Reasoning over visual relationships—spatial, functional, interactional, social, etc.—is considered to be a fundamental component of human cognition. Yet, despite the major advances in visual comprehension in multimodal language models (MLMs), precise reasoning over relationships and their generations remains a challenge. We introduce Robin: an MLM instruction-tuned with densely annotated relationships capable of constructing high-quality dense scene graphs at scale. To train Robin, we curate SVG1, a synthetic scene graph dataset by completing the missing relations of selected objects in existing scene graphs using a teacher MLM and a carefully designed filtering process to ensure high-quality. To generate more accurate and rich scene graphs at scale for any image, we introduce SG-Edit: a self-distillation framework where GPT-4o further refines Robin’s predicted scene graphs by removing unlikely relations and/or suggesting relevant ones. In total, our dataset contains 146K images and 5.6M relationships for 2.6M objects. Results show that our Robin-3B model, despite being trained on less than 3 million instances, outperforms similar-size models trained on over 300 million instances on relationship understanding benchmarks, and even surpasses larger models up to 13B parameters. Notably, it achieves state-of-the-art performance in referring expression comprehension with a score of 88.2, surpassing the previous best of 87.4. Our results suggest that training on the refined scene graph data is crucial to maintaining high performance across diverse visual reasoning tasks2.

Recent research in vision-language models (VLMs) has centered around the possibility of equipping them with implicit long-form chainof-thought reasoning-akin to the success observed in language models-via distillation and reinforcement learning.But what about the non-reasoning models already trained and deployed across the internet?Should we simply abandon them, or is there hope for a search mechanism that can elicit hidden knowledge and induce long reasoning traces-without any additional training or supervision?In this paper, we explore this possibility using a Monte Carlo Tree Search (MCTS)-inspired algorithm, which injects subquestion-subanswer pairs into the model's output stream.We show that framing reasoning as a search process-where subquestions act as latent decisions within a broader inference trajectory-helps the model "connect the dots" between fragmented knowledge and produce extended reasoning traces in nonreasoning models.We evaluate our method across three benchmarks and observe consistent improvements.Notably, our approach yields a 2% overall improvement on MMMU-PRO, including a significant 9% gain in Liberal Arts.Problem (P) What is the position of the blue car in the image?A: parked on the sidewalk B: driving on the road C: parked on the grass D: parked on the road CoT: To determine the position of the blue car in the image, let's analyze the scene: * The image shows a busy urban street with several vehicles.* The blue car is located on the right side of the image.* It is positioned on what appears to be a paved surface, which is consistent with a road.* The car is not on the grass or the sidewalk, as those areas are clearly distinguishable in the image.* The car is stationary, suggesting it is parked.Given these observations, the blue car is parked on the road.

Gender bias in vision-language foundation models (VLMs) raises concerns about their safe deployment and is typically evaluated using benchmarks with gender annotations on real-world images. However, as these benchmarks often contain spurious correlations between gender and non-gender features, such as objects and backgrounds, we identify a critical oversight in gender bias evaluation: Do spurious features distort gender bias evaluation? To address this question, we systematically perturb non-gender features across four widely used benchmarks (COCO-gender, FACET, MIAP, and PHASE) and various VLMs to quantify their impact on bias evaluation. Our findings reveal that even minimal perturbations, such as masking just 10% of objects or weakly blurring backgrounds, can dramatically alter bias scores, shifting metrics by up to 175% in generative VLMs and 43% in CLIP variants. This suggests that current bias evaluations often reflect model responses to spurious features rather than gender bias, undermining their reliability. Since creating spurious feature-free benchmarks is fundamentally challenging, we recommend reporting bias metrics alongside feature-sensitivity measurements to enable a more reliable bias assessment.

