Adaptive Loss Balancing for Noise-Robust GRPO in Generative Recommendation

Reinforcement learning (RL) presents a promising avenue for enhancing generative recommendation beyond supervised imitation, leveraging reward signals to guide policy improvement. However, its efficacy is critically contingent on the trustworthiness of the reward model for the samples it evaluates. In practice, production rankers, the widely adopted reward models, are trained on exposure-biased logs, leading to sample-dependent inaccuracies that violate this assumption. Our stratified analysis uncovers a consistent pattern: reward guidance is most beneficial when the policy exhibits uncertainty and the ranker can effectively discriminate the ground-truth item from rollout negatives. On other samples, the reward signal is either negligible or detrimental, highlighting the risk of uniform RL application. To address such an issue, we introduce AdaGRPO, a novel framework that treats reward-guided optimization as selective admission rather than uniform pressure. Training is anchored in supervised negative log-likelihood, while the GRPO objective is gated by a binary, per-sample clip determined by two rollout diagnostics: policy-side difficulty and reward discriminability. Instances failing either diagnostic default to pure supervision, ensuring stability and mitigating the amplification of noisy gradients. We validate AdaGRPO on a large-scale e-commerce dataset. At the best intermediate checkpoint, it elevates HR@10 from 11.01% to 12.18% while constraining hallucination below 0.22%, and maintains robustness at the final checkpoint (HR@10 11.63%, hallucination 0.27%), outperforming fixed NLL--GRPO mixtures across the retrieval--validity frontier. In production A/B tests, AdaGRPO achieves statistically significant gains in click-through rate and dwell time, confirming its practical utility.

ToolRec: Calibrated Preference Alignment for Query Recommendation in On-Device Assistants

Large Language Models (LLMs) have significantly advanced generative query recommendation. However, existing alignment methods primarily focus on standard chatbot scenarios, falling short in on-device intelligent assistants where users predominantly expect the rapid invocation of system-level tools. Moreover, directly aligning LLMs with real-world click logs introduces severe noise due to varying user activity levels and the failure to emphasize execution-oriented queries. To address these challenges, we propose ToolRec, a calibrated preference alignment framework tailored for on-device query recommendation. To ground query recommendation with executable actions, we first construct SysToolKit, a comprehensive repository of 708 system tools, paired with a context-aware tool retrieval mechanism to ensure recommendation relevance. We then propose a dual-level calibration mechanism to refine raw click data, effectively mitigating user behavioral noise by calibrating signals based on user activity levels, while simultaneously up-weighting click signals on system-level tool-invoking queries. Guided by these refined preference signals, we then align the model using a sample-level weighted Kahneman-Tversky Optimization (KTO). Extensive online A/B tests on our mobile assistant platform OPPO Xiaobu, which has over 150 million monthly active users, demonstrate that ToolRec can significantly improve Click-Through Rate (CTR) and total clicks volume over strong baselines while maintaining high query relevance.

Hacking Generative Perplexity: Why Unconditional Text Evaluation Needs Distributional Metrics

Diffusion and continuous flow-based language models have emerged as the leading non-autoregressive alternatives to language modeling. Progress in both paradigms is overwhelmingly tracked by generative perplexity (gen-PPL): the per-token negative log-likelihood of samples under a frozen autoregressive (AR) scorer such as gpt2-large, typically paired with an empirical-entropy guardrail to rule out low-entropy collapse. We argue that this metric is unsound. By construction, gen-PPL measures only predictability under the scoring AR, not grammaticality or semantic coherence -- and the set of predictable but still low-quality sequences is combinatorially large. To make this concrete, we construct a suite of zero-parameter, deliberately naive samplers that achieve state-of-the-art gen-PPL on LM1B and OpenWebText at non-degenerate entropy, surpassing recently published diffusion and continuous-flow models while producing text that is incoherent by construction. We recommend evaluation suites that directly quantify the distributional divergence between generated and reference text, and use such a suite to re-benchmark recent non-autoregressive models, recovering a more faithful picture of the current state of the art.

Public Machine Learning Solver Framework for Novices in the Machine Learning Domain

Solving machine learning problems is complex and typically reserved for experts. Over the past two decades, systems have emerged to support non-experts. Based on our review, we identify three categories: (1) fully automated AutoML systems, (2) expert cheat sheets for algorithm selection, and (3) decision-support systems using selection criteria (accuracy, transparency, data requirements). We propose a new platform combining categories 2 and 3 to deliver semi-automated, intelligent solution recommendations for non-experts. Unlike existing approaches that recommend a single algorithm, our platform suggests a complete pipeline tailored to the user's problem. It integrates expert-defined selection criteria with transfer learning and automatically extracts data characteristics (e.g., class imbalance, missing values) from user-provided datasets. The platform uses first-order logic to reason over its knowledge base and recommends suitable algorithms ranked by relevance. It features a user-friendly interface and connects to a crowdsourcing platform for ML experts, ensuring continuous updates. The platform is built incrementally, allowing seamless integration of new algorithms, criteria, and domain knowledge. To our knowledge, this is the first free, publicly accessible online framework that systematically captures and operationalizes expert knowledge to guide non-experts in solving ML problems in a structured, transparent manner.

