2026-09-15

Other changes reviewed today:

• Tencent Weixin Moments Ads' ChronicleRec makes it possible to precompress and cache very long behavior sequences. A 7-day A/B test gained GMV +1.61%.
• The X open-source mirror xai-org/x-algorithm had no new commits after the 2026-09-12 commit (6bb4594, Gen4+20).
  • Gen4+N is the number these notes assign in order to X commits after the Gen4 release.
• The Meta engineering and AI blogs and the Netflix TechBlog also had no new sources.

1Picking ad creatives in real time with whole-page gains and losses in view — Radar (Pinterest)

Pinterest has an unusual condition. In its waterfall grid (an infinite-scroll layout where columns align like a waterfall and items have different heights), the vertical length of one ad creative effectively determines the display space of the other ads and organic Pins placed around it. So maximizing only ad CTR locally causes cannibalization that worsens the whole-page metric. Cannibalization is the phenomenon in which the gain in one slot is offset by a loss in another slot. This paper describes, from both the system and the model perspective, how production DCO was designed under these conditions.

A. Creative optimization runs in parallel with ranking to hide latency, and six units finish creative selection verified

PinDCO runs the creative optimization branch in parallel with ranking. The existing ads pipeline is sequential, with ranking after candidate retrieval. The ranking model is heavy, so rerunning ranking for every creative variant is hard, and the parallel branch solves this problem. Because the parallelization is meant to hide ranking latency, end-to-end latency stays within the upper bound set by ranking latency.

The creative optimization branch is a pipeline of six units.

1. Creative Asset Manager: looks up the ad IDs returned by the retrieval stage in a key-value store and expands them into creative metadata. QPS and the number of entries are large, so it keeps metadata to the minimum and puts a cache cluster in front to reduce backend load.
2. Pre-Selection: lightweight models such as rules, a multi-armed bandit, or a linear classifier filter out low-opportunity creatives early.
   • It runs locally inside the ad service, so it finishes without a model server RPC.
   • How aggressively it filters is the dial that trades off performance against cost.
   • In the extreme case where all variants of an ad have low expected value, the lightweight model can leave only one and skip the later stages entirely.
3. Creative Component Fusion Network (CCFN): the ad service requests creative scoring from the model inference server at the same time as the ranking request.
4. Pixel-aware Adjustment Module (PAM): applies a penalty based on vertical size to the CCFN score. This is the stage that reflects whole-page gains and losses.
5. Multi-Armed Bandit: stochastically overrides the predicted winner to collect randomized feedback.
6. Returns the final winner to the ad service.

The order of the six units is in Figure 2 of the paper. The gist is "put the latency budget inside ranking latency, and finish creative scoring, adjustment, and exploration in parallel within it."

B. CCFN has a dedicated tower for each of image, title, and layout, and matches convergence speed with per-component hyperparameters verified

CCFN consists of three layers of building blocks.

• Feature encoding: continuous features are normalized with clipping/whitening/log-transformation, and categorical and ID features are converted to lookup embeddings.
  • Encoded features are split into per-component groups (for example, image, title, layout) and a "Shared" group not tied to any component.
• Component-specific towers: each component group goes into a dedicated tower together with the Shared group to build a component-specific representation.
  • Per-component hyperparameters are used because the image tower has more information to learn than the title and layout towers, so convergence speed differs.
  • If all towers use the same dropout rate and lr, one tower starts to overfit while the others stay underfit.
  • The dropout rate of each tower is tuned separately to match convergence speed.
• Component fusion layer: fuses the tower output embeddings to learn interactions between components and outputs a scalar "creative delta score".
  • This delta is added to the ad-level prediction (logit) produced by the ads ranking model.
  • The sum is supervised by the final engagement label (for example, click).

This delta-score structure matters because CCFN leaves ad-level confounding to the ranking model and only needs to predict the relative differences among creatives within an ad. Ad-level confounding here means how good the ad itself is. So CCFN can be thinner than the ranking model and use a smaller feature set.

There are four feature categories.

• Content Features: vectors precomputed by the Pinterest image and text encoders, and manual features built from domain knowledge.
  • Examples of manual features are foreground-background color distribution, foreground product size, and image aspect ratio.
• User Features: demographics and user understanding signals (for example, shopping intent, user embedding).
• Empirical Features: past performance statistics of a creative variant, aggregated into buckets according to the user and creative slicing scheme.
  • Small buckets are fine-grained but sparse and vulnerable to noise, while large buckets are stable but less specific.
  • Multi-scale aggregation uses fine-grained buckets and coarse-level buckets together.
• Context Features: reflect context dependence. For example, preferred creative brightness varies with hour-of-day, and the best layout differs between mobile and web.

