arXiv subjects

Jonathan May

Publications and source records attributed to Jonathan May.

At least 37 records · Page 2Linked to original sources

arXiv2025

Language Models Can Predict Their Own Behavior

The text produced by language models (LMs) can exhibit specific `behaviors,' such as a failure to follow alignment training, that we hope to detect and react to during deployment. Identifying these behaviors can often only be done post facto, i.e., after the entire text of the output has been generated. We provide evidence that there are times when we can predict how an LM will behave early in computation, before even a single token is generated. We show that probes trained on the internal representation of input tokens alone can predict a wide range of eventual behaviors over the entire output sequence. Using methods from conformal prediction, we provide provable bounds on the estimation error of our probes, creating precise early warning systems for these behaviors. The conformal probes can identify instances that will trigger alignment failures (jailbreaking) and instruction-following failures, without requiring a single token to be generated. An early warning system built on the probes reduces jailbreaking by 91%. Our probes also show promise in pre-emptively estimating how confident the model will be in its response, a behavior that cannot be detected using the output text alone. Conformal probes can preemptively estimate the final prediction of an LM that uses Chain-of-Thought (CoT) prompting, hence accelerating inference. When applied to an LM that uses CoT to perform text classification, the probes drastically reduce inference costs (65% on average across 27 datasets), with negligible accuracy loss. Encouragingly, probes generalize to unseen datasets and perform better on larger models, suggesting applicability to the largest of models in real-world settings.

cs.CL↗

arXiv2025

Can VLMs Recall Factual Associations From Visual References?

Through a controlled study, we identify a systematic deficiency in the multimodal grounding of Vision Language Models (VLMs). While VLMs can recall factual associations when provided a textual reference to an entity; their ability to do so is significantly diminished when the reference is visual instead. Forcing VLMs to rely on image representations of an entity halves their ability to recall factual knowledge, suggesting that VLMs struggle to link their internal knowledge of an entity with its image representation. We show that such linking failures are correlated with the expression of distinct patterns in model internal states, and that probes on these internal states achieve over 92% accuracy at flagging cases where the VLM response is unreliable. These probes can be applied, without retraining, to identify when a VLM will fail to correctly answer a question that requires an understanding of multimodal input. When used to facilitate selective prediction on a visual question answering task, the probes increase coverage by 7.87% (absolute) while also reducing the risk of error by 0.9% (absolute). Addressing the systematic, detectable deficiency is an important avenue in language grounding, and we provide informed recommendations for future directions.

cs.CV↗

arXiv2025

A Little Human Data Goes A Long Way

Faced with an expensive human annotation process, creators of NLP systems increasingly turn to synthetic data generation. While this method shows promise, the extent to which synthetic data can replace human annotation is poorly understood. We investigate the use of synthetic data in Fact Verification (FV) and Question Answering (QA) by studying the effects of incrementally replacing human generated data with synthetic points on eight diverse datasets. Strikingly, replacing up to 90% of the training data only marginally decreases performance, but replacing the final 10% leads to severe declines. We find that models trained on purely synthetic data can be reliably improved by including as few as 125 human generated data points. We show that matching the performance gain of just a little additional human data (only 200 points) requires an order of magnitude more synthetic data and estimate price ratios at which human annotation would be a more cost-effective solution. Our results suggest that even when human annotation at scale is infeasible, there is great value to having a small proportion of the dataset being human generated.

cs.CL↗

arXiv2025

Which Questions Improve Learning the Most? Utility Estimation of Questions with LM-based Simulations

Asking good questions is critical for comprehension and learning, yet evaluating and generating such questions remains a challenging problem. Prior work on inquisitive questions focuses on learner-generated, curiosity-driven queries and evaluates them using indirect metrics, such as salience or information gain, that do not directly capture a question's impact on actual learning outcomes. We introduce QUEST (Question Utility Estimation with Simulated Tests), a framework that uses language models to simulate learners and directly quantify the utility of a question - its contribution to exam performance. QUEST simulates a learner who asks questions and receives answers while studying a textbook chapter, then uses them to take an end-of-chapter exam. Through this simulation, the utility of each question is estimated by its direct effect on exam performance, rather than inferred indirectly based on the underlying content. To support this evaluation, we curate TEXTBOOK-EXAM, a benchmark that aligns textbook sections with end-of-section exam questions across five academic disciplines. Using QUEST, we filter for high-utility questions and fine-tune question generators via rejection sampling. Experiments show that questions generated by QUEST-trained models improve simulated test scores by over 20% compared to strong baselines that are fine-tuned using indirect metrics or leverage prompting methods. Furthermore, utility is only weakly correlated with salience and similarity to exam questions, suggesting that it captures unique signal that benefits downstream performance. QUEST offers a new outcome-driven paradigm for question evaluation and generation - one that moves beyond question-answer content toward measurable improvements in learning outcomes.

cs.CL↗

arXiv2025

Can Vision Language Models Understand Mimed Actions?

