Faeze Brahman

Da Yin, Faeze Brahman, Abhilasha Ravichander et al. · allen-ai, cmu

Closed-source agents suffer from several issues such as a lack of affordability, transparency, and reproducibility, particularly on complex interactive tasks. This motivates the development of open-source alternatives. We introduce LUMOS, one of the first frameworks for training open-source LLM-based agents. LUMOS features a learnable, unified, and modular architecture with a planning module that learns high-level subgoal generation, and a grounding module trained to translate these into actions using various tools in the execution module. The design allows for modular upgrades and wider applicability to diverse interactive tasks. To foster generalizable agent learning, we collect large-scale, unified, and high-quality training annotations derived from diverse ground-truth reasoning rationales across various complex interactive tasks. On 9 datasets, LUMOS exhibits several key advantages: (1) LUMOS excels multiple larger open-source agents on the held-out datasets (unused for training) for each task type. LUMOS even surpasses GPT agents on QA and web tasks; (2) LUMOS outperforms open-source agents produced by chain-of-thoughts and unmodularized integrated training; and (3) LUMOS effectively generalizes to unseen tasks, outperforming 33B-scale agents and domain-specific agents.

Xuhui Zhou, Hyunwoo Kim, Faeze Brahman et al. · allen-ai, cmu

AI agents are increasingly autonomous in their interactions with human users and tools, leading to increased interactional safety risks. We present HAICOSYSTEM, a framework examining AI agent safety within diverse and complex social interactions. HAICOSYSTEM features a modular sandbox environment that simulates multi-turn interactions between human users and AI agents, where the AI agents are equipped with a variety of tools (e.g., patient management platforms) to navigate diverse scenarios (e.g., a user attempting to access other patients' profiles). To examine the safety of AI agents in these interactions, we develop a comprehensive multi-dimensional evaluation framework that uses metrics covering operational, content-related, societal, and legal risks. Through running 1840 simulations based on 92 scenarios across seven domains (e.g., healthcare, finance, education), we demonstrate that HAICOSYSTEM can emulate realistic user-AI interactions and complex tool use by AI agents. Our experiments show that state-of-the-art LLMs, both proprietary and open-sourced, exhibit safety risks in over 50\% cases, with models generally showing higher risks when interacting with simulated malicious users. Our findings highlight the ongoing challenge of building agents that can safely navigate complex interactions, particularly when faced with malicious users. To foster the AI agent safety ecosystem, we release a code platform that allows practitioners to create custom scenarios, simulate interactions, and evaluate the safety and performance of their agents.

Sean Welleck, Ximing Lu, Peter West et al. · allen-ai, uw

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

Peter West, Ximing Lu, Nouha Dziri et al. · allen-ai, cmu

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.

Faeze Brahman, Sachin Kumar, Vidhisha Balachandran et al. · allen-ai, cmu

Chat-based language models are designed to be helpful, yet they should not comply with every user request. While most existing work primarily focuses on refusal of "unsafe" queries, we posit that the scope of noncompliance should be broadened. We introduce a comprehensive taxonomy of contextual noncompliance describing when and how models should not comply with user requests. Our taxonomy spans a wide range of categories including incomplete, unsupported, indeterminate, and humanizing requests (in addition to unsafe requests). To test noncompliance capabilities of language models, we use this taxonomy to develop a new evaluation suite of 1000 noncompliance prompts. We find that most existing models show significantly high compliance rates in certain previously understudied categories with models like GPT-4 incorrectly complying with as many as 30% of requests. To address these gaps, we explore different training strategies using a synthetically-generated training set of requests and expected noncompliant responses. Our experiments demonstrate that while direct finetuning of instruction-tuned models can lead to both over-refusal and a decline in general capabilities, using parameter efficient methods like low rank adapters helps to strike a good balance between appropriate noncompliance and other capabilities.

Zhe Su, Xuhui Zhou, Sanketh Rangreji et al. · allen-ai, cmu

Truthfulness (adherence to factual accuracy) and utility (satisfying human needs and instructions) are both fundamental aspects of Large Language Models, yet these goals often conflict (e.g., sell a car with known flaws), which makes it challenging to achieve both in real-world deployments. We propose AI-LieDar, a framework to study how LLM-based agents navigate these scenarios in an multi-turn interactive setting. We design a set of real-world scenarios where language agents are instructed to achieve goals that are in conflict with being truthful during a multi-turn conversation with simulated human agents. To evaluate the truthfulness at large scale, we develop a truthfulness detector inspired by psychological literature to assess the agents' responses. Our experiment demonstrates that all models are truthful less than 50% of the time, though truthfulness and goal achievement (utility) rates vary across models. We further test the steerability of LLMs towards truthfulness, finding that models can be directed to be truthful or deceptive, and even truth-steered models still lie. These findings reveal the complex nature of truthfulness in LLMs and underscore the importance of further research to ensure the safe and reliable deployment of LLMs and LLM-based agents.

Yufei Tian, Abhilasha Ravichander, Lianhui Qin et al. · cmu

We explore the creative problem-solving capabilities of modern LLMs in a novel constrained setting. To this end, we create MACGYVER, an automatically generated dataset consisting of over 1,600 real-world problems deliberately designed to trigger innovative usage of objects and necessitate out-of-the-box thinking. We then present our collection to both LLMs and humans to compare and contrast their problem-solving abilities. MACGYVER is challenging for both groups, but in unique and complementary ways. For instance, humans excel in tasks they are familiar with but struggle with domain-specific knowledge, leading to a higher variance. In contrast, LLMs, exposed to a variety of specialized knowledge, attempt broader problems but fail by proposing physically-infeasible actions. Finally, we provide a detailed error analysis of LLMs, and demonstrate the potential of enhancing their problem-solving ability with novel prompting techniques such as iterative step-wise reflection and divergent-convergent thinking. This work (1) introduces a fresh arena for intelligent agents focusing on intricate aspects of physical reasoning, planning, and unconventional thinking, which supplements the existing spectrum of machine intelligence; and (2) provides insight into the constrained problem-solving capabilities of both humans and AI.

Jaehun Jung, Lianhui Qin, Sean Welleck et al.

