PaperlayerView exact revision on arXiv

Source-linked AI summary

Advancing Model Research in AgentX: Long-Horizon Autonomy for Industrial Recommender Systems

Shuang Yang, Zijie Zhuang, Changxin Lao, Pengbo Xu, Hanwen Xu, Yusheng Huang, Han Gao, Guanchen Wang, Tianbao Ma, Linxun Chen, Peilin Song, Xuming Wang, Chen Li, Fan Wu, Tao Wang, Zibo Zhao, Xiangyu Wu, An Liu, Fei Pan, Peng Jiang, Chen Yang, Zhaojie Liu, Wenwu Ou

arXiv:2609.30001v1cs.AIcs.IR

TL;DR

Industrial recommendation research needs to use experiment results to decide what to investigate next. AgentX-Model combines reviewed research proposals with multi-round model experiments and returns results to guide further work. In production evaluation, 560 of 636 completed model-changing experiments recorded AUC above their business baselines.

  • Problem

    Industrial recommendation research must decide which implementation, measurements, and open questions from one experiment should guide the next investigation.

    Source: §1 · p. 4

  • Method

    AgentX-Model uses a Research Agent to develop reviewed proposals and a Model Agent to investigate them across rounds, returning code, findings, and unresolved questions for subsequent research.

    Source: §1 · p. 4

  • Results

    560 of 636 completed model-changing experiments recorded AUC above their business baselines, and the five latest online A/B evaluations reported gains across acquisition efficiency, advertising spend, and watch time.

    Source: §1 · p. 4, §7 · p. 24

  • Takeaways & Limitations

    The framework organizes continued research into Reproduce, Follow-up, Composition, and Diagnose, with diagnosis gathering evidence needed to choose a repair.

    Source: §1 · p. 4

    The research is scoped to each business setting’s inputs and prediction tasks; agents do not change upstream feature collection or redefine those tasks.

    Source: §2.1 · p. 6

Abstract

from arXiv · show

Sustaining industrial recommendation research requires using the results of one experiment to decide what to investigate next. We present AgentX-Model, the next generation of AgentX's model research framework, which connects proposal development and model experimentation within sandboxes defined by business inputs and prediction tasks. AgentX-Model adopts a dual-agent architecture comprising a Research Agent and a Model Agent. The Research Agent develops independently reviewed proposals from papers and experimental findings, while the Model Agent conducts multi-round investigations and returns code, measurements, and unresolved questions. Using the returned results, the Research Agent selects a starting implementation and formulates the next research question, allowing subsequent experiments to build on earlier findings. We organize this continuing research around four actions: Reproduce, Follow-up, Composition, and Diagnose. The first three actions drive routine research, while Diagnose acquires the evidence needed to choose a repair, including for issues raised by business feedback and online evaluation, such as prediction bias measured by PCOC. Across the production evaluation, 560 of 636 completed model-changing experiments recorded AUC above their business baselines. As research continued, some experiments recorded AUC above every comparable ancestor in their lineages. The five latest online A/B evaluations across different business settings reported gains including 10-15% in acquisition efficiency, 15-20% in target-segment advertising spend, and 0.3-0.8% in watch time; the watch-time model used approximately 10% fewer FLOPs and parameters. A dependency-aware historical-replay benchmark further evaluates research allocation, with initial results showing no consistent efficiency gain from more complex scheduling when agents already analyze and select concrete candidates.

1. Introduction

Industrial recommendation research depends on using each experiment’s results to choose the next investigation. AgentX-Model connects independently reviewed proposals with multi-round experiments and feeds results into continuing research within business-defined goals and constraints.

