Jianpeng Cheng

Benyu Zhang, Qiang Zhang, Jianpeng Cheng et al.

Large Language Models (LLMs) represent a promising frontier for recommender systems, yet their development has been impeded by the absence of predictable scaling laws, which are crucial for guiding research and optimizing resource allocation. We hypothesize that this may be attributed to the inherent noise, bias, and incompleteness of raw user interaction data in prior continual pre-training (CPT) efforts. This paper introduces a novel, layered framework for generating high-quality synthetic data that circumvents such issues by creating a curated, pedagogical curriculum for the LLM. We provide powerful, direct evidence for the utility of our curriculum by showing that standard sequential models trained on our principled synthetic data significantly outperform ($+130\%$ on recall@100 for SasRec) models trained on real data in downstream ranking tasks, demonstrating its superiority for learning generalizable user preference patterns. Building on this, we empirically demonstrate, for the first time, robust power-law scaling for an LLM that is continually pre-trained on our high-quality, recommendation-specific data. Our experiments reveal consistent and predictable perplexity reduction across multiple synthetic data modalities. These findings establish a foundational methodology for reliable scaling LLM capabilities in the recommendation domain, thereby shifting the research focus from mitigating data deficiencies to leveraging high-quality, structured information.

Cecilia Aas, Hisham Abdelsalam, Irina Belousova et al.

It has recently become feasible to run personal digital assistants on phones and other personal devices. In this paper we describe a design for a natural language understanding system that runs on device. In comparison to a server-based assistant, this system is more private, more reliable, faster, more expressive, and more accurate. We describe what led to key choices about architecture and technologies. For example, some approaches in the dialog systems literature are difficult to maintain over time in a deployment setting. We hope that sharing learnings from our practical experiences may help inform future work in the research community.

Yicheng Fu, Raviteja Anantha, Prabal Vashisht et al.

Generating user intent from a sequence of user interface (UI) actions is a core challenge in comprehensive UI understanding. Recent advancements in multimodal large language models (MLLMs) have led to substantial progress in this area, but their demands for extensive model parameters, computing power, and high latency makes them impractical for scenarios requiring lightweight, on-device solutions with low latency or heightened privacy. Additionally, the lack of high-quality datasets has hindered the development of such lightweight models. To address these challenges, we propose UI-JEPA, a novel framework that employs masking strategies to learn abstract UI embeddings from unlabeled data through self-supervised learning, combined with an LLM decoder fine-tuned for user intent prediction. We also introduce two new UI-grounded multimodal datasets, "Intent in the Wild" (IIW) and "Intent in the Tame" (IIT), designed for few-shot and zero-shot UI understanding tasks. IIW consists of 1.7K videos across 219 intent categories, while IIT contains 914 videos across 10 categories. We establish the first baselines for these datasets, showing that representations learned using a JEPA-style objective, combined with an LLM decoder, can achieve user intent predictions that match the performance of state-of-the-art large MLLMs, but with significantly reduced annotation and deployment resources. Measured by intent similarity scores, UI-JEPA outperforms GPT-4 Turbo and Claude 3.5 Sonnet by 10.0% and 7.2% respectively, averaged across two datasets. Notably, UI-JEPA accomplishes the performance with a 50.5x reduction in computational cost and a 6.6x improvement in latency in the IIW dataset. These results underscore the effectiveness of UI-JEPA, highlighting its potential for lightweight, high-performance UI understanding.

Shlok Mishra, Tsung-Yu Lin, Linda Wang et al.

We present Xray-Visual, a unified vision model architecture for large-scale image and video understanding trained on industry-scale social media data. Our model leverages over 15 billion curated image-text pairs and 10 billion video-hashtag pairs from Facebook and Instagram, employing robust data curation pipelines that incorporate balancing and noise suppression strategies to maximize semantic diversity while minimizing label noise. We introduce a three-stage training pipeline that combines self-supervised MAE, semi-supervised hashtag classification, and CLIP-style contrastive learning to jointly optimize image and video modalities. Our architecture builds on a Vision Transformer backbone enhanced with efficient token reorganization (EViT) for improved computational efficiency. Extensive experiments demonstrate that Xray-Visual achieves state-of-the-art performance across diverse benchmarks, including ImageNet for image classification, Kinetics and HMDB51 for video understanding, and MSCOCO for cross-modal retrieval. The model exhibits strong robustness to domain shift and adversarial perturbations. We further demonstrate that integrating large language models as text encoders (LLM2CLIP) significantly enhances retrieval performance and generalization capabilities, particularly in real-world environments. Xray-Visual establishes new benchmarks for scalable, multimodal vision models, while maintaining superior accuracy and computational efficiency.

