type
Post
status
Published
date
Oct 17, 2026 05:30
slug
rec-weekly-en-2026-W42
summary
Three papers this week converge on the same underlying question: how should a recommender system allocate compute, supervision signal, and feature assets — by actual contribution rather than nominal scale? TransX decomposes sequence transduction into independently scalable components and decouples training from serving. EWM uses explanations to reconstruct credit assignment for delayed rewards. FORTE turns feature deprecation into an engineering pipeline that can run daily. Different methods, same question: which compute, which signals, which features actually earn their keep. Thread 1: Scaling in generative recommendation shifts from "stack more model" to "allocate compute." LinkedIn's TransX reports a counterintuitive finding — the encoder saturates at moderate depth under fixed memory resolution, and adding more supervision should instead move budget toward candidate-conditioned fusion. Online, it delivers +6.02% CTR and +4.45% conversion while cutting serving cost by nearly 80%. Capacity growth and serving cost can be decoupled. Thread 2: Credit assignment for delayed rewards starts borrowing external help. EWM drops temporal-proximity reward propagation. Instead, a frontier language model generates explanations of "which early events caused the current outcome," and a conservative hierarchical operator recursively redistributes reward back along those attributions. On Crafter and WordCraft, the more delayed the reward, the larger the gain over PPO and LaRe. Thread 3: Feature asset lifecycle becomes its own engineering problem. Pinterest's FORTE removes 43.3% of features from a search relevance model and cuts 29.9% of trainable parameters from a large ranking model, lifting serving throughput 19.8–25.5% with no online A/B regression. For the first time, feature deprecation has an offline pipeline that predicts online outcomes.
tags
Recommendation Systems
Weekly
Papers
category
Rec Tech Report
icon
📚
password
priority
1
Weekly Overview
Three papers this week converge on the same underlying question: how should a recommender system allocate compute, supervision signal, and feature assets — by actual contribution rather than nominal scale? TransX decomposes sequence transduction into independently scalable components and decouples training from serving. EWM uses explanations to reconstruct credit assignment for delayed rewards. FORTE turns feature deprecation into an engineering pipeline that can run daily. Different methods, same question: which compute, which signals, which features actually earn their keep.
Thread 1: Scaling in generative recommendation shifts from "stack more model" to "allocate compute." LinkedIn's TransX reports a counterintuitive finding — the encoder saturates at moderate depth under fixed memory resolution, and adding more supervision should instead move budget toward candidate-conditioned fusion. Online, it delivers +6.02% CTR and +4.45% conversion while cutting serving cost by nearly 80%. Capacity growth and serving cost can be decoupled.
Thread 2: Credit assignment for delayed rewards starts borrowing external help. EWM drops temporal-proximity reward propagation. Instead, a frontier language model generates explanations of "which early events caused the current outcome," and a conservative hierarchical operator recursively redistributes reward back along those attributions. On Crafter and WordCraft, the more delayed the reward, the larger the gain over PPO and LaRe.
Thread 3: Feature asset lifecycle becomes its own engineering problem. Pinterest's FORTE removes 43.3% of features from a search relevance model and cuts 29.9% of trainable parameters from a large ranking model, lifting serving throughput 19.8–25.5% with no online A/B regression. For the first time, feature deprecation has an offline pipeline that predicts online outcomes.
Scaling and Credit Assignment in Generative Recommendation
By 2026, generative recommendation faces two concrete bottlenecks. First: how does a sequence transduction model keep growing under a fixed serving budget? Second: when a decision's payoff shows up hundreds or thousands of interactions later, how does RL assign reward back to the step actually responsible? TransX answers from the systems side, EWM from the signal side.
TransX (LinkedIn) — The core move is splitting sequence transduction into three independently scalable components: sequence encoding, behavior memory, and candidate-conditioned fusion. That split drives every scaling conclusion that follows.
Two amortizations on the training side. Causal trajectory computation is amortized over behavior prefixes, avoiding recomputation of the full trajectory for each prefix. Memory state fans out to candidates (memory fanout), so one encoding serves scoring for multiple candidates. Together these lift part of training cost out of the sequence-length × candidate-count product.
Three-stage decoupling on the serving side. Bulk prefill, nearline refresh, and online prediction run separately, connected by a compact behavior-memory KV. Prefill throughput optimization and low-latency online prediction stop fighting each other.
Scaling returns are anisotropic. The paper reports three observations. The encoder saturates past moderate depth under fixed memory resolution and fusion capacity. Adding supervision signal pushes the compute-allocation frontier toward candidate-conditioned fusion. And the memory resolution needed to hold quality stays roughly flat as model capacity grows. That third point matters most — memory doesn't have to inflate with parameter count, which is the precondition for the serving-cost reduction.
