Recommender systems learn almost entirely from items they themselves chose to show. When such a system is retrained greedily on immediate clicks, it narrows what it shows, and it then receives feedback only on that narrower set. Its picture of the catalog gets worse exactly as its exposure concentrates. Epidemiology has studied a structurally similar loop for years, and its central lesson is about timing: in a compounding process, a modest intervention during the early phase does more than a large intervention applied late. One compartmental-model study found that tightening a death-rate target in Lombardy required a 441.7% increase in vaccination effort, and that acting early sharply reduced the effort needed (Xu et al.). Nobody has measured whether recommender feedback collapse has this shape: whether the dose of diversity needed to rescue a recommender grows sharply with how long one waits.
This note proposes to measure that delay-cost curve directly. It also tests a specific way to intervene early. Exposure-level signals such as catalog coverage are lagging indicators, because the logs are extremely sparse. The geometry of the recommender's own item embeddings, measured by how evenly they spread on the unit sphere, should degrade before coverage does. If so, it can serve as an early-warning trigger, much as sentinel nodes give epidemiologists weeks of lead time (Yang et al.). Stated in two sentences: delayed debiasing in a self-retraining recommender should cost more than proportionally, and because sparse logs make behavioral symptoms show up late, the warning sign should be read from the model's representation instead.
Core Claim
Consider a recommender retrained every round on the interactions it caused. Let $\tau$ be the round at which a diversity intervention starts, and $d$ its dose: the fraction of slate slots given to exploration, or the weight on an embedding-uniformity regularizer. The required dose is defined as
$$d^*(\tau) = \min \{ d : C_T(d, \tau) \ge c^* \},$$
where $C_T$ is ground-truth catalog coverage of what users actually consume at horizon $T$, and $c^*$ is a target set to the coverage of a never-collapsed reference run.
The first hypothesis is that $d^*(\tau)$ is convex and superlinear in $\tau$, and becomes infeasible beyond some delay, once simulated users' preferences have themselves narrowed. The second hypothesis is that a trigger based on a drop in embedding uniformity fires measurably earlier than a trigger based on coverage or exposure entropy, and recovers most of the gap between a fixed late schedule and an oracle that knows the true state of user preferences.
Both claims are falsifiable. If $d^*$ is roughly linear in delay, timing is a second-order concern, and an always-on regularizer is the right answer.
The idea is related to CORAL. CORAL models category bursts with a self-exciting Hawkes process, puts an upper confidence bound on the risk of saturation, and penalizes scores step by step. It proves finite-time recovery from saturation. This proposal differs in three ways. First, it asks a global question about delay cost instead of designing a local controller. Second, its trigger comes from representation geometry, not from exposure load. Third, its simulated users carry preferences that drift under exposure. That last point matters because CORAL's recovery theorem concerns the recommender's exposure state. It does not address damage done to the users, which is where a point of no return would come from. If the experiments show no point of no return, that result is consistent with CORAL and counts against hypothesis one.
Research Program
The main constraint is cost. Mapping $d^*(\tau)$ needs thousands of closed-loop runs: a grid over delays and doses, several backbones, and many random seeds. Million-agent LLM platforms such as OASIS take about 15 minutes per step for 10,000 agents on four A100 GPUs. OASIS also reports exaggerated herd behavior in its LLM agents, an artifact that would directly contaminate a collapse measurement. Here, the question of how far agent simulation can scale is therefore answered with a two-tier design.
The workhorse tier is a GPU-vectorized population of stochastic agents, organized like MAgent's batched engine. Their item utilities are anchored to the fully observed KuaiRec small matrix: 1,411 users by 3,327 items at 99.6% density. Every exposure therefore has a real, observed watch-ratio response rather than an imputed one. On top of this anchor, each agent carries two pieces of dynamics. The first is category-level satiation and recovery, modeled after LINUCRL's boredom model. The second is slow preference drift toward what the agent has been exposed to, the mechanism Carroll et al. identify as recommenders reshaping preferences.
Population size and peer effects are added by replicating users with perturbed utilities. The larger KuaiRec matrix (7,176 users, 16.3% density) serves as a held-out check on the calibration. The recommender sees only the logs it generates. Even though the ground truth is dense, its training data is as sparse as a real platform's. This lets us study sparsity and closed-loop feedback together, which Gap 3 in the literature leaves open.
The fidelity tier is a few hundred LLM agents in the style of SimTok and PUB. It is run at a handful of delay settings to check that the shape of the delay-cost curve holds outside the parametric agent model.
The recommenders are matrix factorization, LightGCN, and SimGCL, each retrained every round. The early-warning signal is the uniformity of item embeddings: the log of the average of $\exp(-2\|e_i - e_j\|^2)$ over item pairs, where $e_i$ is item $i$'s normalized embedding. It is tracked together with the effective rank of the embedding matrix, and both are computed at each retraining. SimGCL's analysis motivates this choice: much of the benefit of contrastive learning on sparse graphs comes from keeping embeddings spread out, so a loss of spread should be an early symptom of collapse.
The analysis then proceeds in three steps. It fits the functional form of $d^*(\tau)$. It measures the lead time of each indicator relative to the round at which coverage falls below a threshold. Finally, it compares triggered policies under a matched budget, defined as cumulative short-term engagement given up relative to the greedy policy. Matching that budget is what makes the comparison fair.
