LLM agents have emerged as powerful systems for tackling multi-turn tasks by interleaving internal reasoning and external tool interactions. Agentic Reinforcement Learning has recently drawn significant research attention as a critical post-training paradigm to further refine these capabilities. In this paper, we present AT²PO (**A**gentic **T**urn-based **P**olicy **O**ptimization via **T**ree Search), a unified framework for multi-turn agentic RL that addresses three core challenges: limited exploration diversity, sparse credit assignment, and misaligned policy optimization. AT²PO introduces a turn-level tree structure that jointly enables Entropy-Guided Tree Expansion for strategic exploration and Turn-wise Credit Assignment for fine-grained reward propagation from sparse outcomes. Complementing this, we propose Agentic Turn-based Policy Optimization, a turn-level learning objective that aligns policy updates with the natural decision granularity of agentic interactions. ATPO is orthogonal to tree search and can be readily integrated into any multi-turn RL pipeline. Experiments across seven benchmarks demonstrate consistent improvements over the state-of-the-art baseline by up to 1.84 percentage points in average, with ablation studies validating the effectiveness of each component.

Online advertising governance faces significant challenges due to the non-stationary nature of regulatory policies, where emerging mandates (e.g., restrictions on education or aesthetic anxiety) create severe label inconsistencies and reasoning ambiguities in historical datasets. In this paper, we propose ARGUS, a policy-adaptive governance system that enables evolving reinforcement through multi-agent adversarial umpiring. ARGUS addresses the sparsity of new policy data by employing a three-stage framework: (1) Policy Seeding for initial perception; (2) Adversarial Label Rectification, which utilizes a ”Prosecutor-Defender-Umpire” architecture to resolve conflicts between stale labels and new mandates; and (3) Latent Knowledge Discovery, which employs a tripartite dialectical discussion to unearth sophisticated, “gray-area” violations. By leveraging RAG-enhanced policy knowledge and Chain-of-Thought synthesis as dynamic rewards for reinforcement learning, ARGUS synchronizes its reasoning pathways with evolving regulations. Extensive experiments on both industrial and public datasets demonstrate that ARGUS significantly outperforms traditional fine-tuning baselines, achieving superior policy-adaptive learning with minimal gold data.

Industrial advertising question answering (QA) is a high-stakes task in which hallucinated content, particularly fabricated URLs, can lead to financial loss, compliance violations, and legal risk. Although Retrieval-Augmented Generation (RAG) is widely adopted, deploying it in production remains challenging because industrial knowledge is inherently relational, frequently updated, and insufficiently aligned with generation objectives. We propose a reinforced co-adaptation framework that jointly optimizes retrieval and generation through two components: (1) Graph-aware Retrieval (GraphRAG), which models entity-relation structure over a high-citation knowledge subgraph for multi-hop, domain-specific evidence selection; and (2) evidence-constrained reinforcement learning via Group Relative Policy Optimization (GRPO) with multi-dimensional rewards covering faithfulness, style compliance, safety, and URL validity. Experiments on an internal advertising QA dataset show consistent gains across expert-judged dimensions including accuracy, completeness, and safety, while reducing the hallucination rate by 72%. A two-week online A/B test demonstrates a 28.6% increase in like rate, a 46.2% decrease in dislike rate, and a 92.7% reduction in URL hallucination. The system has been running in production for over half a year and has served millions of QA interactions.

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by incorporating external knowledge, yet traditional single-round retrieval struggles with complex multi-step reasoning.Agentic RAG addresses this by enabling LLMs to dynamically decide when and what to retrieve, but current RL-based training methods suffer from sparse outcome rewards that discard intermediate signals and low sample efficiency where failed samples contribute nothing.We propose Search-P1, a framework that introduces path-centric reward shaping for agentic RAG training, comprising two key components: (1) Path-Centric Reward, which evaluates the structural quality of reasoning trajectories through order-agnostic step coverage and soft scoring that extracts learning signals even from failed samples, and (2) Dual-Track Path Scoring with offline-generated reference planners that assesses paths from both self-consistency and reference-alignment perspectives.Experiments on multiple QA benchmarks demonstrate that Search-P1 achieves significant improvements over Search-R1 and other strong baselines, with an average accuracy gain of 7.7 points.