1. Introduction

  • The Research Agent develops reviewed proposals from papers, business feedback, and experiments, while the Model Agent investigates code and training behavior across multiple rounds.This exchange lets a result become the starting point for another investigation.Source: §1 · p. 4
  • Reproduce, Follow-up, and Composition drive routine research, while Diagnose gathers evidence to choose a repair in response to observations or business feedback.The framework continues within specified goals, constraints, and human-controlled deployment gates without a new request for every experiment.Source: §1 · p. 4
  • 560 of 636 completed model-changing experiments recorded AUC above their business baselines, while some lineages produced results above every comparable ancestor.The five latest online A/B evaluations reported gains across acquisition efficiency, target-segment advertising spend, and watch time; the watch-time model used about 10% fewer FLOPs and parameters.Source: §1 · p. 4
  • Initial historical replay results found no consistent efficiency gain from more complex scheduling when agents already select concrete candidates from evidence.Fixed task-type rotation with agent-based candidate selection remained an effective baseline.Source: §1 · p. 4
  • In Scenario A, 7 of 77 Follow-ups and 5 of 120 Compositions recorded AUC above every comparable ancestor.The report also describes a dependency-aware historical-replay benchmark with 473 experiment nodes across six environments.Source: §1 · p. 5

2. System Overview

AgentX-Model connects experiments into continuing research paths within business-defined sandboxes. Its agents divide cross-experiment planning from within-experiment investigation, using proposals and retained results to coordinate their work.

2.1. A Research Sandbox Defined by the Business Setting

  • Business inputs and required predictions define the sandbox; agents may change model components but do not alter upstream feature collection or redefine prediction tasks.Each experiment records its business baseline version and evaluation conditions, and cross-setting findings must be adapted and tested in the target setting.Source: §2.1 · p. 6

2.2. Two Agents, Two Research Horizons

  • The Model Agent inspects implementations and training behavior, while the Research Agent relates findings to other experiments, papers, and business feedback.Their exchange of proposals and results connects individual experiments into continuing research paths.Source: §2 · p. 5, §2.2 · p. 6
  • A delivered proposal specifies one experiment, and connected experiments form a research path through inheritance or composition.The proposal carries the research question, starting implementation, proposed modification or diagnostic intervention, and evaluation design; shared state retains implementations, measured findings with conditions, and unresolved questions.Source: §2.2 · p. 6

3. Long-Horizon Research Cycle

AgentX-Model conducts reviewed investigations, retains evidence for later experiments, and selects research actions according to the available results and questions. Its cycle separates task-type allocation from the Research Agent’s decision to construct a concrete experiment or defer.

3.1. Conducting an Investigation

  • An Assembler develops a proposal linking the change to starting code and evaluation, while a separate Auditor checks whether the design answers the question.After approval, the Model Agent plans and verifies code, trains and measures the model, and returns each round’s code and findings, retaining useful intermediate models when later revisions lose gains.Source: §3.1 · p. 7

3.2. Continuing from Experimental Evidence

  • Continuations retain round-specific code with measurements, evaluation conditions, and diagnostic observations, rather than overwriting the best measured implementation.The proposal separates starting code from evaluation references: a repair may start from a diagnostic model while retaining an earlier ranking result as its reference.Source: §3.2 · p. 7
  • Baseline, direct-parent, and strongest-comparable-ancestor results answer different questions; path gains require complete comparable ancestry.A repaired descendant can show positive parent gain while its path gain remains negative.Source: §3.2 · pp. 7–8

3.2.1. Four Research Actions

  • Reproduce tests a paper-derived mechanism in the target setting, Follow-up investigates a specific opportunity from measured code, and Composition tests whether distinct changes work together.Composition accounts for shared changes and conflicts, aiming to exceed the best comparable result from either source path rather than merely the business baseline.Source: §3.2.1 · pp. 8–9
  • Diagnose uses observations or targeted interventions to distinguish explanations for a failure and acquire evidence for a subsequent repair.A calibration task can vary loss weight while tracking learned correction and prediction statistics to assess whether stronger calibration pressure addresses observed bias.Source: §3.2.1 · pp. 8–9

3.3. Selecting the Next Investigation

  • Human-designed rules allocate action types, while the Research Agent examines shared results, papers, and business feedback to select sources, formulate a question, and design an experiment.The agent defers when evidence does not support a concrete experiment; resource and eligibility rules also constrain candidates and execution.Source: §3.3 · pp. 9–10

4. Evaluation and Findings

This section evaluates AgentX-Model across production proposal delivery, model outcomes, calibration repair, and cross-setting research, and introduces historical replay to examine experiment selection. It reports outcomes against business baselines and compares allocation policies, candidate availability, and the effects of summarized evidence in replay.

