Wenting Zhao

John X. Morris, Wenting Zhao, Justin T. Chiu et al.

Language models produce a distribution over the next token; can we use this information to recover the prompt tokens? We consider the problem of language model inversion and show that next-token probabilities contain a surprising amount of information about the preceding text. Often we can recover the text in cases where it is hidden from the user, motivating a method for recovering unknown prompts given only the model's current distribution output. We consider a variety of model access scenarios, and show how even without predictions for every token in the vocabulary we can recover the probability vector through search. On Llama-2 7b, our inversion method reconstructs prompts with a BLEU of $59$ and token-level F1 of $78$ and recovers $27\%$ of prompts exactly. Code for reproducing all experiments is available at http://github.com/jxmorris12/vec2text.

Yuntian Deng, Wenting Zhao, Jack Hessel et al. · allen-ai

The increasing availability of real-world conversation data offers exciting opportunities for researchers to study user-chatbot interactions. However, the sheer volume of this data makes manually examining individual conversations impractical. To overcome this challenge, we introduce WildVis, an interactive tool that enables fast, versatile, and large-scale conversation analysis. WildVis provides search and visualization capabilities in the text and embedding spaces based on a list of criteria. To manage million-scale datasets, we implemented optimizations including search index construction, embedding precomputation and compression, and caching to ensure responsive user interactions within seconds. We demonstrate WildVis' utility through three case studies: facilitating chatbot misuse research, visualizing and comparing topic distributions across datasets, and characterizing user-specific conversation patterns. WildVis is open-source and designed to be extendable, supporting additional datasets and customized search and visualization functionalities.

Qianxiong Xu, Wenting Zhao, Guosheng Lin et al.

The key to the success of few-shot segmentation (FSS) lies in how to effectively utilize support samples. Most solutions compress support foreground (FG) features into prototypes, but lose some spatial details. Instead, others use cross attention to fuse query features with uncompressed support FG. Query FG could be fused with support FG, however, query background (BG) cannot find matched BG features in support FG, yet inevitably integrates dissimilar features. Besides, as both query FG and BG are combined with support FG, they get entangled, thereby leading to ineffective segmentation. To cope with these issues, we design a self-calibrated cross attention (SCCA) block. For efficient patch-based attention, query and support features are firstly split into patches. Then, we design a patch alignment module to align each query patch with its most similar support patch for better cross attention. Specifically, SCCA takes a query patch as Q, and groups the patches from the same query image and the aligned patches from the support image as K&V. In this way, the query BG features are fused with matched BG features (from query patches), and thus the aforementioned issues will be mitigated. Moreover, when calculating SCCA, we design a scaled-cosine mechanism to better utilize the support features for similarity calculation. Extensive experiments conducted on PASCAL-5^i and COCO-20^i demonstrate the superiority of our model, e.g., the mIoU score under 5-shot setting on COCO-20^i is 5.6%+ better than previous state-of-the-arts. The code is available at https://github.com/Sam1224/SCCAN.

Shachar Don-Yehiya, Ben Burtenshaw, Ramon Fernandez Astudillo et al. · huggingface, ibm-research

Human feedback on conversations with language language models (LLMs) is central to how these systems learn about the world, improve their capabilities, and are steered toward desirable and safe behaviors. However, this feedback is mostly collected by frontier AI labs and kept behind closed doors. In this work, we bring together interdisciplinary experts to assess the opportunities and challenges to realizing an open ecosystem of human feedback for AI. We first look for successful practices in peer production, open source, and citizen science communities. We then characterize the main challenges for open human feedback. For each, we survey current approaches and offer recommendations. We end by envisioning the components needed to underpin a sustainable and open human feedback ecosystem. In the center of this ecosystem are mutually beneficial feedback loops, between users and specialized models, incentivizing a diverse stakeholders community of model trainers and feedback providers to support a general open feedback pool.

Yang Wu, Yanyan Zhao, Hao Yang et al.

Multimodal sentiment analysis has attracted increasing attention and lots of models have been proposed. However, the performance of the state-of-the-art models decreases sharply when they are deployed in the real world. We find that the main reason is that real-world applications can only access the text outputs by the automatic speech recognition (ASR) models, which may be with errors because of the limitation of model capacity. Through further analysis of the ASR outputs, we find that in some cases the sentiment words, the key sentiment elements in the textual modality, are recognized as other words, which makes the sentiment of the text change and hurts the performance of multimodal sentiment models directly. To address this problem, we propose the sentiment word aware multimodal refinement model (SWRM), which can dynamically refine the erroneous sentiment words by leveraging multimodal sentiment clues. Specifically, we first use the sentiment word position detection module to obtain the most possible position of the sentiment word in the text and then utilize the multimodal sentiment word refinement module to dynamically refine the sentiment word embeddings. The refined embeddings are taken as the textual inputs of the multimodal feature fusion module to predict the sentiment labels. We conduct extensive experiments on the real-world datasets including MOSI-Speechbrain, MOSI-IBM, and MOSI-iFlytek and the results demonstrate the effectiveness of our model, which surpasses the current state-of-the-art models on three datasets. Furthermore, our approach can be adapted for other multimodal feature fusion models easily. Data and code are available at https://github.com/albertwy/SWRM.

Wenting Zhao, Tanya Goyal, Yu Ying Chiu et al. · cmu, uw

While hallucinations of large language models (LLMs) prevail as a major challenge, existing evaluation benchmarks on factuality do not cover the diverse domains of knowledge that the real-world users of LLMs seek information about. To bridge this gap, we introduce WildHallucinations, a benchmark that evaluates factuality. It does so by prompting LLMs to generate information about entities mined from user-chatbot conversations in the wild. These generations are then automatically fact-checked against a systematically curated knowledge source collected from web search. Notably, half of these real-world entities do not have associated Wikipedia pages. We evaluate 118,785 generations from 15 LLMs on 7,919 entities. We find that LLMs consistently hallucinate more on entities without Wikipedia pages and exhibit varying hallucination rates across different domains. Finally, given the same base models, adding a retrieval component only slightly reduces hallucinations but does not eliminate hallucinations.