Nonverbal communication (NVC) plays an integral role in human language, but studying NVC in general is challenging because of its broad scope and high variance in interpretation among individuals and cultures. However, mime -- the theatrical technique of suggesting intent using only gesture, expression, and movement -- is a subset of NVC that consists of explicit and embodied actions with much lower human interpretation variance. We argue that a solid understanding of mimed actions is a crucial prerequisite for vision-language models capable of interpreting and commanding more subtle aspects of NVC. Hence, we propose Mime Identification Multimodal Evaluation (MIME), a novel video-based question answering benchmark comprising of 86 mimed actions. Constructed with motion capture data, MIME consists of variations of each action with perturbations applied to the character, background, and viewpoint for evaluating recognition robustness. We find that both open-weight and API-based vision-language models perform significantly worse than humans on MIME, motivating the need for increased research for instilling more robust understanding of human gestures.

cs.CL↗

arXiv2025

Teaching Language Models To Gather Information Proactively

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

cs.AI↗

arXiv2025

R2D2: Remembering, Replaying and Dynamic Decision Making with a Reflective Agentic Memory

The proliferation of web agents necessitates advanced navigation and interaction strategies within complex web environments. Current models often struggle with efficient navigation and action execution due to limited visibility and understanding of web structures. Our proposed R2D2 framework addresses these challenges by integrating two paradigms: Remember and Reflect. The Remember paradigm uses a replay buffer that aids agents in reconstructing the web environment dynamically, thus enabling the formulation of a detailed "map" of previously visited pages. This helps in reducing navigational errors and optimizing the decision-making process during web interactions. Conversely, the Reflect paradigm allows agents to learn from past mistakes by providing a mechanism for error analysis and strategy refinement, enhancing overall task performance. We evaluate R2D2 using the WebArena benchmark, demonstrating substantial improvements over existing methods, including a 50% reduction in navigation errors and a threefold increase in task completion rates. Our findings suggest that a combination of memory-enhanced navigation and reflective learning promisingly advances the capabilities of web agents, potentially benefiting various applications such as automated customer service and personal digital assistants.

cs.AI↗

arXiv2025

Should I Trust You? Detecting Deception in Negotiations using Counterfactual RL

An increasingly common socio-technical problem is people being taken in by offers that sound ``too good to be true'', where persuasion and trust shape decision-making. This paper investigates how \abr{ai} can help detect these deceptive scenarios. We analyze how humans strategically deceive each other in \textit{Diplomacy}, a board game that requires both natural language communication and strategic reasoning. This requires extracting logical forms of proposed agreements in player communications and computing the relative rewards of the proposal using agents' value functions. Combined with text-based features, this can improve our deception detection. Our method detects human deception with a high precision when compared to a Large Language Model approach that flags many true messages as deceptive. Future human-\abr{ai} interaction tools can build on our methods for deception detection by triggering \textit{friction} to give users a chance of interrogating suspicious proposals.

cs.CL↗

arXiv2025

Tuning-Free Personalized Alignment via Trial-Error-Explain In-Context Learning

Language models are aligned to the collective voice of many, resulting in generic outputs that do not align with specific users' styles. In this work, we present Trial-Error-Explain In-Context Learning (TICL), a tuning-free method that personalizes language models for text generation tasks with fewer than 10 examples per user. TICL iteratively expands an in-context learning prompt via a trial-error-explain process, adding model-generated negative samples and explanations that provide fine-grained guidance towards a specific user's style. TICL achieves favorable win rates on pairwise comparisons with LLM-as-a-judge up to 91.5% against the previous state-of-the-art and outperforms competitive tuning-free baselines for personalized alignment tasks of writing emails, essays and news articles. Both lexical and qualitative analyses show that the negative samples and explanations enable language models to learn stylistic context more effectively and overcome the bias towards structural and formal phrases observed in their zero-shot outputs. By front-loading inference compute to create a user-specific in-context learning prompt that does not require extra generation steps at test time, TICL presents a novel yet simple approach for personalized alignment.