Despite their impressive capabilities, large pre-trained language models (LMs) struggle with consistent reasoning; recently, prompting LMs to generate explanations that self-guide the inference has emerged as a promising direction to amend this. However, these approaches are fundamentally bounded by the correctness of explanations, which themselves are often noisy and inconsistent. In this work, we develop Maieutic Prompting, which infers a correct answer to a question even from the noisy and inconsistent generations of LM. Maieutic Prompting induces a tree of explanations abductively (e.g. X is true, because ...) and recursively, then frames the inference as a satisfiability problem over these explanations and their logical relations. We test Maieutic Prompting for true/false QA on three challenging benchmarks that require complex commonsense reasoning. Maieutic Prompting achieves up to 20% better accuracy than state-of-the-art prompting methods, and as a fully unsupervised approach, performs competitively with supervised models. We also show that Maieutic Prompting improves robustness in inference while providing interpretable rationales.

Faeze Brahman, Baolin Peng, Michel Galley et al. · microsoft-research

Large pre-trained language models have recently enabled open-ended generation frameworks (e.g., prompt-to-text NLG) to tackle a variety of tasks going beyond the traditional data-to-text generation. While this framework is more general, it is under-specified and often leads to a lack of controllability restricting their real-world usage. We propose a new grounded keys-to-text generation task: the task is to generate a factual description about an entity given a set of guiding keys, and grounding passages. To address this task, we introduce a new dataset, called EntDeGen. Inspired by recent QA-based evaluation measures, we propose an automatic metric, MAFE, for factual correctness of generated descriptions. Our EntDescriptor model is equipped with strong rankers to fetch helpful passages and generate entity descriptions. Experimental result shows a good correlation (60.14) between our proposed metric and human judgments of factuality. Our rankers significantly improved the factual correctness of generated descriptions (15.95% and 34.51% relative gains in recall and precision). Finally, our ablation study highlights the benefit of combining keys and groundings.

Chao Zhao, Faeze Brahman, Kaiqiang Song et al.

Narrative summarization aims to produce a distilled version of a narrative to describe its most salient events and characters. Summarizing a narrative is challenging as it requires an understanding of event causality and character behaviors. To encourage research in this direction, we propose NarraSum, a large-scale narrative summarization dataset. It contains 122K narrative documents, which are collected from plot descriptions of movies and TV episodes with diverse genres, and their corresponding abstractive summaries. Experiments show that there is a large performance gap between humans and the state-of-the-art summarization models on NarraSum. We hope that this dataset will promote future research in summarization, as well as broader studies of natural language understanding and generation. The dataset is available at https://github.com/zhaochaocs/narrasum.

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

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

Hanjie Chen, Faeze Brahman, Xiang Ren et al.

Generating free-text rationales is a promising step towards explainable NLP, yet evaluating such rationales remains a challenge. Existing metrics have mostly focused on measuring the association between the rationale and a given label. We argue that an ideal metric should focus on the new information uniquely provided in the rationale that is otherwise not provided in the input or the label. We investigate this research problem from an information-theoretic perspective using conditional V-information (Hewitt et al., 2021). More concretely, we propose a metric called REV (Rationale Evaluation with conditional V-information), to quantify the amount of new, label-relevant information in a rationale beyond the information already available in the input or the label. Experiments across four benchmarks with reasoning tasks, including chain-of-thought, demonstrate the effectiveness of REV in evaluating rationale-label pairs, compared to existing metrics. We further demonstrate REV is consistent with human judgments on rationale evaluations and provides more sensitive measurements of new information in free-text rationales. When used alongside traditional performance metrics, REV provides deeper insights into models' reasoning and prediction processes.

Teng Xiao, Yige Yuan, Hamish Ivison et al.

This paper introduces MR-Search, an in-context meta reinforcement learning (RL) formulation for agentic search with self-reflection. Instead of optimizing a policy within a single independent episode with sparse rewards, MR-Search trains a policy that conditions on past episodes and adapts its search strategy across episodes. MR-Search learns to learn a search strategy with self-reflection, allowing search agents to improve in-context exploration at test-time. Specifically, MR-Search performs cross-episode exploration by generating explicit self-reflections after each episode and leveraging them as additional context to guide subsequent attempts, thereby promoting more effective exploration during test-time. We further introduce a multi-turn RL algorithm that estimates a dense relative advantage at the turn level, enabling fine-grained credit assignment on each episode. Empirical results across various benchmarks demonstrate the advantages of MR-Search over baselines based RL, showing strong generalization and relative improvements of 9.2% to 19.3% across eight benchmarks. Our code and data are available at https://github.com/tengxiao1/MR-Search.

Kavel Rao, Liwei Jiang, Valentina Pyatkin et al. · allen-ai

Moral or ethical judgments rely heavily on the specific contexts in which they occur. Understanding varying shades of defeasible contextualizations (i.e., additional information that strengthens or attenuates the moral acceptability of an action) is critical to accurately represent the subtlety and intricacy of grounded human moral judgment in real-life scenarios. We introduce defeasible moral reasoning: a task to provide grounded contexts that make an action more or less morally acceptable, along with commonsense rationales that justify the reasoning. To elicit high-quality task data, we take an iterative self-distillation approach that starts from a small amount of unstructured seed knowledge from GPT-3 and then alternates between (1) self-distillation from student models; (2) targeted filtering with a critic model trained by human judgment (to boost validity) and NLI (to boost diversity); (3) self-imitation learning (to amplify the desired data quality). This process yields a student model that produces defeasible contexts with improved validity, diversity, and defeasibility. From this model we distill a high-quality dataset, δ-Rules-of-Thumb, of 1.2M entries of contextualizations and rationales for 115K defeasible moral actions rated highly by human annotators 85.9% to 99.8% of the time. Using δ-RoT we obtain a final student model that wins over all intermediate student models by a notable margin.