Wenting Zhao, Ye Liu, Yao Wan et al. · salesforce

Question answering on tabular data (a.k.a TableQA), which aims at generating answers to questions grounded on a provided table, has gained significant attention recently. Prior work primarily produces concise factual responses through information extraction from individual or limited table cells, lacking the ability to reason across diverse table cells. Yet, the realm of free-form TableQA, which demands intricate strategies for selecting relevant table cells and the sophisticated integration and inference of discrete data fragments, remains mostly unexplored. To this end, this paper proposes a generalized three-stage approach: Table-to- Graph conversion and cell localizing, external knowledge retrieval, and the fusion of table and text (called TAG-QA), to address the challenge of inferring long free-form answers in generative TableQA. In particular, TAG-QA (1) locates relevant table cells using a graph neural network to gather intersecting cells between relevant rows and columns, (2) leverages external knowledge from Wikipedia, and (3) generates answers by integrating both tabular data and natural linguistic information. Experiments showcase the superior capabilities of TAG-QA in generating sentences that are both faithful and coherent, particularly when compared to several state-of-the-art baselines. Notably, TAG-QA surpasses the robust pipeline-based baseline TAPAS by 17% and 14% in terms of BLEU-4 and PARENT F-score, respectively. Furthermore, TAG-QA outperforms the end-to-end model T5 by 16% and 12% on BLEU-4 and PARENT F-score, respectively.

Wenting Zhao, Justin T Chiu, Jena D. Hwang et al. · allen-ai, berkeley

Language technologies that accurately model the dynamics of events must perform commonsense reasoning. Existing work evaluating commonsense reasoning focuses on making inferences about common, everyday situations. To instead investigate the ability to model unusual, unexpected, and unlikely situations, we explore the task of uncommonsense abductive reasoning. Given a piece of context with an unexpected outcome, this task requires reasoning abductively to generate an explanation that makes the unexpected outcome more likely in the context. To this end, we curate and release a new English language corpus called UNcommonsense. We characterize the performance differences between human explainers and the best-performing large language models, finding that model-enhanced human-written explanations achieve the highest quality by trading off between specificity and diversity. Finally, we experiment with several imitation learning algorithms to train open and accessible language models on this task. When compared with the vanilla supervised fine-tuning approach, these methods consistently reduce lose rates on both common and uncommonsense abductive reasoning judged by human evaluators.

Liangwei Yang, Jielin Qiu, Zixiang Chen et al.

Many decision-support settings require systems that adapt to individual users, but evaluation data for this problem remain limited. Existing benchmarks for user understanding often rely on simulated users or model-generated behavior, even though recent work cautions that model-based simulations can diverge systematically from human behavior. We introduce \textsc{BehaviorBench}, a benchmark for evaluating personalized decision modeling from real-world behavioral traces. \textsc{BehaviorBench} reconstructs wallet-level decision histories from observed public prediction-market and on-chain records, and organizes them into two complementary task layers: \emph{Belief prediction}, which predicts a user's final revealed stance and confidence in a market, and \emph{Trade prediction}, which predicts the direction and amount of individual transactions. Across 2,000 evaluation wallets, the benchmark contains 141,445 Belief instances and 1,485,972 Trade instances, with disjoint support pools for retrieval-based evaluation. We evaluate frontier and open-weight generative models under four history interfaces: no personalization, direct recent history, generated user profiles, and retrieved support-wallet evidence. Personalization improves Belief prediction more consistently than Trade prediction, model rankings change across task layers and metrics, and different history interfaces expose different failure modes. \textsc{BehaviorBench} provides an evaluation setting for studying whether personalized methods can use real-world behavioral evidence rather than simulated users alone.

Huihan Li, Yuting Ning, Zeyi Liao et al. · uw

To effectively use large language models (LLMs) for real-world queries, it is imperative that they generalize to the long-tail distribution, i.e. rare examples where models exhibit low confidence. In this work, we take the first step towards evaluating LLMs in the long-tail distribution of inferential knowledge. We exemplify long-tail evaluation on the Natural Language Inference task. First, we introduce Logic-Induced-Knowledge-Search (LINK), a systematic long-tail data generation framework, to obtain factually-correct yet long-tail inferential statements. LINK uses variable-wise prompting grounded on symbolic rules to seek low-confidence statements while ensuring factual correctness. We then use LINK to curate Logic-Induced-Long-Tail (LINT), a large-scale long-tail inferential knowledge dataset that contains 108K statements spanning four domains. We evaluate popular LLMs on LINT; we find that state-of-the-art LLMs show significant performance drop (21% relative drop for GPT4) on long-tail data as compared to on head distribution data, and smaller models show even more generalization weakness. These results further underscore the necessity of long-tail evaluation in developing generalizable LLMs.

Pranjal Aggarwal, Marjan Ghazvininejad, Seungone Kim et al. · meta-ai

The ability to precisely derive mathematical objects is a core requirement for downstream STEM applications, including mathematics, physics, and chemistry, where reasoning must culminate in formally structured expressions. Yet, current LM evaluations of mathematical and scientific reasoning rely heavily on simplified answer formats such as numerical values or multiple choice options due to the convenience of automated assessment. In this paper we provide three contributions for improving reasoning over mathematical objects: (i) we build and release training data and benchmarks for deriving mathematical objects, the Principia suite; (ii) we provide training recipes with strong LLM-judges and verifiers, where we show that on-policy judge training boosts performance; (iii) we show how on-policy training can also be used to scale test-time compute via aggregation. We find that strong LMs such as Qwen3-235B and o3 struggle on Principia, while our training recipes can bring significant improvements over different LLM backbones, while simultaneously improving results on existing numerical and MCQA tasks, demonstrating cross-format generalization of reasoning abilities.

Yibo Wang, Wenting Zhao, Yao Wan et al.

Named Entity Recognition (NER) aims to extract and classify entity mentions in the text into pre-defined types (e.g., organization or person name). Recently, many works have been proposed to shape the NER as a machine reading comprehension problem (also termed MRC-based NER), in which entity recognition is achieved by answering the formulated questions related to pre-defined entity types through MRC, based on the contexts. However, these works ignore the label dependencies among entity types, which are critical for precisely recognizing named entities. In this paper, we propose to incorporate the label dependencies among entity types into a multi-task learning framework for better MRC-based NER. We decompose MRC-based NER into multiple tasks and use a self-attention module to capture label dependencies. Comprehensive experiments on both nested NER and flat NER datasets are conducted to validate the effectiveness of the proposed Multi-NER. Experimental results show that Multi-NER can achieve better performance on all datasets.