How Small Can You Go? LoRA Fine-Tuning 270M-8B Models for Merchant Information Extraction in Financial Transactions

Financial transaction processing requires extracting structured merchant information from noisy, abbreviated bank transaction strings at scale. Our current production system, a LoRA-fine-tuned LLaMA 3.1-8B, achieves 96.95% F1 on this task, but deploying 8-billion-parameter models imposes prohibitive memory, latency, and cost constraints. To identify more efficient alternatives, we conduct a deployment-focused study of 24 model variants spanning four model families: Gemma 3 (270M, 1B, 4B), Qwen 3.5 (0.8B, 2B, 4B), Aya (3.35B), and LLaMA 3.1-8B, systematically evaluating accuracy, inference throughput, training cost, and hardware behavior to assess production suitability. Our findings show that: (1) reproducing the LLaMA 3.1-8B fine-tune with a LoRA rank of 8 achieves 96.75% F1, only 0.20 points below the rank-32 baseline; (2) Qwen 3.5 4B with JSON-only prompting reaches 96.60% F1, within 0.35 points of the 8B baseline while using roughly half the parameters; (3) the 0.8B Qwen 3.5 model achieves 94.75% F1, matching models 2.5-4x larger and offering an attractive latency-accuracy trade-off; (4) chain-of-thought fine-tuning generally improves F1 by 0.3-1.8 points across most models, although Qwen 3.5 4B performs best with direct JSON-only prompting; and (5) Qwen 3.5 Think and Nothink training templates produce nearly identical results (F1 differences <0.004), indicating that explicit reasoning supervision is unnecessary for structured extraction tasks. We further deploy all 14 fine-tuned sub-8B models as Databricks Model Serving endpoints and observe that benchmark performance transfers reliably to production, with an average F1 change of only 0.8 points. Aya 3.35B, based on the Cohere2 architecture, is the sole exception, exhibiting a 3-5 point decline under serving conditions. Based on these results, we provide deployment recommendations across accuracy and latency requirements, ...

DeRes: Decoupling Residual Stability and Adaptivity for Scalable CTR Prediction

Transformer-based CTR models face a growing bottleneck at the residual connection: under Pre-Norm, early user-interest signals are diluted layer by layer; the identity skip cannot forget stale interests; and each layer sees only its immediate predecessor, losing long-range cross-layer dependencies. Recent attention-based residual variants (AttnRes) address parts of this in language models, but drop the protective identity skip and have not been tried in recommendation. Drawing on Dual Path Networks (DPN) and the HORNN view of residuals, we present DeRes, which routes each layer through two parallel paths -- an Identity residual path that preserves first-order feature reuse and gradient flow, and a Block Attention Residual path that attends over compressed outputs of all earlier blocks for high-order recall. A vector-wise gate decides, per hidden dimension, the weight given to each path. We further propose Pointwise AttnRes, replacing the Softmax in the cross-layer attention with SiLU so that multiple past blocks can be activated simultaneously and irrelevant ones receive negative (forgetting) weights -- better aligned with CTR's parallel multi-interest patterns. On a large-scale industrial dataset (331M interactions from a major social-media platform), Criteo (45M), and Avazu (40M), DeRes outperforms twelve baselines including OneTrans, TokenMixer-Large, UniMixer, mHC, and AttnRes, achieving up to +0.32% AUC at under 5% extra FLOPs. Beyond a single operating point, DeRes fits a markedly steeper compute-AUC scaling law (gamma=0.118 vs. 0.071 for OneTrans, a 1.66x gap), so an 8-layer DeRes matches a 16-layer OneTrans -- about 2x compute saving at equivalent AUC. Ablations confirm that the dual-path design outperforms either single path, Identity beats learnable residuals, and SiLU beats Softmax.