Skyler Hallinan, Faeze Brahman, Ximing Lu et al. · uw

While text style transfer has many applications across natural language processing, the core premise of transferring from a single source style is unrealistic in a real-world setting. In this work, we focus on arbitrary style transfer: rewriting a text from an arbitrary, unknown style to a target style. We propose STEER: Unified Style Transfer with Expert Reinforcement, a unified frame-work developed to overcome the challenge of limited parallel data for style transfer. STEER involves automatically generating a corpus of style-transfer pairs using a product of experts during decoding. The generated offline data is then used to pre-train an initial policy before switching to online, off-policy reinforcement learning for further improvements via fine-grained reward signals. STEER is unified and can transfer to multiple target styles from an arbitrary, unknown source style, making it particularly flexible and efficient. Experimental results on a challenging dataset with text from a diverse set of styles demonstrate state-of-the-art results compared to competitive baselines. Remarkably, STEER outperforms the 175B parameter instruction-tuned GPT-3 on overall style transfer quality, despite being 226 times smaller in size. We also show STEER is robust, maintaining its style transfer capabilities on out-of-domain data, and surpassing nearly all baselines across various styles. The success of our method highlights the potential of RL algorithms when augmented with controllable decoding to overcome the challenge of limited data supervision.

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

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

Tuhin Chakrabarty, Vishakh Padmakumar, Faeze Brahman et al.

The development of large language models (LLMs) capable of following instructions and engaging in conversational interactions sparked increased interest in their utilization across various support tools. We investigate the utility of modern LLMs in assisting professional writers via an empirical user study (n=30). The design of our collaborative writing interface is grounded in the cognitive process model of writing that views writing as a goal-oriented thinking process encompassing non-linear cognitive activities: planning, translating, and reviewing. Participants are asked to submit a post-completion survey to provide feedback on the potential and pitfalls of LLMs as writing collaborators. Upon analyzing the writer-LLM interactions, we find that while writers seek LLM's help across all three types of cognitive activities, they find LLMs more helpful in translation and reviewing. Our findings from analyzing both the interactions and the survey responses highlight future research directions in creative writing assistance using LLMs.

Tenghao Huang, Ehsan Qasemi, Bangzheng Li et al.

Storytelling's captivating potential makes it a fascinating research area, with implications for entertainment, education, therapy, and cognitive studies. In this paper, we propose Affective Story Generator (AffGen) for generating interesting narratives. AffGen introduces "intriguing twists" in narratives by employing two novel techniques-Dynamic Beam Sizing and Affective Reranking. Dynamic Beam Sizing encourages less predictable, more captivating word choices using a contextual multi-arm bandit model. Affective Reranking prioritizes sentence candidates based on affect intensity. Our empirical evaluations, both automatic and human, demonstrate AffGen's superior performance over existing baselines in generating affectively charged and interesting narratives. Our ablation study and analysis provide insights into the strengths and weaknesses of AffGen.

Jaehun Jung, Faeze Brahman, Yejin Choi

We present a principled approach to provide LLM-based evaluation with a rigorous guarantee of human agreement. We first propose that a reliable evaluation method should not uncritically rely on model preferences for pairwise evaluation, but rather assess the confidence of judge models and selectively decide when to trust its judgement. We then show that under this selective evaluation framework, human agreement can be provably guaranteed -- such that the model evaluation aligns with that of humans to a user-specified agreement level. As part of our framework, we also introduce Simulated Annotators, a novel confidence estimation method that significantly improves judge calibration and thus enables high coverage of evaluated instances. Finally, we propose Cascaded Selective Evaluation, where we use cheaper models as initial judges and escalate to stronger models only when necessary -- again, while still providing a provable guarantee of human agreement. Experimental results show that Cascaded Selective Evaluation guarantees strong alignment with humans, far beyond what LLM judges could achieve without selective evaluation. For example, on a subset of Chatbot Arena where GPT-4 almost never achieves 80% human agreement, our method, even while employing substantially cost-effective models such as Mistral-7B, guarantees over 80% human agreement with almost 80% test coverage.

Chao Zhao, Faeze Brahman, Tenghao Huang et al.

Pre-trained models (PTMs) have lead to great improvements in natural language generation (NLG). However, it is still unclear how much commonsense knowledge they possess. With the goal of evaluating commonsense knowledge of NLG models, recent work has proposed the problem of generative commonsense reasoning, e.g., to compose a logical sentence given a set of unordered concepts. Existing approaches to this problem hypothesize that PTMs lack sufficient parametric knowledge for this task, which can be overcome by introducing external knowledge or task-specific pre-training objectives. Different from this trend, we argue that PTM's inherent ability for generative commonsense reasoning is underestimated due to the order-agnostic property of its input. In particular, we hypothesize that the order of the input concepts can affect the PTM's ability to utilize its commonsense knowledge. To this end, we propose a pre-ordering approach to elaborately manipulate the order of the given concepts before generation. Experiments show that our approach can outperform the more sophisticated models that have access to a lot of external data and resources.

Anvesh Rao Vijjini, Faeze Brahman, Snigdha Chaturvedi

In this paper, we introduce the task of modeling interpersonal relationships for story generation. For addressing this task, we propose Relationships as Latent Variables for Story Generation, (ReLiSt). ReLiSt generates stories sentence by sentence and has two major components - a relationship selector and a story continuer. The relationship selector specifies a latent variable to pick the relationship to exhibit in the next sentence and the story continuer generates the next sentence while expressing the selected relationship in a coherent way. Our automatic and human evaluations demonstrate that ReLiSt is able to generate stories with relationships that are more faithful to desired relationships while maintaining the content quality. The relationship assignments to sentences during inference bring interpretability to ReLiSt.

Yash Kumar Lal, Li Zhang, Faeze Brahman et al.

How-to procedures, such as how to plant a garden, are now used by millions of users, but sometimes need customizing to meet a user's specific needs, e.g., planting a garden without pesticides. Our goal is to measure and improve an LLM's ability to perform such customization. Our approach is to test several simple multi-LLM-agent architectures for customization, as well as an end-to-end LLM, using a new evaluation set, called CustomPlans, of over 200 WikiHow procedures each with a customization need. We find that a simple architecture with two LLM agents used sequentially performs best, one that edits a generic how-to procedure and one that verifies its executability, significantly outperforming (10.5% absolute) an end-to-end prompted LLM. This suggests that LLMs can be configured reasonably effectively for procedure customization. This also suggests that multi-agent editing architectures may be worth exploring further for other customization applications (e.g. coding, creative writing) in the future.