Jianpeng Cheng, Xian Wu, Jiangfan Zhang et al.

Recent research has demonstrated that Universal Multimodal Embedding (UME) benefits significantly from Chain-of-Thought (CoT) reasoning. In this paradigm, a generative model produces explicit reasoning traces for a multimodal query, with the final representation extracted from an <eos> embedding token attending to both the query and the reasoning. Despite its effectiveness, the computational overhead of generating explicit CoT traces is often prohibitive. In this work, we propose replacing explicit CoT with latent think tokens, which are interpreted as latent variables that can produce explicit CoT traces as observed variables. By optimizing think tokens using CoT generation loss and subsequent embedding tokens using contrastive loss, we produce high-performance, reasoning-aware representations at a constant inference cost. Our study investigates two key architectural designs: 1) how think and embeddings tokens should be extracted from the same LLM backbone. 2) how the tokens should be trained as two dependent tasks. We introduce TTE-Flash-2B, a reasoning-aware multimodal representation model that outperforms its explicit-CoT counterpart on the MMEB-v2 benchmark, while producing latent think tokens that are interpretable both textually and visually. Furthermore, zero-shot evaluation across 15 video datasets reveals scaling behavior as the number of think tokens increases, and motivating a pilot study of adaptive think budget allocation based on task requirements.

Xuanming Cui, Jianpeng Cheng, Hong-you Chen et al.

There is a growing interest in Universal Multimodal Embeddings (UME), where models are required to generate task-specific representations. While recent studies show that Multimodal Large Language Models (MLLMs) perform well on such tasks, they treat MLLMs solely as encoders, overlooking their generative capacity. However, such an encoding paradigm becomes less effective as instructions become more complex and require compositional reasoning. Inspired by the proven effectiveness of chain-of-thought reasoning, we propose a general Think-Then-Embed (TTE) framework for UME, composed of a reasoner and an embedder. The reasoner MLLM first generates reasoning traces that explain complex queries, followed by an embedder that produces representations conditioned on both the original query and the intermediate reasoning. This explicit reasoning step enables more nuanced understanding of complex multimodal instructions. Our contributions are threefold. First, by leveraging a powerful MLLM reasoner, we achieve state-of-the-art performance on the MMEB-V2 benchmark, surpassing proprietary models trained on massive in-house datasets. Second, to reduce the dependency on large MLLM reasoners, we finetune a smaller MLLM reasoner using high-quality embedding-centric reasoning traces, achieving the best performance among open-source models with a 7% absolute gain over recently proposed models. Third, we investigate strategies for integrating the reasoner and embedder into a unified model for improved efficiency without sacrificing performance.

Joe Stacey, Jianpeng Cheng, John Torr et al.

Spurred by recent advances in Large Language Models (LLMs), virtual assistants are poised to take a leap forward in terms of their dialogue capabilities. Yet a major bottleneck to achieving genuinely transformative task-oriented dialogue capabilities remains the scarcity of high quality data. Existing datasets, while impressive in scale, have limited domain coverage and contain few genuinely challenging conversational phenomena; those which are present are typically unlabelled, making it difficult to assess the strengths and weaknesses of models without time-consuming and costly human evaluation. Moreover, creating high quality dialogue data has until now required considerable human input, limiting both the scale of these datasets and the ability to rapidly bootstrap data for a new target domain. We aim to overcome these issues with LUCID, a modularised and highly automated LLM-driven data generation system that produces realistic, diverse and challenging dialogues. We use LUCID to generate a seed dataset of 4,277 conversations across 100 intents to demonstrate its capabilities, with a human review finding consistently high quality labels in the generated data.

