Searcharxiv⌕ Search

arXiv subjects

Gedi Zhou

Publications and source records attributed to Gedi Zhou.

6 recordsLinked to original sources

What Unlinked Aggregates Still Prove: Ranker Comparisons from Exact and Differentially Private Item Counts

Privacy controls can prevent an evaluator from linking purchases to the recommendations users received, while a merchant can still release complete item-level purchase counts for the evaluation cohort. We ask when these counts identify which recommender performs better. The counts constrain who could have bought each item. We bound the performance difference over all assignments consistent with the counts. A reduced transportation network handles full catalogs and can certify attainable endpoints. If the interval excludes zero, it identifies the better recommender for that cohort. With differentially private counts, one confidence guarantee applies jointly to all comparisons from the same release. In two e-commerce cohorts of about 20,000 and 100,000 users, exact counts narrowed every interval compared with bounds that ignored the counts. Most decisions involved a recommender that favors popular items. For NDCG@10, which rewards ranking purchased items higher, each setting combined one of three personalized pairs with a pattern and rate of missing links. Exact counts settled 14 of 168 settings in the smaller cohort and none of 81 in the larger cohort. In the smaller cohort, exact counts broken down by user group settled two personalized pairs when 90% of users remained linked, at the cost of revealing more about individual purchases. Whole-cohort private releases with one count per item retained much of the narrowing but settled no personalized pair in either cohort. Grouping less popular items into bins let private counts settle one personalized pair in the smaller cohort.

cs.IR↗

Where the Evidence Lives: Measuring Privacy and Diagnosability of Exported AI-Agent Traces

Execution traces help diagnose AI-agent failures, but sharing them can expose customer identities. For example, detecting that an agent handled the wrong customer's records requires knowing whether the records belong to the requested customer, but not necessarily knowing who that customer is. Our executable benchmark measures what a trace export reveals and whether a proposed repair restores the correct output in synthetic multi-tenant workflows. Conventional redaction can remove evidence needed for repair while leaving encoded identifiers recoverable. Two exports preserve ownership comparisons by reporting the match as a comparison fact or by using pseudonyms consistent within each execution. Both preserve the original diagnostic program's repairs. The tested identifier matchers identify no requested customer from either export. An independently written program produces correct outputs on 39 of 42 faulty cases with comparison facts and on all 42 with execution-scoped pseudonyms. Row counts still expose identities. With prior knowledge of customers' row counts, an attacker uniquely identifies the requested customer in 97 of 240 executions under each of the 15 export policies. Removing selected counters from the comparison-fact export stops this attack and preserves the original program's repairs. The benchmark lets engineers test what their diagnostic program can repair and what an attacker can still infer from the same export.

cs.CR↗

Baseline and Incremental: User, Item, and Request Debiasing in Instagram Feed Ranking

A feed ranker scores candidates with a value model, a weighted sum of engagement predictions. Each prediction conflates a baseline, what this viewer, item or request would produce anyway, with the incremental amount by which this candidate exceeds it. The literature mostly removes the baseline. We argue that a multi-signal value model should keep the raw prediction and add the incremental part as a separately weighted term. On Instagram Feed, three debiasing families deployed independently, for viewer habit, item popularity and request context, are one parameterised operator, and a design procedure picks a point in it. Media-type (format) bias goes instead to a debiased label chosen by the same procedure. In a controlled four-arm ablation of deployed Feed terms (neither term, each alone, both), the two-term arm beats both single-term arms at production weights on point estimates in 11 of 13 judged (term, metric) pairs. Of the 26 direct contrasts, 17 are significantly for it and only 1 against it, on impressions. The ablation also suggests that the offline diagnostic should screen signals, not rank them. Such terms have served Instagram Feed production for most of a year. Beyond Feed, Meta, whose apps serve billions of people, has adopted the operator across surfaces, apps and both ranking stages. Its launches there drive sessions and daily active users, the core visitation and retention metrics.

cs.IR↗

Persistent Negatives for Adversarial Black-Box On-Policy Distillation

Black-box On-Policy Distillation (OPD) seeks to improve a student from its own generations when the teacher provides sampled responses but not token probabilities. Adversarial distillation offers one route: it learns a discriminator over prompt-matched teacher and student responses and uses its score as the policy reward. However, sampling discriminator negatives from the latest student at each step couples the learned reward to a negative distribution that changes after every policy update. We address this moving-target problem with persistent-negative adversarial distillation, a live-pool method that replaces a fraction of each discriminator batch with historical, prompt-matched teacher--student comparisons. Under matched discriminator compute, historical comparisons train the discriminator, while GRPO remains on-policy with fresh student responses. Our analysis identifies the Bayes-optimal reward as a teacher-to-negative log-density ratio and, under explicit assumptions, shows how persistent negatives anchor the discriminator and reduce reward-estimation MSE relative to fresh-negative training. Across two student families, three judges, and four judged-chat benchmarks, persistent-negative adversarial distillation consistently improves performance over current methods at matched discriminator compute. It also yields smoother fresh-policy discriminator trajectories, with fewer below-chance dips. These findings identify the discriminator's negative distribution as an important design axis in black-box on-policy distillation.

cs.CL↗

LLM-Driven Reasoning for Constraint-Aware Feature Selection in Industrial Systems

Feature selection is a crucial step in large-scale industrial machine learning systems, directly affecting model accuracy, efficiency, and maintainability. Traditional feature selection methods rely on labeled data and statistical heuristics, making them difficult to apply in production environments where labeled data are limited and multiple operational constraints must be satisfied. To address this, we propose Model Feature Agent (MoFA), a model-driven framework that performs sequential, reasoning-based feature selection using both semantic and quantitative feature information. MoFA incorporates feature definitions, importance scores, correlations, and metadata (e.g., feature groups or types) into structured prompts and selects features through interpretable, constraint-aware reasoning. We evaluate MoFA in three real-world industrial applications: (1) True Interest and Time-Worthiness Prediction, where it improves accuracy while reducing feature group complexity, (2) Value Model Enhancement, where it discovers high-order interaction terms that yield substantial engagement gains in online experiments, and (3) Notification Behavior Prediction, where it selects compact, high-value feature subsets that improve both model accuracy and inference efficiency. Together, these results demonstrate the practicality and effectiveness of LLM-based reasoning for feature selection in real production systems.

cs.CL↗

EBPO: Empirical Bayes Shrinkage for Stabilizing Group-Relative Policy Optimization

Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for enhancing the reasoning capabilities of Large Language Models (LLMs). However, dominant approaches like Group Relative Policy Optimization (GRPO) face critical stability challenges: they suffer from high estimator variance under computational constraints (small group sizes) and vanishing gradient signals in saturated failure regimes where all responses yield identical zero rewards. To address this, we propose Empirical Bayes Policy Optimization (EBPO), a novel framework that regularizes local group-based baselines by borrowing strength from the policy's accumulated global statistics. Instead of estimating baselines in isolation, EBPO employs a shrinkage estimator that dynamically balances local group statistics with a global prior updated via Welford's online algorithm. Theoretically, we demonstrate that EBPO guarantees strictly lower Mean Squared Error (MSE), bounded entropy decay, and non-vanishing penalty signals in failure scenarios compared to GRPO. Empirically, EBPO consistently outperforms GRPO and other established baselines across diverse benchmarks, including AIME and OlympiadBench. Notably, EBPO exhibits superior training stability, achieving high-performance gains even with small group sizes, and benefits significantly from difficulty-stratified curriculum learning.

cs.LG↗