Wenting Zhao, Konstantine Arkoudas, Weiqi Sun et al.

Task-oriented parsing (TOP) aims to convert natural language into machine-readable representations of specific tasks, such as setting an alarm. A popular approach to TOP is to apply seq2seq models to generate linearized parse trees. A more recent line of work argues that pretrained seq2seq models are better at generating outputs that are themselves natural language, so they replace linearized parse trees with canonical natural-language paraphrases that can then be easily translated into parse trees, resulting in so-called naturalized parsers. In this work we continue to explore naturalized semantic parsing by presenting a general reduction of TOP to abstractive question answering that overcomes some limitations of canonical paraphrasing. Experimental results show that our QA-based technique outperforms state-of-the-art methods in full-data settings while achieving dramatic improvements in few-shot settings.

Wenting Zhao, Ge Gao, Claire Cardie et al.

When seeking information from unfamiliar documents, users frequently pose questions that cannot be answered by the documents. While existing large language models (LLMs) identify these unanswerable questions, they do not assist users in reformulating their questions, thereby reducing their overall utility. We curate CouldAsk, an evaluation benchmark composed of existing and new datasets for document-grounded question answering, specifically designed to study reformulating unanswerable questions. We evaluate state-of-the-art open-source and proprietary LLMs on CouldAsk. The results demonstrate the limited capabilities of these models in reformulating questions. Specifically, GPT-4 and Llama2-7B successfully reformulate questions only 26% and 12% of the time, respectively. Error analysis shows that 62% of the unsuccessful reformulations stem from the models merely rephrasing the questions or even generating identical questions. We publicly release the benchmark and the code to reproduce the experiments.

Yi Lu, Jing Nathan Yan, Songlin Yang et al.

Broad textual understanding and in-context learning require language models that utilize full document contexts. Due to the implementation challenges associated with directly training long-context models, many methods have been proposed for extending models to handle long contexts. However, owing to differences in data and model classes, it has been challenging to compare these approaches, leading to uncertainty as to how to evaluate long-context performance and whether it differs from standard evaluation. We implement a controlled protocol for extension methods with a standardized evaluation, utilizing consistent base models and extension data. Our study yields several insights into long-context behavior. First, we reaffirm the critical role of perplexity as a general-purpose performance indicator even in longer-context tasks. Second, we find that current approximate attention methods systematically underperform across long-context tasks. Finally, we confirm that exact fine-tuning based methods are generally effective within the range of their extension, whereas extrapolation remains challenging. All codebases, models, and checkpoints will be made available open-source, promoting transparency and facilitating further research in this critical area of AI development.

Wenting Zhao, Ye Liu, Tong Niu et al.

Large Language Models (LLMs) have exhibited impressive generation capabilities, but they suffer from hallucinations when solely relying on their internal knowledge, especially when answering questions that require less commonly known information. Retrieval-augmented LLMs have emerged as a potential solution to ground LLMs in external knowledge. Nonetheless, recent approaches have primarily emphasized retrieval from unstructured text corpora, owing to its seamless integration into prompts. When using structured data such as knowledge graphs, most methods simplify it into natural text, neglecting the underlying structures. Moreover, a significant gap in the current landscape is the absence of a realistic benchmark for evaluating the effectiveness of grounding LLMs on heterogeneous knowledge sources (e.g., knowledge base and text). To fill this gap, we have curated a comprehensive dataset that poses two unique challenges: (1) Two-hop multi-source questions that require retrieving information from both open-domain structured and unstructured knowledge sources; retrieving information from structured knowledge sources is a critical component in correctly answering the questions. (2) The generation of symbolic queries (e.g., SPARQL for Wikidata) is a key requirement, which adds another layer of challenge. Our dataset is created using a combination of automatic generation through predefined reasoning chains and human annotation. We also introduce a novel approach that leverages multiple retrieval tools, including text passage retrieval and symbolic language-assisted retrieval. Our model outperforms previous approaches by a significant margin, demonstrating its effectiveness in addressing the above-mentioned reasoning challenges.

Wenting Zhao, Ye Liu, Yao Wan et al.

Few-shot table-to-text generation is a task of composing fluent and faithful sentences to convey table content using limited data. Despite many efforts having been made towards generating impressive fluent sentences by fine-tuning powerful pre-trained language models, the faithfulness of generated content still needs to be improved. To this end, this paper proposes a novel approach Attend, Memorize and Generate (called AMG), inspired by the text generation process of humans. In particular, AMG (1) attends over the multi-granularity of context using a novel strategy based on table slot level and traditional token-by-token level attention to exploit both the table structure and natural linguistic information; (2) dynamically memorizes the table slot allocation states; and (3) generates faithful sentences according to both the context and memory allocation states. Comprehensive experiments with human evaluation on three domains (i.e., humans, songs, and books) of the Wiki dataset show that our model can generate higher qualified texts when compared with several state-of-the-art baselines, in both fluency and faithfulness.

Shangyi Geng, Wenting Zhao, Alexander M Rush

$K$-nearest neighbor language models ($k$NN-LMs), which integrate retrieval with next-word prediction, have demonstrated strong performance in language modeling as well as downstream NLP benchmarks. These results have led researchers to argue that models trained on poor quality or outdated data could perform well by employing a $k$NN extension that has access to a higher-quality datastore. In this work, we ask whether this improved ability to recall information really translates into downstream abilities. We extensively evaluate $k$NN-LMs on a diverse set of tasks, ranging from sentiment classification and commonsense reasoning to multi-hop reasoning. Results show that $k$NN-LMs excel at memory-intensive tasks, where utilizing the patterns in the input is sufficient for determining the output, but struggle with reasoning tasks that require integrating multiple pieces of information to derive new knowledge. We further demonstrate through oracle experiments and qualitative analysis that even with perfect retrieval, $k$NN-LMs still fail to determine the correct answers, placing an upper bound on their reasoning performance. Code and datastores are released at https://github.com/GSYfate/knnlm-limits/.

Yibo Wang, Yanbing Xue, Bo Liu et al.

