Pei Zhou

Hyunwoo Kim, Jack Hessel, Liwei Jiang et al. · allen-ai, cmu

Data scarcity has been a long standing issue in the field of open-domain social dialogue. To quench this thirst, we present SODA: the first publicly available, million-scale high-quality social dialogue dataset. By contextualizing social commonsense knowledge from a knowledge graph, we are able to distill an exceptionally broad spectrum of social interactions from a large language model. Human evaluation shows that conversations in SODA are more consistent, specific, and (surprisingly) natural than those in prior human-authored datasets. Using SODA, we train COSMO: a generalizable conversation model that is significantly more natural and consistent on unseen datasets than best-performing conversation models (e.g., GODEL, BlenderBot-1, Koala, Vicuna). Experiments reveal COSMO is sometimes even preferred to the original human-written gold responses. Additionally, our results shed light on the distinction between knowledge-enriched conversations and natural social chitchats. We plan to make our data, model, and code public.

Pei Zhou, Aman Madaan, Srividya Pranavi Potharaju et al. · cmu

"Thinking is for Doing." Humans can infer other people's mental states from observations--an ability called Theory-of-Mind (ToM)--and subsequently act pragmatically on those inferences. Existing question answering benchmarks such as ToMi ask models questions to make inferences about beliefs of characters in a story, but do not test whether models can then use these inferences to guide their actions. We propose a new evaluation paradigm for large language models (LLMs): Thinking for Doing (T4D), which requires models to connect inferences about others' mental states to actions in social scenarios. Experiments on T4D demonstrate that LLMs such as GPT-4 and PaLM 2 seemingly excel at tracking characters' beliefs in stories, but they struggle to translate this capability into strategic action. Our analysis reveals the core challenge for LLMs lies in identifying the implicit inferences about mental states without being explicitly asked about as in ToMi, that lead to choosing the correct action in T4D. To bridge this gap, we introduce a zero-shot prompting framework, Foresee and Reflect (FaR), which provides a reasoning structure that encourages LLMs to anticipate future challenges and reason about potential actions. FaR boosts GPT-4's performance from 50% to 71% on T4D, outperforming other prompting methods such as Chain-of-Thought and Self-Ask. Moreover, FaR generalizes to diverse out-of-distribution story structures and scenarios that also require ToM inferences to choose an action, consistently outperforming other methods including few-shot in-context learning.

Pranjal Aggarwal, Aman Madaan, Ankit Anand et al. · cmu

Large language models (LLMs) are now available from cloud API providers in various sizes and configurations. While this diversity offers a broad spectrum of choices, effectively leveraging the options to optimize computational cost and performance remains challenging. In this work, we present Automix, an approach that strategically routes queries to larger LMs, based on the approximate correctness of outputs from a smaller LM. Central to Automix are two key technical contributions. First, it has a few-shot self-verification mechanism, which estimates the reliability of its own outputs without requiring extensive training. Second, given that self-verification can be noisy, it employs a POMDP based router that can effectively select an appropriately sized model, based on answer confidence. Experiments across five language models and five challenging datasets show that Automix consistently surpasses strong baselines, reducing computational cost by over 50% for comparable performance.

Pei Zhou, Hyundong Cho, Pegah Jandaghi et al. · allen-ai

Human communication relies on common ground (CG), the mutual knowledge and beliefs shared by participants, to produce coherent and interesting conversations. In this paper, we demonstrate that current response generation (RG) models produce generic and dull responses in dialogues because they act reflexively, failing to explicitly model CG, both due to the lack of CG in training data and the standard RG training procedure. We introduce Reflect, a dataset that annotates dialogues with explicit CG (materialized as inferences approximating shared knowledge and beliefs) and solicits 9k diverse human-generated responses each following one common ground. Using Reflect, we showcase the limitations of current dialogue data and RG models: less than half of the responses in current data are rated as high quality (sensible, specific, and interesting) and models trained using this data have even lower quality, while most Reflect responses are judged high quality. Next, we analyze whether CG can help models produce better-quality responses by using Reflect CG to guide RG models. Surprisingly, we find that simply prompting GPT3 to "think" about CG generates 30% more quality responses, showing promising benefits to integrating CG into the RG process.

Taiwei Shi, Zhuoer Wang, Longqi Yang et al. · amazon-science, microsoft-research

As large language models (LLMs) continue to advance, aligning these models with human preferences has emerged as a critical challenge. Traditional alignment methods, relying on human or LLM annotated datasets, are limited by their resource-intensive nature, inherent subjectivity, misalignment with real-world user preferences, and the risk of feedback loops that amplify model biases. To overcome these limitations, we introduce WildFeedback, a novel framework that leverages in-situ user feedback during conversations with LLMs to create preference datasets automatically. Given a corpus of multi-turn user-LLM conversation, WildFeedback identifies and classifies user feedback to LLM responses between conversation turns. The user feedback is then used to create examples of preferred and dispreferred responses according to users' preference. Our experiments demonstrate that LLMs fine-tuned on WildFeedback dataset exhibit significantly improved alignment with user preferences, as evidenced by both traditional benchmarks and our proposed checklist-guided evaluation. By incorporating in-situ feedback from actual users, WildFeedback addresses the scalability, subjectivity, and bias challenges that plague existing approaches, marking a significant step toward developing LLMs that are more responsive to the diverse and evolving needs of their users.