4.1. Experimental Setting

  • Production evaluation covered Scenarios A–D, while historical replay also included E; fixed-seed repeated-run AUC standard deviations varied by setting.The pooled standard deviations were 0.002208 for Scenario A, 0.001010 for Scenarios B/C/E, and 0.001099 for Scenario D. The implementation used Pi and Claude Code with privately deployed GLM-5.2 in production, and GLM-5.3 for subsequent replay processing and assessment.Source: §4.1 · p. 10, §4.1.1 · p. 11

4.2. Sustained Operation in Production

  • In a sampled 10-day window, 262 of 886 proposal attempts were delivered, and 177 of 200 frozen-subset proposals yielded verifiable measurements.The Model Agent returned outputs for 198 of the 200 frozen-subset proposals. Producing them took a median 33.6 minutes and a 90th percentile of 49.9 minutes, excluding queueing and training.Source: §4.2 · p. 11, Table 3 · p. 11

4.3.1. Performance Gains Across Evaluation Settings

  • 560 of 636 completed model-changing experiments exceeded their business baselines, and all four evaluation settings produced models above baseline.The five latest Launch Reviews report relative gains in acquisition efficiency, target-segment advertising spend, and user engagement; LR 3 also reduced cost by pruning auxiliary training objectives unused in online prediction.Source: §4.3.1 · p. 12, Table 4 · p. 12, Table 5 · p. 13

4.3.2. Evolution Through Follow-up and Composition

  • In Scenario A, the cohort best AUC rose only 0.000185 after day T+6, as proposals shifted toward loss, optimization, and simplification questions.Among comparable Follow-ups, 46 of 91 exceeded their direct parent, while 7 of 77 with complete comparable ancestry exceeded every ancestor. Of 189 multi-round experiments, 111 improved after their first valid round and 91 ended below an earlier best, motivating preservation of best-round code and evaluation context.Source: §4.3.2 · pp. 13–15, Figure 3 · p. 14, Table 7 · p. 15
  • A Scenario A Composition reached 0.779965 after non-affine LayerNorm reduced diffusion-loss spikes, exceeding its first round by 0.005842 and its strongest ancestor by 0.001059.The first round scored 0.774123; the LayerNorm was the only implementation change between rounds. The source experiments adapted LGCD [20] and AFTM from FuXi-𝛽[21].Source: §4.3.2 · pp. 15–16, Figure 5 · p. 16

4.3.3. From Ranking Gains to Calibration Repair

  • After business feedback flagged prediction bias, a Diagnose increased calibration-loss weight from 1 to 10 but left test PCOC at 0.953836.A subsequent Follow-up estimated a multiplicative correction from training calibration statistics; its two healthy runs reduced test calibration error by 22.8% and 22.7%, with ranking AUC still above baseline. The path motivates staged optimization that retains ranking gains while meeting business constraints.Source: §4.3.3 · pp. 16–17

4.4. Knowledge Transfer Across Settings

  • In an exploratory matched-baseline Scenario D cohort, Knowledge Transfer had five above-baseline results among 15 experiments, versus none among seven paper-derived Reproduce experiments.The largest increment was 0.000789; the cohorts came from different periods without matched experiment budgets.Source: §4.4 · p. 18, Table 9 · p. 18

