2026-09-22

Other changes reviewed today:

• The two papers share authors and use the same metric, GSRR (Good Search Result Rate).
• The automated experiment protocol paper is the first to state, with production detail, that VDD, a video discovery surface, has its own retrieval stack.
• Confirmed that all authors of ANGLE, Tencent's ads retrieval paper already on the radar, are affiliated with Tencent.

1A GPU path and a CPU path co-serve ultra-large-scale personalized search — Meta

The paper calls this the "personalization–scale paradox": the full serving inventory is too large for GPU memory, and an interaction-heavy model is hard to fit into a latency budget on CPU. Few infrastructure papers on Meta search have been published. In the cited literature, there has been no production architecture statement for a long time after Embedding-based retrieval in Facebook search (EBR, KDD 2020) in 2020 and Que2Search (KDD 2021) in 2021. This source claims to resolve the paradox through co-serving orchestration rather than a new model class, and fills that gap with production A/B figures.

• The source corpus is multi-trillion scale, and the inventory loaded for serving is selected from it separately verified.
  • The 4-trillion figure is on a "before selection" basis and is not the target of the ANN (approximate nearest neighbor search) scan.
  • After passing policy, quality, language, freshness, and duplicate checks, hardware-specific selectors build the GPU snapshot and the CPU snapshot independently.
• The GPU path runs two-tower retrieval and interaction pre-ranking fused in one process over a billion-scale inventory verified.
  • Pre-ranking is a stage that scores the candidates retrieval picked once more and trims them before handing them to the main ranker.
• The CPU path handles a tens-of-billions inventory, roughly 20× the GPU side, and uses only the dot product verified.
  • It combines a dedicated dense embedding index, eager term-at-a-time scoring, bulk scanning, and finer clustering.
  • No cross features are used between user and document vectors.
• Per request, the two paths can run one only, both, or each be turned off independently. Candidates are deduped at the top-level aggregator before moving to the downstream ranker verified.

• On serving latency, the GPU path is faster at P50 and P95 and 16% slower only at P99 verified.
  • This is because it feeds 224 candidates, including interaction pre-ranking, into the shared ranker; it is not a pure kernel comparison.
  • The dense CPU path passes 59 after dedupe.
• From a capacity plan viewpoint, the CPU path delivers a 4× throughput improvement over a certain reference dense retrieval implementation. However, this compares only vector ops, excluding shared ranking, networking, and the GPU path verified.

• In the full-system A/B, model-scored relevance and substantive engagement both rise against the legacy CPU-only configuration verified.
  • When the two paths were A/B tested separately, each was also positive within its own deployment scope.
• The metric is GSRR (Good Search Result Rate), which the paper itself defines as a retention-oriented substantive engagement metric.
  • The daily assigned population is in the millions, randomly assigned per account and split roughly in half.

"We present a hybrid GPU–CPU co-serving system that resolves the paradox through orchestration rather than a new model class. A high-depth GPU pathway fuses retrieval and interaction pre-ranking over a curated online pool on the order of a billion documents, while a high-breadth CPU pathway searches an independently selected online inventory roughly twenty times larger with lightweight personalized scoring." — arXiv 2609.21281, 2026-09-18

"The system is deployed in production. A full-system A/B test against the legacy CPU-only configuration improves model-scored relevance and substantive engagement, while separate pathway experiments show positive value at their own deployment scopes." — Same source

• However, taken together with the context of the VDD paper, this "Search" spans an integrated search surface within Facebook where search leads to video consumption, rather than pure web search. The evidence is that both papers use GSRR.
• Its architecture differs from candidate generation in Facebook Feed, which this repository tracks, but it suggests that the two surfaces share a good part of their infrastructure. The evidence is that Xue et al., 2026 is cited as the "existing GPU platform".

↑ Contents

2A human-gated LLM automated experiment ran for 37 days on video retrieval — Meta

An autoresearch loop in which an LLM agent cycles through code changes, execution, and evaluation already exists, and running minutes-long iterations inside a self-contained script is well known, so that is not what this source addresses. The subject is the failure modes that arise when an automated agent tries to run hours-long executions and weeks-long campaigns in an online system where training, indexing, and serving run hourly across multiple subsystems. Examples are a code change that is actually a no-op, a leaked data window, silently changed evaluator semantics, or control and treatment passing through different serving funnels; even then the campaign can "finish" successfully, but its conclusion is not valid. The source claims to guard against this fragility with a human-gated verification protocol (verification with a gate where a person checks) called EvoPilot, and demonstrates it with a 37-day campaign on VDD's model-based retrieval system.

• A request is a triple (s, q, u), made of the seed video s, the machine-generated pivot query q, and the user/context profile u verified.
• MBR (model-based retrieval) is a dense candidate generator, not a downstream ranker, and approximates f(s,q,u,d) in two stages verified.
  • The first stage is ANN based on the two-tower dot product.
  • The second stage is conditional non-linear scoring and is optional.
• The request tower Q takes the request-side features of the pivot query, user context, and seed video, and the candidate tower D takes only candidate video features verified.
  • Both vectors are normalized, so the dot product and cosine induce the same order.
• The primary training loss is out-of-session in-batch InfoNCE. To keep co-engagement from contaminating the negatives, examples from the same session within a batch are excluded from the negatives verified.
  • InfoNCE is a contrastive loss that trains the model to distinguish the positive pair from many negative candidates.
• The interaction head G(h_r, h_d, a(d)) is a candidate-specific function that outputs a weighted binary cross-entropy on click and engagement-derived auxiliary labels verified.
  • The total loss is a weighted sum of InfoNCE and BCE.
• The relevance output of the interaction head remains for interface compatibility, but its training weight is 0 in the VDD setting verified.
• Only the embeddings and the index are republished every hour, and the weights themselves are not retrained every hour verified.
  • The D tower is applied to the servable inventory to build an ANN index at the scale of "hundreds of millions of videos" on the "existing GPU platform (Xue et al., 2026)", and the index is refreshed every hour.
• At request time, ANN picks the top P by affinity a(d), and a deterministic weighted sum of task scores keeps the top-K. The baseline configuration is P = K = 3000 with the interaction head inactive, which gives pure two-tower ANN results verified.
  • When candidate-conditioned scoring is on, the interaction head scores each of the P candidates in between.
• Above MBR sit the downstream retrieval funnel and a ranker on top of it. This source does not cover the layers above.