Pei Zhou, Andrew Zhu, Jennifer Hu et al.

We propose a novel task, G4C, to study teacher-student natural language interactions in a goal-driven and grounded environment. Dungeons and Dragons (D&D), a role-playing game, provides an ideal setting to investigate such interactions. Here, the Dungeon Master (DM), i.e., the teacher, guides the actions of several players -- students, each with their own personas and abilities -- to achieve shared goals grounded in a fantasy world. Our approach is to decompose and model these interactions into (1) the DM's intent to guide players toward a given goal; (2) the DM's guidance utterance to the players expressing this intent; and (3) a theory-of-mind (ToM) model that anticipates the players' reaction to the guidance one turn into the future. We develop a novel reinforcement learning (RL) method for training a DM that generates guidance for players by rewarding utterances where the intent matches the ToM-anticipated player actions. Human and automated evaluations show that a DM trained to explicitly model intents and incorporate ToM of the players using RL generates better-quality guidance that is 3x more likely to fulfill the DM's intent than a vanilla natural language generation (NLG) approach.

Zhiping Xiao, Jeffrey Zhu, Yining Wang et al. · uw

Ideological divisions in the United States have become increasingly prominent in daily communication. Accordingly, there has been much research on political polarization, including many recent efforts that take a computational perspective. By detecting political biases in a corpus of text, one can attempt to describe and discern the polarity of that text. Intuitively, the named entities (i.e., the nouns and the phrases that act as nouns) and hashtags in text often carry information about political views. For example, people who use the term "pro-choice" are likely to be liberal, whereas people who use the term "pro-life" are likely to be conservative. In this paper, we seek to reveal political polarities in social-media text data and to quantify these polarities by explicitly assigning a polarity score to entities and hashtags. Although this idea is straightforward, it is difficult to perform such inference in a trustworthy quantitative way. Key challenges include the small number of known labels, the continuous spectrum of political views, and the preservation of both a polarity score and a polarity-neutral semantic meaning in an embedding vector of words. To attempt to overcome these challenges, we propose the Polarity-aware Embedding Multi-task learning (PEM) model. This model consists of (1) a self-supervised context-preservation task, (2) an attention-based tweet-level polarity-inference task, and (3) an adversarial learning task that promotes independence between an embedding's polarity dimension and its semantic dimensions. Our experimental results demonstrate that our PEM model can successfully learn polarity-aware embeddings that perform well classification tasks. We examine a variety of applications and we thereby demonstrate the effectiveness of our PEM model. We also discuss important limitations of our work and encourage caution when applying the it to real-world scenarios.

Hanxiang Ren, Pei Zhou, Xunzhe Zhou et al.

Language-conditioned manipulation policies typically process instructions and observations through shared network parameters. This task-state entanglement provides a pathway for observation leakage -- networks learn scene-to-action shortcuts that bypass language grounding entirely. DISC eliminates this failure structurally. Rather than conditioning a universal policy on language, DISC uses a hypernetwork to generate the entire parameter set of a task-specific visuomotor policy from the instruction alone. The generated policy never directly accesses language; therefore, its task-awareness must come from the language. Consequently, observation leakage has no pathway to emerge. On the other hand, generating coherent high-dimensional policy weights is itself a challenging problem. We address it with a two-stage hypernetwork whose refinement stage embeds the structure of gradient-based optimization as a feed-forward inductive bias, producing globally consistent parameters without actual gradient computation. Trained entirely from scratch on standard data budgets, DISC outperforms all entangled baselines on LIBERO-90 and Meta-World, with advantages that widen on complex, long-horizon tasks -- and surpasses the large-scale pretrained $π_0$ despite using no external pretraining data. On a real-world benchmark where all tasks share identical visual context, DISC substantially outperforms entangled alternatives, directly confirming that language-generated policy parameters, not visual shortcuts, drive behavior. The hypernetwork further learns a semantically structured parameter manifold that enables few-shot adaptation from minimal demonstrations and robust generalization across paraphrased instructions. Our code is available at: {https://github.com/ReNginx/DISC}.

Pan Wang, Yang Liu, Guile Wu et al.

