit@10 from 0.0032 to 0.0220 with targeted augmentation. This test asks whether the coupling holds outside the WebShop catalog. Failure there would limit the claim to structured product search.