Feasibility to detect rapid change and disappearance of seagrass: Lessons from nearly 80 years of vegetation change in the Ako, Seto Inland Sea, Japan

This study analyses the Ako tidal flat in the Seto Inland Sea, Japan, where nearly all Zostera marina disappeared within a single year in 2025. Using aerial photographs from the 1940s onward, high-resolution satellite imagery, GRUS images (2.5-5 m), and monthly Sentinel-2 composites (10 m), we reconstructed approximately 80 years of seagrass distribution. YOLO-based segmentation using deep learning achieved high accuracy (overall accuracy >= 0.9) across these datasets; although species could not be discriminated, the models captured the major temporal dynamics in vegetation area. The long-term mean seagrass area was 6.8 ha, but values fluctuated widely, from 3.5 ha in 1974 to 41.3 ha in 1989 except 0.2 ha in 2025. Sentinel-2 composites from 2019 to 2026 revealed clear seasonality, with vegetation increasing in early summer and declining from autumn. In 2025, however, the area decreased sharply after summer and remained anomalously low throughout the winter of 2025-2026. Our results, indicating that the 2025 event was not a normal fluctuation but a rapid ecosystem shift involving the loss of the dominant canopy-forming species, most plausibly driven by regionally elevated summer water temperatures. The findings also have implications for seagrass Essential Ocean Variables (EOVs) and the State of Nature (SoN) metrics used in TNFD-aligned nature-related disclosures. Unlike forests, seagrass meadows require finer temporal resolution because both pronounced seasonality and abrupt collapse strongly influence area-based indicators. Therefore, in addition to previously noted issues such as species-level classification accuracy, we recommend that (1) baselines be defined over the longest available record and justified ecologically, (2) seasonal standardization be applied before inter-annual comparisons, and (3) years with extreme area anomalies be flagged rather than used as reference points.

OneFeed: A Unified Generative Framework for Feed Content Enhancement and Query Generation

Modern feed recommendation and search systems are deeply connected in user behavior butare usually modeled by separate architectures. Feed recommendation mainly captures implicitinterests from browsing interactions, while search systems rely on explicit user queries to retrieveintent-matched content. This separation causes fragmented user understanding and missedopportunities for using feed interactions to improve query generation and using generated queriesto enhance feed candidate retrieval.In this paper, we propose OneFeed, a unified generative framework for jointly modelingfeed content enhancement and query generation. OneFeed encodes heterogeneous user behaviorsequences with a shared behavior encoder and employs two generative heads: a Feed SemanticID Generator that produces content semantic IDs for recommendation retrieval, and an IntentQuery Generator that produces natural-language queries for search-based candidate expansion.To bridge the semantic gap between recommendation content and search queries, we introduce aSID-Query alignment objective that learns a shared semantic space for content semantic IDs andquery representations. We further design a closed-loop self-enhancement paradigm that leveragesimplicit user feedback from generated content and search-retrieved results to improve bothgeneration tasks. We provide a detailed experimental protocol using public recommendationdatasets with weakly supervised query construction, define a comprehensive set of evaluationmetrics, report expected performance estimates grounded in known baseline values, and validatethe executability of the proposed pipeline through a minimal local prototype. OneFeed providesa practical and extensible direction for unifying search and recommendation through generativemodeling.

Bradley-Terry Rankings for Recommender Systems Across Dataset Taxonomies

The ranking of recommendation algorithms is a challenging problem since model performance is sensitive to dataset characteristics such as sparsity, sequential structure, and scale. This drives a demand for a proper methodology for fair comparison between algorithms. Naive aggregation of performance metrics (e.g., averaging NDCG over benchmarks) can yield misleading rankings, undermining practical selection. To address this problem, we introduce a novel, data-driven ranking methodology based on Bradley-Terry (BT) model. We demonstrate that the obtained ranking depends on key dataset statistics. Additionally, we propose a novel metric for evaluating ranking consistency and demonstrate robustness of our ranking to incomplete data. Finally, we introduce a dataset-specific methodology for ranking algorithms on unseen datasets without running the models, relying on extensions of the Bradley-Terry framework, including BT trees and BT models with covariates.

When Large Language Models Fail in Healthcare: Evaluating Sensitivity to Prompt Variations

Large Language Models (LLMs) are increasingly used in healthcare for tasks such as clinical question answering, diagnosis support, and report summarization. Despite their promise, these models remain highly sensitive to subtle prompt perturbations, both lexical and syntactic, posing serious risks in safety-critical clinical applications. In this study, we conduct a systematic sensitivity analysis to evaluate the robustness of both general-purpose (e.g., GPT-3.5, Llama3) and medical-specific LLMs (e.g., ClinicalBERT, BioLlama3, BioBERT) using the MedMCQA benchmark. We categorize perturbations into natural and adversarial types and examine their effect on model consistency, accuracy, and reliability in clinical reasoning tasks. Our findings reveal that medical LLMs are not intrinsically safe. Even minor variations in phrasing can alter clinical advice, and targeted adversarial prompts can provoke harmful outputs. In high-stakes settings like healthcare, such unpredictability is unacceptable-models that change diagnoses due to reworded inputs or hallucinate medications when slightly rephrased cannot be reliably trusted by clinicians. While models tend to show resilience to simple lexical substitutions or paraphrasing, they often break down under syntactic reordering or misleading contextual cues. This fragility is evident across both general-purpose and domain-specific LLMs. Notably, adversarial manipulations can lead to clinically dangerous outputs, such as recommending incorrect dosages or omitting critical findings.