4D spatial intelligence involves perceiving and processing how objects move or change over time. Humans naturally possess 4D spatial intelligence, supporting a broad spectrum of spatial reasoning abilities. To what extent can Multimodal Large Language Models (MLLMs) achieve human-level 4D spatial intelligence? In this work, we present Spatial4D-Bench, a versatile 4D spatial intelligence benchmark designed to comprehensively assess the 4D spatial reasoning abilities of MLLMs. Unlike existing spatial intelligence benchmarks that are often small-scale or limited in diversity, Spatial4D-Bench provides a large-scale, multi-task evaluation benchmark consisting of ~40,000 question-answer pairs covering 18 well-defined tasks. We systematically organize these tasks into six cognitive categories: object understanding, scene understanding, spatial relationship understanding, spatiotemporal relationship understanding, spatial reasoning and spatiotemporal reasoning. Spatial4D-Bench thereby offers a structured and comprehensive benchmark for evaluating the spatial cognition abilities of MLLMs, covering a broad spectrum of tasks that parallel the versatility of human spatial intelligence. We benchmark various state-of-the-art open-source and proprietary MLLMs on Spatial4D-Bench and reveal their substantial limitations in a wide variety of 4D spatial reasoning aspects, such as route plan, action recognition, and physical plausibility reasoning. We hope that the findings provided in this work offer valuable insights to the community and that our benchmark can facilitate the development of more capable MLLMs toward human-level 4D spatial intelligence. More resources can be found on our project page.

Weiwei Sun, Xuhui Zhou, Jiarui Liu et al.

Humans learn social norms and behaviors from verbal feedback (e.g., a parent saying "that was rude" or a friend explaining "here's why that hurt"). Yet, learning from feedback for LLMs has largely focused on domains like code and math, where RL rewards are directly verifiable and condensed into scalar values. As LLMs are increasingly used to simulate human behavior, e.g., standing in for users, patients, students, and other personas, there is a pressing need to make them more human-like, which requires embracing a fundamentally different kind of signal: feedback that is verbal, subjective, and multi-faceted. We present DITTO, a model trained by treating verbal feedback as a first-class signal in reinforcement learning. After each rollout, DITTO receives verbal feedback and generates a feedback-conditioned improved rollout; both outputs are jointly optimized with GRPO, distilling verbal guidance into the base policy without requiring feedback at test time. We also introduce SOUL (Simulation gym Of hUman-Like behavior), a unified benchmark and training data suite spanning 10 tasks across six categories: Theory of Mind, character role play, social skill, learner simulation, user simulation, and persona simulation. DITTO achieves an average 36% improvement over the base model and exceeds GPT-5.4 on 6 of 10 SOUL benchmarks, demonstrating that RL with verbal feedback is a promising direction for training LLMs to simulate human behavior.

Feng Chen, Tianzhe Chu, Li Sun et al.

Evaluating embodied systems on real dexterous hardware requires more than isolated primitive skills: an agent must perceive a changing tabletop scene, choose a context-appropriate action, execute it with a dexterous hand, and leave the scene usable for later decisions. We introduce DexHoldem, a real-world system-level benchmark built around Texas Hold'em dexterous manipulation with a ShadowHand. DexHoldem provides 1,470 teleoperated demonstrations across 14 Texas Hold'em manipulation primitives, a standardized physical policy benchmark, and an agentic perception benchmark that tests whether agents can recover the structured game state needed for embodied decision making. On primitive execution, $π_{0.5}$ obtains the highest task completion rate ($61.2\%$), while $π_{0.5}$ and $π_0$ tie on scene-preserving success rate ($47.5\%$). On agentic perception, Opus 4.7 obtains the best strict problem-level accuracy ($34.3\%$), while GPT 5.5 obtains the best average field-wise accuracy ($66.8\%$), exposing a gap between isolated visual sub-capabilities and complete routing-relevant state recovery. Finally, we instantiate the full embodied-agent loop in three case studies, where waiting, recovery dispatches, human-help requests, and repeated primitive execution reveal how perception and policy errors accumulate during closed-loop deployment. DexHoldem therefore evaluates dexterous tabletop execution, agentic perception, and embodied decision routing in a shared physical setting. Project page: https://dexholdem.github.io/Dexholdem/.

Bowen Jiang, Taiwei Shi, Ryo Kamoi et al.

This paper introduces OMAR: One Model, All Roles, a reinforcement learning framework that enables AI to develop social intelligence through multi-turn, multi-agent conversational self-play. Unlike traditional paradigms that rely on static, single-turn optimizations, OMAR allows a single model to role-play all participants in a conversation simultaneously, learning to achieve long-term goals and complex social norms directly from dynamic social interaction. To ensure training stability across long dialogues, we implement a hierarchical advantage estimation that calculates turn-level and token-level advantages. Evaluations in the SOTOPIA social environment and Werewolf strategy games show that our trained models develop fine-grained, emergent social intelligence, such as empathy, persuasion, and compromise seeking, demonstrating the effectiveness of learning collaboration even under competitive scenarios. While we identify practical challenges like reward hacking, our results show that rich social intelligence can emerge without human supervision. We hope this work incentivizes further research on AI social intelligence in group conversations.

Ryo Kamoi, Ameya Godbole, Longqi Yang et al.