The ability to acknowledge the inevitable uncertainty in their knowledge and reasoning is a prerequisite for AI systems to be truly truthful and reliable. In this paper, we present a taxonomy of uncertainty specific to vision-language AI systems, distinguishing between epistemic uncertainty (arising from a lack of information) and aleatoric uncertainty (due to inherent unpredictability), and further explore finer categories within. Based on this taxonomy, we synthesize a benchmark dataset, CertainlyUncertain, featuring 178K visual question answering (VQA) samples as contrastive pairs. This is achieved by 1) inpainting images to make previously answerable questions into unanswerable ones; and 2) using image captions to prompt large language models for both answerable and unanswerable questions. Additionally, we introduce a new metric confidence-weighted accuracy, that is well correlated with both accuracy and calibration error, to address the shortcomings of existing metrics. Despite the recent rapid progress in vision-language models (VLMs), evaluations on our benchmark show that they perform poorly in uncertain scenarios. Further experiments demonstrate that supervised fine-tuning with CertainlyUncertain enhances the performance of VLMs, and reduces the calibration error. These improvements extend beyond our benchmark to existing refusal-oriented datasets and show positive results on reducing hallucinations, while maintaining performance on standard VQA benchmarks. Our work underscores the importance of addressing uncertainty in vision-language AI systems to improve their reliability and trustworthiness in real-world applications.

Creativity has long been considered one of the most difficult aspect of human intelligence for AI to mimic. However, the rise of Large Language Models (LLMs), like ChatGPT, has raised questions about whether AI can match or even surpass human creativity. We present CREATIVITY INDEX as the first step to quantify the linguistic creativity of a text by reconstructing it from existing text snippets on the web. CREATIVITY INDEX is motivated by the hypothesis that the seemingly remarkable creativity of LLMs may be attributable in large part to the creativity of human-written texts on the web. To compute CREATIVITY INDEX efficiently, we introduce DJ SEARCH, a novel dynamic programming algorithm that can search verbatim and near-verbatim matches of text snippets from a given document against the web. Experiments reveal that the CREATIVITY INDEX of professional human authors is on average 66.2% higher than that of LLMs, and that alignment reduces the CREATIVITY INDEX of LLMs by an average of 30.1%. In addition, we find that distinguished authors like Hemingway exhibit measurably higher CREATIVITY INDEX compared to other human writers. Finally, we demonstrate that CREATIVITY INDEX can be used as a surprisingly effective criterion for zero-shot machine text detection, surpassing the strongest existing zero-shot system, DetectGPT, by a significant margin of 30.2%, and even outperforming the strongest supervised system, GhostBuster, in five out of six domains.

Abhilasha Ravichander, Jillian Fisher, Taylor Sorensen, Ximing Lu, Maria Antoniak, Bill Yuchen Lin, Niloofar Mireshghallah, Chandra Bhagavatula, Yejin Choi. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Data diversity is crucial for training a strong language model. Yet metrics of diversity often diverge from this goal, measuring variations in heuristic features—like n-grams or embeddings—that are detached from how the model actually performs on a target task. This motivates us to ask: *Can we redefine data diversity—beyond measuring variations in heuristic features—in a way that better predicts model generalization?* Through large-scale empirical analyses spanning over 300 training runs, carefully controlled for data scale and quality, we show that data diversity can be a strong predictor of generalization in LLM reasoning—as measured by average model performance on unseen out-of-distribution benchmarks. We introduce **G-Vendi**, a metric that quantifies diversity via the entropy of model-induced loss gradients. G-Vendi scales to million-sample datasets and yet consistently outperforms heuristic alternatives, achieving strong correlation ($\text{Spearman's } \rho \approx 0.9$) with out-of-distribution (OOD) performance across both natural language inference (NLI) and math reasoning tasks. Building on this insight, we present **Prismatic Synthesis**, a framework for generating diverse synthetic data by targeting underrepresented regions in gradient space. Experimental results show that Prismatic Synthesis consistently improves model performance as we scale synthetic data—not just on in-distribution test but across unseen, out-of-distribution benchmarks—significantly outperforming state-of-the-art models in both domains. For example, PrismMath-7B, our model distilled from a 32B LLM without human verification, outperforms R1-Distill-Qwen-7B—trained on proprietary data generated by 671B R1—on 6 out of 7 challenging math benchmarks.