Position bias, the phenomenon whereby users tend to focus on higher-ranked items of the search result list regardless of the actual relevance to queries, is prevailing in many ranking systems. Position bias in training data biases the ranking model, leading to increasingly unfair item rankings, click-through-rate (CTR), and conversion rate (CVR) predictions. To jointly mitigate position bias in both item CTR and CVR prediction, we propose two position-bias-free CTR and CVR prediction models: Position-Aware Click-Conversion (PACC) and PACC via Position Embedding (PACC-PE). PACC is built upon probability decomposition and models position information as a probability. PACC-PE utilizes neural networks to model product-specific position information as embedding. Experiments on the E-commerce sponsored product search dataset show that our proposed models have better ranking effectiveness and can greatly alleviate position bias in both CTR and CVR prediction.

Zhongfen Deng, Seunghyun Yoon, Trung Bui et al.

Aspect-based meeting transcript summarization aims to produce multiple summaries, each focusing on one aspect of content in a meeting transcript. It is challenging as sentences related to different aspects can mingle together, and those relevant to a specific aspect can be scattered throughout the long transcript of a meeting. The traditional summarization methods produce one summary mixing information of all aspects, which cannot deal with the above challenges of aspect-based meeting transcript summarization. In this paper, we propose a two-stage method for aspect-based meeting transcript summarization. To select the input content related to specific aspects, we train a sentence classifier on a dataset constructed from the AMI corpus with pseudo-labeling. Then we merge the sentences selected for a specific aspect as the input for the summarizer to produce the aspect-based summary. Experimental results on the AMI corpus outperform many strong baselines, which verifies the effectiveness of our proposed method.

Guangbu Liu, Tong Zhang, Xudong Wang et al.

Graph structure expression plays a vital role in distinguishing various graphs. In this work, we propose a Structure-Sensitive Graph Dictionary Embedding (SS-GDE) framework to transform input graphs into the embedding space of a graph dictionary for the graph classification task. Instead of a plain use of a base graph dictionary, we propose the variational graph dictionary adaptation (VGDA) to generate a personalized dictionary (named adapted graph dictionary) for catering to each input graph. In particular, for the adaptation, the Bernoulli sampling is introduced to adjust substructures of base graph keys according to each input, which increases the expression capacity of the base dictionary tremendously. To make cross-graph measurement sensitive as well as stable, multi-sensitivity Wasserstein encoding is proposed to produce the embeddings by designing multi-scale attention on optimal transport. To optimize the framework, we introduce mutual information as the objective, which further deduces to variational inference of the adapted graph dictionary. We perform our SS-GDE on multiple datasets of graph classification, and the experimental results demonstrate the effectiveness and superiority over the state-of-the-art methods.

Jielin Qiu, Ming Zhu, Wenting Zhao et al.

Audio-native large language models (audio-LLMs) commonly use Whisper as their audio encoder. However, Whisper was trained exclusively on speech data, producing weak representations for music and environmental sound. This forces downstream audio-LLMs to compensate through extensive training on large-scale non-speech data. We present Whisper-AuT, a domain-adapted audio encoder obtained by fine-tuning Whisper-large-v3 on a curated mixture of speech (80%), environmental sound (10%), and music (10%) totaling approximately 20M samples. The full encoder-decoder is trained end-to-end with a seq2seq captioning objective; the decoder is then discarded and only the encoder is retained. Linear probe evaluations show that Whisper-AuT achieves +23.0% on ESC-50 (environmental sound), +5.0% on GTZAN (music genre), and +0.7% on Speech Commands (keyword spotting) compared to the original Whisperlarge-v3 encoder. Whisper-AuT is designed as a drop-in replacement for Whisper in audio-LLM architectures, with the goal of reducing downstream training cost by providing stronger initial audio representations for non-speech domains.

Mouxiang Chen, Lei Zhang, Yunlong Feng et al.

We propose SWE-Universe, a scalable and efficient framework for automatically constructing real-world software engineering (SWE) verifiable environments from GitHub pull requests (PRs). To overcome the prevalent challenges of automatic building, such as low production yield, weak verifiers, and prohibitive cost, our framework utilizes a building agent powered by an efficient custom-trained model. This agent employs iterative self-verification and in-loop hacking detection to ensure the reliable generation of high-fidelity, verifiable tasks. Using this method, we scale the number of real-world multilingual SWE environments to a million scale (807,693). We demonstrate the profound value of our environments through large-scale agentic mid-training and reinforcement learning. Finally, we applied this technique to Qwen3-Max-Thinking and achieved a score of 75.3% on SWE-Bench Verified. Our work provides both a critical resource and a robust methodology to advance the next generation of coding agents.

Wenting Zhao, Ibrahim Abdelaziz, Julian Dolby et al.

Dynamically typed languages such as Python have become very popular. Among other strengths, Python's dynamic nature and its straightforward linking to native code have made it the de-facto language for many research areas such as Artificial Intelligence. This flexibility, however, makes static analysis very hard. While creating a sound, or a soundy, analysis for Python remains an open problem, we present in this work Serenity, a framework for static analysis of Python that turns out to be sufficient for some tasks. The Serenity framework exploits two basic mechanisms: (a) reliance on dynamic dispatch at the core of language translation, and (b) extreme abstraction of libraries, to generate an abstraction of the code. We demonstrate the efficiency and usefulness of Serenity's analysis in two applications: code completion and automated machine learning. In these two applications, we demonstrate that such analysis has a strong signal, and can be leveraged to establish state-of-the-art performance, comparable to neural models and dynamic analysis respectively.

Jielin Qiu, Liangwei Yang, Ming Zhu et al.

During live sales calls, customers frequently ask detailed product questions that require representatives to manually search internal databases and CRM systems. This process typically takes 25-65 seconds per query, creating awkward pauses that hurt customer experience and reduce sales efficiency. We present SalesCopilot, a real-time AI-powered assistant that eliminates this bottleneck by automatically detecting customer questions, retrieving relevant information from the product database, and displaying concise answers on the representative's dashboard in seconds. The system integrates streaming speech-to-text transcription, large language model (LLM)-based question detection, and retrieval-augmented generation (RAG) over a structured product database into a unified real-time pipeline. We demonstrate SalesCopilot on an insurance sales scenario with 50 products spanning 10 categories (2,490 FAQs, 290 coverage details, and 162 pricing tiers). In our benchmark evaluation, SalesCopilot achieves a measured mean response time of 2.8 seconds with 100% question detection rate, representing a 14xspeedup compared to manual CRM search in an internal study. The system is domain-agnostic and can be adapted to any enterprise sales domain by replacing the product database.

Jielin Qiu, Zixiang Chen, Liangwei Yang et al.