Simulating human conversations using large language models (LLMs) has emerged as a scalable methodology for modeling human social interaction. However, simulating human conversations is challenging because they inherently involve inconsistent and uncollaborative behaviors, such as misunderstandings and interruptions. Analysis comparing inconsistent and uncollaborative behaviors in human- and LLM-generated conversations remains limited, although reproducing these behaviors is integral to simulating human-like and complex social interaction. In this work, we introduce CoCoEval, an evaluation framework that analyzes LLM-simulated conversations by detecting 10 types of inconsistent and uncollaborative behaviors at the turn level using an LLM-as-a-Judge. Using CoCoEval, we evaluate GPT-4.1, GPT-5.1, and Claude Opus 4 by comparing the frequencies of detected behaviors in conversations simulated by each model and in human conversations across academic, business, and governmental meetings, as well as debates. Our analysis shows that (1) under vanilla prompting, LLM-simulated conversations exhibit far fewer inconsistent and uncollaborative behaviors than human conversations; (2) prompt engineering does not provide reliable control over these behaviors, as our results show that different prompts lead to their under- or overproduction; and (3) supervised fine-tuning on human conversations can lead LLMs to overproduce a narrow set of behaviors, such as repetition. Our findings highlight the difficulty of simulating human conversations, raising concerns about the use of LLMs as a proxy for human social interaction.

Pei Zhou, Wanting Yao, Qian Luo et al.

Goal-conditioned policy learning for robotic manipulation presents significant challenges in maintaining performance across diverse objectives and environments. We introduce Hyper-GoalNet, a framework that generates task-specific policy network parameters from goal specifications using hypernetworks. Unlike conventional methods that simply condition fixed networks on goal-state pairs, our approach separates goal interpretation from state processing -- the former determines network parameters while the latter applies these parameters to current observations. To enhance representation quality for effective policy generation, we implement two complementary constraints on the latent space: (1) a forward dynamics model that promotes state transition predictability, and (2) a distance-based constraint ensuring monotonic progression toward goal states. We evaluate our method on a comprehensive suite of manipulation tasks with varying environmental randomization. Results demonstrate significant performance improvements over state-of-the-art methods, particularly in high-variability conditions. Real-world robotic experiments further validate our method's robustness to sensor noise and physical uncertainties. Code is available at: https://github.com/wantingyao/hyper-goalnet.

Dawei Wang, Lingping Gao, Ziquan Lan et al.

Recently, there have been many advances in autonomous driving society, attracting a lot of attention from academia and industry. However, existing works mainly focus on cars, extra development is still required for self-driving truck algorithms and models. In this paper, we introduce an intelligent self-driving truck system. Our presented system consists of three main components, 1) a realistic traffic simulation module for generating realistic traffic flow in testing scenarios, 2) a high-fidelity truck model which is designed and evaluated for mimicking real truck response in real-world deployment, 3) an intelligent planning module with learning-based decision making algorithm and multi-mode trajectory planner, taking into account the truck's constraints, road slope changes, and the surrounding traffic flow. We provide quantitative evaluations for each component individually to demonstrate the fidelity and performance of each part. We also deploy our proposed system on a real truck and conduct real world experiments which shows our system's capacity of mitigating sim-to-real gap. Our code is available at https://github.com/InceptioResearch/IITS

Pei Zhou, Jay Pujara, Xiang Ren et al.

We introduce SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. Core to the framework is a self-discovery process where LLMs select multiple atomic reasoning modules such as critical thinking and step-by-step thinking, and compose them into an explicit reasoning structure for LLMs to follow during decoding. SELF-DISCOVER substantially improves GPT-4 and PaLM 2's performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT). Furthermore, SELF-DISCOVER outperforms inference-intensive methods such as CoT-Self-Consistency by more than 20%, while requiring 10-40x fewer inference compute. Finally, we show that the self-discovered reasoning structures are universally applicable across model families: from PaLM 2-L to GPT-4, and from GPT-4 to Llama2, and share commonalities with human reasoning patterns.

Pei Zhou, Yanchao Yang

We aim to discover manipulation concepts embedded in the unannotated demonstrations, which are recognized as key physical states. The discovered concepts can facilitate training manipulation policies and promote generalization. Current methods relying on multimodal foundation models for deriving key states usually lack accuracy and semantic consistency due to limited multimodal robot data. In contrast, we introduce an information-theoretic criterion to characterize the regularities that signify a set of physical states. We also develop a framework that trains a concept discovery network using this criterion, thus bypassing the dependence on human semantics and alleviating costly human labeling. The proposed criterion is based on the observation that key states, which deserve to be conceptualized, often admit more physical constraints than non-key states. This phenomenon can be formalized as maximizing the mutual information between the putative key state and its preceding state, i.e., Maximal Mutual Information (MaxMI). By employing MaxMI, the trained key state localization network can accurately identify states of sufficient physical significance, exhibiting reasonable semantic compatibility with human perception. Furthermore, the proposed framework produces key states that lead to concept-guided manipulation policies with higher success rates and better generalization in various robotic tasks compared to the baselines, verifying the effectiveness of the proposed criterion.