4.5. Selecting the Next Experiment

  • The historical-replay benchmark compares experiment-allocation policies on dependency graphs where selecting a candidate reveals results and can unlock later experiments.Fixed routing, bandit routing, and joint selection retain agent-based candidate selection; replay outcomes are limited to previously executed branches. The supplied passages describe the benchmark and run setup, not comparative policy outcomes.Source: §4.5 · pp. 18–19
  • All 60 runs on Scenarios A, B, C, and E reached the baseline-plus-0.001 target, including 46 on the first selection; neither Scenario D graph reached it.Scenario D’s Transfer graph gained 0.000789 (0.810445 from a 0.809656 baseline); its Reproduce graph had no result above baseline, and all 12 runs exhausted their candidate pools.Source: §4.5 · p. 19
  • Final AUC rankings varied: Uniform Random’s Scenario A mean exceeded Fixed, Greedy, and Joint, while Bandit led on the two larger graphs; no policy led consistently.Bandit’s Scenario A mean was strongly influenced by one run reaching 0.788360; its sample standard deviation across three runs was 0.005092. Fixed rotation with agent-based candidate selection remained competitive.Source: §4.5 · pp. 19–20, Table 10 · p. 20
  • 0.819780 was reached by all five policies in all three Scenario C repeats after 7–20 selections; the Composition combined results of 0.818920 and 0.819011, both above the 0.816849 baseline.The 21-node graph’s 20-selection budget covered nearly the entire graph, and the minimum number of selections needed was three.Source: §4.5 · p. 20, Table 11 · p. 20
  • Joint reached the shared Scenario C endpoint after an average 10.67 selections (range 7–16), versus Fixed’s 14.67 (12–16), across three repeats.Both policies reached the same model through different sequences of prerequisite and other experiments.Source: §4.5 · p. 20, Table 11 · p. 20
  • In Scenario A, none of 15 runs selected any of 83 Compositions; all nine agent runs lacked an eligible Composition because its source experiments had not all been selected.Unlike Scenario C, these runs never unlocked a combination, motivating evaluation of experiments by the research opportunities they enable.Source: §4.5 · p. 20
  • At 110 of 120 checkpoints, a better recorded result remained reachable within budget, yet 19 were followed by no higher AUC; the other 10 had no better reachable result.None of those 10 checkpoints was limited solely by unfinished-dependency costs, so a flat search curve could indicate either an unused opportunity or the recorded pool’s limit.Source: §4.5 · p. 21
  • The 36 agent runs took approximately 430 minutes and used 97.5 million metered tokens, including 86.7 million cache-read tokens; Uniform Random and Greedy required no language-model calls.The agent-run totals were about 124 minutes for Fixed, 157 for Bandit, and 149 for Joint; these replay costs are separate from historical training costs.Source: §4.5 · p. 21
  • Summaries reduced non-positive-gain selections in both Scenario E batches, but Scenario A positive selections were unchanged or lower, and summaries did not consistently improve best AUC or reach the same result sooner.In one comparative-summary Scenario A pair, the summary condition’s best AUC was 0.017284 higher; other pairs showed mixed outcomes, and Scenario E had mixed timing and final-AUC results.Source: §4.5 · p. 21, Table 12 · p. 22

5. Related Work

Prior work explores agentic recommendation research, reuse of experimental experience, and policies for allocating research effort. AgentX-Model builds on these lines by studying how measured implementations and findings can guide continued model research and how replay can compare allocation choices.

Agentic research for recommendation

  • Agentic recommendation work spans prototype-based study implementation, semantically verified architecture changes, trial allocation, online configuration, and human-gated evaluation.Examples include AutoRecLab [7], NOVA [3], RecHarness [4], CORAL, AutoLR [9], and RecSys Factory [29].Source: §5 · p. 22

Long-horizon research and experience reuse

  • AgentX-Model studies sustained research paths by carrying returned implementations and findings into the next proposal’s question, starting code, and evaluation target.Its foundation includes AgentX’s paper-driven exploration, multi-round experimentation, and cross-paper combination [1], alongside evidence records linking conclusions to code, measurements, and conditions [2].Source: §5 · p. 22
  • Industrial recommendation systems reuse operational experience through asynchronous experimentation, shared memory, and feedback loops for execution and idea development.Auto-RecSys [6] uses outcomes to guide proposals and experiment history to filter duplicates and develop combinations of partial successes.Source: §5 · p. 22
  • Cumulative-search systems preserve evaluated solutions or shared research state so earlier results can inform later investigation.Examples include evolutionary program improvement [12], empirical-software tree search, and Git-backed contributions with dependencies and verification records [8].Source: §5 · pp. 22–23