cs.CL↗

arXiv2024

NewsEdits 2.0: Learning the Intentions Behind Updating News

As events progress, news articles often update with new information: if we are not cautious, we risk propagating outdated facts. In this work, we hypothesize that linguistic features indicate factual fluidity, and that we can predict which facts in a news article will update using solely the text of a news article (i.e. not external resources like search engines). We test this hypothesis, first, by isolating fact-updates in large news revisions corpora. News articles may update for many reasons (e.g. factual, stylistic, narrative). We introduce the NewsEdits 2.0 taxonomy, an edit-intentions schema that separates fact updates from stylistic and narrative updates in news writing. We annotate over 9,200 pairs of sentence revisions and train high-scoring ensemble models to apply this schema. Then, taking a large dataset of silver-labeled pairs, we show that we can predict when facts will update in older article drafts with high precision. Finally, to demonstrate the usefulness of these findings, we construct a language model question asking (LLM-QA) abstention task. We wish the LLM to abstain from answering questions when information is likely to become outdated. Using our predictions, we show, LLM absention reaches near oracle levels of accuracy.

cs.CL↗

arXiv2024

Explaining Mixtures of Sources in News Articles

Human writers plan, then write. For large language models (LLMs) to play a role in longer-form article generation, we must understand the planning steps humans make before writing. We explore one kind of planning, source-selection in news, as a case-study for evaluating plans in long-form generation. We ask: why do specific stories call for specific kinds of sources? We imagine a generative process for story writing where a source-selection schema is first selected by a journalist, and then sources are chosen based on categories in that schema. Learning the article's plan means predicting the schema initially chosen by the journalist. Working with professional journalists, we adapt five existing schemata and introduce three new ones to describe journalistic plans for the inclusion of sources in documents. Then, inspired by Bayesian latent-variable modeling, we develop metrics to select the most likely plan, or schema, underlying a story, which we use to compare schemata. We find that two schemata: stance and social affiliation best explain source plans in most documents. However, other schemata like textual entailment explain source plans in factually rich topics like "Science". Finally, we find we can predict the most suitable schema given just the article's headline with reasonable accuracy. We see this as an important case-study for human planning, and provides a framework and approach for evaluating other kinds of plans. We release a corpora, NewsSources, with annotations for 4M articles.

cs.CL↗

arXiv2024

WinoQueer: A Community-in-the-Loop Benchmark for Anti-LGBTQ+ Bias in Large Language Models

We present WinoQueer: a benchmark specifically designed to measure whether large language models (LLMs) encode biases that are harmful to the LGBTQ+ community. The benchmark is community-sourced, via application of a novel method that generates a bias benchmark from a community survey. We apply our benchmark to several popular LLMs and find that off-the-shelf models generally do exhibit considerable anti-queer bias. Finally, we show that LLM bias against a marginalized community can be somewhat mitigated by finetuning on data written about or by members of that community, and that social media text written by community members is more effective than news text written about the community by non-members. Our method for community-in-the-loop benchmark development provides a blueprint for future researchers to develop community-driven, harms-grounded LLM benchmarks for other marginalized communities. Note: This version corrects a bug found in evaluation code after publication. General findings have not changed, but tables 5 and 6 and figure 1 have been corrected.

cs.CL↗

arXiv2024

FoodPuzzle: Developing Large Language Model Agents as Flavor Scientists

Flavor development in the food industry is increasingly challenged by the need for rapid innovation and precise flavor profile creation. Traditional flavor research methods typically rely on iterative, subjective testing, which lacks the efficiency and scalability required for modern demands. This paper presents three contributions to address the challenges. Firstly, we define a new problem domain for scientific agents in flavor science, conceptualized as the generation of hypotheses for flavor profile sourcing and understanding. To facilitate research in this area, we introduce the FoodPuzzle, a challenging benchmark consisting of 978 food items and 1,766 flavor molecules profiles. We propose a novel Scientific Agent approach, integrating in-context learning and retrieval augmented techniques to generate grounded hypotheses in the domain of food science. Experimental results indicate that our model significantly surpasses traditional methods in flavor profile prediction tasks, demonstrating its potential to transform flavor development practices.

cs.CL↗

arXiv2024

Are Large Language Models Capable of Generating Human-Level Narratives?