Lei Sun, Hang Guo, Bin Ren et al.

This paper presents an overview of the NTIRE 2025 Image Denoising Challenge (σ = 50), highlighting the proposed methodologies and corresponding results. The primary objective is to develop a network architecture capable of achieving high-quality denoising performance, quantitatively evaluated using PSNR, without constraints on computational complexity or model size. The task assumes independent additive white Gaussian noise (AWGN) with a fixed noise level of 50. A total of 290 participants registered for the challenge, with 20 teams successfully submitting valid results, providing insights into the current state-of-the-art in image denoising.

Jie He, Jennifer Neville, Mengting Wan et al.

Large Language Models (LLMs) can enhance their capabilities as AI assistants by integrating external tools, allowing them to access a wider range of information. While recent LLMs are typically fine-tuned with tool usage examples during supervised fine-tuning (SFT), questions remain about their ability to develop robust tool-usage skills and can effectively generalize to unseen queries and tools. In this work, we present GenTool, a novel training framework that prepares LLMs for diverse generalization challenges in tool utilization. Our approach addresses two fundamental dimensions critical for real-world applications: Zero-to-One Generalization, enabling the model to address queries initially lacking a suitable tool by adopting and utilizing one when it becomes available, and Weak-to-Strong Generalization, allowing models to leverage enhanced versions of existing tools to solve queries. To achieve this, we develop synthetic training data simulating these two dimensions of tool usage and introduce a two-stage fine-tuning approach: optimizing tool ranking, then refining tool selection. Through extensive experiments across four generalization scenarios, we demonstrate that our method significantly enhances the tool-usage capabilities of LLMs ranging from 1B to 8B parameters, achieving performance that surpasses GPT-4o. Furthermore, our analysis also provides valuable insights into the challenges LLMs encounter in tool generalization.

Shuaiqi Wang, Vikas Raunak, Arturs Backurs et al.

Differentially private (DP) synthetic data generation is a promising technique for utilizing private datasets that otherwise cannot be exposed for model training or other analytics. While much research literature has focused on generating private unstructured text and image data, in enterprise settings, structured data (e.g., tabular) is more common, often including natural language fields or components. Existing synthetic data evaluation techniques (e.g., FID) struggle to capture the structural properties and correlations of such datasets. In this work, we propose Struct-Bench, a framework and benchmark for evaluating synthetic datasets derived from structured datasets that contain natural language data. The Struct-Bench framework requires users to provide a representation of their dataset structure as a Context-Free Grammar (CFG). Our benchmark comprises 5 real-world and 2 synthetically generated datasets, each annotated with CFGs. We show that these datasets demonstrably present a great challenge even for state-of-the-art DP synthetic data generation methods. Struct-Bench also includes reference implementations of different metrics and a leaderboard, thereby providing researchers a standardized evaluation platform to benchmark and investigate privacy-preserving synthetic data generation methods. Further, we also present a case study showing how to use Struct-Bench to improve the synthetic data quality of Private Evolution (PE) on structured data. The benchmark and the leaderboard have been publicly made available at https://struct-bench.github.io.

Ziyi Liu, Bahar Sarrafzadeh, Pei Zhou et al.

While Large Language Models (LLMs) are increasingly used in agentic frameworks to assist individual users, there is a growing need for agents that can proactively manage complex, multi-party collaboration. Systematic evaluation methods for such proactive agents remain scarce, limiting progress in developing AI that can effectively support multiple people together. Negotiation offers a demanding testbed for this challenge, requiring socio-cognitive intelligence to navigate conflicting interests between multiple participants and multiple topics and build consensus. Here, we present ProMediate, the first framework for evaluating proactive AI mediator agents in complex, multi-topic, multi-party negotiations. ProMediate consists of two core components: (i) a simulation testbed based on realistic negotiation cases and theory-driven difficulty levels (ProMediate-Easy, ProMediate-Medium, and ProMediate-Hard), with a plug-and-play proactive AI mediator grounded in socio-cognitive mediation theories, capable of flexibly deciding when and how to intervene; and (ii) a socio-cognitive evaluation framework with a new suite of metrics to measure consensus changes, intervention latency, mediator effectiveness, and intelligence. Together, these components establish a systematic framework for assessing the socio-cognitive intelligence of proactive AI agents in multi-party settings. Our results show that a socially intelligent mediator agent outperforms a generic baseline, via faster, better-targeted interventions. In the ProMediate-Hard setting, our social mediator increases consensus change by 3.6 percentage points compared to the generic baseline (10.65\% vs 7.01\%) while being 77\% faster in response (15.98s vs. 3.71s). In conclusion, ProMediate provides a rigorous, theory-grounded testbed to advance the development of proactive, socially intelligent agents.