C. PAM reduces the score of vertically long creatives according to vertical size to protect the whole-page metric verified

PAM multiplies the CCFN output score by a vertical-size penalty to reflect cannibalization. In the waterfall grid, a vertically long ad creative can raise the ad's own CTR, but it pushes out the display space of other ads and organic Pins that would be placed around it. It assumes the grid width is constant and uses aspect ratio as a proxy for vertical size.

• $ar_{rel} = ar_{cand} / ar_{orig}$: the relative value of the candidate creative's aspect ratio divided by the ad's original aspect ratio.
• $P(ar_{rel}) = \mathrm{clip}(1 - \tanh(k(ar_{rel} - 1)), 0, 1)$: the penalty function. It is 1 when the aspect ratio equals the original, shrinks via tanh as it grows, and is clipped to the range [0, 1].
• $S = O \cdot P(ar_{rel})$: the final score is the CCFN output multiplied by the penalty.

$k$ is a tunable penalty strength. It is set to match a target average aspect ratio using offline replay, which simulates winner selection from logged model scores and candidate aspect ratios. The paper states that the tanh form was chosen arbitrarily. It adds that other functions can be used as long as they are monotonically decreasing and have values in the range [0, 1].

In the ablation comparing with and without PAM, PAM suppresses the aspect ratio increase by about 2 pp while sacrificing only about 0.3 pp of the CTR gain. The results are limited to DCO ads (ads with multiple variants).

• Without PAM (w/o PAM), DCO ad CTR is +9.8% and aspect ratio is +7.21% relative to the No-CR baseline.
• PinDCO with PAM has DCO ad CTR of +9.5% and aspect ratio of +5.22% relative to the No-CR baseline.

That this trade is favorable from a whole-page view is confirmed by the online metrics in Section 4.2 of the paper (E below).

D. Epsilon-greedy exploration also observes creatives other than the winner, and caching and dynamic batching reduce serving latency verified

• ε-greedy exploration: instead of always greedily picking the highest score, it picks one of the remaining candidates uniformly with probability ε.
  • It explores this way because filling in the Empirical Features requires observations for creatives that are not the winner as well.
  • Also, under conditions where new creatives arrive in large numbers and user preferences keep shifting, exploration reduces selection bias.
• Serving efficiency: it uses caching and dynamic batching for serving efficiency.
  • Caching: a local cache on the model server reduces the load on the feature store. To raise the cache hit rate, scoring requests are sharded by creative id so that the same creative always goes to the same machine.
  • Dynamic batching: one ad request contains hundreds of ad candidates, and each ad has multiple creative variants. Simply concatenating them into one request is vulnerable to stragglers and lowers the parallelism of request/response handling. Stragglers are the problem where one slow candidate delays the whole. Instead, the variants are flattened, split into batches of a tuned size, and scored in parallel.
  • Ablation: removing dynamic batching increases P99 latency by +87% and P90 latency by +114%. Removing the cache increases P99 by +9.6% and P90 by +12.8%. This means dynamic batching is the far more important optimization for latency.

E. PinDCO beat the baselines on every axis of the offline and online experiments and kept the whole-page metric positive verified

CCFN was best on both offline metrics verified (Offline table below). Online, PinDCO produced the largest improvement with ad CTR +3.09% while also keeping the whole-page metric Successful Session positive verified (Online table below).

Offline (relative % change versus the No-CR baseline):

Lightweight-Only (a configuration using only the pre-selection model) also improves. However, judging from the size of its ROC-AUC improvement, it falls short for fine-grained ranking. Peri-CR is an existing baseline that trains and serves the ads ranking module and the creative module in parallel.

Successful Session is the number of user sessions that contain an action defined as a "success", such as a click, a save, or a social interaction. It is a whole-page-level metric that covers both ads and organic content.

Online (relative % change versus the No-CR baseline, two platform-level metrics):

Lightweight-Only and Peri-CR improved ad CTR but reduced Successful Session by −0.24% and −0.12%, respectively. In contrast, PinDCO raised ad CTR by +3.09% while keeping Successful Session positive at +0.04%. The authors state that this difference comes from the combination of "a better creative scoring model + PAM + exploration-driven training data". The conclusion section states the deployment explicitly: "We have deployed PinDCO in the Pinterest Advertising Platform".