We present a technical tutorial for building enterprise-grade realtime voice agents from first principles. While end-to-end speech-to-speech models may ultimately provide the best latency for voice agents, fully self-hosted end-to-end solutions are not yet available. We evaluate the closest candidate, Qwen3-Omni, across three configurations: its cloud-only DashScope Realtime API achieves $\sim$702ms audio-to-audio latency with streaming, but is not self-hostable; its local vLLM deployment supports only the Thinker (text generation from audio, 516ms), not the Talker (audio synthesis); and its local Transformers deployment runs the full pipeline but at $\sim$146s -- far too slow for realtime. The cascaded streaming pipeline (STT $\rightarrow$ LLM $\rightarrow$ TTS) therefore remains the practical architecture for self-hosted realtime voice agents, and the focus of this tutorial. We build a complete voice agent using Deepgram (streaming STT), vLLM-served LLMs with function calling (streaming text generation), and ElevenLabs (streaming TTS), achieving a measured time-to-first-audio of 755ms (best case 729ms) with full function calling support. We release the full codebase as a 9-chapter progressive tutorial with working, tested code for every component.

Liangwei Yang, Shiyu Wang, Haolin Chen et al.

As large language models (LLMs) transition from research prototypes to real-world systems, customization has emerged as a central bottleneck. While text prompts can already customize LLM behavior, we argue that text-only prompting does not constitute a suitable control interface for scalable, stable, and inference-only customization. This position paper argues that model providers should expose \emph{vector prompt inputs} as part of the public interface for customizing LLMs. We support this position with diagnostic evidence showing that vector prompt tuning continues to improve with increasing supervision whereas text-based prompt optimization saturates early, and that vector prompts exhibit dense, global attention patterns indicative of a distinct control mechanism. We further discuss why inference-only customization is increasingly important under realistic deployment constraints, and why exposing vector prompts need not fundamentally increase model leakage risk under a standard black-box threat model. We conclude with a call to action for the community to rethink prompt interfaces as a core component of LLM customization.

Justin T. Chiu, Wenting Zhao, Derek Chen et al.

Large language models (LLMs) excel at processing and generating both text and code. However, LLMs have had limited applicability in grounded task-oriented dialogue as they are difficult to steer toward task objectives and fail to handle novel grounding. We present a modular and interpretable grounded dialogue system that addresses these shortcomings by composing LLMs with a symbolic planner and grounded code execution. Our system consists of a reader and planner: the reader leverages an LLM to convert partner utterances into executable code, calling functions that perform grounding. The translated code's output is stored to track dialogue state, while a symbolic planner determines the next appropriate response. We evaluate our system's performance on the demanding OneCommon dialogue task, involving collaborative reference resolution on abstract images of scattered dots. Our system substantially outperforms the previous state-of-the-art, including improving task success in human evaluations from 56% to 69% in the most challenging setting.

Revanth Gangi Reddy, Ye Liu, Wenting Zhao et al.

Maintaining large-scale, multilingual codebases hinges on accurately localizing issues, which requires mapping natural-language error descriptions to the relevant functions that need to be modified. However, existing ranking approaches are often Python-centric and perform a single-pass search over the codebase. This work introduces SweRank+, a framework that couples SweRankMulti, a cross-lingual code ranking tool, with SweRankAgent, an agentic search setup, for iterative, multi-turn reasoning over the code repository. SweRankMulti comprises a code embedding retriever and a listwise LLM reranker, and is trained using a carefully curated large-scale issue localization dataset spanning multiple popular programming languages. SweRankAgent adopts an agentic search loop that moves beyond single-shot localization with a memory buffer to reason and accumulate relevant localization candidates over multiple turns. Our experiments on issue localization benchmarks spanning various languages demonstrate new state-of-the-art performance with SweRankMulti, while SweRankAgent further improves localization over single-pass ranking.

Hiroaki Hayashi, Bo Pang, Wenting Zhao et al.

Large language model (LLM) based agents are increasingly used to tackle software engineering tasks that require multi-step reasoning and code modification, demonstrating promising yet limited performance. However, most existing LLM agents typically operate within static execution frameworks, lacking a principled mechanism to learn and self-improve from their own experience and past rollouts. As a result, their performance remains bounded by the initial framework design and the underlying LLM's capabilities. We propose Self-Abstraction from Grounded Experience (SAGE), a framework that enables agents to learn from their own task executions and refine their behavior through self-abstraction. After an initial rollout, the agent induces a concise plan abstraction from its grounded experience, distilling key steps, dependencies, and constraints. This learned abstraction is then fed back as contextual guidance, refining the agent's policy and supporting more structured, informed subsequent executions. Empirically, SAGE delivers consistent performance gains across diverse LLM backbones and agent architectures. Notably, it yields a 7.2% relative performance improvement over the strong Mini-SWE-Agent baseline when paired with the GPT-5 (high) backbone. SAGE further achieves strong overall performance on SWE-Bench Verified benchmark, reaching 73.2% and 74% Pass@1 resolve rates with the Mini-SWE-Agent and OpenHands CodeAct agent framework, respectively.

Zhongfen Deng, Hao Peng, Tao Zhang et al.

Product attribute value extraction is an important task in e-Commerce which can help several downstream applications such as product search and recommendation. Most previous models handle this task using sequence labeling or question answering method which rely on the sequential position information of values in the product text and are vulnerable to data discrepancy between training and testing. This limits their generalization ability to real-world scenario in which each product can have multiple descriptions across various shopping platforms with different composition of text and style. They also have limited zero-shot ability to new values. In this paper, we propose a multi-task learning model with value generation/classification and attribute prediction called JPAVE to predict values without the necessity of position information of values in the text. Furthermore, the copy mechanism in value generator and the value attention module in value classifier help our model address the data discrepancy issue by only focusing on the relevant part of input text and ignoring other information which causes the discrepancy issue such as sentence structure in the text. Besides, two variants of our model are designed for open-world and closed-world scenarios. In addition, copy mechanism introduced in the first variant based on value generation can improve its zero-shot ability for identifying unseen values. Experimental results on a public dataset demonstrate the superiority of our model compared with strong baselines and its generalization ability of predicting new values.

Arnav Kumar Jain, Gonzalo Gonzalez-Pumariega, Wayne Chen et al.

