Why Agentic Creative Production Wins the Generative-Retrieval Era

In February, Kuaishou published GR4AD, the generative recommender now fully deployed in its advertising system, serving over 400 million users. (If the name is unfamiliar: Kuaishou is China’s second short-video giant after Douyin, TikTok’s domestic sibling, with an ads business around ten billion dollars a year. You may know it as the company behind the Kling video model.) Headline number: up to 4.2% more ad revenue in online A/B against their production DLRM stack. That is an enormous lift for a mature ads system, but the headline isn’t the interesting part. The interesting part is the architecture, because Meta is publicly converging on the same one, and the rest of the industry is publishing in the same direction. GEM, Meta’s LLM-scale ads foundation model, shipped in 2025 with a reported 5% lift in Instagram conversions. Andromeda, its retrieval engine, has a stated roadmap toward autoregressive generation. And in July 2026 Meta published hierarchical interest representation, an exploratory foundation layer that discretizes LLM-enriched user and advertiser embeddings into “Bag-of-Meaning” tokens for Andromeda’s retrieval index: semantic IDs in all but name, now being built on both sides of the auction.

The short version: ad retrieval is becoming generative. The system stops scoring (user, ad) pairs and starts generating the ad it believes the user will convert on, token by token, the way a language model generates the next word. The tokens are discrete semantic IDs derived from the creative itself.

It restructures the economics of who wins on these platforms, and GR4AD is worth reading closely precisely because it is unusually candid about production details the big platforms only gesture at.

From scoring to generating

The DLRM-era stack (deep learning recommendation models, whose reference implementation Meta open-sourced back in 2019), compressed: candidate retrieval runs approximate-nearest-neighbor search over embedding indexes, then a cascade of rankers scores (user, ad) pairs with increasing precision. The embeddings are substantially behavioral. An ad’s representation is learned from who engaged with it, which means a new ad is a cold-start problem by construction: the system can’t know what an ad is until enough people have seen it, and the index it lives in is rebuilt on a minutes-level cadence.

Generative retrieval inverts this. Encode the user’s context as a sequence, then decode the recommendation directly, generating a semantic ID (a short sequence of discrete codes quantized from a content embedding) via beam search. The lineage runs from semantic-ID retrieval (TIGER) through Meta’s HSTU generative recommenders to Kuaishou’s OneRec line for its organic feed. GR4AD is the ads-specific, fully deployed version, and its production details are the payload:

• UA-SID tokenization. A fine-tuned multimodal LLM ingests the creative (video, product, advertiser metadata) into a Unified Ad Embedding, then a multi-granularity residual quantizer turns it into a coarse-to-fine token sequence: large codebooks first (16,384 entries) for dominant semantic factors, smaller ones (1,024) for residuals. An ad is its content tokens.
• The last token level isn’t semantic at all. It’s a hash of business features: item and account IDs, conversion type. Two identical creatives in different accounts, optimizing to different events, are different token sequences with different retrieval trajectories. Account structure is inside the representation.
• Business value is generated, not re-ranked. GR4AD discretizes eCPM into equiprobable buckets and appends the eCPM token as one more generation step after the semantic sequence, trained with the same cross-entropy as everything else, plus value-aware sample weighting (VSL) and a ranking-aligned RL stage (RSPO) that optimizes the generated list against NDCG directly.
• Serving is engineered for wide beams. A “LazyAR” decoder runs the first two-thirds of the network without conditioning on the previous token (parallel across levels, autoregressive only at the top), roughly doubling QPS. Beam-shared KV caching and traffic-adaptive beam schedules do the rest. Production: 500+ QPS per L20 GPU at under 100ms.
• The index is bidirectional (semantic ID to item and back) and updates in seconds, not the minutes-level rebuild of an embedding index.

Four properties fall out of this design, and each one moves money.

Cold start collapses

In a behavioral-embedding world, a new creative buys its way into retrieval with exposure. In a content-token world, the semantic ID positions the ad the moment it’s computed: indexed in seconds, retrievable on its first impression, inheriting the retrieval behavior of its semantic neighbors instead of starting from zero.

The distributional evidence that this is real: alongside the revenue lift, Kuaishou reports 17.5% more ad delivery for small and medium advertisers and 7.28% higher conversion among less-active users. Generative retrieval redistributes delivery away from accumulated-history advantage toward semantic coverage. The moat of “we’ve spent $50M so the system knows us” thins. What the system knows is your content.

For creative production this means every new creative is a live experiment from minute one. Freshness compounds. Velocity stops being throttled by warm-up.

Value is generated, not re-ranked

The eCPM token is the quiet bombshell. The retriever isn’t finding relevant ads and letting a separate ranker worry about money. It is trained end to end to generate toward high-value regions of semantic space. Predicted business value is part of the token stream.

Which means creative quality, in the retrievable sense, and monetization are no longer separate axes. Where your creative sits in semantic space is its economics. A slate that occupies many distinct high-eCPM neighborhoods gives the decoder many paths that terminate in your ads. A slate that occupies two neighborhoods, however many assets deep, gives it two.

The retrievable surface is on a compute dial

GR4AD’s scaling results are unusually clean. At a fixed 0.16B parameters, widening the beam from 128 to 1024 moves the revenue lift from +2.33% to +4.21%. Scaling the model from 0.03B to 0.32B moves it from +2.13% to +4.43%. Monotonic, near-log-linear: an inference-time scaling law for ad revenue.

Platforms now own a dial that converts GPU spend into revenue by exploring more semantic neighborhoods per request. They will turn it every quarter, because the trade is mechanical. And a wider beam is only worth something to an advertiser whose slate spans neighborhoods. The platform is buying more exploration of a space you may or may not have bothered to occupy.