Yicheng Fu, Raviteja Anantha, Jianpeng Cheng

While server-side Large Language Models (LLMs) demonstrate proficiency in function calling and complex reasoning, deploying Small Language Models (SLMs) directly on devices brings opportunities to improve latency and privacy but also introduces unique challenges for accuracy and memory. We introduce CAMPHOR, an innovative on-device SLM multi-agent framework designed to handle multiple user inputs and reason over personal context locally, ensuring privacy is maintained. CAMPHOR employs a hierarchical architecture where a high-order reasoning agent decomposes complex tasks and coordinates expert agents responsible for personal context retrieval, tool interaction, and dynamic plan generation. By implementing parameter sharing across agents and leveraging prompt compression, we significantly reduce model size, latency, and memory usage. To validate our approach, we present a novel dataset capturing multi-agent task trajectories centered on personalized mobile assistant use-cases. Our experiments reveal that fine-tuned SLM agents not only surpass closed-source LLMs in task completion F1 by~35\% but also eliminate the need for server-device communication, all while enhancing privacy.

Seanie Lee, Jianpeng Cheng, Joris Driesen et al.

Few-shot dialogue state tracking (DST) with Large Language Models (LLM) relies on an effective and efficient conversation retriever to find similar in-context examples for prompt learning. Previous works use raw dialogue context as search keys and queries, and a retriever is fine-tuned with annotated dialogues to achieve superior performance. However, the approach is less suited for scaling to new domains or new annotation languages, where fine-tuning data is unavailable. To address this problem, we handle the task of conversation retrieval based on text summaries of the conversations. A LLM-based conversation summarizer is adopted for query and key generation, which enables effective maximum inner product search. To avoid the extra inference cost brought by LLM-based conversation summarization, we further distill a light-weight conversation encoder which produces query embeddings without decoding summaries for test conversations. We validate our retrieval approach on MultiWOZ datasets with GPT-Neo-2.7B and LLaMA-7B/30B. The experimental results show a significant improvement over relevant baselines in real few-shot DST settings.

Alexandru Coca, Bo-Hsiang Tseng, Pete Boothroyd et al.

Programmable task-oriented dialogue (TOD) agents enable language models to follow structured dialogue policies, but their effectiveness hinges on accurate state tracking. We present PyTOD, an agent that generates executable code to track dialogue state and uses policy and execution feedback for efficient error correction. To this end, PyTOD employs a simple constrained decoding approach, using a language model instead of grammar rules to follow API schemata. This leads to state-of-the-art state tracking performance on the challenging SGD benchmark. Our experiments show that PyTOD surpasses strong baselines in both accuracy and robust user goal estimation as the dialogue progresses, demonstrating the effectiveness of execution-aware state tracking.

Alexandru Coca, Mark Gaynor, Zhenxing Zhang et al.

This work evaluates the potential of large language models (LLMs) to power digital assistants capable of complex action execution. These assistants rely on pre-trained programming knowledge to execute multi-step goals by composing objects and functions defined in assistant libraries into action execution programs. To achieve this, we develop ASPERA, a framework comprising an assistant library simulation and a human-assisted LLM data generation engine. Our engine allows developers to guide LLM generation of high-quality tasks consisting of complex user queries, simulation state and corresponding validation programs, tackling data availability and evaluation robustness challenges. Alongside the framework we release Asper-Bench, an evaluation dataset of 250 challenging tasks generated using ASPERA, which we use to show that program generation grounded in custom assistant libraries is a significant challenge to LLMs compared to dependency-free code generation.

Jianpeng Cheng, Devang Agrawal, Hector Martinez Alonso et al.

We consider a new perspective on dialog state tracking (DST), the task of estimating a user's goal through the course of a dialog. By formulating DST as a semantic parsing task over hierarchical representations, we can incorporate semantic compositionality, cross-domain knowledge sharing and co-reference. We present TreeDST, a dataset of 27k conversations annotated with tree-structured dialog states and system acts. We describe an encoder-decoder framework for DST with hierarchical representations, which leads to 20% improvement over state-of-the-art DST approaches that operate on a flat meaning space of slot-value pairs.

Bo-Hsiang Tseng, Jianpeng Cheng, Yimai Fang et al.

Natural language understanding (NLU) and natural language generation (NLG) are two fundamental and related tasks in building task-oriented dialogue systems with opposite objectives: NLU tackles the transformation from natural language to formal representations, whereas NLG does the reverse. A key to success in either task is parallel training data which is expensive to obtain at a large scale. In this work, we propose a generative model which couples NLU and NLG through a shared latent variable. This approach allows us to explore both spaces of natural language and formal representations, and facilitates information sharing through the latent space to eventually benefit NLU and NLG. Our model achieves state-of-the-art performance on two dialogue datasets with both flat and tree-structured formal representations. We also show that the model can be trained in a semi-supervised fashion by utilising unlabelled data to boost its performance.