"PinDCO achieves the largest gain at +3.09%, substantially outperforming both baselines. … PinDCO achieves a positive +0.04% lift in Successful Session while also delivering the strongest Ad CTR improvement. This result suggests that improving local ad engagement alone is insufficient, and that creative selection should account for broader whole-page-level effects on user experience." — PinDCO: Whole-Page Aware Dynamic Creative Optimization at Scale, 2026-07-21

1. It is a production claim that treats DCO as a ranking problem. Until now, the repository corpus had no source that treated ad creative optimization as a separate system. Meta's ads ranking lineage (From User Sequences to Scaling Laws, etc.) and X's connection of a sixth ad format, header-image-prompt-ad, in the 2026-09-09 commit (Gen4+18) covered ranking and format expansion of ads themselves. PinDCO, in contrast, takes head-on the question of which creative to pick within an ad. The signal for the trend is that, under the condition where the number of creative variants exploded after generative AI, this role emerged as a production system.
2. It is the first production claim of whole-page-aware creative ranking. On surfaces such as Pinterest, Instagram, and Threads, where the organic feed uses a waterfall grid, the vertical size of a creative directly determines the display space of neighboring slots. PAM states this interaction explicitly as an aspect ratio penalty. It also states that in the ablation it sacrifices only 0.3 pp of CTR and reduces the aspect ratio increase by about 2 pp. The signal is that an approach in the same direction as slate optimization in organic ranking (optimizing the whole set of items placed together on one screen) is also required for ad creative scoring.
3. It is a production case that separates the creative model from the ads ranking model through a delta-score structure. CCFN predicts only the delta that is summed with the logit of the ads ranking model. Ad-level confounding is delegated to the ranking model, and the creative model learns only the relative differences within an ad. Thanks to this separation, CCFN can be thin with a small feature set, and it can iterate independently of updates to the ads ranking pipeline. Netflix Foundation Model separated surfaces by attaching heads on top of a foundation. PinDCO keeps the ranking model as the base and attaches creative scoring as a residual head, so its role itself differs.

↑ Contents

2Precompressing and caching very long behavior sequences and feeding them into ad conversion prediction — Radar (Tencent)

The recently emerged target-independent compression builds a summary once per user and caches it. However, most of these methods append a query token to the end of the sequence and use bidirectional encoding. As a result, the summary can lose order information and be redundant, and it misses the heterogeneous temporal structure of "what happened recently".

ChronicleRec in this paper is a pre-train-and-transfer framework that makes it possible to precompress ultra-long user behavior sequences into target-independent Chronicle Tokens and cache them. To improve the target-independent family, it combines the following four elements.

• Recency-aware multi-granularity merge: recent events are kept fine-grained, and older events are merged coarsely.
• Interleaved causal encoding: learnable query tokens are distributed through the merged sequence and causal attention is used. Each query summarizes only the history before its own temporal anchor.
• Multi-horizon parallel branches: parallel branches that mask different recent-history windows learn complementary long-range interests.
• Mask-and-predict pre-training and Chronicle Alignment: pre-training reconstructs held-out recent behavior from older history, aligning the historical signal with near-present intent.

• In the ablation on Tencent AdLive that varied query placement and attention direction, the interleaved-causal combination had the highest GAUC (the mean of per-user AUC) at 0.8034 verified.
  • The other combinations are end-causal 0.7890, interleaved-bidirectional 0.7930, and end-bidirectional 0.7951.
  • The paper makes two conditional comparisons. Under causal encoding, interleaved placement has a GAUC 0.0144 higher than end placement, and under interleaved placement, causal is 0.0104 higher than bidirectional.
  • With end placement, bidirectional (0.7951) is higher than causal (0.7890), and no significance test is presented for this comparison.
• In the ablation on the same Tencent AdLive that removed components one at a time, GAUC dropped the most when the Chronicle Alignment stage was removed verified.
  • Removing Chronicle Alignment lowers GAUC by −0.0086, removing recency-aware merge by −0.0076, removing the causal mask by −0.0074, and removing multi-branch by −0.0072.
• On KuaiRand, even when recent history was hidden and only the distant past was left, GAUC was much higher than the baseline that sees only recent history verified.
  • When the most recent 1,000 events were fully hidden and only the older history was left, GAUC was 0.5499. The Short-Attn baseline that sees only recent history scored 0.5307.
  • This result supports, with a controlled experiment, the statement that the distant past alone still carries a self-sufficient interest signal.
• On the pCVR (predicted conversion rate) surface of Weixin Moments Ads, Chronicle Tokens were added while the ranking backbone was left as is, and a 7-day A/B test gained GMV +1.61% verified.
  • The MixFormer ranking architecture was already in place on this surface.
  • Chronicle Tokens were fed into both the NS-Token path and the S-Token path. NS-Token handles non-sequential feature interactions, and S-Token handles sequential behavior.
  • The A/B result is GMV +1.61%, 95% CI [0.678%, 2.547%], p<0.05. The confidence interval excludes 0.