Recent advances in reasoning-centric language models have highlighted reinforcement learning (RL) as a promising method for aligning models with verifiable rewards. However, it remains contentious whether RL truly expands a model’s reasoning capabilities or merely amplifies high-reward outputs already latent in the base model’s distribution, and whether continually scaling up RL compute reliably leads to improved reasoning performance. In this work, we challenge prevailing assumptions by demonstrating that prolonged RL (ProRL) training can uncover novel reasoning strategies that are inaccessible to base models, even under extensive sampling. We introduce ProRL, a novel training methodology that incorporates KL divergence control, reference policy resetting, and a diverse suite of tasks. Our empirical analysis reveals that RL-trained models consistently outperform base models across a wide range of pass@$k$ evaluations, including scenarios where base models fail entirely regardless of the number of attempts. We further show that reasoning boundary improvements correlates strongly with task competence of base model and training duration, suggesting that RL can explore and populate new regions of solution space over time. These findings offer new insights into the conditions under which RL meaningfully expands reasoning boundaries in language models and establish a foundation for future work on long-horizon RL for reasoning. We will release model weights and data to support further research.

Human values are crucial to human decision-making. Value pluralism is the view that multiple correct values may be held in tension with one another (e.g., when considering lying to a friend to protect their feelings, how does one balance honesty with friendship?). As statistical learners, AI systems fit to averages by default, washing out these potentially irreducible value conflicts. To improve AI systems to better reflect value pluralism, the first-order challenge is to explore the extent to which AI systems can model pluralistic human values, rights, and duties as well as their interaction. We introduce ValuePrism, a large-scale dataset of 218k values, rights, and duties connected to 31k human-written situations. ValuePrism’s contextualized values are generated by GPT-4 and deemed high-quality by human annotators 91% of the time. We conduct a large-scale study with annotators across diverse social and demographic backgrounds to try to understand whose values are represented. With ValuePrism, we build Value Kaleidoscope (or Kaleido), an open, light-weight, and structured language-based multi-task model that generates, explains, and assesses the relevance and valence (i.e., support or oppose) of human values, rights, and duties within a specific context. Humans prefer the sets of values output by our system over the teacher GPT- 4, finding them more accurate and with broader coverage. In addition, we demonstrate that Kaleido can help explain variability in human decision-making by outputting contrasting values. Finally, we show that Kaleido’s representations transfer to other philosophical frameworks and datasets, confirming the benefit of an explicit, modular, and interpretable approach to value pluralism. We hope that our work will serve as a step to making more explicit the implicit values behind human decision-making and to steering AI systems to make decisions that are more in accordance with them.

Huihan Li, Yuting Ning, Zeyi Liao, Siyuan Wang, Xiang Lorraine Li, Ximing Lu, Wenting Zhao, Faeze Brahman, Yejin Choi, Xiang Ren. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Authorship obfuscation, rewriting a text to intentionally obscure the identity of the author, is an important but challenging task.Current methods using large language models (LLMs) lack interpretability and controllability, often ignoring author-specific stylistic features, resulting in less robust performance overall.To address this, we develop STYLEREMIX, an adaptive and interpretable obfuscation method that perturbs specific, fine-grained style elements of the original input text.STYLEREMIX uses pre-trained Low Rank Adaptation (LoRA) modules to rewrite an input specifically along various stylistic axes (e.g., formality and length) while maintaining low computational cost.STYLEREMIX outperforms state-of-theart baselines and much larger LLMs in a variety of domains as assessed by both automatic and human evaluation.Additionally, we release AUTHORMIX, a large set of 30K high-quality, long-form texts from a diverse set of 14 authors and 4 domains, and DISC, a parallel corpus of 1,500 texts spanning seven style axes in 16 unique directions 1 .

