Yisi Sang

Minhua Lin, Juncheng Wu, Zijun Wang et al.

LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.

Ying Xu, Dakuo Wang, Mo Yu et al. · cmu

Question answering (QA) is a fundamental means to facilitate assessment and training of narrative comprehension skills for both machines and young children, yet there is scarcity of high-quality QA datasets carefully designed to serve this purpose. In particular, existing datasets rarely distinguish fine-grained reading skills, such as the understanding of varying narrative elements. Drawing on the reading education research, we introduce FairytaleQA, a dataset focusing on narrative comprehension of kindergarten to eighth-grade students. Generated by educational experts based on an evidence-based theoretical framework, FairytaleQA consists of 10,580 explicit and implicit questions derived from 278 children-friendly stories, covering seven types of narrative elements or relations. Our dataset is valuable in two folds: First, we ran existing QA models on our dataset and confirmed that this annotation helps assess models' fine-grained learning skills. Second, the dataset supports question generation (QG) task in the education domain. Through benchmarking with QG models, we show that the QG model trained on FairytaleQA is capable of asking high-quality and more diverse questions.

Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran et al. · amazon-science, cmu

Evaluation in machine learning is usually informed by past choices, for example which datasets or metrics to use. This standardization enables the comparison on equal footing using leaderboards, but the evaluation choices become sub-optimal as better alternatives arise. This problem is especially pertinent in natural language generation which requires ever-improving suites of datasets, metrics, and human evaluation to make definitive claims. To make following best model evaluation practices easier, we introduce GEMv2. The new version of the Generation, Evaluation, and Metrics Benchmark introduces a modular infrastructure for dataset, model, and metric developers to benefit from each others work. GEMv2 supports 40 documented datasets in 51 languages. Models for all datasets can be evaluated online and our interactive data card creation and rendering tools make it easier to add new datasets to the living benchmark.

Zewen Liu, Zhan Shi, Yisi Sang et al.

Auto-harness systems such as A-Evolve, GEPA, and Meta-Harness improve LLM agents by optimizing prompts, skills, tools, memories, and supporting infrastructure from execution feedback, but they are typically evaluated on fixed offline benchmarks. Real deployments instead present open-ended task streams: histories grow without a fixed endpoint, heterogeneous tasks require different harnesses, and problem distributions shift over time. These challenges make a single repeatedly and densely updated harness brittle, causing performance degradation as accuracy peaks early and then declines. This motivates sustained harness construction with task-wise adaptation. We introduce Adaptive Auto-Harness, a framework and system for such streams. The framework decomposes the gap to an oracle harness into evolution loss and adaptation loss. The system addresses these losses with a stateful multi-agent evolver, a harness tree with solve-time routing, and human-steering hooks for cases where history lacks the needed signal. Across prediction-market, security-competition, and event-forecasting streams, Adaptive Auto-Harness outperforms five existing auto-harness baselines and ablations attribute gains to better construction, routing, or targeted human steering. Code is available in https://github.com/A-EVO-Lab/AdaptiveHarness .

Bingsheng Yao, Guiming Chen, Ruishi Zou et al.

While most existing works on LLM prompting techniques focus only on how to select a better set of data samples inside one single prompt input (In-Context Learning or ICL), why can not we design and leverage multiple prompts together to further improve the LLM's performance? In this work, we propose In-Context Sampling (ICS), a low-resource LLM prompting technique to produce confident predictions by optimizing the construction of multiple ICL prompt inputs. Extensive experiments with three open-source LLMs (FlanT5-XL, Mistral-7B, and Mixtral-8x7B) on four NLI datasets (e-SNLI, Multi-NLI, ANLI, and Contract-NLI) and one QA dataset (CommonsenseQA) illustrate that ICS can consistently enhance LLMs' performance. An in-depth evaluation with three data similarity-based ICS strategies suggests that these strategies can further elevate LLM's performance, which sheds light on a new yet promising future research direction.