Team Olmo, Allyson Ettinger, Amanda Bertsch et al. · uw

We introduce Olmo 3, a family of state-of-the-art, fully-open language models at the 7B and 32B parameter scales. Olmo 3 model construction targets long-context reasoning, function calling, coding, instruction following, general chat, and knowledge recall. This release includes the entire model flow, i.e., the full lifecycle of the family of models, including every stage, checkpoint, data point, and dependency used to build it. Our flagship model, Olmo 3 Think 32B, is the strongest fully-open thinking model released to-date.

Avinandan Bose, Shuyue Stella Li, Faeze Brahman et al.

Cold-start personalization requires inferring user preferences through interaction when no user-specific historical data is available. The core challenge is a routing problem: each task admits dozens of preference dimensions, yet individual users care about only a few, and which ones matter depends on who is asking. With a limited question budget, asking without structure will miss the dimensions that matter. Reinforcement learning is the natural formulation, but in multi-turn settings its terminal reward fails to exploit the factored, per-criterion structure of preference data, and in practice learned policies collapse to static question sequences that ignore user responses. We propose decomposing cold-start elicitation into offline structure learning and online Bayesian inference. Pep (Preference Elicitation with Priors) learns a structured world model of preference correlations offline from complete profiles, then performs training-free Bayesian inference online to select informative questions and predict complete preference profiles, including dimensions never asked about. The framework is modular across downstream solvers and requires only simple belief models. Across medical, mathematical, social, and commonsense reasoning, Pep achieves 80.8% alignment between generated responses and users' stated preferences versus 68.5% for RL, with 3-5x fewer interactions. When two users give different answers to the same question, Pep changes its follow-up 39-62% of the time versus 0-28% for RL. It does so with ~10K parameters versus 8B for RL, showing that the bottleneck in cold-start elicitation is the capability to exploit the factored structure of preference data.

Zishuo Zheng, Vidhisha Balachandran, Chan Young Park et al.

As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources (e.g., model developers, users, and tools) within a single prompt context. Thus, enforcing an instruction hierarchy (IH) in LLMs, where higher-level directives override lower-priority requests, is critical for the reliability and controllability of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. Specifically, the model must first "think" about the relationship between a given user prompt and higher-priority (system) instructions before generating a response. To enable this capability via training, we construct VerIH, an instruction hierarchy dataset of constraint-following tasks with verifiable answers. This dataset comprises both aligned and conflicting system-user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our finetuned models achieve consistent improvements on instruction following and instruction hierarchy benchmarks. This reasoning ability also generalizes to safety-critical settings beyond the training distribution. By treating safety issues as resolving conflicts between adversarial user inputs and predefined higher-priority policies, our trained model enhances robustness against jailbreak and prompt injection attacks. These results demonstrate that reasoning over instruction hierarchies provides a practical path to reliable LLMs, where updates to system prompts yield controllable and robust changes in model behavior.

Hamish Ivison, Muru Zhang, Faeze Brahman et al. · uw

Selecting high-quality training data from a larger pool is a crucial step when instruction-tuning language models, as carefully curated datasets often produce models that outperform those trained on much larger, noisier datasets. Automated data selection approaches for instruction-tuning are typically tested by selecting small datasets (roughly 10k samples) from small pools (100-200k samples). However, popular deployed instruction-tuned models often train on hundreds of thousands to millions of samples, subsampled from even larger data pools. We present a systematic study of how well data selection methods scale to these settings, selecting up to 2.5M samples from pools of up to 5.8M samples and evaluating across 7 diverse tasks. We show that many recently proposed methods fall short of random selection in this setting (while using more compute), and even decline in performance when given access to larger pools of data to select over. However, we find that a variant of representation-based data selection (RDS+), which uses weighted mean pooling of pretrained LM hidden states, consistently outperforms more complex methods across all settings tested -- all whilst being more compute-efficient. Our findings highlight that the scaling properties of proposed automated selection methods should be more closely examined. We release our code, data, and models at https://github.com/hamishivi/automated-instruction-selection.

Nathan Lambert, Jacob Morrison, Valentina Pyatkin et al. · allen-ai, cambridge

Language model post-training is applied to refine behaviors and unlock new skills across a wide range of recent language models, but open recipes for applying these techniques lag behind proprietary ones. The underlying training data and recipes for post-training are simultaneously the most important pieces of the puzzle and the portion with the least transparency. To bridge this gap, we introduce Tulu 3, a family of fully-open state-of-the-art post-trained models, alongside its data, code, and training recipes, serving as a comprehensive guide for modern post-training techniques. Tulu 3, which builds on Llama 3.1 base models, achieves results surpassing the instruct versions of Llama 3.1, Qwen 2.5, Mistral, and even closed models such as GPT-4o-mini and Claude 3.5-Haiku. The training algorithms for our models include supervised finetuning (SFT), Direct Preference Optimization (DPO), and a novel method we call Reinforcement Learning with Verifiable Rewards (RLVR). With Tulu 3, we introduce a multi-task evaluation scheme for post-training recipes with development and unseen evaluations, standard benchmark implementations, and substantial decontamination of existing open datasets on said benchmarks. We conclude with analysis and discussion of training methods that did not reliably improve performance. In addition to the Tulu 3 model weights and demo, we release the complete recipe -- including datasets for diverse core skills, a robust toolkit for data curation and evaluation, the training code and infrastructure, and, most importantly, a detailed report for reproducing and further adapting the Tulu 3 approach to more domains.