Reinforcement Learning with Human Feedback (RLHF) is the most prominent method for Language Model (LM) alignment. However, RLHF is an unstable and data-hungry process that continually requires new high-quality LM-generated data for finetuning. We introduce Advantage-Leftover Lunch RL (A-LoL), a new class of offline policy gradient algorithms that enable RL training on any pre-existing data. By assuming the entire LM output sequence as a single action, A-LoL allows incorporating sequence-level classifiers or human-designed scoring functions as rewards. Subsequently, by using LM’s value estimate, A-LoL only trains on positive advantage (leftover) data points, making it resilient to noise. Overall, A-LoL is an easy-to-implement, sample-efficient, and stable LM training recipe. We demonstrate the effectiveness of A-LoL and its variants with a set of four different language generation tasks. We compare against both online RL (PPO) and recent preference-based (DPO, PRO) and reward-based (GOLD) offline RL baselines. On the commonly-used RLHF benchmark, Helpful and Harmless Assistant (HHA), LMs trained with A-LoL methods achieve the highest diversity while also being rated more safe and helpful than the baselines according to humans. Additionally, in the remaining three tasks, A-LoL could optimize multiple distinct reward functions even when using noisy or suboptimal training data.

Alignment tuning has become the de facto standard practice for enabling base large language models (LLMs) to serve as open-domain AI assistants. The alignment tuning process typically involves instruction learning through supervised fine-tuning (SFT) and preference tuning via reinforcement learning from human feedback (RLHF). A recent study, LIMA (Zhou et al., 2023), shows that using merely 1K examples for SFT can achieve significant alignment performance as well, suggesting that the effect of alignment tuning might be "superficial." This raises questions about how exactly the alignment tuning transforms a base LLM. We analyze the effect of alignment tuning by examining the token distribution shift between base LLMs and their aligned counterparts (e.g., Llama-2 and Llama-2-chat). Our findings reveal that base LLMs and their alignment-tuned versions perform nearly identically in decoding on the majority of token positions (i.e., they share the top-ranked tokens). Most distribution shifts occur with stylistic tokens (e.g., discourse markers, safety disclaimers). This direct evidence strongly supports the hypothesis that alignment tuning primarily learns to adopt the language style of AI assistants, and that the knowledge required for answering user queries predominantly comes from the base LLMs themselves. Based on these findings, we rethink the alignment of LLMs by posing the research question: how effectively can we align base LLMs without SFT or RLHF? To address this, we introduce a simple, tuning-free alignment method, URIAL (Untuned LLMs with Restyled In-context Alignment). URIAL achieves effective alignment purely through in-context learning (ICL) with base LLMs, requiring as few as three constant stylistic examples and a system prompt. We conduct a fine-grained and interpretable evaluation on a diverse set of examples, named just-eval-instruct. Results demonstrate that base LLMs with URIAL can match or even surpass the performance of LLMs aligned with SFT (Mistral-7b-Instruct) or SFT+RLHF (Llama-2-70b-chat). We show that the gap between tuning-free and tuning-based alignment methods can be significantly reduced through strategic prompting and ICL. Our findings on the superficial nature of alignment tuning and results with URIAL suggest that deeper analysis and theoretical understanding of alignment is crucial to future LLM research.

The ability to derive underlying principles from a handful of observations and then generalize to novel situations---known as inductive reasoning---is central to human intelligence. Prior work suggests that language models (LMs) often fall short on inductive reasoning, despite achieving impressive success on research benchmarks. In this work, we conduct a systematic study of the inductive reasoning capabilities of LMs through $\textit{iterative hypothesis refinement}$, a technique that more closely mirrors the human inductive process than standard input-output prompting. Iterative hypothesis refinement employs a three-step process: proposing, selecting, and refining hypotheses in the form of textual rules. By examining the intermediate rules, we observe that LMs are phenomenal $\textit{hypothesis proposers}$ (i.e., generating candidate rules), and when coupled with a (task-specific) symbolic interpreter that is able to systematically filter the proposed set of rules, this hybrid approach achieves strong results across inductive reasoning benchmarks that require inducing causal relations, language-like instructions, and symbolic concepts. However, they also behave as puzzling $\textit{inductive reasoners}$, showing notable performance gaps between rule induction (i.e., identifying plausible rules) and rule application (i.e., applying proposed rules to instances), suggesting that LMs are proposing hypotheses without being able to actually apply the rules. Through empirical and human analyses, we further reveal several discrepancies between the inductive reasoning processes of LMs and humans, shedding light on both the potentials and limitations of using LMs in inductive reasoning tasks.