Mo Yu, Qiujing Wang, Shunchi Zhang et al. · ibm-research

When reading a story, humans can quickly understand new fictional characters with a few observations, mainly by drawing analogies to fictional and real people they already know. This reflects the few-shot and meta-learning essence of humans' inference of characters' mental states, i.e., theory-of-mind (ToM), which is largely ignored in existing research. We fill this gap with a novel NLP dataset, ToM-in-AMC, the first assessment of machines' meta-learning of ToM in a realistic narrative understanding scenario. Our dataset consists of ~1,000 parsed movie scripts, each corresponding to a few-shot character understanding task that requires models to mimic humans' ability of fast digesting characters with a few starting scenes in a new movie. We propose a novel ToM prompting approach designed to explicitly assess the influence of multiple ToM dimensions. It surpasses existing baseline models, underscoring the significance of modeling multiple ToM dimensions for our task. Our extensive human study verifies that humans are capable of solving our problem by inferring characters' mental states based on their previously seen movies. In comparison, our systems based on either state-of-the-art large language models (GPT-4) or meta-learning algorithms lags >20% behind, highlighting a notable limitation in existing approaches' ToM capabilities.

Yisi Sang, Xiangyang Mou, Mo Yu et al. · ibm-research

We propose a new task for assessing machines' skills of understanding fictional characters in narrative stories. The task, TVShowGuess, builds on the scripts of TV series and takes the form of guessing the anonymous main characters based on the backgrounds of the scenes and the dialogues. Our human study supports that this form of task covers comprehension of multiple types of character persona, including understanding characters' personalities, facts and memories of personal experience, which are well aligned with the psychological and literary theories about the theory of mind (ToM) of human beings on understanding fictional characters during reading. We further propose new model architectures to support the contextualized encoding of long scene texts. Experiments show that our proposed approaches significantly outperform baselines, yet still largely lag behind the (nearly perfect) human performance. Our work serves as a first step toward the goal of narrative character comprehension.

Yisi Sang, Xiangyang Mou, Jing Li et al. · ibm-research

As the body of research on machine narrative comprehension grows, there is a critical need for consideration of performance assessment strategies as well as the depth and scope of different benchmark tasks. Based on narrative theories, reading comprehension theories, as well as existing machine narrative reading comprehension tasks and datasets, we propose a typology that captures the main similarities and differences among assessment tasks; and discuss the implications of our typology for new task design and the challenges of narrative reading comprehension.

Yisi Sang, Xiangyang Mou, Mo Yu et al.

An NLP model that understands stories should be able to understand the characters in them. To support the development of neural models for this purpose, we construct a benchmark, Story2Personality. The task is to predict a movie character's MBTI or Big 5 personality types based on the narratives of the character. Experiments show that our task is challenging for the existing text classification models, as none is able to largely outperform random guesses. We further proposed a multi-view model for personality prediction using both verbal and non-verbal descriptions, which gives improvement compared to using only verbal descriptions. The uniqueness and challenges in our dataset call for the development of narrative comprehension techniques from the perspective of understanding characters.

Yuxuan Lu, Jing Huang, Yan Han et al.

Recent research shows that LLM Agents can generate ``believable'' human behaviors via prompt-only methods, and such agents have been increasingly adopted in downstream applications. However, existing evaluation of these agents only focuses on qualitative believability (whether human raters think they are accurate), leaving open questions of whether LLM agents can accurately generate step-by-step actions mimicking a particular human's behavior in a multi-turn interaction task. In this work, we take shopping as a case study and present the first large-scale quantitative evaluation of state-of-the-art LLMs' ability to accurately simulate human behavior. Using real-world data from 31,865 online shopping sessions containing 230,965 user actions, our evaluation reveals that prompt-based LLMs (DeepSeek-R1, Llama, Claude) achieve only 11.86% accuracy in generating human actions, highlighting a substantial gap in actual behavioral accuracy. Through experiments, we also showcase that strategies as simple as fine-tuning LLMs on real human click-through data augmented with synthesized reasoning traces can greatly enhance models' performance. The fine-tuned Qwen2.5-7B achieves 17.26% action generation accuracy and 33.86% F1 score on final purchase prediction, representing substantial improvements of 5.4% and 13.85% over prompt-only baselines. This work establishes the first rigorous benchmark for human behavior simulation and provides actionable insights for developing more accurate LLM agents for future downstream applications.

Kun Qian, Yisi Sang, Farima Fatahi Bayat et al.