Rigorous content moderation is crucial for online advertising but leads to millions of daily rejections. This scale renders manual rectification infeasible, particularly for video advertisements.However, existing safety-driven methods often suffer from aggressive over-editing, which compromises the advertiser’s original semantic intent merely to satisfy compliance.In this work, we target the rectification of textual violations in video ads, covering both speech transcripts and on-screen text. We propose ℛ³, a novel framework designed to harmonize compliance with original semantic intent preservation.Our approach integrates three key innovations: (1) an experience-driven data synthesis framework that bootstraps high-quality supervision via group-**R**elative compliance experience extractor; (2) a curriculum **R**einforcement learning strategy with hierarchical rewards designed to enforce compliance while maximizing semantic consistency;and (3) a comprehensive video **R**ectification framework seamlessly integrating text recognition, rewriting, and re-rendering for industrial deployment. Extensive experiments on industrial datasets and online A/B testing demonstrate that ℛ³ significantly outperforms state-of-the-art baselines, achieving an optimal trade-off between violation rectification and intent preservation.

While advertising is a cornerstone of commercial growth, it is constrained by online violation detection systems that reject non-compliant content at a million-scale daily. Advertisers urgently require automated solutions to rectify these advertisements, especially visual ads, as manual fixing is unscalable. Although recent safety-driven methods can achieve compliance, they typically suffer from over-editing, destroying the original commercial intent and perceptual similarity.To address this, we present SSR-A, a framework tailored for the minimalist rectification of non-compliant image ads.Instead of fine-tuning image editing models directly, SSR-A focuses on translating violation policies into targeted editing instructions.We first introduce a Spatial- and Semantic-Aware Instruction Synthesis Pipeline, where MLLMs synthesize candidate instructions—incorporating spatial grounding and semantic guidance—and select the optimal instruction via multi-dimensional evaluation. Furthermore, we align the model using Curriculum Reinforcement Learning, employing GRPO with multi-faceted rewards to progressively navigate the trade-off between compliance and visual preservation. Extensive experiments and online A/B tests show that SSR-A significantly outperforms state-of-the-art baselines in both compliance and preservation of visual and commercial consistency.

While large pre-trained visual-language models have shown promising results on traditional visual question answering benchmarks, it is still challenging for them to answer complex VQA problems which requires diverse world knowledge. Motivated by the research of retrieval-augmented generation in the field of natural language processing, we use Dense Passage Retrieval (DPR) to retrieve related knowledge to help the model answer questions. However, DPR conduct retrieving in natural language space, which may not ensure comprehensive acquisition of image information. Thus, the retrieved knowledge is not truly conducive to helping answer the question, affecting the performance of the overall system. To address this issue, we propose a novel framework that leverages the visual-language model to select the key knowledge retrieved by DPR and answer questions. The framework consists of two modules: Selector and Answerer, where both are initialized by the MLLM and parameter-efficiently finetuned by self-bootstrapping: find key knowledge in the retrieved knowledge documents using the Selector, and then use them to finetune the Answerer to predict answers; obtain the pseudo-labels of key knowledge documents based on the predictions of the Answerer and weak supervision labels, and then finetune the Selector to select key knowledge; repeat. Our framework significantly enhances the performance of the baseline on the challenging open-domain Knowledge-based VQA benchmark, OK-VQA, achieving a state-of-the-art accuracy of 62.83%.

Deploying large-scale pre-trained models in the prompt-tuning paradigm has demonstrated promising performance in few-shot learning. Particularly, vision-language pre-training models (VL-PTMs) have been intensively explored in various few-shot downstream tasks. However, most existing works only apply VL-PTMs to visual tasks like image classification, with few attempts being made on language tasks like text classification. In few-shot text classification, a feasible paradigm for deploying VL-PTMs is to align the input samples and their category names via the text encoders. However, it leads to the waste of visual information learned by the image encoders of VL-PTMs. To overcome this drawback, we propose a novel method named Visual Prompt Tuning (VPT). To our best knowledge, this method is the first attempt to deploy VL-PTM in few-shot text classification task. The main idea is to generate the image embeddings w.r.t. category names as visual prompt and then add them to the aligning process. Extensive experiments show that our VPT can achieve significant improvements under both zero-shot and few-shot settings. Importantly, our VPT even outperforms the most recent prompt-tuning methods on five public text classification datasets.

This article describes a large-scale evaluation of the use of Statistical Machine Translation for professional subtitling. The work was carried out within the FP7 EU-funded project SUMAT and involved two rounds of evaluation: a quality evaluation and a measure of productivity gain/loss. We present the SMT systems built for the project and the corpora they were trained on, which combine professionally created and crowd-sourced data. Evaluation goals, methodology and results are presented for the eleven translation pairs that were evaluated by professional subtitlers. Overall, a majority of the machine translated subtitles received good quality ratings. The results were also positive in terms of productivity, with a global gain approaching 40%. We also evaluated the impact of applying quality estimation and filtering of poor MT output, which resulted in higher productivity gains for filtered files as opposed to fully machine-translated files. Finally, we present and discuss feedback from the subtitlers who participated in the evaluation, a key aspect for any eventual adoption of machine translation technology in professional subtitling.