Yushi Du, Yixuan Li, Baoxiong Jia et al. · pku

Human-humanoid collaboration shows significant promise for applications in healthcare, domestic assistance, and manufacturing. While compliant robot-human collaboration has been extensively developed for robotic arms, enabling compliant human-humanoid collaboration remains largely unexplored due to humanoids' complex whole-body dynamics. In this paper, we propose a proprioception-only reinforcement learning approach, COLA, that combines leader and follower behaviors within a single policy. The model is trained in a closed-loop environment with dynamic object interactions to predict object motion patterns and human intentions implicitly, enabling compliant collaboration to maintain load balance through coordinated trajectory planning. We evaluate our approach through comprehensive simulator and real-world experiments on collaborative carrying tasks, demonstrating the effectiveness, generalization, and robustness of our model across various terrains and objects. Simulation experiments demonstrate that our model reduces human effort by 24.7%. compared to baseline approaches while maintaining object stability. Real-world experiments validate robust collaborative carrying across different object types (boxes, desks, stretchers, etc.) and movement patterns (straight-line, turning, slope climbing). Human user studies with 23 participants confirm an average improvement of 27.4% compared to baseline models. Our method enables compliant human-humanoid collaborative carrying without requiring external sensors or complex interaction models, offering a practical solution for real-world deployment.

Liangjin Liu, Haoyang Zheng, Zhengzhong Zhu et al.

Isolated Sign Language Recognition (ISLR) is challenged by gestures that are morphologically similar yet semantically distinct, a problem rooted in the complex interplay between hand shape and motion trajectory. Existing methods, often relying on a single reference frame, struggle to resolve this geometric ambiguity. This paper introduces Dual-SignLanguageNet (DSLNet), a dual-reference, dual-stream architecture that decouples and models gesture morphology and trajectory in separate, complementary coordinate systems. The architecture processes these streams through specialized networks: a topology-aware graph convolution models the view-invariant shape from a wrist-centric frame, while a Finsler geometry-based encoder captures the context-aware trajectory from a facial-centric frame. These features are then integrated via a geometry-driven optimal transport fusion mechanism. DSLNet sets a new state-of-the-art, achieving 93.70%, 89.97%, and 99.79% accuracy on the challenging WLASL-100, WLASL-300, and LSA64 datasets, respectively, with significantly fewer parameters than competing models.

Tenghao Huang, Sihao Chen, Muhao Chen et al.

Large language models (LLMs) are increasingly expected to function as collaborative partners, engaging in back-and-forth dialogue to solve complex, ambiguous problems. However, current LLMs often falter in real-world settings, defaulting to passive responses or narrow clarifications when faced with incomplete or under-specified prompts, falling short of proactively gathering the missing information that is crucial for high-quality solutions. In this work, we introduce a new task paradigm: proactive information gathering, where LLMs must identify gaps in the provided context and strategically elicit implicit user knowledge through targeted questions. To systematically study and train this capability, we design a scalable framework that generates partially specified, real-world tasks, masking key information and simulating authentic ambiguity. Within this setup, our core innovation is a reinforcement finetuning strategy that rewards questions that elicit genuinely new, implicit user information -- such as hidden domain expertise or fine-grained requirements -- that would otherwise remain unspoken. Experiments demonstrate that our trained Qwen-2.5-7B model significantly outperforms o3-mini by 18% on automatic evaluation metrics. More importantly, human evaluation reveals that clarification questions and final outlines generated by our model are favored by human annotators by 42% and 28% respectively. Together, these results highlight the value of proactive clarification in elevating LLMs from passive text generators to genuinely collaborative thought partners.

Ziyi Liu, Abhishek Anand, Pei Zhou et al.

Large language models (LLMs) have demonstrated the potential to mimic human social intelligence. However, most studies focus on simplistic and static self-report or performance-based tests, which limits the depth and validity of the analysis. In this paper, we developed a novel framework, InterIntent, to assess LLMs' social intelligence by mapping their ability to understand and manage intentions in a game setting. We focus on four dimensions of social intelligence: situational awareness, self-regulation, self-awareness, and theory of mind. Each dimension is linked to a specific game task: intention selection, intention following, intention summarization, and intention guessing. Our findings indicate that while LLMs exhibit high proficiency in selecting intentions, achieving an accuracy of 88%, their ability to infer the intentions of others is significantly weaker, trailing human performance by 20%. Additionally, game performance correlates with intention understanding, highlighting the importance of the four components towards success in this game. These findings underline the crucial role of intention understanding in evaluating LLMs' social intelligence and highlight the potential of using social deduction games as a complex testbed to enhance LLM evaluation. InterIntent contributes a structured approach to bridging the evaluation gap in social intelligence within multiplayer games.

Lee Kezar, Pei Zhou

There is little prior work on quantifying the relationships between facial expressions and emotionality in American Sign Language. In this final report, we provide two methods for studying these relationships through probability and prediction. Using a large corpus of natural signing manually annotated with facial features paired with lexical emotion datasets, we find that there exist many relationships between emotionality and the face, and that a simple classifier can predict what someone is saying in terms of broad emotional categories only by looking at the face.