We address the problem of code generation from multi-turn execution feedback. Existing methods either generate code without feedback or use complex, hierarchical reinforcement learning to optimize multi-turn rewards. We propose a simple yet scalable approach, $μ$Code, that solves multi-turn code generation using only single-step rewards. Our key insight is that code generation is a one-step recoverable MDP, where the correct code can be recovered from any intermediate code state in a single turn. $μ$Code iteratively trains both a generator to provide code solutions conditioned on multi-turn execution feedback and a verifier to score the newly generated code. Experimental evaluations show that our approach achieves significant improvements over the state-of-the-art baselines. We provide analysis of the design choices of the reward models and policy, and show the efficacy of $μ$Code at utilizing the execution feedback. Our code is available at https://github.com/portal-cornell/muCode.

Zeyao Ma, Jing Zhang, Xiaokang Zhang et al.

Large language models (LLMs) have demonstrated strong coding capabilities but still struggle to solve competitive programming problems correctly in a single attempt. Execution-based re-ranking offers a promising test-time scaling strategy, yet existing methods are constrained by either difficult test case generation or inefficient random input sampling. To address this limitation, we propose Agentic Verifier, an execution-based agent that actively reasons about program behaviors and searches for highly discriminative test inputs that expose behavioral discrepancies among candidate solutions. Through multi-turn interaction with code execution environments, the verifier iteratively refines the candidate input generator and produces targeted counterexamples rather than blindly sampling inputs. We train the verifier to acquire this discriminative input generation capability via a scalable pipeline combining large-scale data synthesis, rejection fine-tuning, and agentic reinforcement learning. Extensive experiments across five competitive programming benchmarks demonstrate consistent improvements over strong execution-based baselines, achieving up to +10-15% absolute gains in Best@K accuracy. Further analysis reveals clear test-time scaling behavior and highlights the verifier's broader potential beyond reranking.

Wenting Zhao, Xiang Ren, Jack Hessel et al.

Chatbots such as GPT-4 and ChatGPT are now serving millions of users. Despite their widespread use, there remains a lack of public datasets showcasing how these tools are used by a population of users in practice. To bridge this gap, we offered free access to ChatGPT for online users in exchange for their affirmative, consensual opt-in to anonymously collect their chat transcripts and request headers. From this, we compiled WildChat, a corpus of 1 million user-ChatGPT conversations, which consists of over 2.5 million interaction turns. We compare WildChat with other popular user-chatbot interaction datasets, and find that our dataset offers the most diverse user prompts, contains the largest number of languages, and presents the richest variety of potentially toxic use-cases for researchers to study. In addition to timestamped chat transcripts, we enrich the dataset with demographic data, including state, country, and hashed IP addresses, alongside request headers. This augmentation allows for more detailed analysis of user behaviors across different geographical regions and temporal dimensions. Finally, because it captures a broad range of use cases, we demonstrate the dataset's potential utility in fine-tuning instruction-following models. WildChat is released at https://wildchat.allen.ai under AI2 ImpACT Licenses.

Yibo Wang, Congying Xia, Wenting Zhao et al.

Unit test generation has become a promising and important use case of LLMs. However, existing evaluation benchmarks for assessing LLM unit test generation capabilities focus on function- or class-level code rather than more practical and challenging project-level codebases. To address such limitation, we propose ProjectTest, a project-level benchmark for unit test generation covering Python, Java, and JavaScript. ProjectTest features 20 moderate-sized and high-quality projects per language. We evaluate nine frontier LLMs on ProjectTest and the results show that all frontier LLMs tested exhibit moderate performance on ProjectTest on Python and Java, highlighting the difficulty of ProjectTest. We also conduct a thorough error analysis, which shows that even frontier LLMs, such as Claude-3.5-Sonnet, have significant basic yet critical errors, including compilation and cascade errors. Motivated by this observation, we further evaluate all frontier LLMs under manual error-fixing and self-error-fixing scenarios to assess their potential when equipped with error-fixing mechanisms. Our code and dataset is available at \href{https://github.com/YiboWANG214/ProjectTest}{ProjectTest}.

Yating Wang, Wenting Zhao, Yaqi Zhao et al.

Large language models store not only isolated facts but also rules that support reasoning across symbolic expressions, natural language explanations, and concrete instances. Yet most model editing methods are built for fact-level knowledge, assuming that a target edit can be achieved through a localized intervention. This assumption does not hold for rule-level knowledge, where a single rule must remain consistent across multiple interdependent forms. We investigate this problem through a mechanistic study of rule-level knowledge editing. To support this study, we extend the RuleEdit benchmark from 80 to 200 manually verified rules spanning mathematics and physics. Fine-grained causal tracing reveals a form-specific organization of rule knowledge in transformer layers: formulas and descriptions are concentrated in earlier layers, while instances are more associated with middle layers. These results suggest that rule knowledge is not uniformly localized, and therefore cannot be reliably edited by a single-layer or contiguous-block intervention. Based on this insight, we propose Distributed Multi-Layer Editing (DMLE), which applies a shared early-layer update to formulas and descriptions and a separate middle-layer update to instances. While remaining competitive on standard editing metrics, DMLE achieves substantially stronger rule-level editing performance. On average, it improves instance portability and rule understanding by 13.91 and 50.19 percentage points, respectively, over the strongest baseline across GPT-J-6B, Qwen2.5-7B, Qwen2-7B, and LLaMA-3-8B. The code is available at https://github.com/Pepper66/DMLE.

Wenting Zhao, Nan Jiang, Celine Lee et al.

With the goal of benchmarking generative systems beyond expert software development ability, we introduce Commit0, a benchmark that challenges AI agents to write libraries from scratch. Agents are provided with a specification document outlining the library's API as well as a suite of interactive unit tests, with the goal of producing an implementation of this API accordingly. The implementation is validated through running these unit tests. As a benchmark, Commit0 is designed to move beyond static one-shot code generation towards agents that must process long-form natural language specifications, adapt to multi-stage feedback, and generate code with complex dependencies. Commit0 also offers an interactive environment where models receive static analysis and execution feedback on the code they generate. Our experiments demonstrate that while current agents can pass some unit tests, none can yet fully reproduce full libraries. Results also show that interactive feedback is quite useful for models to generate code that passes more unit tests, validating the benchmarks that facilitate its use.

Wenting Zhao, Pranjal Aggarwal, Swarnadeep Saha et al.

