Training with Missed Targets in Generative Recommendation: Separating Supervision from Probability Competition

Abstract:Generative recommenders return a limited candidate set and may omit observed targets before reranking. A training strategy appends these missed targets to reranker training lists, although inference still ranks only original candidates. This operation simultaneously changes retrieved-target weight, adds supervision over appended targets, and makes the two groups compete for probability. An append/no-append comparison therefore cannot explain changes in returned-item rankings. We construct three matched losses that hold retrieved-target weight fixed while introducing appended-target supervision and group competition separately. The intermediate loss trains within both groups but normalizes them separately, preventing training-only targets from competing with inference candidates. Experiments with a released OneRec model and locally trained Amazon generators show that this competition can harm returned-item ranking. In four prespecified Amazon Video Games comparisons, removing it improved full-target normalized discounted cumulative gain (FT-NDCG) by 7.8--22.2\%; 95\% intervals over users and three of four intervals over training runs excluded zero. A conservative development-set rule selected appended-target training for two of three generators in one held-out category and rejected it for all three in another, avoiding a 1.7\% loss. Candidate completion should therefore be evaluated for each generator rather than applied automatically.

Information Retrieval (cs.IR); Machine Learning (cs.LG)

arXiv:2610.10124 [cs.IR]

(or arXiv:2610.10124v1 [cs.IR] for this version)

https://doi.org/10.48550/arXiv.2610.10124 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

configs/www_openonerec_native_pretrained_v1.json

configs/www_openonerec_pro_ads_native_v1.json

reproducibility/main12_onerec_seeds_20260915/SEALED_PROTOCOL.md

reproducibility/main12_onerec_seeds_20260915/readout.json

reproducibility/main12_onerec_seeds_20260915/run.py

reproducibility/main12_onerec_seeds_20260915/sealed_manifest.json

reproducibility/main12_rating_sensitivity_20260915/SEALED_PROTOCOL.md

reproducibility/main12_rating_sensitivity_20260915/label_summary.json

reproducibility/main12_rating_sensitivity_20260915/readout.json

reproducibility/main12_rating_sensitivity_20260915/run.py

reproducibility/main12_rating_sensitivity_20260915/sealed_manifest.json

reproducibility/main13_amazon_rating_20260915/PROTOCOL.md

reproducibility/main13_amazon_rating_20260915/readout.json

reproducibility/main13_amazon_rating_20260915/run.py

reproducibility/main13_onerec_optimizer_20260915/PROTOCOL.md

reproducibility/main13_onerec_optimizer_20260915/manifest.json

reproducibility/main13_onerec_optimizer_20260915/readout.json

reproducibility/main13_onerec_optimizer_20260915/run.py

reproducibility/main13_selection_repeated_20260915/analyze.py

reproducibility/main13_selection_repeated_20260915/readout.json

reproducibility/main14_product_20260915/SEALED_PROTOCOL.md

reproducibility/main14_product_20260915/final_readout.json

reproducibility/main14_product_20260915/fit_summary.json

reproducibility/main14_product_20260915/selection_commit.json

reproducibility/main14_product_20260915/standalone_audit.json

reproducibility/main15_onerec_convergence_20260915/PROTOCOL.md

reproducibility/main15_onerec_convergence_20260915/analyze_selection.py

reproducibility/main15_onerec_convergence_20260915/manifest.json

reproducibility/main15_onerec_convergence_20260915/readout.json

reproducibility/main15_onerec_convergence_20260915/run.py

reproducibility/main15_onerec_convergence_20260915/selection_readout.json

reproducibility/main15_product_diagnostic_20260915/analyze.py

reproducibility/main15_product_diagnostic_20260915/diagnostic.json

reproducibility/main16_gradient_pathway_20260915/PROTOCOL.md

reproducibility/main16_gradient_pathway_20260915/readout.json

reproducibility/main16_gradient_pathway_20260915/run.py

reproducibility/main16_gradient_pathway_20260915/test_probe.py

reproducibility/main17_construct_identifiability_20260916/PROTOCOL.md

reproducibility/main17_construct_identifiability_20260916/readout.json

reproducibility/main17_construct_identifiability_20260916/run.py

reproducibility/main20_policy_gate_20260916/PROTOCOL.md

reproducibility/main20_policy_gate_20260916/audit.py

reproducibility/main20_policy_gate_20260916/crossfit_home.py

reproducibility/main20_policy_gate_20260916/crossfit_home_readout.json

reproducibility/main20_policy_gate_20260916/readout.json

reproducibility/main20_policy_gate_20260916/run.py

reproducibility/main20_policy_gate_20260916/unit_results.csv

reproducibility/main23_evidence_portfolio_20260916/PROTOCOL.md

reproducibility/main23_evidence_portfolio_20260916/audit.py

reproducibility/main23_evidence_portfolio_20260916/readout.json

reproducibility/main23_evidence_portfolio_20260916/run.py

reproducibility/main24_cell_confirmation_20260916/PROTOCOL.md

reproducibility/main24_cell_confirmation_20260916/primary.json

reproducibility/main24_cell_confirmation_20260916/readout_complete.json

reproducibility/main24_cell_confirmation_20260916/selection_commit.json

reproducibility/main26_health_confirmation_20260916/PROTOCOL.md

reproducibility/main26_health_confirmation_20260916/primary.json

reproducibility/main26_health_confirmation_20260916/readout_complete.json

reproducibility/main26_health_confirmation_20260916/selection_commit.json

reproducibility/verify_packaged_results.py

requirements-hf.txt

requirements.txt

research/frozen_feedback_20260911/build_mature_dataset.py

research/frozen_feedback_20260911/mature_mask_control_core.py

research/frozen_feedback_20260911/mature_replication_core.py

research/frozen_feedback_20260911/mature_robustness_core.py

research/frozen_feedback_20260911/prepare_mature_replication.py

research/frozen_feedback_20260911/run_mature_heads.py

research/frozen_feedback_20260911/run_mature_missing_features.py

research/main11_onerec_profile_20260915/core.py

research/main11_onerec_profile_20260915/run.py

research/main12_onerec_seeds_20260915/run.py

research/main6_confirmation_20260914/common.py

research/main6_followup_20260914/__init__.py

research/main6_followup_20260914/train.py

research/sidecar_residual_20260909/ads_exact_score_prototype.py

research/sidecar_residual_20260909/ads_native_inputs.py

research/sidecar_www_revision_20260911/packed_ads_hidden.py

research/sidecar_www_revision_20260911/qualify_ads_hidden.py

research/source_mass_mechanism_20260912/core.py

research/source_mass_revision_20260914/common.py

research/source_mass_revision_20260914/data.py

research/source_mass_revision_20260914/readout.py

research/source_mass_validation_20260912/common.py

research/source_mass_validation_20260912/heads.py

src/c3id/openonerec_native.py

(87 additional files not shown) You must enabled JavaScript to view entire file list.

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.