Large language models (LMs) are capable of generating free-text rationales to aid question answering. However, prior work 1) suggests that useful self-rationalization is emergent only at significant scales (e.g., 175B parameter GPT-3); and 2) focuses largely on downstream performance, ignoring the semantics of the rationales themselves, e.g., are they faithful, true, and helpful for humans? In this work, we enable small-scale LMs (∼200x smaller than GPT-3) to generate rationales that not only improve downstream task performance, but are also more plausible, consistent, and diverse, assessed both by automatic and human evaluation. Our method, MaRio (Multi-rewArd RatIOnalization), is a multi-reward conditioned self-rationalization algorithm that optimizes multiple distinct properties like plausibility, diversity and consistency. Results on three difficult question-answering datasets StrategyQA, QuaRel and OpenBookQA show that not only does MaRio improve task accuracy, but it also improves the self-rationalization quality of small LMs across the aforementioned axes better than a supervised fine-tuning (SFT) baseline. Extensive human evaluations confirm that MaRio rationales are preferred vs. SFT rationales, as well as qualitative improvements in plausibility and consistency.

The recent wave of generative AI has sparked unprecedented global attention, with both excitement and concern over potentially superhuman levels of artificial intelligence: models now take only seconds to produce outputs that would challenge or exceed the capabilities even of expert humans. At the same time, models still show basic errors in understanding that would not be expected even in non-expert humans. This presents us with an apparent paradox: how do we reconcile seemingly superhuman capabilities with the persistence of errors that few humans would make? In this work, we posit that this tension reflects a divergence in the configuration of intelligence in today's generative models relative to intelligence in humans. Specifically, we propose and test the **Generative AI Paradox** hypothesis: generative models, having been trained directly to reproduce expert-like outputs, acquire generative capabilities that are not contingent upon---and can therefore exceed---their ability to understand those same types of outputs. This contrasts with humans, for whom basic understanding almost always precedes the ability to generate expert-level outputs. We test this hypothesis through controlled experiments analyzing generation vs.~understanding in generative models, across both language and image modalities. Our results show that although models can outperform humans in generation, they consistently fall short of human capabilities in measures of understanding, as well as weaker correlation between generation and understanding performance, and more brittleness to adversarial inputs. Our findings support the hypothesis that models' generative capability may not be contingent upon understanding capability, and call for caution in interpreting artificial intelligence by analogy to human intelligence.

With increased power and prevalence of AI systems, it is ever more critical that AI systems are designed to serve all, i.e., people with diverse values and perspectives. However, aligning models to serve pluralistic human values remains an open research question. In this piece, we propose a roadmap to pluralistic alignment, specifically using large language models as a test bed. We identify and formalize three possible ways to define and operationalize pluralism in AI systems: 1) Overton pluralistic models that present a spectrum of reasonable responses; 2) Steerably pluralistic models that can steer to reflect certain perspectives; and 3) Distributionally pluralistic models that are well-calibrated to a given population in distribution. We also formalize and discuss three possible classes of pluralistic benchmarks: 1) Multi-objective benchmarks, 2) Trade-off steerable benchmarks that incentivize models to steer to arbitrary trade-offs, and 3) Jury-pluralistic benchmarks that explicitly model diverse human ratings. We use this framework to argue that current alignment techniques may be fundamentally limited for pluralistic AI; indeed, we highlight empirical evidence, both from our own experiments and from other work, that standard alignment procedures might reduce distributional pluralism in models, motivating the need for further research on pluralistic alignment.

Jillian Fisher, Ximing Lu, Jaehun Jung, Liwei Jiang, Zaid Harchaoui, Yejin Choi. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Jaehun Jung, Peter West, Liwei Jiang, Faeze Brahman, Ximing Lu, Jillian Fisher, Taylor Sorensen, Yejin Choi. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