Scaling up test-time compute, by generating multiple independent solutions and selecting or aggregating among them, has become a central paradigm for improving large language models (LLMs) on challenging reasoning tasks. While most prior work relies on simple majority voting or reward model ranking to aggregate solutions, these approaches may only yield limited benefits. In this work, we propose to learn aggregation as an explicit reasoning skill: given a set of candidate solutions, we train an aggregator model to review, reconcile, and synthesize a final, correct answer using reinforcement learning from verifiable rewards. A key ingredient is careful balancing of easy and hard training examples, allowing the model to learn both to recover minority-but-correct answers as well as easy majority-correct answers. Empirically, we find our method, AggLM, outperforms both strong rule-based and reward-model baselines, across multiple benchmarks. Furthermore, it generalizes effectively to solutions from differing models, including stronger ones than contained in the training data, all while requiring substantially fewer tokens than majority voting with larger numbers of solutions.

Wei Xiong, Wenting Zhao, Weizhe Yuan et al.

As models increasingly leverage multi-step reasoning strategies to solve complex problems, supervising the logical validity of these intermediate steps has become a critical research challenge. Process reward models address this by providing step-by-step feedback, but current approaches have two major drawbacks: they typically function as classifiers without providing explanations, and their reliance on supervised fine-tuning with static datasets limits generalization. Inspired by recent advances, we reframe stepwise reward modeling from a classification task to a reasoning task itself. We thus propose a generative judge that reasons about the policy model's reasoning steps (i.e., meta-reasons), outputting thinking tokens before delivering a final verdict. Our model, StepWiser, is trained by reinforcement learning using relative outcomes of rollouts. We show it provides (i) better judgment accuracy on intermediate steps than existing methods; (ii) can be used to improve the policy model at training time; and (iii) improves inference-time search.

Wenting Zhao, Ye Liu, Yao Wan et al.

Task-Oriented Parsing (TOP) enables conversational assistants to interpret user commands expressed in natural language, transforming them into structured outputs that combine elements of both natural language and intent/slot tags. Recently, Large Language Models (LLMs) have achieved impressive performance in synthesizing computer programs based on a natural language prompt, mitigating the gap between natural language and structured programs. Our paper focuses on harnessing the capabilities of LLMs for semantic parsing tasks, addressing the following three key research questions: 1) How can LLMs be effectively utilized for semantic parsing tasks? 2) What defines an effective prompt? and 3) How can LLM overcome the length constraint and streamline prompt design by including all examples as prompts? We introduce k Nearest Neighbor In-Context Learning(kNN-ICL), which simplifies prompt engineering by allowing it to be built on top of any design strategy while providing access to all demo examples. Extensive experiments show that: 1)Simple ICL without kNN search can achieve a comparable performance with strong supervised models on the TOP tasks, and 2) kNN-ICL significantly improves the comprehension of complex requests by seamlessly integrating ICL with a nearest-neighbor approach. Notably, this enhancement is achieved without the need for additional data or specialized prompts.

Bo Liu, Chuanyang Jin, Seungone Kim et al. · meta-ai

Self-improving systems require environmental interaction for continuous adaptation. We introduce SPICE (Self-Play In Corpus Environments), a reinforcement learning framework where a single model acts in two roles: a Challenger that mines documents from a large corpus to generate diverse reasoning tasks, and a Reasoner that solves them. Through adversarial dynamics, the Challenger creates an automatic curriculum at the frontier of the Reasoner's capability, while corpus grounding provides the rich, near-inexhaustible external signal necessary for sustained improvement. Unlike existing ungrounded self-play methods that offer more limited benefits, SPICE achieves consistent gains across mathematical (+8.9%) and general reasoning (+9.8%) benchmarks on multiple model families. Our analysis reveals how document grounding is a key ingredient in SPICE to continuously generate its own increasingly challenging goals and achieve them, enabling sustained self-improvement.

Ming Zhu, Juntao Tan, Rithesh Murthy et al.

LLM-based user simulation is the primary mechanism for end-to-end agent evaluation, yet simulated users are poor proxies for real humans: unconstrained LLM defaults produce a Formalism Ceiling (style match rates of 6-8% against real users), while hand-crafted behavioral directives trigger Directive Amplification, where models hyper-interpret instructions into unnatural behavioral extremes that vary dramatically across simulator models. We present RealUserSim, the first user simulation framework grounded in real behavioral data. From 14,000+ authentic human-LLM conversations (WildChat), we extract 7,275 executable behavioral profiles and use them to ground LLM simulators. A fidelity benchmark (PT3) on 600 conversations across 71+ domains with anti-leakage controls shows that grounded simulation raises match rate from 24.2% to 45.3% across five behavioral dimensions. Agent evaluation on TauBench with 6 simulator models and extensive analysis shows that grounded simulation acts as a realistic stress test, surfacing three failure mechanisms invisible to cooperative simulators (mean -3.2% to -3.5% task success degradation), while Directive Amplification in existing benchmarks produces unrealistic behavior that compromises the validity of agent evaluation.

Chuanyang Jin, Jing Xu, Bo Liu et al.

We posit that to achieve continual model improvement and multifaceted alignment, future models must learn from natural human interaction. Current conversational models are aligned using pre-annotated, expert-generated human feedback. In this work, we introduce Reinforcement Learning from Human Interaction (RLHI), a paradigm that learns directly from in-the-wild user conversations. We develop two complementary methods: (1) RLHI with User-Guided Rewrites, which revises unsatisfactory model outputs based on users' natural-language follow-up responses, (2) RLHI with User-Based Rewards, which learns via a reward model conditioned on knowledge of the user's long-term interaction history (termed persona). Together, these methods link long-term user personas to turn-level preferences via persona-conditioned preference optimization. Trained on conversations derived from WildChat, both RLHI variants outperform strong baselines in personalization and instruction-following, and similar feedback enhances performance on reasoning benchmarks. These results suggest organic human interaction offers scalable, effective supervision for personalized alignment.

Chia Hsiang Kao, Wenting Zhao, Shreelekha Revankar et al.