Account structure became a semantic lever

Because the final token level hashes account ID and conversion type, the same creative under three conversion objectives is literally three token sequences: three separate retrieval experiments, not redundancy. Objective and structure sweeps stop being cargo cult and become structural exploration of the representation itself. (Kuaishou’s mechanism won’t be Meta’s mechanism, but the direction, business context entering the ad’s identity, is the shared design choice.)

Compose them, and throughput’s value rises every quarter

These are not four observations; they are one system. Cold-start collapse makes velocity cheap. Value-conditioned generation makes semantic diversity pay. Inference-time scaling grows the retrievable surface mechanically. Objective-aware identity multiplies the experiment space.

Add it up: the platforms are training retrievers whose loss functions reward high-volume, semantically diverse, objective-tagged creative output. More of it every quarter, encoded in models that are already in production.

What the quantizer does to “volume”

Here is the subtlety that kills the naive read of everything above: this is not an argument for spamming variants.

The front of the pipeline is a quantizer whose entire job is to map similar content to the same tokens. Thirty permutations of one concept (new hook frame, same claim, same visual grammar) land on the same or adjacent semantic prefixes. From the decoder’s point of view you haven’t widened your slate; you’ve deepened a single token. In the auction you are competing with yourself. In the portfolio, one fatigue event takes out the whole block.

The unit of exploration is the semantic neighborhood, not the asset. Diversity that counts is diversity the encoder can see: a different angle, a different visual system, a different voice, a different claimed value. This is measurable on your own slate today. Embed your ads, cluster them, count the clusters. Many “300-creative” accounts are nine neighborhoods deep. The platform’s model knows this even when the spreadsheet doesn’t.

What this breaks

Targeting micro-management. The system finds the pocket. Your job is to give it enough semantically distinct material to have pockets to find. The residual skill isn’t audience selection; it’s coverage.

Winner-squeezing. Concentrating spend on last quarter’s champion was rational when learning was slow and each proven creative was precious. When cold start is free and the surface widens quarterly, concentration is a structural under-allocation of exploration.

The agency cadence. Twenty hand-crafted concepts per quarter was a deliverable for the DLRM world. The job is now hundreds of objective-tagged, semantically distinct variants per week with feedback loops measured in days. A boutique shop can’t reach that cadence. An in-house creative factory can, but it’s built on payroll, and it has a ceiling and a burn rate.

What it doesn’t break is the need for judgment. It relocates judgment from “which ad is best” to “which regions of the space do we probe next, and what did the last probe teach us.” That second question has budget physics of its own, covered in the testing-economics companion to this piece, because a probe you can’t afford to read is not an experiment.

The loop that wins is agent-shaped

The work that compounds in this regime is a tight cycle:

1. Read the delivery and conversion signal from the ad account.
2. Form a hypothesis about why: which semantic region is working, for whom.
3. Produce creatives that probe the hypothesis: script, storyboard, generation, edit, render.
4. Tag and structure them (account, objective, audience hint), knowing structure is representation.
5. Launch, read the deltas, update the hypothesis. Repeat.

Every step is now machine-executable. Reading signal is SQL over modeled ad data. Hypothesis formation is an LLM with the right context. Production is a generation-and-edit stack. Structuring and launching are idempotent API writes. The bottleneck used to be human attention dragged across all five steps. Now it’s spec clarity and judgment about which loops to run.

And there is a deeper symmetry. The platform’s model reasons about your creatives as points in a semantic space. A production system that reasons over the same space (which neighborhoods it occupies, which are empty, which the market has already validated) is playing the same game as the retriever, from the supply side. Human intuition approximates this. An agent system does it literally, embedding its own ads and its market’s, and planning each round as coverage over the space. Humans cap the loop at attention. Agents cap it at budget.

Meta’s hierarchical-interest work makes the symmetry explicit on the demand side too. The platform is learning to represent buyers as multi-level interest hierarchies, connected to advertisers, products and conversion pixels in one graph, with distillation tricks to keep reasoning where deep-funnel signal is sparse. That is a persona system at platform scale. A production loop that plans coverage over angles and buyers, not just assets, is mirroring both halves of the retriever’s world model.

That is the thesis HyperScale is built on: an agent team that runs the loop end to end. It reads the account, plans each round as a portfolio of hypotheses across angles, audiences, and formats, produces the ads, launches with your consent, and writes every verdict into a belief ledger the next round plans from. We aren’t building against a prediction of where platforms are going. We’re building against what they have already published and deployed.

The next two years

For technical founders and investors in growth infrastructure: the durable assets are the creative-production substrate (multimodal models, production agents, tight feedback loops) and per-tenant semantic understanding: what works for this brand, against which audiences, in which lifecycle phase. Orchestration scaffolding is necessary and commoditizing. And the substrate transfers. “Produce semantically diverse variants, observe the value distribution, promote winners” is the same loop whether the surface is an ad, a paywall, or an onboarding flow. Teams that build it once compound across surfaces.

For growth teams running real budgets: the question for your agency or your in-house team is no longer “how do we beat our best creative.” It’s “how many semantically distinct, objective-tagged variants shipped last week, and how long is our feedback cycle.” If the answers are “a dozen” and “a month,” the platform’s beam is widening into neighborhoods someone else’s slate will occupy.

The shorter version

Ad retrieval is becoming a generative model that wants more high-quality, semantically diverse, objective-tagged creative than any human team can produce, and the platforms’ own scaling laws say their appetite grows every quarter. Matching that appetite is a throughput problem. Throughput at this cadence is an agent problem. The platforms are publishing the answer key.

The question is who has the production line to ship against it.