Research selection and historical replay

  • Research-allocation methods separate solution quality from the policy that chooses which opportunity receives the next experiment budget.Prior approaches compare search policies [14], predict candidates worth executing [15], or replay recorded discovery trees [16].Source: §5 · p. 23
  • The replay benchmark compares allocation among Reproduce, Follow-up, and Composition while respecting experiment dependencies.Fixed task-type rotation is the baseline, with agents choosing candidates within the allocated type; bandit and joint selection vary allocation.Source: §5 · p. 23

6. Lessons from Long-Horizon Model Research

The lessons organize long-horizon research around preserving the state needed to continue, letting evidence refine the research question, and matching decisions to evidence and dependencies. They connect concrete practices for retaining implementations and evaluation context with guidance for selecting and evaluating future experiments.

L1. Preserve the best measured implementation alongside the latest research state

  • Retain the best measured implementation alongside the latest research state because later rounds may finish below an earlier best.Keep the latest diagnosis and failed attempts too, since they explain what has been tried since that model was obtained.Source: §6 · p. 23

L2. Distinguish parent-level AUC gains from a new path-best result

  • Distinguish gains over a direct parent from a new path-best result, and specify starting code separately from the reference used to judge progress.In Scenario A, continuations more often exceeded direct parents than all comparable ancestors; a repair may need diagnostic code rather than the strongest ancestor.Source: §6 · p. 23

L3. Carry evaluation context together with implementations

  • Carry predictions, labels, sample population, metric, and comparison results with the code because these details determine what a continuation tests.In the calibration case, review caught a draft targeting a different prediction task from the intended repair.Source: §6 · p. 24

L4. Let business feedback update the research objective

  • Let business feedback update the research objective while specifying both the new requirement and the performance to preserve.In the calibration path, feedback redirected research from ranking gains toward prediction bias, requiring diagnosis and a repair that checked retained ranking performance.Source: §6 · p. 24

L5. Treat candidate availability as part of experiment selection

  • Account for candidate availability when selecting experiments, since Composition candidates may remain unavailable until their source experiments are selected.In the Scenario A replay, preferring Composition alone did not make its source experiments available.Source: §6 · p. 24

L6. Evaluate experience reuse according to the decision it supports

  • Evaluate experience reuse against the specific decision it supports, separating target-setting model ideas, implementation guidance, and selection guidance.Transferred experience produced proposals evaluated in Scenario D, while summaries did not consistently improve best AUC or shorten selection paths; instruction revisions improved first-draft implementation plans.Source: §6 · p. 24

7. Conclusion

AgentX-Model carries experiment results into subsequent model research within business-defined sandboxes. Production records and online evaluations report model improvements, while the research paths illustrate how results and business feedback guide further work.

7. Conclusion

  • Online A/B tests reported gains in acquisition efficiency, target-segment advertising spend, and engagement, with task pruning reducing model cost.Production records document sustained operation and offline model improvements within business-defined sandboxes.Source: §7 · p. 24
  • Continuing research entails retaining useful measured implementations, choosing references that distinguish inherited performance from new progress, and pursuing questions raised by results.Business feedback redirected one path from ranking gains to diagnosis and offline calibration improvement.Source: §7 · p. 24
  • Fixed task-type rotation with agent-based candidate selection remained competitive in historical replay, while missing source experiments could leave useful combinations unavailable.The conclusion identifies pursuing experiments that make useful continuations possible as a next selection question.Source: §7 · p. 24

Appendices

The appendix describes an independently reviewed proposal loop in which an Assembler develops and revises a proposal and an Auditor assesses it before delivery. Approval and automated checks gate delivery, while the process retains review history for later analysis.