Earth Observation (EO) provides critical planetary data for environmental monitoring, disaster management, climate science, and other scientific domains. Here we ask: Are AI systems ready for reliable Earth Observation? We introduce \datasetnamenospace, a benchmark of 140 yes/no questions from NASA Earth Observatory articles across 13 topics and 17 satellite sensors. Using Google Earth Engine API as a tool, LLM agents can only achieve an accuracy of 33% because the code fails to run over 58% of the time. We improve the failure rate for open models by fine-tuning synthetic data, allowing much smaller models (Llama-3.1-8B) to achieve comparable accuracy to much larger ones (e.g., DeepSeek-R1). Taken together, our findings identify significant challenges to be solved before AI agents can automate earth observation, and suggest paths forward. The project page is available at https://iandrover.github.io/UnivEarth.

Ziga Kovacic, Justin T. Chiu, Celine Lee et al.

Maintainable and general software allows developers to build robust applications efficiently, yet achieving these qualities often requires refactoring specialized solutions into reusable components. This challenge becomes particularly relevant as code agents become used to solve isolated one-off programming problems. We investigate code agents' capacity to refactor code in ways that support growth and reusability. We first investigate what makes a good refactoring, finding via simulation results and a human study that Minimum Description Length best correlates with preferable refactorings. We then present both a benchmark and a method for refactoring: MiniCode, a benchmark where multiple files must be refactored into a shared library, and Librarian, a sample-and-rerank method for generating reusable libraries. We compare Librarian to state-of-the-art library generation methods, and study it on real-world code bases.

Shaden Shaar, Wayne Chen, Maitreyi Chatterjee et al. · cmu

Most existing work on event extraction has focused on sentence-level texts and presumes the identification of a trigger-span -- a word or phrase in the input that evokes the occurrence of an event of interest. Event arguments are then extracted with respect to the trigger. Indeed, triggers are treated as integral to, and trigger detection as an essential component of, event extraction. In this paper, we provide the first investigation of the role of triggers for the more difficult and much less studied task of document-level event extraction. We analyze their usefulness in multiple end-to-end and pipelined transformer-based event extraction models for three document-level event extraction datasets, measuring performance using triggers of varying quality (human-annotated, LLM-generated, keyword-based, and random). We find that whether or not systems benefit from explicitly extracting triggers depends both on dataset characteristics (i.e. the typical number of events per document) and task-specific information available during extraction (i.e. natural language event schemas). Perhaps surprisingly, we also observe that the mere existence of triggers in the input, even random ones, is important for prompt-based in-context learning approaches to the task.

Jiangshu Du, Yibo Wang, Wenting Zhao et al.

This work is motivated by two key trends. On one hand, large language models (LLMs) have shown remarkable versatility in various generative tasks such as writing, drawing, and question answering, significantly reducing the time required for many routine tasks. On the other hand, researchers, whose work is not only time-consuming but also highly expertise-demanding, face increasing challenges as they have to spend more time reading, writing, and reviewing papers. This raises the question: how can LLMs potentially assist researchers in alleviating their heavy workload? This study focuses on the topic of LLMs assist NLP Researchers, particularly examining the effectiveness of LLM in assisting paper (meta-)reviewing and its recognizability. To address this, we constructed the ReviewCritique dataset, which includes two types of information: (i) NLP papers (initial submissions rather than camera-ready) with both human-written and LLM-generated reviews, and (ii) each review comes with "deficiency" labels and corresponding explanations for individual segments, annotated by experts. Using ReviewCritique, this study explores two threads of research questions: (i) "LLMs as Reviewers", how do reviews generated by LLMs compare with those written by humans in terms of quality and distinguishability? (ii) "LLMs as Metareviewers", how effectively can LLMs identify potential issues, such as Deficient or unprofessional review segments, within individual paper reviews? To our knowledge, this is the first work to provide such a comprehensive analysis.

Wenting Zhao, Justin T. Chiu, Claire Cardie et al.

Abductive reasoning aims to find plausible explanations for an event. This style of reasoning is critical for commonsense tasks where there are often multiple plausible explanations. Existing approaches for abductive reasoning in natural language processing (NLP) often rely on manually generated annotations for supervision; however, such annotations can be subjective and biased. Instead of using direct supervision, this work proposes an approach for abductive commonsense reasoning that exploits the fact that only a subset of explanations is correct for a given context. The method uses posterior regularization to enforce a mutual exclusion constraint, encouraging the model to learn the distinction between fluent explanations and plausible ones. We evaluate our approach on a diverse set of abductive reasoning datasets; experimental results show that our approach outperforms or is comparable to directly applying pretrained language models in a zero-shot manner and other knowledge-augmented zero-shot methods.

Wenting Zhao, Justin T. Chiu, Claire Cardie et al.

Explainable multi-hop question answering (QA) not only predicts answers but also identifies rationales, i. e. subsets of input sentences used to derive the answers. This problem has been extensively studied under the supervised setting, where both answer and rationale annotations are given. Because rationale annotations are expensive to collect and not always available, recent efforts have been devoted to developing methods that do not rely on supervision for rationales. However, such methods have limited capacities in modeling interactions between sentences, let alone reasoning across multiple documents. This work proposes a principled, probabilistic approach for training explainable multi-hop QA systems without rationale supervision. Our approach performs multi-hop reasoning by explicitly modeling rationales as sets, enabling the model to capture interactions between documents and sentences within a document. Experimental results show that our approach is more accurate at selecting rationales than the previous methods, while maintaining similar accuracy in predicting answers.

Di Chen, Yiwei Bai, Sebastian Ament et al.

Crystal-structure phase mapping is a core, long-standing challenge in materials science that requires identifying crystal structures, or mixtures thereof, in synthesized materials. Materials science experts excel at solving simple systems but cannot solve complex systems, creating a major bottleneck in high-throughput materials discovery. Herein we show how to automate crystal-structure phase mapping. We formulate phase mapping as an unsupervised pattern demixing problem and describe how to solve it using Deep Reasoning Networks (DRNets). DRNets combine deep learning with constraint reasoning for incorporating scientific prior knowledge and consequently require only a modest amount of (unlabeled) data. DRNets compensate for the limited data by exploiting and magnifying the rich prior knowledge about the thermodynamic rules governing the mixtures of crystals with constraint reasoning seamlessly integrated into neural network optimization. DRNets are designed with an interpretable latent space for encoding prior-knowledge domain constraints and seamlessly integrate constraint reasoning into neural network optimization. DRNets surpass previous approaches on crystal-structure phase mapping, unraveling the Bi-Cu-V oxide phase diagram, and aiding the discovery of solar-fuels materials.