We introduce WildTeaming, an automatic red-teaming framework that mines in-the-wild user-chatbot interactions to discover 5.7K unique clusters of novel jailbreak tactics, and then composes selections of multiple mined tactics for systematic exploration of novel and even more challenging jailbreaks. Compared to prior work that performed red-teaming via recruited human workers, gradient-based optimization, or iterative revision with large language models (LLMs), our work investigates jailbreaks from chatbot users in-the-wild who were not specifically instructed to break the system. WildTeaming reveals previously unidentified vulnerabilities of frontier LLMs, resulting in more diverse and successful adversarial attacks compared to state-of-the-art jailbreaking methods. While there exist many datasets for jailbreak evaluation, very few open-source datasets exist for jailbreak training, as safety training data has been closed among all frontier models even when their weights are open. Therefore, with WildTeaming we create WildJailbreak, a large-scale open-source synthetic safety dataset with 262K vanilla (direct request) and adversarial (complex jailbreak) prompt-response pairs. In order to mitigate exaggerated safety behaviors, WildJailbreak provides two contrastive types of queries: 1) harmful queries (both vanilla and adversarial) and 2) benign queries that resemble harmful queries in form but contain no harmful intent. As WildJailbreak considerably upgrades the quality and scale of existing safety resources, it uniquely enables us to examine the scaling effects of data and the interplay of data properties and model capabilities during safety training. Through extensive model training and evaluations, we identify the training properties that enable an ideal balance of safety behaviors: appropriate safeguarding without over-refusal, effective handling of both vanilla and adversarial queries, and minimal, if any, decrease in general capabilities. All the components of WildJailbreak contribute to achieving balanced safety behaviors of models

Chandra Bhagavatula, Jena D. Hwang, Doug Downey, Ronan Le Bras, Ximing Lu, Lianhui Qin, Keisuke Sakaguchi, Swabha Swayamdipta, Peter West, Yejin Choi. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Valentina Pyatkin, Jena D. Hwang, Vivek Srikumar, Ximing Lu, Liwei Jiang, Yejin Choi, Chandra Bhagavatula. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Language models are capable of commonsense reasoning: while domain-specific models can learn from explicit knowledge (e.g. commonsense graphs [6] ethical norms [25]), and larger models like GPT-3 [7] mani-fest broad commonsense reasoning capacity. Can their knowledge be extended to multimodal inputs such as images and audio without paired domain data? In this work, we propose‡ESPER (Extending Sensory PErception with Reinforcement learning) which enables text-only pretrained models to address multimodal tasks such as visual commonsense reasoning. Our key novelty is to use rein-forcement learning to align multimodal inputs to language model generations without direct supervision: for example, our reward optimization relies only on cosine similarity derived from CLIP [52] and requires no additional paired (image, text) data. Experiments demonstrate that ESPER outperforms baselines and prior work on a variety of multimodal text generation tasks ranging from captioning to commonsense reasoning; these include a new benchmark we collect and release, the ESP dataset, which tasks models with generating the text of several different domains for each image. Our code and data are publicly released at https://github.com/JiwanChung/esper.

Hyunwoo Kim, Jack Hessel, Liwei Jiang, Peter West, Ximing Lu, Youngjae Yu, Pei Zhou, Ronan Bras, Malihe Alikhani, Gunhee Kim, Maarten Sap, Yejin Choi. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Ximing Lu, Faeze Brahman, Peter West, Jaehun Jung, Khyathi Chandu, Abhilasha Ravichander, Prithviraj Ammanabrolu, Liwei Jiang, Sahana Ramnath, Nouha Dziri, Jillian Fisher, Bill Lin, Skyler Hallinan, Lianhui Qin, Xiang Ren, Sean Welleck, Yejin Choi. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Transformer large language models (LLMs) have sparked admiration for their exceptional performance on tasks that demand intricate multi-step reasoning. Yet, these models simultaneously show failures on surprisingly trivial problems. This begs the question: Are these errors incidental, or do they signal more substantial limitations? In an attempt to demystify transformer LLMs, we investigate the limits of these models across three representative compositional tasks---multi-digit multiplication, logic grid puzzles, and a classic dynamic programming problem. These tasks require breaking problems down into sub-steps and synthesizing these steps into a precise answer. We formulate compositional tasks as computation graphs to systematically quantify the level of complexity, and break down reasoning steps into intermediate sub-procedures. Our empirical findings suggest that transformer LLMs solve compositional tasks by reducing multi-step compositional reasoning into linearized subgraph matching, without necessarily developing systematic problem-solving skills. To round off our empirical study, we provide theoretical arguments on abstract multi-step reasoning problems that highlight how autoregressive generations' performance can rapidly decay with increased task complexity.