• The campaign ran experiments for a total of 37 days over 7 directions on the VDD MBR pipeline, varying the interaction head, data, objective function, and more verified.
• The primitive autoresearch attempt concluded that the interaction head lowers the offline hit-rate by 22 percentage points, but the conclusion was wrong verified.
  • The interaction head had originally been turned off because manual experiments confirmed no improvement, and this attempt retried that direction.
• EvoPilot traced this drop through human-gated verification; the cause was not the interaction head but a pre-existing defect where the evaluation code computed the output depth as 600 instead of 3,000 verified.
  • After the defect was fixed, the offline metric improved by 3.20 percentage points in the matched comparison.
• In a 7-day randomized online experiment instrumented on a US·Canada slice at the scale of tens of millions of accounts, GSRR on the VDD slice rose significantly by a relative +0.66% verified.
• After the experiment review, the interaction head treatment was adopted as the default serving configuration of VDD MBR and remained enabled at the manuscript cutoff verified.
• The post-study replay and mutation test filtered out invalid comparisons and let valid comparisons pass. This verifies that the controls added this time hold up under post-hoc reproduction.
• Recovering interrupted rounds from durable state and saving about 5 GPU-hours of artifact regeneration cost is a separate empirical result.

"We formalize a VDD request as a triple (s, q, u): the opened seed video s, the machine-generated pivot query q, and a user/context profile u. From a servable inventory 𝒟, the system must retrieve a short ranked list of videos worth watching next. Retrieval is conditioned jointly on (s, q, u). This differs from intent-driven search, which uses a user query alone, and from next-item recommendation, which uses s and u without a query, search relevance, or a retrieval funnel." — arXiv 2609.21257, 2026-09-18

"Publishing applies D to the servable inventory and builds an ANN index of hundreds of millions of videos on an existing GPU platform (Xue et al., 2026). Candidate embeddings and the index are refreshed hourly, but the neural weights are not retrained every hour." — Same source

"After the experiment review, the interaction-head treatment became the default VDD MBR online configuration and remained enabled at manuscript cutoff." — Same source

• It is the first statement, with production detail, that a surface inside Meta that mixes "search + recommendation" and uses a seed video and a pivot query together has its own retrieval stack.
  • It is a discovery surface on a different axis from Facebook Feed, whose axis is the friend graph, and Instagram Reels, whose axis is the playback sequence.
• This source is the first case study showing that when industry LLM automated experiments actually run on a production pipeline, a specification alone is not enough and a human-gated verification layer that catches even evaluator code defects is needed.
  • In radar.md, this angle differs from the spread of LLM-native ranking thread being watched. That thread puts the LLM inside the model architecture, and this source puts the LLM on top of experiment automation.
  • This source is also the first large-scale production statement of the latter angle.

↑ Contents

3Confirmed that all authors of the ads retrieval paper ANGLE are affiliated with Tencent — Tencent

↑ Contents

Radar

Of the new arXiv cs.IR papers triaged today, 30 were seen and 2 passed. The count is based on 30 parsed <entry> elements, with a date range of 2026-09-16–09-18. The 2 that passed are Meta's production papers posted together on the same day. The verdict used three requirements (a triangle): platform-company authors, a production A/B statement, and novelty in mechanism or scale.

• Passed (decided to fold into the radar)
  • Hybrid GPU-CPU Retrieval for Personalized Search at Ultra-Large Scale (Meta), arXiv 2609.21281. Covered in detail as source 1 today, and folded into new this month in radar.md.
  • Verify, Don't Trust (Meta, EvoPilot and VDD), arXiv 2609.21257. Covered in detail as source 2 today, and folded into new this month in radar.md.
• Rejected (not watched)
  • Auto-Bidding with Disentangled Advertiser Profiles and Train-Free Adaptation (ADAPT) (Alibaba Taobao), arXiv 2609.21308. The Taobao affiliation is confirmed, but the source is limited to auto-bidding (a system where advertisers decide bid prices) and is not novel in ranking or retrieval mechanisms. No production A/B figures are stated at the abstract level. It is not watched.
  • For the earlier papers by the Hybrid GPU-CPU Retrieval author group, only the related citations were checked.
  • The remaining papers, which are academic increments or narrow in domain, do not satisfy the requirement triangle.
    • Topics: multi-hop retrieval score confidence, autolab, news personalization interview, fashion matching, subtractive unlearning (SURF), KDD Cup 2026 UniRec challenge solution (academic), text re-ranking, retrieval-QA confidence, RTL verification, self-evolving search index, RAG evidence, agent web search, encrypted keyword search, material prediction RAG, local service AI provider recommendation, sim-to-real person retrieval, algebraic composable search, AI privacy landscape, visual document retrieval MAGIC, EHR longitudinal mental health, reusable RAG visualization, semantic layer telemetry, LLM embedding reasoning quality, financial SEC filing QA, zero-token graph RAG

Industry blog skimming is done only on Mondays. Today (Tuesday) is not a target.

↑ Contents

What to check next

↑ Contents