Baselines
The naive reference point is greedy retraining with no intervention. It sets the collapse trajectory and the upper bound on short-term engagement, and every intervention's cost is measured against it.
The second baseline is an always-on intervention given the same total budget. One version spreads exploration slots uniformly across rounds. The other trains SimGCL's noise-based uniformity regularizer from the first round. This baseline is the most direct test of whether timing matters at all, so the proposal fails if triggered policies cannot beat it at matched budget.
The strongest existing method is CORAL. It is run on the same MF, LightGCN, and SimGCL backbones, with its penalty scale tuned on the same validation seeds and with the same number of tuning trials as our trigger thresholds. The file reports no closed-loop numbers for CORAL. No reported result was found on Lacuna for CORAL's coverage or recall in a KuaiRec-anchored loop, so it will be run from scratch. We also include EXPLORE's copula-based re-ranker, which models session weariness. The file reports no closed-loop number for it either.
For offline reference only, SimGCL reports more than a 25% Recall@20 improvement over LightGCN on Amazon-Book. LINUCRL's farsighted policy reports a 66.7% relative improvement over a fixed baseline on Criteo, against 61.3% for an oracle greedy policy that knows the boredom state. That gap is the kind of margin from timing that this proposal expects to see at the population level.
The upper bound is an oracle trigger that fires on true preference drift, which is observable only in simulation. The key ablation keeps the dose schedule fixed and swaps the trigger signal among three options: embedding uniformity, exposure entropy, and coverage. A second ablation applies the best fixed front-loaded schedule with no trigger, to separate "act early" from "act on a signal."
Our method beats the field if, at matched engagement cost, it reaches target coverage with a lower total dose than both always-on and CORAL, with nonoverlapping 95% confidence intervals across at least 50 seeds. Its trigger must also lead coverage by a consistent number of rounds.
Why not just…?
Why not just regularize for uniformity from the start and forget timing? This is the most serious threat. If SimGCL-style regularization keeps embeddings spread out at every round, it may prevent collapse at a lower total cost than any trigger. The matched-budget comparison settles the question. If always-on reaches $c^*$ with an equal or smaller cumulative engagement sacrifice, the timing claim fails for this regime, and we will report that. There is a reason to expect otherwise. Front-loaded vaccination beats uniform rollout to the same final coverage (Wang et al.), and uniform spending puts budget into rounds where collapse is not yet underway.
Why not just use CORAL's step-wise saturation bound as the trigger? CORAL's Hawkes intensity is computed from exposure counts. Under sparse logs, those counts reflect what the model chose to show, not what users would prefer. Its signal may therefore be a lagging indicator, like travel bans triggered by already-visible outbreaks (Chinazzi et al.). The indicator-swap ablation measures this directly. If CORAL's bound leads coverage as early as embedding uniformity does, the representation-based trigger adds nothing beyond CORAL, and the contribution shrinks to the delay-cost curve itself.
Why not just run a small LLM simulation, as SimTok does? SimTok already varies interventions such as cold-start matching ratios. It cannot afford the dense grid over delay and dose needed to establish convexity. Its users are also LLM agents, whose herd-like artifacts in OASIS could make a tipping point appear where none exists. The LLM tier here is a check on shape, not the main instrument. If the parametric and LLM tiers disagree about the curve's shape, that disagreement is itself a finding about how fragile simulator-based conclusions are.
Why not just evaluate offline on logged data? KuaiRec shows that sparse logs can invert method rankings and inflate popularity baselines. More fundamentally, the effect of delaying an intervention is a counterfactual about future exposures, and no static log contains it.
Decisive Checks
The first check is the shape of $d^*(\tau)$ in the KuaiRec-anchored loop. We fit linear, power-law, and threshold models and compare them by held-out likelihood across seeds. A linear fit kills hypothesis one. The second check is whether the uniformity trigger leads the coverage threshold in more than 80% of seeds, by a median of at least several retraining rounds, across all three backbones. The third check is the matched-budget comparison against always-on SimGCL and CORAL described above. The fourth is whether the curve's shape, convex or not, matches between the parametric and LLM tiers at three or more delay points. Clearing all four would establish that, for recommenders, when to intervene is a first-class design variable, and that the place to look for the warning is inside the model rather than in its sparse logs.
References
- Control strategies for COVID-19 epidemic with vaccination, shield immunity and quarantine: A metric temporal logic approach
- Optimizing sentinel surveillance in temporal network epidemiology
- CORAL: Uncertainty-Aware Regulation of Exposure Concentration in Recommender Systems
- OASIS: Open Agents Social Interaction Simulations on a Large Scale
- MAgent: A Many-Agent Reinforcement Learning Platform for Artificial Collective Intelligence
- KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systems
- Fighting Boredom in Recommender Systems with Linear Reinforcement Learning
- Estimating and Penalizing Preference Shift in Recommender Systems
- Simulating Filter Bubble on Short-video Recommender System with Large Language Model Agents
- PUB: An LLM-Enhanced Personality-Driven User Behaviour Simulator for Recommender System Evaluation
- Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation
- Relevance Meets Diversity: A User-Centric Framework for Knowledge Exploration Through Recommendations
- Population-Level Effectiveness of COVID-19 Vaccination Program in the United States: Causal Analysis Based on Structural Nested Mean Model
- Cryptic transmission of SARS-CoV-2 and the first COVID-19 wave