A.1. Proposal Refinement Mechanism

  • Proposal refinement addresses concrete design errors, including a calibration draft aimed at the wrong prediction task and a Composition that repeated one source.These cases motivate review before committing training resources.Source: §A.1 · p. 28
  • The Figure 6 loop takes assigned actions and shared evidence as inputs, then constructs, independently reviews, and revises a proposal before approved delivery.Execution limits are enforced by the harness.Source: Figure 6 · p. 28
  • The Assembler develops the proposal while a separate Auditor returns issues that must be addressed before delivery.Both roles receive baseline inputs, prediction tasks, model structure, and supporting code references, but retain distinct responsibilities.Source: §A.1 · p. 28
  • The Auditor checks whether the proposed change and evaluation can answer the research question, including source contributions in Compositions and potentially contradictory observations in Diagnoses.Its assessment considers supporting evidence, expected observations, and unresolved assumptions.Source: §A.1 · p. 28
  • After review, the Assembler addresses the most important issue and resubmits, while time, revision, and no-progress limits stop attempts that cannot secure approval.The agent can consult further when revisions reach a new code path or conflicting evidence, or defer if support remains insufficient.Source: §A.1 · pp. 28–29
  • Delivery requires Auditor approval plus automated checks for required information, source references, and matching model versions.Only the exact reviewed version is delivered; later edits require another review, and drafts, feedback, and revisions are retained.Source: §A.1 · p. 29

A.2. Proposal Revision Cases

Independent review revised proposals against source experiments, while the research workflow carries evidence through model iterations, process repair, and reusable implementation memory.

A.2. Proposal Revision Cases

  • Auditor review corrected calibration targets or composition designs and added outcome-aligned checks, even though both initial drafts passed automated validation.The reviews checked proposals against parent reports and code.Source: §A.2 · p. 29

A.3. Workflow Evolution and Development Observations

  • The Research Agent chooses evidence and revises proposals within a task type, while the harness retains control of permissions, budgets, checks, and delivery.The agent-driven workflow replaces code-prescribed investigation and repair choices while retaining the same independent Auditor and execution boundaries.Source: §A.3 · p. 29, Figure 7 · p. 30

B.1. Multi-Round Model Research

  • The Model Agent runs repeated planning-to-evaluation cycles within the reviewed question’s scope, using results and unresolved questions to guide later rounds.Round code and measurements are retained, including useful intermediate models.Source: §B.1 · pp. 30–31, Figure 8 · p. 31
  • Structural and fidelity checks, smoke tests, and independent evidence verification gate implementations before training; exhausted repair budgets leave a failure record.First-attempt check results remain available after successful repair to track recurring defects.Source: §B.1 · p. 31

B.2. Self-Evolution Through Process Repair

  • Recurring failures are grouped by deterministic signatures, then routed to the responsible component; a missing-library error can be classified as an environment failure eligible for retry.The retry classification changes the system’s response, not the external environment itself.Source: §B.2 · p. 32
  • Tool patches replay archived cases, while instruction changes generate fresh outputs under isolated versions and held-out acceptance cases before human approval.Replay checks target decisions and preserved controls; fresh-output tests check targeted defects and other errors.Source: §B.2 · p. 32
  • After deployment, later execution records are checked for recurrence, which can reopen the repair investigation.The test split and acceptance rules remain outside the repair agent’s control.Source: §B.2 · p. 32

B.3. Agent-Initiated Memory Extraction and Reuse

  • The Model Agent can extract implementation memories for human review; approved guidance records its source tasks, supporting evidence, and applicability conditions.For later tasks, the agent checks whether the memory applies before using it in planning and coding.Source: §B.3 · p. 33

B.4. Instruction Revision Evaluation

A fixed-batch regression evaluated shared instruction revisions on first-draft plans, finding higher structural pass rate and semantic fidelity with no unacknowledged omissions in the batch.