Pei Zhou, Karthik Gopalakrishnan, Behnam Hedayatnia et al.

Implicit knowledge, such as common sense, is key to fluid human conversations. Current neural response generation (RG) models are trained to generate responses directly, omitting unstated implicit knowledge. In this paper, we present Think-Before-Speaking (TBS), a generative approach to first externalize implicit commonsense knowledge (think) and use this knowledge to generate responses (speak). We expect that externalizing implicit knowledge allows more efficient learning, produces more informative responses, and enables more explainable models. We analyze different choices to collect knowledge-aligned dialogues, represent implicit knowledge, and transition between knowledge and dialogues. Empirical results show TBS models outperform end-to-end and knowledge-augmented RG baselines on most automatic metrics and generate more informative, specific, and commonsense-following responses, as evaluated by human annotators. TBS also generates knowledge that makes sense and is relevant to the dialogue around 85\% of the time.

Pei Zhou, Karthik Gopalakrishnan, Behnam Hedayatnia et al.

Smooth and effective communication requires the ability to perform latent or explicit commonsense inference. Prior commonsense reasoning benchmarks (such as SocialIQA and CommonsenseQA) mainly focus on the discriminative task of choosing the right answer from a set of candidates, and do not involve interactive language generation as in dialogue. Moreover, existing dialogue datasets do not explicitly focus on exhibiting commonsense as a facet. In this paper, we present an empirical study of commonsense in dialogue response generation. We first auto-extract commonsensical dialogues from existing dialogue datasets by leveraging ConceptNet, a commonsense knowledge graph. Furthermore, building on social contexts/situations in SocialIQA, we collect a new dialogue dataset with 25K dialogues aimed at exhibiting social commonsense in an interactive setting. We evaluate response generation models trained using these datasets and find that models trained on both extracted and our collected data produce responses that consistently exhibit more commonsense than baselines. Finally we propose an approach for automatic evaluation of commonsense that relies on features derived from ConceptNet and pre-trained language and dialog models, and show reasonable correlation with human evaluation of responses' commonsense quality. We are releasing a subset of our collected data, Commonsense-Dialogues, containing about 11K dialogs.

Ting-Yun Chang, Yang Liu, Karthik Gopalakrishnan et al.

Pretrained language models have demonstrated outstanding performance in many NLP tasks recently. However, their social intelligence, which requires commonsense reasoning about the current situation and mental states of others, is still developing. Towards improving language models' social intelligence, we focus on the Social IQA dataset, a task requiring social and emotional commonsense reasoning. Building on top of the pretrained RoBERTa and GPT2 models, we propose several architecture variations and extensions, as well as leveraging external commonsense corpora, to optimize the model for Social IQA. Our proposed system achieves competitive results as those top-ranking models on the leaderboard. This work demonstrates the strengths of pretrained language models, and provides viable ways to improve their performance for a particular task.

Ting-Yun Chang, Yang Liu, Karthik Gopalakrishnan et al.

Pretrained language models have excelled at many NLP tasks recently; however, their social intelligence is still unsatisfactory. To enable this, machines need to have a more general understanding of our complicated world and develop the ability to perform commonsense reasoning besides fitting the specific downstream tasks. External commonsense knowledge graphs (KGs), such as ConceptNet, provide rich information about words and their relationships. Thus, towards general commonsense learning, we propose two approaches to \emph{implicitly} and \emph{explicitly} infuse such KGs into pretrained language models. We demonstrate our proposed methods perform well on SocialIQA, a social commonsense reasoning task, in both limited and full training data regimes.

Pei Zhou, Pegah Jandaghi, Bill Yuchen Lin et al.

Humans use commonsense reasoning (CSR) implicitly to produce natural and coherent responses in conversations. Aiming to close the gap between current response generation (RG) models and human communication abilities, we want to understand why RG models respond as they do by probing RG model's understanding of commonsense reasoning that elicits proper responses. We formalize the problem by framing commonsense as a latent variable in the RG task and using explanations for responses as textual form of commonsense. We collect 6k annotated explanations justifying responses from four dialogue datasets and ask humans to verify them and propose two probing settings to evaluate RG models' CSR capabilities. Probing results show that models fail to capture the logical relations between commonsense explanations and responses and fine-tuning on in-domain data and increasing model sizes do not lead to understanding of CSR for RG. We hope our study motivates more research in making RG models emulate the human reasoning process in pursuit of smooth human-AI communication.

Ninareh Mehrabi, Pei Zhou, Fred Morstatter et al.