Team OLMo, Pete Walsh, Luca Soldaini et al. · allen-ai, cambridge

We present OLMo 2, the next generation of our fully open language models. OLMo 2 includes a family of dense autoregressive language models at 7B, 13B and 32B scales with fully released artifacts -- model weights, full training data, training code and recipes, training logs and thousands of intermediate checkpoints. In this work, we describe our modified model architecture and training recipe, focusing on techniques for achieving better training stability and improved per-token efficiency. Our updated pretraining data mixture introduces a new, specialized data mix called Dolmino Mix 1124, which significantly improves model capabilities across many downstream task benchmarks when introduced via late-stage curriculum training (i.e. specialized data during the annealing phase of pretraining). Finally, we incorporate best practices from Tülu 3 to develop OLMo 2-Instruct, focusing on permissive data and extending our final-stage reinforcement learning with verifiable rewards (RLVR). Our OLMo 2 base models sit at the Pareto frontier of performance to training compute, often matching or outperforming open-weight only models like Llama 3.1, Qwen 2.5, and Gemma 2 while using fewer FLOPs and with fully transparent training data, code, and recipe. Our fully open OLMo 2-Instruct models are competitive with open-weight only models of comparable size and even some proprietary models like GPT-3.5 Turbo and GPT 4o Mini.

Zhaoyi Joey Hou, Bowei Alvin Zhang, Yining Lu et al.

Creativity is often seen as a hallmark of human intelligence. While large language models (LLMs) are increasingly perceived as producing creative text, there is still no holistic framework to evaluate their creativity across diverse scenarios. Existing evaluation methods remain fragmented, with dramatic variation across domains and tasks, largely due to differing definitions and measurements of creativity. Inspired by the hypothesis that creativity is not one fixed idea, we propose CreativityPrism, an evaluation analysis framework that decomposes creativity into three dimensions: quality, novelty, and diversity. CreativityPrism incorporates nine tasks, three domains, i.e., divergent thinking, creative writing, and logical reasoning, and twenty evaluation metrics, which measure each dimension in task-specific, unique ways. We evaluate 17 state-of-the-art (SoTA) proprietary and open-sourced LLMs on CreativityPrism and analyze the performance correlations among different metrics and task domains. Our results reveal a notable gap between proprietary and open-source models. Overall, model performance tends to be highly correlated across tasks within the same domain and less so across different domains. Among evaluation dimensions, diversity and quality metrics show strong correlations - models that perform well on one often excel on the other - whereas novelty exhibits much weaker correlation with either. These findings support our hypothesis that strong performance in one creativity task or dimension does not necessarily generalize to others, underscoring the need for a holistic evaluation of LLM creativity.

Liwei Jiang, Kavel Rao, Seungju Han et al.

We introduce WildTeaming, an automatic LLM safety red-teaming framework that mines in-the-wild user-chatbot interactions to discover 5.7K unique clusters of novel jailbreak tactics, and then composes multiple tactics for systematic exploration of novel jailbreaks. Compared to prior work that performed red-teaming via recruited human workers, gradient-based optimization, or iterative revision with LLMs, our work investigates jailbreaks from chatbot users who were not specifically instructed to break the system. WildTeaming reveals previously unidentified vulnerabilities of frontier LLMs, resulting in up to 4.6x more diverse and successful adversarial attacks compared to state-of-the-art jailbreak methods. While many datasets exist for jailbreak evaluation, very few open-source datasets exist for jailbreak training, as safety training data has been closed even when model weights are open. 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. To mitigate exaggerated safety behaviors, WildJailbreak provides two contrastive types of queries: 1) harmful queries (vanilla & adversarial) and 2) benign queries that resemble harmful queries in form but contain no harm. 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 experiments, we identify the training properties that enable an ideal balance of safety behaviors: appropriate safeguarding without over-refusal, effective handling of vanilla and adversarial queries, and minimal, if any, decrease in general capabilities. All components of WildJailbeak contribute to achieving balanced safety behaviors of models.

Ashutosh Baheti, Ximing Lu, Faeze Brahman et al.

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. We also release our experimental code. https://github.com/abaheti95/LoL-RL

Shuyue Stella Li, Rui Xin, Teng Xiao et al.

Language models encode substantial evaluative knowledge from pretraining, yet current post-training methods rely on external supervision (human annotations, proprietary models, or scalar reward models) to produce reward signals. Each imposes a ceiling. Human judgment cannot supervise capabilities beyond its own, proprietary APIs create dependencies, and verifiable rewards cover only domains with ground-truth answers. Self-improvement from a model's own evaluative capacity is a reward source that scales with the model itself, yet remains largely untapped by current methods. We introduce EVOLM, a post-training method that structures this capacity into explicit discriminative rubrics and uses them as training signal. EVOLM trains two capabilities within a single language model in alternation: (1) a rubric generator producing instance-specific evaluation criteria optimized for discriminative utility, which maximizes a small frozen judge's ability to distinguish preferred from dispreferred responses; and (2) a policy trained using those rubric-conditioned scores as reward. All preference signals are constructed from the policy's own outputs via temporal contrast with earlier checkpoints, requiring no human annotation or external supervision. EVOLM trains a Qwen3-8B model to generate rubrics that outperform GPT-4.1 on RewardBench-2 by 25.7%. The co-trained policy achieves 69.3% average on the OLMo3-Adapt suite, outperforming policies trained with GPT-4.1 prompted rubrics by 3.9% and with the state-of-the-art 8B reward model SkyWork-RM by 16%. Overall, EVOLM demonstrates that structuring a model's evaluative capacity into co-evolving discriminative rubrics enables self-improvement without external supervision.

Lester James V. Miranda, Yizhong Wang, Yanai Elazar et al. · allen-ai, cambridge

Learning from human feedback has enabled the alignment of language models (LMs) with human preferences. However, collecting human preferences is expensive and time-consuming, with highly variable annotation quality. An appealing alternative is to distill preferences from LMs as a source of synthetic annotations, offering a cost-effective and scalable alternative, albeit susceptible to other biases and errors. In this work, we introduce HyPER, a Hybrid Preference routER that defers an annotation to either humans or LMs, achieving better annotation quality while reducing the cost of human-only annotation. We formulate this as an optimization problem: given a preference dataset and an evaluation metric, we (1) train a performance prediction model (PPM) to predict a reward model's (RM) performance on an arbitrary combination of human and LM annotations and (2) employ a routing strategy that selects a combination that maximizes the predicted performance. We train the PPM on MultiPref, a new preference dataset with 10k instances paired with humans and LM labels. We show that the selected hybrid mixture of synthetic and direct human preferences using HyPER achieves better RM performance compared to using either one exclusively by 7-13% on RewardBench and generalizes across unseen preference datasets and other base models. We also observe the same trend in other benchmarks using Best-of-N reranking, where the hybrid mix has 2-3% better performance. Finally, we analyze features from HyPER and find that prompts with moderate safety concerns or complexity benefit the most from human feedback.