Jianpeng Cheng, Siva Reddy, Mirella Lapata

Semantic parsing is the task of converting natural language utterances into machine interpretable meaning representations which can be executed against a real-world environment such as a database. Scaling semantic parsing to arbitrary domains faces two interrelated challenges: obtaining broad coverage training data effectively and cheaply; and developing a model that generalizes to compositional utterances and complex intentions. We address these challenges with a framework which allows to elicit training data from a domain ontology and bootstrap a neural parser which recursively builds derivations of logical forms. In our framework meaning representations are described by sequences of natural language templates, where each template corresponds to a decomposed fragment of the underlying meaning representation. Although artificial, templates can be understood and paraphrased by humans to create natural utterances, resulting in parallel triples of utterances, meaning representations, and their decompositions. These allow us to train a neural semantic parser which learns to compose rules in deriving meaning representations. We crowdsource training data on six domains, covering both single-turn utterances which exhibit rich compositionality, and sequential utterances where a complex task is procedurally performed in steps. We then develop neural semantic parsers which perform such compositional tasks. In general, our approach allows to deploy neural semantic parsers quickly and cheaply from a given domain ontology.

Bowen Li, Jianpeng Cheng, Yang Liu et al.

Dependency grammar induction is the task of learning dependency syntax without annotated training data. Traditional graph-based models with global inference achieve state-of-the-art results on this task but they require $O(n^3)$ run time. Transition-based models enable faster inference with $O(n)$ time complexity, but their performance still lags behind. In this work, we propose a neural transition-based parser for dependency grammar induction, whose inference procedure utilizes rich neural features with $O(n)$ time complexity. We train the parser with an integration of variational inference, posterior regularization and variance reduction techniques. The resulting framework outperforms previous unsupervised transition-based dependency parsers and achieves performance comparable to graph-based models, both on the English Penn Treebank and on the Universal Dependency Treebank. In an empirical comparison, we show that our approach substantially increases parsing speed over graph-based models.

Jianpeng Cheng, Mirella Lapata

Weakly-supervised semantic parsers are trained on utterance-denotation pairs, treating logical forms as latent. The task is challenging due to the large search space and spuriousness of logical forms. In this paper we introduce a neural parser-ranker system for weakly-supervised semantic parsing. The parser generates candidate tree-structured logical forms from utterances using clues of denotations. These candidates are then ranked based on two criterion: their likelihood of executing to the correct denotation, and their agreement with the utterance semantics. We present a scheduled training procedure to balance the contribution of the two objectives. Furthermore, we propose to use a neurally encoded lexicon to inject prior domain knowledge to the model. Experiments on three Freebase datasets demonstrate the effectiveness of our semantic parser, achieving results within the state-of-the-art range.

Jianpeng Cheng, Siva Reddy, Vijay Saraswat et al.

This paper describes a neural semantic parser that maps natural language utterances onto logical forms which can be executed against a task-specific environment, such as a knowledge base or a database, to produce a response. The parser generates tree-structured logical forms with a transition-based approach which combines a generic tree-generation algorithm with domain-general operations defined by the logical language. The generation process is modeled by structured recurrent neural networks, which provide a rich encoding of the sentential context and generation history for making predictions. To tackle mismatches between natural language and logical form tokens, various attention mechanisms are explored. Finally, we consider different training settings for the neural semantic parser, including a fully supervised training where annotated logical forms are given, weakly-supervised training where denotations are provided, and distant supervision where only unlabeled sentences and a knowledge base are available. Experiments across a wide range of datasets demonstrate the effectiveness of our parser.

Ben Krause, Marco Damonte, Mihai Dobre et al.

We present Edina, the University of Edinburgh's social bot for the Amazon Alexa Prize competition. Edina is a conversational agent whose responses utilize data harvested from Amazon Mechanical Turk (AMT) through an innovative new technique we call self-dialogues. These are conversations in which a single AMT Worker plays both participants in a dialogue. Such dialogues are surprisingly natural, efficient to collect and reflective of relevant and/or trending topics. These self-dialogues provide training data for a generative neural network as well as a basis for soft rules used by a matching score component. Each match of a soft rule against a user utterance is associated with a confidence score which we show is strongly indicative of reply quality, allowing this component to self-censor and be effectively integrated with other components. Edina's full architecture features a rule-based system backing off to a matching score, backing off to a generative neural network. Our hybrid data-driven methodology thus addresses both coverage limitations of a strictly rule-based approach and the lack of guarantees of a strictly machine-learning approach.