Prompt engineering is an iterative procedure often requiring extensive manual effort to formulate suitable instructions for effectively directing large language models (LLMs) in specific tasks. Incorporating few-shot examples is a vital and effective approach to providing LLMs with precise instructions, leading to improved LLM performance. Nonetheless, identifying the most informative demonstrations for LLMs is labor-intensive, frequently entailing sifting through an extensive search space. In this demonstration, we showcase a human-in-the-loop tool called APE (Active Prompt Engineering) designed for refining prompts through active learning. Drawing inspiration from active learning, APE iteratively selects the most ambiguous examples for human feedback, which will be transformed into few-shot examples within the prompt. The demo recording can be found with the submission or be viewed at https://youtu.be/OwQ6MQx53-Y.

Ronak Pradeep, Daniel Lee, Ali Mousavi et al.

The rapid advancement of Large Language Models (LLMs) and conversational assistants necessitates dynamic, scalable, and configurable conversational datasets for training and evaluation. These datasets must accommodate diverse user interaction modes, including text and voice, each presenting unique modeling challenges. Knowledge Graphs (KGs), with their structured and evolving nature, offer an ideal foundation for current and precise knowledge. Although human-curated KG-based conversational datasets exist, they struggle to keep pace with the rapidly changing user information needs. We present ConvKGYarn, a scalable method for generating up-to-date and configurable conversational KGQA datasets. Qualitative psychometric analyses confirm our method can generate high-quality datasets rivaling a popular conversational KGQA dataset while offering it at scale and covering a wide range of human-interaction configurations. We showcase its utility by testing LLMs on diverse conversations - exploring model behavior on conversational KGQA sets with different configurations grounded in the same KG fact set. Our results highlight the ability of ConvKGYarn to improve KGQA foundations and evaluate parametric knowledge of LLMs, thus offering a robust solution to the constantly evolving landscape of conversational assistants.

Kun Qian, Anton Belyi, Fei Wu et al.

The quality of a knowledge graph directly impacts the quality of downstream applications (e.g. the number of answerable questions using the graph). One ongoing challenge when building a knowledge graph is to ensure completeness and freshness of the graph's entities and facts. In this paper, we introduce ODKE, a scalable and extensible framework that sources high-quality entities and facts from open web at scale. ODKE utilizes a wide range of extraction models and supports both streaming and batch processing at different latency. We reflect on the challenges and design decisions made and share lessons learned when building and deploying ODKE to grow an industry-scale open domain knowledge graph.

Farima Fatahi Bayat, Kun Qian, Benjamin Han et al.

Detecting factual errors in textual information, whether generated by large language models (LLM) or curated by humans, is crucial for making informed decisions. LLMs' inability to attribute their claims to external knowledge and their tendency to hallucinate makes it difficult to rely on their responses. Humans, too, are prone to factual errors in their writing. Since manual detection and correction of factual errors is labor-intensive, developing an automatic approach can greatly reduce human effort. We present FLEEK, a prototype tool that automatically extracts factual claims from text, gathers evidence from external knowledge sources, evaluates the factuality of each claim, and suggests revisions for identified errors using the collected evidence. Initial empirical evaluation on fact error detection (77-85\% F1) shows the potential of FLEEK. A video demo of FLEEK can be found at https://youtu.be/NapJFUlkPdQ.

Yuxuan Lu, Ziyi Wang, Yingzhou Lu et al.

Training tool-calling agents requires large-scale trajectory data with verifiable labels, yet existing approaches either synthesize environments that diverge from real API behavior or generate tasks without ground-truth outcomes for verification. We present FireFly, a pipeline for generating verified tool-call data from real-world MCP servers. Our key insight is to invert the standard synthesis pipeline: rather than generating tasks and hoping they are solvable, we first let a strong LLM explore real APIs along graph-guided DAG structures, then synthesize tasks backward from observed outcomes, guaranteeing label correctness by construction. To handle the scale of real-world tool spaces (${\sim}$1,000 tools), we build a pairwise tool graph and sample sub-DAGs to focus exploration on semantically coherent workflows. To address environment drift in live APIs, we construct a retrieval-augmented simulator that caches all exploration results and replays them during training and evaluation, enabling fully offline and reproducible RL. Applying this pipeline yields 5,144 verified tasks spanning 240 servers and 993 tools. A 4B-parameter model trained with GRPO on FireFly matches Claude Sonnet 4.6 on our held-out test set and shows improvements on multiple tool-calling benchmarks including Tau2-Bench, MCPMark, and MCP-Atlas.