Paul Röttger, Musashi Hinck, Valentin Hofmann et al. · allen-ai

Large language models (LLMs) are helping millions of users write texts about diverse issues, and in doing so expose users to different ideas and perspectives. This creates concerns about issue bias, where an LLM tends to present just one perspective on a given issue, which in turn may influence how users think about this issue. So far, it has not been possible to measure which issue biases LLMs manifest in real user interactions, making it difficult to address the risks from biased LLMs. Therefore, we create IssueBench: a set of 2.49m realistic English-language prompts to measure issue bias in LLM writing assistance, which we construct based on 3.9k templates (e.g. "write a blog about") and 212 political issues (e.g. "AI regulation") from real user interactions. Using IssueBench, we show that issue biases are common and persistent in 10 state-of-the-art LLMs. We also show that biases are very similar across models, and that all models align more with US Democrat than Republican voter opinion on a subset of issues. IssueBench can easily be adapted to include other issues, templates, or tasks. By enabling robust and realistic measurement, we hope that IssueBench can bring a new quality of evidence to ongoing discussions about LLM biases and how to address them.

Shuyue Stella Li, Jimin Mun, Faeze Brahman et al. · allen-ai, cmu

Large language models (LLMs) often fail to ask effective questions under uncertainty, making them unreliable in domains where proactive information-gathering is essential for decision-making. We present ALignment via Fine-grained Attributes, (ALFA) a framework that improves LLM question-asking by (i) decomposing the notion of a "good" question into a set of theory-grounded attributes (e.g., clarity, relevance), (ii) controllably synthesizing attribute-specific question variations, and (iii) aligning models via preference-based optimization to explicitly learn to ask better questions along these fine-grained attributes. Focusing on clinical reasoning as a case study, we introduce the MediQ-AskDocs dataset, composed of 17k real-world clinical interactions augmented with 80k attribute-specific preference pairs of follow-up questions, as well as a novel expert-annotated interactive healthcare QA task to evaluate question-asking abilities. Models aligned with ALFA reduce diagnostic errors by 56.6% on MediQ-AskDocs compared to SoTA instruction-tuned LLMs, with a question-level win-rate of 64.4% and strong generalizability. Our findings suggest that explicitly guiding question-asking with structured, fine-grained attributes offers a scalable path to improve LLMs, especially in expert application domains.

Jaehun Jung, Ximing Lu, Liwei Jiang et al. · allen-ai, uw

The current winning recipe for automatic summarization is using proprietary large-scale language models (LLMs) such as ChatGPT as is, or imitation learning from them as teacher models. While increasingly ubiquitous dependence on such large-scale language models is convenient, there remains an important question of whether small-scale models could have achieved competitive results, if we were to seek an alternative learning method -- that allows for a more cost-efficient, controllable, yet powerful summarizer. We present InfoSumm, a novel framework to distill a powerful summarizer based on the information-theoretic objective for summarization, without relying on either the LLM's capability or human-written references. To achieve this, we first propose a novel formulation of the desiderata of summarization (saliency, faithfulness and brevity) through the lens of mutual information between the original document and the summary. Based on this formulation, we start off from Pythia-2.8B as the teacher model, which is not yet capable of summarization, then self-train the model to optimize for the information-centric measures of ideal summaries. Distilling from the improved teacher, we arrive at a compact but powerful summarizer with only 568M parameters that performs competitively against ChatGPT, without ever relying on ChatGPT's capabilities. Extensive analysis demonstrates that our approach outperforms in-domain supervised models in human evaluation, let alone state-of-the-art unsupervised methods, and wins over ChatGPT in controllable summarization.

Tong Chen, Faeze Brahman, Jiacheng Liu et al.

Language models (LMs) can memorize and reproduce segments from their pretraining data verbatim even in non-adversarial settings, raising concerns about copyright, plagiarism, privacy, and creativity. We introduce Paraphrase Preference Optimization (ParaPO), a post-training method that fine-tunes LMs to reduce unintentional regurgitation while preserving their overall utility. ParaPO trains LMs to prefer paraphrased versions of memorized segments over the original verbatim content from the pretraining data. To maintain the ability to recall famous quotations when appropriate, we develop a variant of ParaPO that uses system prompts to control regurgitation behavior. In our evaluation on Llama3.1-8B, ParaPO consistently reduces regurgitation across all tested datasets (e.g., reducing the regurgitation metric from 17.3 to 12.9 in creative writing), whereas unlearning methods used in prior work to mitigate regurgitation are less effective outside their targeted unlearned domain (from 17.3 to 16.9). When applied to the instruction-tuned Tulu3-8B model, ParaPO with system prompting successfully preserves famous quotation recall while reducing unintentional regurgitation (from 8.7 to 6.3 in creative writing) when prompted not to regurgitate. In contrast, without ParaPO tuning, prompting the model not to regurgitate produces only a marginal reduction (8.7 to 8.4).

Shuyue Stella Li, Avinandan Bose, Faeze Brahman et al.

Current large language model (LLM) development treats task-solving and preference alignment as separate challenges, optimizing first for objective correctness, then for alignment to aggregated human preferences. This paradigm fails in human-facing applications where solving a problem correctly is insufficient if the response mismatches the user's needs. This challenge intensifies in just-in-time scenarios where no prior user interaction history exists due to cold-start conditions or privacy constraints. LLMs need to identify what they don't know about user preferences, strategically elicit preference values through questioning, then adapt their reasoning processes and responses accordingly -- a complicated chain of cognitive processes which we term personalized reasoning. We introduce PREFDISCO, an evaluation methodology that transforms static benchmarks into interactive personalization tasks using psychologically-grounded personas with sparse preferences. Our framework creates scenarios where identical questions require different reasoning chains depending on user context, as optimal explanation approaches vary by individual expertise and preferences while maintaining factual accuracy. Evaluation of 21 frontier models across 10 tasks reveals 29.0% of naive personalization attempts produce worse preference alignment than generic responses, yet generic responses also fail to serve individual user needs effectively. These findings suggest personalized reasoning requires dedicated development rather than emerging naturally. PREFDISCO establishes personalized reasoning as a measurable research frontier and reveals fundamental limitations in current LLMs' interactive capabilities, providing a foundation for developing systems that can adapt to individual users in education, healthcare, and technical domains where personalization is critical.