Jianpeng Cheng, Adam Lopez, Mirella Lapata

Generative models defining joint distributions over parse trees and sentences are useful for parsing and language modeling, but impose restrictions on the scope of features and are often outperformed by discriminative models. We propose a framework for parsing and language modeling which marries a generative model with a discriminative recognition model in an encoder-decoder setting. We provide interpretations of the framework based on expectation maximization and variational inference, and show that it enables parsing and language modeling within a single implementation. On the English Penn Treen-bank, our framework obtains competitive performance on constituency parsing while matching the state-of-the-art single-model language modeling score.

Jianpeng Cheng, Siva Reddy, Vijay Saraswat et al.

We introduce a neural semantic parser that converts natural language utterances to intermediate representations in the form of predicate-argument structures, which are induced with a transition system and subsequently mapped to target domains. The semantic parser is trained end-to-end using annotated logical forms or their denotations. We obtain competitive results on various datasets. The induced predicate-argument structures shed light on the types of representations useful for semantic parsing and how these are different from linguistically motivated ones.

Xingxing Zhang, Jianpeng Cheng, Mirella Lapata

Conventional graph-based dependency parsers guarantee a tree structure both during training and inference. Instead, we formalize dependency parsing as the problem of independently selecting the head of each word in a sentence. Our model which we call \textsc{DeNSe} (as shorthand for {\bf De}pendency {\bf N}eural {\bf Se}lection) produces a distribution over possible heads for each word using features obtained from a bidirectional recurrent neural network. Without enforcing structural constraints during training, \textsc{DeNSe} generates (at inference time) trees for the overwhelming majority of sentences, while non-tree outputs can be adjusted with a maximum spanning tree algorithm. We evaluate \textsc{DeNSe} on four languages (English, Chinese, Czech, and German) with varying degrees of non-projectivity. Despite the simplicity of the approach, our parsers are on par with the state of the art.

Jianpeng Cheng, Mirella Lapata

Traditional approaches to extractive summarization rely heavily on human-engineered features. In this work we propose a data-driven approach based on neural networks and continuous sentence features. We develop a general framework for single-document summarization composed of a hierarchical document encoder and an attention-based extractor. This architecture allows us to develop different classes of summarization models which can extract sentences or words. We train our models on large scale corpora containing hundreds of thousands of document-summary pairs. Experimental results on two summarization datasets demonstrate that our models obtain results comparable to the state of the art without any access to linguistic annotation.

Jianpeng Cheng, Li Dong, Mirella Lapata

In this paper we address the question of how to render sequence-level networks better at handling structured input. We propose a machine reading simulator which processes text incrementally from left to right and performs shallow reasoning with memory and attention. The reader extends the Long Short-Term Memory architecture with a memory network in place of a single memory cell. This enables adaptive memory usage during recurrence with neural attention, offering a way to weakly induce relations among tokens. The system is initially designed to process a single sequence but we also demonstrate how to integrate it with an encoder-decoder architecture. Experiments on language modeling, sentiment analysis, and natural language inference show that our model matches or outperforms the state of the art.

Jianpeng Cheng, Dimitri Kartsaklis

Deep compositional models of meaning acting on distributional representations of words in order to produce vectors of larger text constituents are evolving to a popular area of NLP research. We detail a compositional distributional framework based on a rich form of word embeddings that aims at facilitating the interactions between words in the context of a sentence. Embeddings and composition layers are jointly learned against a generic objective that enhances the vectors with syntactic information from the surrounding context. Furthermore, each word is associated with a number of senses, the most plausible of which is selected dynamically during the composition process. We evaluate the produced vectors qualitatively and quantitatively with positive results. At the sentence level, the effectiveness of the framework is demonstrated on the MSRPar task, for which we report results within the state-of-the-art range.

Jianpeng Cheng, Dimitri Kartsaklis, Edward Grefenstette

This paper aims to explore the effect of prior disambiguation on neural network- based compositional models, with the hope that better semantic representations for text compounds can be produced. We disambiguate the input word vectors before they are fed into a compositional deep net. A series of evaluations shows the positive effect of prior disambiguation for such deep models.