This paper investigates the capability of LLMs in storytelling, focusing on narrative development and plot progression. We introduce a novel computational framework to analyze narratives through three discourse-level aspects: i) story arcs, ii) turning points, and iii) affective dimensions, including arousal and valence. By leveraging expert and automatic annotations, we uncover significant discrepancies between the LLM- and human- written stories. While human-written stories are suspenseful, arousing, and diverse in narrative structures, LLM stories are homogeneously positive and lack tension. Next, we measure narrative reasoning skills as a precursor to generative capacities, concluding that most LLMs fall short of human abilities in discourse understanding. Finally, we show that explicit integration of aforementioned discourse features can enhance storytelling, as is demonstrated by over 40% improvement in neural storytelling in terms of diversity, suspense, and arousal.

cs.CL↗

arXiv2024

Speechworthy Instruction-tuned Language Models

Current instruction-tuned language models are exclusively trained with textual preference data and thus are often not aligned with the unique requirements of other modalities, such as speech. To better align language models with the speech domain, we explore (i) prompting strategies grounded in radio-industry best practices and (ii) preference learning using a novel speech-based preference data of 20K samples, generated with a wide spectrum of prompts that induce varying dimensions of speech-suitability and labeled by annotators who listen to response pairs. Both human and automatic evaluation show that both prompting and preference learning increase the speech-suitability of popular instruction-tuned LLMs. Interestingly, we find that prompting and preference learning can be additive; combining them achieves the best win rates in head-to-head comparison, resulting in responses that are preferred or tied to the base model in 76.2% of comparisons on average. Lastly, we share lexical, syntactical, and qualitative analyses to showcase how each method contributes to improving the speech-suitability of generated responses.

cs.AI↗

arXiv2024

Authorship Style Transfer with Policy Optimization

Authorship style transfer aims to rewrite a given text into a specified target while preserving the original meaning in the source. Existing approaches rely on the availability of a large number of target style exemplars for model training. However, these overlook cases where a limited number of target style examples are available. The development of parameter-efficient transfer learning techniques and policy optimization (PO) approaches suggest lightweight PO is a feasible approach to low-resource style transfer. In this work, we propose a simple two-stage tune-and-optimize technique for low-resource textual style transfer. We apply our technique to authorship transfer as well as a larger-data native language style task and in both cases find it outperforms state-of-the-art baseline models.

cs.CL↗

arXiv2024

Style Transfer with Multi-iteration Preference Optimization

Numerous recent techniques for text style transfer characterize their approaches as variants of reinforcement learning and preference optimization. In this work, we consider the relationship between these approaches and a class of optimization approaches developed primarily for (non-neural) statistical machine translation, formerly known as `tuning'. Inspired by these techniques from the past, we improve upon established preference optimization approaches, incorporating multiple iterations of exploration and optimization, and choosing contrastive examples by following a `hope' vs `fear' sampling strategy. Cognizant of the difference between machine translation and style transfer, however, we further tailor our framework with a new pseudo-parallel generation method and a dynamic weighted reward aggregation method to tackle the lack of parallel data and the need for a multi-objective reward. We evaluate our model on two commonly used text style transfer datasets. Through automatic and human evaluation results we show the effectiveness and the superiority of our model compared to state-of-the-art baselines.

cs.CL↗

arXiv2024

BotEval: Facilitating Interactive Human Evaluation

Following the rapid progress in natural language processing (NLP) models, language models are applied to increasingly more complex interactive tasks such as negotiations and conversation moderations. Having human evaluators directly interact with these NLP models is essential for adequately evaluating the performance on such interactive tasks. We develop BotEval, an easily customizable, open-source, evaluation toolkit that focuses on enabling human-bot interactions as part of the evaluation process, as opposed to human evaluators making judgements for a static input. BotEval balances flexibility for customization and user-friendliness by providing templates for common use cases that span various degrees of complexity and built-in compatibility with popular crowdsourcing platforms. We showcase the numerous useful features of BotEval through a study that evaluates the performance of various chatbots on their effectiveness for conversational moderation and discuss how BotEval differs from other annotation tools.

cs.CL↗