Keivan Rezaei, Khyathi Chandu, Soheil Feizi et al. · cmu

Large language models trained on web-scale corpora can memorize undesirable data containing misinformation, copyrighted material, or private or sensitive information. Recently, several machine unlearning algorithms have been proposed to eliminate the effect of such datapoints from trained models -- that is, to approximate a model that had never been trained on these datapoints in the first place. However, evaluating the effectiveness of unlearning algorithms remains an open challenge. Previous work has relied on heuristics -- such as verifying that the model can no longer reproduce the specific information targeted for removal while maintaining accuracy on unrelated test data. These approaches inadequately capture the complete effect of reversing the influence of datapoints on a trained model. In this work, we propose the RESTOR framework for machine unlearning evaluation, which assesses the ability of unlearning algorithms for targeted data erasure, by evaluating the ability of models to forget the knowledge introduced in these datapoints, while simultaneously recovering the model's knowledge state had it never encountered these datapoints. RESTOR helps uncover several novel insights about popular unlearning algorithms, and the mechanisms through which they operate -- for instance, identifying that some algorithms merely emphasize forgetting but not recovering knowledge, and that localizing unlearning targets can enhance unlearning performance.

Rulin Shao, Akari Asai, Shannon Zejiang Shen et al.

Deep research models perform multi-step research to produce long-form, well-attributed answers. However, most open deep research models are trained on easily verifiable short-form QA tasks via reinforcement learning with verifiable rewards (RLVR), which does not extend to realistic long-form tasks. We address this with Reinforcement Learning with Evolving Rubrics (RLER), in which we construct and maintain rubrics that co-evolve with the policy model during training; this allows the rubrics to incorporate information that the model has newly explored and to provide discriminative, on-policy feedback. Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research. Across four long-form deep research benchmarks in science, healthcare and general domains, DR Tulu substantially outperforms existing open deep research models, and matches or exceeds proprietary deep research systems, while being significantly smaller and cheaper per query. To facilitate future research, we release all data, models, and code, including our new MCP-based agent infrastructure for deep research systems.

Tong Chen, Akari Asai, Luke Zettlemoyer et al.

Language models often generate factually incorrect information unsupported by their training data, a phenomenon known as extrinsic hallucination. Existing mitigation approaches often degrade performance on open-ended generation and downstream tasks, limiting their practical utility. We propose an online reinforcement learning method using a novel binary retrieval-augmented reward (RAR) to address this tradeoff. Unlike continuous reward schemes, our approach assigns a reward of one only when the model's output is entirely factually correct, and zero otherwise. We evaluate our method on Qwen3 reasoning models across diverse tasks. For open-ended generation, binary RAR achieves a 39.3% reduction in hallucination rates, substantially outperforming both supervised training and continuous-reward RL baselines. In short-form question answering, the model learns calibrated abstention, strategically outputting "I don't know" when faced with insufficient parametric knowledge. This yields 44.4% and 21.7% fewer incorrect answers on PopQA and GPQA, respectively. Crucially, these factuality gains come without performance degradation on instruction following, math, or code, whereas continuous-reward RL, despite improving factuality, induces quality regressions.

Mingqian Zheng, Wenjia Hu, Patrick Zhao et al. · allen-ai, cmu

Current LLMs are trained to refuse potentially harmful input queries regardless of whether users actually had harmful intents, causing a tradeoff between safety and user experience. Through a study of 480 participants evaluating 3,840 query-response pairs, we examine how different refusal strategies affect user perceptions across varying motivations. Our findings reveal that response strategy largely shapes user experience, while actual user motivation has negligible impact. Partial compliance -- providing general information without actionable details -- emerges as the optimal strategy, reducing negative user perceptions by over 50% to flat-out refusals. Complementing this, we analyze response patterns of 9 state-of-the-art LLMs and evaluate how 6 reward models score different refusal strategies, demonstrating that models rarely deploy partial compliance naturally and reward models currently undervalue it. This work demonstrates that effective guardrails require focusing on crafting thoughtful refusals rather than detecting intent, offering a path toward AI safety mechanisms that ensure both safety and sustained user engagement.

Ashutosh Baheti, Debanjana Chakraborty, Faeze Brahman et al. · allen-ai, cmu

Obeying precise constraints on top of multiple external attributes is a common computational problem underlying seemingly different domains, from controlled text generation to protein engineering. Existing language model (LM) controllability methods for multi-attribute constraint satisfaction often rely on specialized architectures or gradient-based classifiers, limiting their flexibility to work with arbitrary black-box evaluators and pretrained models. Current general-purpose large language models, while capable, cannot achieve fine-grained multi-attribute control over external attributes. Thus, we create Multi-Attribute Constraint Satisfaction (MACS), a generalized method capable of finetuning language models on any sequential domain to satisfy user-specified constraints on multiple external real-value attributes. Our method trains LMs as editors by sampling diverse multi-attribute edit pairs from an initial set of paraphrased outputs. During inference, LM iteratively improves upon its previous solution to satisfy constraints for all attributes by leveraging our designed constraint satisfaction reward. We additionally experiment with reward-weighted behavior cloning to further improve the constraint satisfaction rate of LMs. To evaluate our approach, we present a new Fine-grained Constraint Satisfaction (FineCS) benchmark, featuring two challenging tasks: (1) Text Style Transfer, where the goal is to simultaneously modify the sentiment and complexity of reviews, and (2) Protein Design, focusing on modulating fluorescence and stability of Green Fluorescent Proteins (GFP). Our empirical results show that MACS achieves the highest threshold satisfaction in both FineCS tasks, outperforming strong domain-specific baselines. Our work opens new avenues for generalized and real-value multi-attribute control, with implications for diverse applications spanning NLP and bioinformatics.

Jaeyoung Lee, Ximing Lu, Jack Hessel et al.