B.4. Instruction Revision Evaluation

  • The regression used the same 40 paper-reproduction inputs whose earlier failures had informed the instruction revision.These shared instructions generate plans across tasks.Source: §B.4 · p. 33
  • Each input generated one new plan under each instruction version, enabling a first-draft comparison before per-plan repair.The revised instructions improved first drafts before any per-plan repair.Source: §B.4 · p. 33
  • The Model Agent proposed changes to its shared plan-generation skill and constraint checklist, and a human reviewed and approved them.Source: §B.4 · p. 33
  • Evaluation combined deterministic checks of required plan elements with an LLM judge’s assessment of fidelity to input mechanisms.It also counted omissions of required mechanisms that the plan did not acknowledge.Source: §B.4 · p. 33
  • Structural pass rate increased by 57.5 percentage points and semantic fidelity by 22.5 points, with no remaining unacknowledged omissions in this batch.Table 13 compares first drafts before and after instruction revision on the same 40 paper-reproduction inputs.Source: Table 13 · p. 33, §B.4 · p. 33

C. Selected Findings from AgentX-Model Research

C.1. Mem-GF: Parameter-Free, Sample-Adaptive Feature Weighting

The section reports a parameter-free, sample-adaptive feature-weighting branch and a separate quantization adaptation that combines learned prototypes with continuous features.

C.1. Mem-GF: Parameter-Free, Sample-Adaptive Feature Weighting

  • Mem-GF adapts graph filtering to feature tokens within each sample, using token relationships to adjust ranker inputs without an added weighting network.The branch applies an implicit similarity operator through matrix–vector products and modulates tokens before residual fusion.Source: §C.1 · p. 34, Figure 9 · p. 34
  • In Scenario E, AUC increased from the business baseline’s 0.815403 to 0.820512, a gain of 0.005109, with no trainable parameters added by the branch.The branch adjusts feature representations using sample-dependent weighting and coordinate rescaling.Source: §C.1 · p. 34

C.2. TokenMinds-Inspired Quantization: Pairing Features with Shared Prototypes

  • The quantization adaptation retains continuous features while directly quantizing intermediate tokens into learned prototypes, whose fused vectors join original tokens in the prediction network.Continuous features serve as queries and keys, and quantized features as values; the prototypes are learned vectors rather than predefined user categories.Source: §C.2 · p. 35, Figure 10 · p. 35

C.3. LGCD with AFTM: Generating Missing Semantic Profiles

The Scenario A implementation combines generated semantic representations with AFTM-inspired mixing in the denoiser, and outperformed both the business baseline and strongest direct parent on the reported metric.

C.3. LGCD with AFTM: Generating Missing Semantic Profiles

  • The adaptation generates semantic vectors from business features using existing embeddings as training targets, then supplies generated representations for users with or without profiles.The ranking path uses an initial-state predictor followed by a short reverse diffusion chain; the generated representation is fused back into the ranker.Source: §C.3 · pp. 35–36, Figure 11 · p. 36
  • Because LGCD replaces the vector-quantization path, the composition moves AFTM-inspired mixing into the denoiser, using noisy state and diffusion time to gate the user condition.With one condition token, the adapted gate modulates each condition dimension rather than relying on standard cross-attention’s unit softmax weight.Source: §C.3 · p. 36
  • The complete implementation reached 0.779965, versus the business baseline’s 0.773683 and strongest direct parent’s 0.778906.The recorded margins were 0.006282 over the business baseline and 0.001059 over the strongest direct parent.Source: §C.3 · p. 36

C.4. ORQ with DIGER: Changing How a Quantizer Selects and Updates Prototypes

This composition changes prototype selection and updates within an existing quantizer while retaining its progressive compression structure. Its best complete implementation exceeded both direct parents and the business baseline.

C.4. ORQ with DIGER: Changing How a Quantizer Selects and Updates Prototypes

  • The composition retains ORQ’s progressive compression structure while inserting DIGER-inspired exploration into existing code-assignment steps, motivated by concentrated code usage.ORQ rotates features, selects dimensions for quantization, and passes remaining dimensions and residuals to the next level.Source: §C.4 · p. 36
  • Training adds Gumbel noise to frequently selected codes; gradients follow soft assignments, which also inform codebook updates, while inference removes noise and selects one vector.Dimension selection uses a hard mask that does not pass gradients to its scoring network through that operation.Source: §C.4 · p. 36
  • 0.778186 exceeded the stronger ORQ parent’s 0.775863 by 0.002323 and the 0.773683 business baseline by 0.004503.The DIGER parent reached 0.775681.Source: §C.4 · p. 37
Loading 2609.30001v1…