Instruction following vision-language (VL) models offer a flexible interface that supports a broad range of multimodal tasks in a zero-shot fashion. However, interfaces that operate on full images do not directly enable the user to “point to" and access specific regions within images. This capability is important not only to support reference-grounded VL benchmarks, but also, for practical applications that require precise within-image reasoning. We build Localized Visual Commonsense model which allows users to specify (multiple) regions- as-input. We train our model by sampling localized commonsense knowledge from a large language model (LLM): specifically, we prompt a LLM to collect commonsense knowledge given a global literal image description and a local literal region description automatically generated by a set of VL models. This pipeline is scalable and fully automatic, as no aligned or human-authored image and text pairs are required. With a separately trained critic model that selects high quality examples, we find that training on the localized commonsense corpus expanded solely from images can successfully distill existing VL models to support a reference-as-input interface. Empirical results and human evaluations in zero-shot settings demonstrate that our distillation method results in more precise VL models of reasoning compared to a baseline of passing a generated referring expression.

Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, Hannaneh Hajishirzi. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

As humans, we navigate a multimodal world, building a holistic understanding from all our senses. We introduce @MERLOT RESERVE, a model that represents videos jointly over time - through a new training objective that learns from audio, subtitles, and video frames. Given a video, we replace snippets of text and audio with a MASK token; the model learns by choosing the correct masked-out snippet. Our objective learns faster than alternatives, and performs well at scale: we pretrain on 20 million YouTube videos. Empirical results show that @MERLOT RESERVE learns strong multimodal representations. When finetuned, it sets state-of-the-art on Visual Commonsense Reasoning (VCR), TVQA, and Kinetics-600; outperforming prior work by 5%, 7%, and 1.5% respectively. Ablations show that these tasks benefit from audio pretraining - even VCR, a QA task centered around images (without sound). Moreover, our objective enables out-of-the-box prediction, revealing strong multimodal commonsense understanding. In a fully zero-shot setting, our model obtains competitive results on four video tasks, even outperforming supervised approaches on the recently proposed Situated Reasoning (STAR) benchmark. We analyze why audio enables better vision-language representations, suggesting significant opportunities for future research. We conclude by discussing ethical and societal implications of multimodal pretraining.

Most existing dialogue systems fail to respond properly to potentially unsafe user utterances by either ignoring or passively agreeing with them.To address this issue, we introduce PROSOCIALDIALOG, the first large-scale multi-turn dialogue dataset to teach conversational agents to respond to problematic content following social norms.Covering diverse unethical, problematic, biased, and toxic situations, PROSOCIALDIALOG contains responses that encourage prosocial behavior, grounded in commonsense social rules (i.e., rules-ofthumb, RoTs).Created via a human-AI collaborative framework, PROSOCIALDIALOG consists of 58K dialogues, with 331K utterances, 160K unique RoTs, and 497K dialogue safety labels accompanied by free-form rationales.With this dataset, we introduce a dialogue safety detection module, Canary, capable of generating RoTs given conversational context, and a socially-informed dialogue agent, Prost.Empirical results show that Prost generates more socially acceptable dialogues compared to other state-of-the-art language and dialogue models in both in-domain and out-of-domain settings.Additionally, Canary effectively guides off-the-shelf language models to generate significantly more prosocial responses.Our work highlights the promise and importance of creating and steering conversational AI to be socially responsible.

Knowledge underpins reasoning.Recent research demonstrates that when relevant knowledge is provided as additional context to commonsense question answering (QA), it can substantially enhance the performance even on top of state-of-the-art.The fundamental challenge is where and how to find such knowledge that is high quality and on point with respect to the question; knowledge retrieved from knowledge bases are incomplete and knowledge generated from language models are inconsistent.We present RAINIER 1 , or Reinforced Knowledge Introspector, that learns to generate contextually relevant knowledge in response to given questions.Our approach starts by imitating knowledge generated by GPT-3, then learns to generate its own knowledge via reinforcement learning where rewards are shaped based on the increased performance on the resulting question answering.RAINIER demonstrates substantial and consistent performance gains when tested over 9 different commonsense benchmarks: including 5 datasets that are seen during model training, as well as 4 datasets that are kept unseen.Our work is the first to report that knowledge generated by models that are orders of magnitude smaller than GPT-3, even without direct supervision on the knowledge itself, can exceed the quality of commonsense knowledge elicited from GPT-3.

Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Hao Peng, Ximing Lu, Dragomir Radev, Yejin Choi, Noah A. Smith. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah Smith, Yejin Choi. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Peter West, Chandra Bhagavatula, Jack Hessel, Jena Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, Yejin Choi. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Yanpeng Zhao, Jack Hessel, Youngjae Yu, Ximing Lu, Rowan Zellers, Yejin Choi. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Large-scale language models often learn behaviors that are misaligned with user expectations. Generated text may contain offensive or toxic language, contain significant repetition, or be of a different sentiment than desired by the user. We consider the task of unlearning these misalignments by fine-tuning the language model on signals of what not to do. We introduce Quantized Reward Konditioning (Quark), an algorithm for optimizing a reward function that quantifies an (un)wanted property, while not straying too far from the original model. Quark alternates between (i) collecting samples with the current language model, (ii) sorting them into quantiles based on reward, with each quantile identified by a reward token prepended to the language model’s input, and (iii) using a standard language modeling loss on samples from each quantile conditioned on its reward token, while remaining nearby the original language model via a KL-divergence penalty. By conditioning on a high-reward token at generation time, the model generates text that exhibits less of the unwanted property. For unlearning toxicity, negative sentiment, and repetition, our experiments show that Quark outperforms both strong baselines and state-of-the-art reinforcement learning methods like PPO, while relying only on standard language modeling primitives.

Theorem proving in natural mathematical language – the mixture of symbolic and natural language used by humans – plays a central role in mathematical advances and education, and tests aspects of reasoning that are core to intelligence. Yet it has remained underexplored with modern generative models. We study large-scale language models on two new generation tasks: suggesting the next step in a mathematical proof, and full proof generation. We develop NaturalProver, a language model that generates proofs by conditioning on background references (e.g. theorems and definitions that are either retrieved or human-provided), and optionally enforces their presence with constrained decoding. On theorems from the NaturalProofs benchmark, NaturalProver improves the quality of next-step suggestions and generated proofs over fine-tuned GPT-3, according to human evaluations from university-level mathematics students. NaturalProver is capable of proving some theorems that require short (2-6 step) proofs, and providing next-step suggestions that are rated as correct and useful over 40% of the time, which is to our knowledge the first demonstration of these capabilities using neural language models.

Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A. Smith, Yejin Choi. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Peter West, Ximing Lu, Ari Holtzman, Chandra Bhagavatula, Jena D. Hwang, Yejin Choi. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Ximing Lu, Peter West, Rowan Zellers, Ronan Le Bras, Chandra Bhagavatula, Yejin Choi. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

As humans, we understand events in the visual world contextually, performing multimodal reasoning across time to make inferences about the past, present, and future. We introduce MERLOT, a model that learns multimodal script knowledge by watching millions of YouTube videos with transcribed speech -- in an entirely label-free, self-supervised manner. By pretraining with a mix of both frame-level (spatial) and video-level (temporal) objectives, our model not only learns to match images to temporally corresponding words, but also to contextualize what is happening globally over time. As a result, MERLOT exhibits strong out-of-the-box representations of temporal commonsense, and achieves state-of-the-art performance on 12 different video QA datasets when finetuned. It also transfers well to the world of static images, allowing models to reason about the dynamic context behind visual scenes. On Visual Commonsense Reasoning, MERLOT~answers questions correctly with 80.6\% accuracy, outperforming state-of-the-art models of similar size by over 3\%, even those that make heavy use of auxiliary supervised data (like object bounding boxes).Ablation analyses demonstrate the complementary importance of: 1) training on videos versus static images; 2) scaling the magnitude and diversity of the pretraining video corpus; and 3) using diverse objectives that encourage full-stack multimodal reasoning, from the recognition to cognition level.