Given the growing influx of misinformation across news and social media, there is a critical need for systems that can provide effective real-time verification of news claims. Large language or multimodal model based verification has been proposed to scale up online policing mechanisms for mitigating spread of false and harmful content. While these can potentially reduce burden on human fact-checkers, such efforts may be hampered by foundation model training data becoming outdated. In this work, we test the limits of improving foundation model performance without continual updating through an initial study of knowledge transfer using either existing intra- and inter- domain benchmarks or explanations generated from large language models (LLMs). We evaluate on 12 public benchmarks for fact-checking and misinformation detection as well as two other tasks relevant to content moderation -- toxicity and stance detection. Our results on two recent multi-modal fact-checking benchmarks, Mocheg and Fakeddit, indicate that knowledge transfer strategies can improve Fakeddit performance over the state-of-the-art by up to 1.7% and Mocheg performance by up to 2.9%.

Bill Yuchen Lin, Yuntian Deng, Khyathi Chandu et al.

We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs. For automated evaluation with WildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo. WildBench evaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments. WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie. Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation. Additionally, we propose a simple method to mitigate length bias, by converting outcomes of ``slightly better/worse'' to ``tie'' if the winner response exceeds the loser one by more than $K$ characters. WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric. WildBench results demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard's 0.91 and AlpacaEval2.0's 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates.

Faeze Brahman, Chandra Bhagavatula, Valentina Pyatkin et al.

Procedural planning, which entails decomposing a high-level goal into a sequence of temporally ordered steps, is an important yet intricate task for machines. It involves integrating common-sense knowledge to reason about complex and often contextualized situations, e.g. ``scheduling a doctor's appointment without a phone''. While current approaches show encouraging results using large language models (LLMs), they are hindered by drawbacks such as costly API calls and reproducibility issues. In this paper, we advocate planning using smaller language models. We present PlaSma, a novel two-pronged approach to endow small language models with procedural knowledge and (constrained) language planning capabilities. More concretely, we develop symbolic procedural knowledge distillation to enhance the commonsense knowledge in small language models and an inference-time algorithm to facilitate more structured and accurate reasoning. In addition, we introduce a new related task, Replanning, that requires a revision of a plan to cope with a constrained situation. In both the planning and replanning settings, we show that orders-of-magnitude smaller models (770M-11B parameters) can compete and often surpass their larger teacher models' capabilities. Finally, we showcase successful application of PlaSma in an embodied environment, VirtualHome.

Bill Yuchen Lin, Yicheng Fu, Karina Yang et al.

We introduce SwiftSage, a novel agent framework inspired by the dual-process theory of human cognition, designed to excel in action planning for complex interactive reasoning tasks. SwiftSage integrates the strengths of behavior cloning and prompting large language models (LLMs) to enhance task completion performance. The framework comprises two primary modules: the Swift module, representing fast and intuitive thinking, and the Sage module, emulating deliberate thought processes. The Swift module is a small encoder-decoder LM fine-tuned on the oracle agent's action trajectories, while the Sage module employs LLMs such as GPT-4 for subgoal planning and grounding. We develop a heuristic method to harmoniously integrate the two modules, resulting in a more efficient and robust problem-solving process. In 30 tasks from the ScienceWorld benchmark, SwiftSage significantly outperforms other methods such as SayCan, ReAct, and Reflexion, demonstrating its effectiveness in solving complex interactive tasks.

Jaehun Jung, Peter West, Liwei Jiang et al.

We present Impossible Distillation, a novel framework for paraphrasing and sentence summarization, that distills a high-quality dataset and model from a low-quality teacher that itself cannot perform these tasks. Unlike prior works that rely on an extreme-scale teacher model (e.g., GPT3) or task-specific architecture, we hypothesize and verify the paraphrastic proximity intrinsic to pre-trained LMs (e.g., GPT2), where paraphrases occupy a proximal subspace in the LM distribution. By identifying and distilling generations from these subspaces, Impossible Distillation produces a high-quality dataset and model even from GPT2-scale LMs. We evaluate our method on multiple benchmarks spanning unconstrained / syntax-controlled paraphrase generation and sentence summarization. Our model with 770M parameters consistently outperforms strong baselines, including models distilled from ChatGPT, and sometimes, even ChatGPT itself. Also, we find that our distilled dataset from 1.5B LMs exhibits higher diversity and fidelity than up to 13 times larger datasets.

Ximing Lu, Faeze Brahman, Peter West et al.

While extreme-scale language models have demonstrated exceptional performance on a variety of language tasks, the degree of control over these language models through pure prompting can often be limited. Directly fine-tuning such language models can be effective for tailoring them, but it can be either extremely costly (e.g., GPT-3) or not even feasible for the broader community (e.g., GPT-4). We propose Inference-time Policy Adapters (IPA), which efficiently tailors a language model such as GPT-3 without fine-tuning it. IPA guides a large base model during decoding time through a lightweight policy adapter trained to optimize an arbitrary user objective with reinforcement learning. On five challenging text generation tasks, such as toxicity reduction and lexically constrained generation, IPA consistently brings significant improvements over off-the-shelf language models. It outperforms competitive baseline methods, sometimes even including expensive fine-tuning. In particular, tailoring GPT-2 with IPA can outperform GPT-3, while tailoring GPT-3 with IPA brings a major performance boost over GPT-3 (and sometimes even over GPT-4). Our promising results highlight the potential of IPA as a lightweight alternative to tailoring extreme-scale language models.

Tenghao Huang, Faeze Brahman, Vered Shwartz et al.

Pre-trained language models learn socially harmful biases from their training corpora, and may repeat these biases when used for generation. We study gender biases associated with the protagonist in model-generated stories. Such biases may be expressed either explicitly ("women can't park") or implicitly (e.g. an unsolicited male character guides her into a parking space). We focus on implicit biases, and use a commonsense reasoning engine to uncover them. Specifically, we infer and analyze the protagonist's motivations, attributes, mental states, and implications on others. Our findings regarding implicit biases are in line with prior work that studied explicit biases, for example showing that female characters' portrayal is centered around appearance, while male figures' focus on intellect.