Online numbers. The core TransX model delivers +6.02% CTR and +4.45% conversion. The post-trained retrieval variant gains +2.96% CTR; the preference-ranking variant gains +1.33% conversion. Training infrastructure lifts supervision-target throughput by 84% under equal conditions. Decoupled serving cuts total serving cost by nearly 80% versus colocated serving at equal traffic and SLO, delivers roughly 5× the model FLOPs per request at equal total cost, and at matched FLOPs per request, a single H200 handles about 4× the request throughput of the baseline.
Placed in the context of the past six months, TransX addresses a problem adjacent to but distinct from RelayGR. RelayGR uses cross-stage relay inference to fit long sequences into a fixed P99 SLO, achieving 1.5× sequence length and 3.6× throughput. TransX moves the question upstream: should compute go to the encoder or the fusion layer? On deployment, OneLoc at Kuaishou Local Life uses geo-aware semantic IDs and RL multi-objective balancing to reach +21.016% GMV and +17.891% orders — evidence that generative recommendation can carry primary metrics in vertical domains. GRLM uses structured Term IDs to gain 5–10% CTR on e-commerce datasets, taking the identifier-design route. TransX's difference: it reframes "scaling" from a model dimension into a budget-allocation dimension — under fixed serving cost, which component's growth yields the highest marginal return.
EWM — The problem statement is direct. Standard RL propagates reward by temporal proximity, so a decision whose value only materializes 800 interactions later gets its signal attenuated to near zero by the discount factor. No discount-rate tuning fixes this, because what decays is relevance, not time.
Method. Explanatory World Models use explanations to identify "which early events likely caused the realized outcome." A frontier language model generates the explanation, the explanation serves as a learning signal, and a conservative hierarchical operator recursively redistributes reward back to the implicated early actions. The constraint: preserve the undiscounted episode return. That conservatism matters — it stops the explanation model from shuffling reward around and drifting the policy optimization objective.
Three explanation variants. The paper operationalizes "what counts as a useful explanation" into three versions, corresponding to Halpern's actual causation, Lombrozo's explanatory simplicity, and Deutsch's hard-to-vary. Their relative performance in experiments is itself an ablation on explanation quality.
Experiments. On two controlled long-horizon RL benchmarks, Crafter and WordCraft, EWM substantially outperforms PPO and LaRe, and the gain grows with reward delay — exactly the method's design assumption. On the public sequential recommendation benchmark KuaiRand, it beats SFT and LaRe offline and A2C online. In offline evaluation on industrial generative recommendation, Recall improves notably and behavioral proxy metrics also improve. Note the missing online A/B — the industrial conclusion stops at offline.
Comparing against recent credit-assignment work shows the divergence in approach. ITPO takes the implicit process reward model route, deriving turn-level process rewards from sparse outcome signals and improving PPO/GRPO/RLOO convergence on math tutoring, document writing, and medical recommendation. SafeCRS uses Safe-SFT plus Safe-GDPO to cut safety violations by 96.5% while preserving recommendation quality — a preference-alignment problem, not delayed attribution. ClaHF ports RLHF's preference-ranking idea to text classification. EWM differs by not learning a process reward model at all — it has the language model do causal attribution directly, treating the explanation as reusable structured credit.
Both papers point the same way: the bottleneck in generative recommendation has moved from "can the model run" to "how do compute and signal get allocated by real contribution." TransX proves at the systems layer that serving cost need not scale linearly with capacity. EWM proves at the signal layer that delayed rewards can be re-attributed. Both require an explicit account of which part is doing the work.
Feature Lifecycle Governance in Production Recommender Systems
Industrial recommender systems carry a structural asymmetry: adding a feature must show a gain on some metric, while removing a feature must show no regression on any metric. The former has a mature offline-evaluation-plus-A/B pipeline. The latter runs on luck and manual effort.
FORTE (Pinterest) — Splits feature deprecation into discovery, validation, and removal, and delivers a model-agnostic framework.
Why deprecation is hard. The paper lists six concrete cases. A feature interacts with others, so removing it in isolation misjudges its value. Importance shifts day to day. It can't be evaluated before the production model updates or data drifts. A/B test slots compete with other launches. Shipping directly causes regressions. And a feature consumed by multiple models requires cross-model coordination to remove. These six cover essentially every reason features "can't be deleted" in practice.