Ziyi Wang, Yuxuan Lu, Yimeng Zhang et al.

Tool-calling agents are increasingly deployed in real-world customer-facing workflows. Yet most studies on tool-calling agents focus on idealized settings with general, fixed, and well-specified tasks. In real-world applications, user requests are often (1) ambiguous, (2) changing over time, or (3) infeasible due to policy constraints, and training and evaluation data that cover these diverse, complex interaction patterns remain under-represented. To bridge the gap, we present Trajectory2Task, a verifiable data generation pipeline for studying tool use at scale under three realistic user scenarios: ambiguous intent, changing intent, and infeasible intents. The pipeline first conducts multi-turn exploration to produce valid tool-call trajectories. It then converts these trajectories into user-facing tasks with controlled intent adaptations. This process yields verifiable task that support closed-loop evaluation and training. We benchmark seven state-of-the-art LLMs on the generated complex user scenario tasks and observe frequent failures. Finally, using successful trajectories obtained from task rollouts, we fine-tune lightweight LLMs and find consistent improvements across all three conditions, along with better generalization to unseen tool-use domains, indicating stronger general tool-calling ability.

Samira Khorshidi, Azadeh Nikfarjam, Suprita Shankar et al.

Knowledge graphs (KGs) are foundational to many AI applications, but maintaining their freshness and completeness remains costly. We present ODKE+, a production-grade system that automatically extracts and ingests millions of open-domain facts from web sources with high precision. ODKE+ combines modular components into a scalable pipeline: (1) the Extraction Initiator detects missing or stale facts, (2) the Evidence Retriever collects supporting documents, (3) hybrid Knowledge Extractors apply both pattern-based rules and ontology-guided prompting for large language models (LLMs), (4) a lightweight Grounder validates extracted facts using a second LLM, and (5) the Corroborator ranks and normalizes candidate facts for ingestion. ODKE+ dynamically generates ontology snippets tailored to each entity type to align extractions with schema constraints, enabling scalable, type-consistent fact extraction across 195 predicates. The system supports batch and streaming modes, processing over 9 million Wikipedia pages and ingesting 19 million high-confidence facts with 98.8% precision. ODKE+ significantly improves coverage over traditional methods, achieving up to 48% overlap with third-party KGs and reducing update lag by 50 days on average. Our deployment demonstrates that LLM-based extraction, grounded in ontological structure and verification workflows, can deliver trustworthiness, production-scale knowledge ingestion with broad real-world applicability. A recording of the system demonstration is included with the submission and is also available at https://youtu.be/UcnE3_GsTWs.

Yisi Sang, Jeffrey Stanton

Human annotated data is the cornerstone of today's artificial intelligence efforts, yet data labeling processes can be complicated and expensive, especially when human labelers disagree with each other. The current work practice is to use majority-voted labels to overrule the disagreement. However, in the subjective data labeling tasks such as hate speech annotation, disagreement among individual labelers can be difficult to resolve. In this paper, we explored why such disagreements occur using a mixed-method approach - including interviews with experts, concept mapping exercises, and self-reporting items - to develop a multidimensional scale for distilling the process of how annotators label a hate speech corpus. We tested this scale with 170 annotators in a hate speech annotation task. Results showed that our scale can reveal facets of individual differences among annotators (e.g., age, personality, etc.), and these facets' relationships to an annotator's final label decision of an instance. We suggest that this work contributes to the understanding of how humans annotate data. The proposed scale can potentially improve the value of the currently discarded minority-vote labels.

Qunfang Wu, Yisi Sang, Dakuo Wang et al.

Due to the limitations imposed by the COVID-19 pandemic, many users have shifted their shopping patterns from offline to online. Livestream shopping has become popular as one of the online shopping media. However, many streamers' malicious selling behaviors have been reported. In this research, we sought to explore streamers' malicious selling strategies and understand how viewers perceive these strategies. First, we recorded 40 livestream shopping sessions from two popular livestream platforms in China -- Taobao and TikTok (or "Douyin" in Chinese). We identified 16 malicious selling strategies and found that platform designs enhanced these malicious selling strategies. Second, through an interview study with 13 viewers, we provide a rich description of viewers' awareness of malicious selling strategies and the challenges they encountered while trying to overcome malicious selling. We conclude by discussing the policy and design implications of countering malicious selling.