"ChronicleRec increases GMV by +1.61%, with a 95% confidence interval of [0.678%, 2.547%] that excludes zero, i.e., a statistically significant improvement." — ChronicleRec, 2026-09-11

1. It confirmed order-aware compression within the target-independent family with a production A/B test. In earlier target-independent summaries, a query appended to the end of the sequence (end-of-sequence query) and bidirectional attention were standard. ChronicleRec changes these to interleaved and causal, keeping temporal order and position information inside the summary vector. On Tencent AdLive, the GAUC gain is +0.0144 for interleaved placement under the causal condition and +0.0104 for causal encoding under the interleaved condition, and the result led to production GMV +1.61%.
2. The ability to cache ultra-long sequences actually helps reduce online cost. A Chronicle Token can be computed once per user and cached, so it is not coupled to candidate scoring. The Meta HSTU and Wukong lineage (Meta recommendation models that reported scaling laws) and Netflix Foundation Model were approaches that scaled up the sequence encoder itself. ChronicleRec, in contrast, leaves the ranker backbone as is and feeds only the lifelong representation in through a cache, so the adoption cost is low.
3. The independence experiment on the distant-history signal is a data point comparable with other companies' claims. On KuaiRand, even with the most recent 1,000 events fully masked, GAUC beat the recent-window baseline by a wide margin. This observation contrasts with the design of X Phoenix (the ranking Transformer of X For You), which sets 4,096 recent actions as the cap of its attention window. It also contrasts with the design in which Netflix Foundation Model encodes the lifetime history. Cases that confirmed the statement "the distant past alone carries a self-sufficient predictive signal" with a controlled experiment are uncommon.

↑ Contents

Corrections

None.

↑ Contents

Radar

Of the 23 arXiv cs.IR items triaged in this batch (12 new, 6 cross-lists, 5 replacements), two passed the acceptance criteria.

• PinDCO is a Pinterest paper that was accepted to RecSys 2026 and launched in production.
• ChronicleRec reported a 7-day A/B test on Tencent Weixin Moments Ads.

Both were added as entries to the "New this month" section of radar.md.

Items worth mentioning in triage but not folded in are as follows.

• OneLA is a systems paper that optimizes recurrent linear-attention serving for large-beam decoding in generative recommendation.
  • Source: arXiv 2609.12399, 2026-09-11, author email domain @kuaishou.com
  • It states that it represents all beam states as a shared prompt state and compact append-only divergent transition records verified.
  • It states that, by handling this with a lightweight ancestry index and a fused GPU kernel, it speeds up end-to-end decode by 1.54–2.46× and greatly reduces memory and data movement verified.
  • It has only an offline speedup and no online A/B test, so it lacks the production claim axis of the acceptance criteria.
  • However, it is worth watching later as a serving-extension thread of the OneRec (Kuaishou's generative recommendation model) and OneRec-Thinking line.
• MIMA is a framework that adds Hungarian matching-based multi-positive exclusive assignment to the interest collapse problem of multi-interest recommendation.
  • Source: arXiv 2609.12842, 2026-09-11, author email domain @alibaba-inc.com
  • Interest collapse is the problem where multiple interest representations become similar to each other and indistinguishable.
  • It groups items that co-occur in one request into a positive set and assigns each positive exclusively to a different interest.
  • It states "significant business gains" in offline experiments on three public datasets and an industrial dataset and in an online A/B test, but the abstract has no concrete uplift numbers verified.
  • It also does not state which Alibaba surface was used. We wait for a follow-up that releases concrete figures.
• Who Are We Recommending To? Recommender Systems in the Agentic Web is a position paper by five Spotify Research authors.
  • Source: arXiv 2609.11945, 2026-07-22 v1
  • It argues that in a world where LLM agents consume and transact items as user proxies, ranking splits in two depending on whether the recipient of the recommendation is a person or an agent.
  • It appeared in the same IR listing and points in the same direction as PAMR (arXiv 2609.11942), which has a different group of authors.
  • It has no production A/B test or new mechanism, so it falls outside the acceptance criteria. We keep it as reference material for piecing together how the Spotify serving stack may change.

↑ Contents

What to check next

↑ Contents