Three components. An efficient perturbation implementation makes daily full-feature evaluation feasible. Grouped ablation constructs candidate feature sets from all low-rank features, specifically handling interactions and avoiding the underestimation of one-at-a-time removal. Lineage tracing tracks every downstream consuming model — the precondition for realizing full savings, since a feature deleted from the training set but still computed at serving saves nothing.
The key validation is offline predictive power for online outcomes. FORTE's offline results are good enough to predict online A/B, which is what lets candidate sets go straight into online testing. On the search relevance model it produces three nested candidate sets of 32, 48, and 59 features. The online guardrail rejects the largest, 59-feature set, and the 48-feature set ships, removing 43.3% of the model's features. On the large ranking model it identifies 8 removable candidates; retraining plus A/B confirms no regression before shipping, cutting trainable parameters 29.9%, lifting serving throughput 19.8–25.5%, and saving hundreds of thousands of dollars in annual training cost. Note the search relevance case: the guardrail vetoing the largest set shows offline prediction still has limits — the framework's value is catching failures before they reach production.
This approach complements several recent efficiency papers. UG-Separation at ByteDance uses masking to decouple user-side and group-side computation, making user-side compute reusable and cutting online inference latency 20% — the "compute less of the redundant stuff" route. LLM-I2I goes the other way, using an LLM to synthesize and filter long-tail interactions, gaining +6.02% recall and +1.22% GMV — the "fill in what's missing" route. Factorized Transport Alignment uses a lightweight optimal-transport approximation to unify multimodal multi-view inputs, fusing them into a single cached embedding at inference and improving Recall@500 by 7.9%. FORTE's angle is rarer: it treats features as assets under depreciation management and asks which investments no longer pay returns. In an industry where feature counts only go up, this reverse operation may be worth more than another model architecture.
Directions to Watch
Decoupling serving cost from model capacity. TransX's evidence chain is complete: memory resolution flat as capacity grows, 80% serving-cost reduction from decoupling, 5× FLOPs at equal cost, 4× H200 throughput at equal FLOPs. If this holds on other architectures, the constraint on "bigger models" shifts from budget to training data and organizational capability. What to track: whether other teams reproduce the "encoder saturates at moderate depth" observation, and where candidate-conditioned fusion's capacity ceiling sits. Beyond LinkedIn, the relay-inference line around RelayGR and Kuaishou's OneLoc both push systems-side optimization, but public full cost ledgers remain scarce.
Delayed-reward attribution shifts from "learn a process reward" to "attribute via explanation." EWM's gain grows with reward delay on Crafter and WordCraft, a scaling relationship that contrasts with ITPO's process-reward-model route. The real question is controllability of explanation quality — if the language model's attributions systematically favor recent events, the conservative hierarchical operator can preserve total return but not allocation correctness. The missing online A/B in the industrial setting is the biggest gap; KuaiRand's online results count only as partial evidence.
Feature governance as its own engineering discipline. FORTE's 43.3% and 29.9% show substantial inefficient investment sitting in existing features. The pipeline transfers better than any specific model because it's model-agnostic. Worth watching: how grouped ablation's candidate-set construction controls combinatorial explosion at tens of thousands of features, and whether lineage tracing lands in organizations spanning multiple teams and model repositories. Beyond Pinterest, ByteDance's UG-Separation attacks the same efficiency problem from compute reuse. The two lines may merge into a unified "existing-asset audit" toolchain.
Paper Roundup
Generative Recommendation and Credit Assignment
TransX — LinkedIn proposes a sequence transduction framework decoupling sequence encoding, behavior memory, and candidate-conditioned fusion, with amortized causal trajectories in training and decoupled prefill/refresh/prediction in serving. Online: +6.02% CTR, +4.45% conversion, +84% training throughput, nearly 80% serving-cost reduction, and roughly 5× the production baseline's FLOPs at equal cost.
EWM — Proposes Explanatory World Models: a frontier language model generates explanations identifying early contributing events, and a conservative hierarchical operator recursively redistributes reward while preserving undiscounted episode return. Outperforms PPO and LaRe on Crafter/WordCraft; beats SFT/LaRe offline and A2C online on KuaiRand; improves offline Recall in industrial generative recommendation.
Production Systems and Feature Engineering
FORTE — Pinterest proposes a model-agnostic feature evaluation and deprecation framework, using efficient perturbation, grouped ablation, and lineage tracing to support daily evaluation. Removes 43.3% of features from a search relevance model; cuts 29.9% of trainable parameters from a ranking model, lifts serving throughput 19.8–25.5%, and saves hundreds of thousands of dollars in annual training cost.