Warning: this paper contains content that may be offensive or upsetting. Numerous natural language processing models have tried injecting commonsense by using the ConceptNet knowledge base to improve performance on different tasks. ConceptNet, however, is mostly crowdsourced from humans and may reflect human biases such as "lawyers are dishonest." It is important that these biases are not conflated with the notion of commonsense. We study this missing yet important problem by first defining and quantifying biases in ConceptNet as two types of representational harms: overgeneralization of polarized perceptions and representation disparity. We find that ConceptNet contains severe biases and disparities across four demographic categories. In addition, we analyze two downstream models that use ConceptNet as a source for commonsense knowledge and find the existence of biases in those models as well. We further propose a filtered-based bias-mitigation approach and examine its effectiveness. We show that our mitigation approach can reduce the issues in both resource and models but leads to a performance drop, leaving room for future work to build fairer and stronger commonsense models.

Pei Zhou, Rahul Khanna, Seyeon Lee et al.

Pre-trained language models (PTLMs) have achieved impressive performance on commonsense inference benchmarks, but their ability to employ commonsense to make robust inferences, which is crucial for effective communications with humans, is debated. In the pursuit of advancing fluid human-AI communication, we propose a new challenge, RICA: Robust Inference capability based on Commonsense Axioms, that evaluates robust commonsense inference despite textual perturbations. To generate data for this challenge, we develop a systematic and scalable procedure using commonsense knowledge bases and probe PTLMs across two different evaluation settings. Extensive experiments on our generated probe sets with more than 10k statements show that PTLMs perform no better than random guessing on the zero-shot setting, are heavily impacted by statistical biases, and are not robust to perturbation attacks. We also find that fine-tuning on similar statements offer limited gains, as PTLMs still fail to generalize to unseen inferences. Our new large-scale benchmark exposes a significant gap between PTLMs and human-level language understanding and offers a new challenge for PTLMs to demonstrate commonsense.

Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen et al.

Recently, large-scale pre-trained language models have demonstrated impressive performance on several commonsense-reasoning benchmark datasets. However, building machines with commonsense to compose realistically plausible sentences remains challenging. In this paper, we present a constrained text generation task, CommonGen associated with a benchmark dataset, to explicitly test machines for the ability of generative commonsense reasoning. Given a set of common concepts (e.g., {dog, frisbee, catch, throw}); the task is to generate a coherent sentence describing an everyday scenario using these concepts (e.g., "a man throws a frisbee and his dog catches it"). The CommonGen task is challenging because it inherently requires 1) relational reasoning with background commonsense knowledge, and 2) compositional generalization ability to work on unseen concept combinations. Our dataset, constructed through a combination of crowdsourced and existing caption corpora, consists of 79k commonsense descriptions over 35k unique concept-sets. Experiments show that there is a large gap between state-of-the-art text generation models (e.g., T5) and human performance. Furthermore, we demonstrate that the learned generative commonsense reasoning capability can be transferred to improve downstream tasks such as CommonsenseQA by generating additional context.

Weijia Shi, Muhao Chen, Pei Zhou et al.

Contextualized word embedding models, such as ELMo, generate meaningful representations of words and their context. These models have been shown to have a great impact on downstream applications. However, in many cases, the contextualized embedding of a word changes drastically when the context is paraphrased. As a result, the downstream model is not robust to paraphrasing and other linguistic variations. To enhance the stability of contextualized word embedding models, we propose an approach to retrofitting contextualized embedding models with paraphrase contexts. Our method learns an orthogonal transformation on the input space, which seeks to minimize the variance of word representations on paraphrased contexts. Experiments show that the retrofitted model significantly outperforms the original ELMo on various sentence classification and language inference tasks.

Pei Zhou, Weijia Shi, Jieyu Zhao et al.

Recent studies have shown that word embeddings exhibit gender bias inherited from the training corpora. However, most studies to date have focused on quantifying and mitigating such bias only in English. These analyses cannot be directly extended to languages that exhibit morphological agreement on gender, such as Spanish and French. In this paper, we propose new metrics for evaluating gender bias in word embeddings of these languages and further demonstrate evidence of gender bias in bilingual embeddings which align these languages with English. Finally, we extend an existing approach to mitigate gender bias in word embeddings under both monolingual and bilingual settings. Experiments on modified Word Embedding Association Test, word similarity, word translation, and word pair translation tasks show that the proposed approaches effectively reduce the gender bias while preserving the utility of the embeddings.

Pei Zhou, Muhao Chen, Kai-Wei Chang et al.

Quantifying differences in terminologies from various academic domains has been a longstanding problem yet to be solved. We propose a computational approach for analyzing linguistic variation among scientific research fields by capturing the semantic change of terms based on a neural language model. The model is trained on a large collection of literature in five computer science research fields, for which we obtain field-specific vector representations for key terms, and global vector representations for other words. Several quantitative approaches are introduced to identify the terms whose semantics have drastically changed, or remain unchanged across different research fields. We also propose a metric to quantify the overall linguistic variation of research fields. After quantitative evaluation on human annotated data and qualitative comparison with other methods, we show that our model can improve cross-disciplinary data collaboration by identifying terms that potentially induce confusion during interdisciplinary studies.