The paper presents a novel technique for speech translation using hierarchical phrased-based statistical machine translation (HPB-SMT). The system is based on translation of speech from phone sequences as opposed to conventional approach of speech translation from word sequences. The technique facilitates speech translation by allowing a machine translation (MT) system to access to phonetic information. This enables the MT system to act as both a word recognition and a translation component. This results in better performance than conventional speech translation approaches by recovering from recognition error with help of a source language model, translation model and target language model. For this purpose, the MT translation models are adopted to work on source language phones using a grapheme-to-phoneme component. The source-side phonetic confusions are handled using a confusion network. The result on IWLST'10 English- Chinese translation task shows a significant improvement in translation quality. In this paper, results for HPB-SMT are compared with previously published results of phrase-based statistical machine translation (PB-SMT) system (Baseline). The HPB-SMT system outperforms PB-SMT in this regard.

Content localisation via machine translation (MT) is a sine qua non, especially for international online business. While most applications utilise rule-based solutions due to the lack of suitable in-domain parallel corpora for statistical MT (SMT) training, in this paper we investigate the possibility of applying SMT where huge amounts of monolingual content only are available. We describe a case study where an analysis of a very large amount of monolingual online trading data from eBay is conducted by ALS with a view to reducing this corpus to the most representative sample in order to ensure the widest possible coverage of the total data set. Furthermore, minimal yet optimal sets of sentences/words/terms are selected for generation of initial translation units for future SMT system-building.

This paper presents a case-study of work done by Applied Language Solutions (ALS) for a large social networking provider who claim to have built the world’s first multi-language social network, where Internet users from all over the world can communicate in languages that are available in the system. In an initial phase, the social networking provider contracted ALS to build Machine Translation (MT) engines for twelve language-pairs: Russian⇔English, Russian⇔Turkish, Russian⇔Arabic, Turkish⇔English, Turkish⇔Arabic and Arabic⇔English. All of the input data is user-generated content, so we faced a number of problems in building large-scale, robust, high-quality engines. Primarily, much of the source-language data is of ‘poor’ or at least ‘non-standard’ quality. This comes in many forms: (i) content produced by non-native speakers, (ii) content produced by native speakers containing non-deliberate typos, or (iii) content produced by native speakers which deliberately departs from spelling norms to bring about some linguistic effect. Accordingly, in addition to the ‘regular’ pre-processing techniques used in the building of our statistical MT systems, we needed to develop routines to deal with all these scenarios. In this paper, we describe how we handle shortforms, acronyms, typos, punctuation errors, non-dictionary slang, wordplay, censor avoidance and emoticons. We demonstrate automatic evaluation scores on the social network data, together with insights from the the social networking provider regarding some of the typical errors made by the MT engines, and how we managed to correct these in the engines.

In this paper, we provide a description of the Dublin City University’s (DCU) submissions in the IWSLT 2011 evaluationcampaign.1 WeparticipatedintheArabic-Englishand Chinese-English Machine Translation(MT) track translation tasks. We use phrase-based statistical machine translation (PBSMT) models to create the baseline system. Due to the open-domain nature of the data to be translated, we use domain adaptation techniques to improve the quality of translation. Furthermore, we explore target-side syntactic augmentation for an Hierarchical Phrase-Based (HPB) SMT model. Combinatory Categorial Grammar (CCG) is used to extract labels for target-side phrases and non-terminals in the HPB system. Combining the domain adapted language models with the CCG-augmented HPB system gave us the best translations for both language pairs providing statistically significant improvements of 6.09 absolute BLEU points (25.94% relative) and 1.69 absolute BLEU points (15.89% relative) over the unadapted PBSMT baselines for the Arabic-English and Chinese-English language pairs, respectively.

In this paper, we present a novel approach to incorporate source-side syntactic reordering patterns into phrase-based SMT. The main contribution of this work is to use the lattice scoring approach to exploit and utilize reordering information that is favoured by the baseline PBSMT system. By referring to the parse trees of the training corpus, we represent the observed reorderings with source-side syntactic patterns. The extracted patterns are then used to convert the parsed inputs into word lattices, which contain both the original source sentences and their potential reorderings. Weights of the word lattices are estimated from the observations of the syntactic reordering patterns in the training corpus. Finally, the PBSMT system is tuned and tested on the generated word lattices to show the benefits of adding potential source-side reorderings in the inputs. We confirmed the effectiveness of our proposed method on a medium-sized corpus for Chinese-English machine translation task. Our method outperformed the baseline system by 1.67% relative on a randomly selected testset and 8.56% relative on the NIST 2008 testset in terms of BLEU score.