arXiv subjects

Jonathan May

Publications and source records attributed to Jonathan May.

At least 19 recordsLinked to original sources

arXiv2026

FlashMap: Flow Map Speculative Decoding

Speculative decoding accelerates large language model (LLM) inference by using a lightweight drafter to propose a block of tokens that the target model verifies in one parallel pass. Autoregressive drafters generate this block one token at a time, whereas block diffusion drafters such as DFlash generate it in one pass, making drafting faster. However, these diffusion drafters predict each token independently, ignoring dependencies between tokens in the block. We introduce FlashMap, a continuous flow drafter that conditions a flow map language model on the target model's hidden states and generates the whole block in one forward pass while capturing inter-token dependencies. On Qwen3-8B, FlashMap is faster than a matched DFlash baseline on every benchmark, with 7% higher average speedup and up to 6.1x lossless acceleration.

cs.CL↗

arXiv2026

How To Train Your World Model: Fine-tuning vs RAG for LM-based World Modeling

World models (WMs) simulate the transition dynamics of environments, enabling agents to plan over the consequences of their actions. In text-based environments, fine-tuning a Language Model (LM) to serve as a WM has emerged as a dominant paradigm. However, despite the widespread success of non-parametric approaches such as Retrieval Augmented Generation (RAG), retrieval for LM-based world modelling remains underexplored. We conduct a systematic evaluation across five diverse environments spanning embodied, web navigation and social settings, comparing fine-tuning and RAG-based approaches for LM-based world modelling. Our study reveals that fine-tuning often outperforms RAG, with fine-tuned WMs enabling agents to obtain higher rewards on 15/20 settings. While both construction paradigms benefit from additional and more diverse exploration, RAG-based approaches prove more data-efficient, and fine-tuning approaches disproportionately benefit from scaling the amount of experience collected. With a focus on RAG-based WMs, we devise a procedure that uses counterfactual intervention to estimate the error rate of the retrieval stage, and show that retrievers consistently surface suboptimal transitions from the experience buffer. Hoping to address this failing, we study a variety of query reformulation strategies, demonstrating that a hierarchical approach outperforms the traditional retrieval pipeline. Finally, we compose our findings into a hybrid world modelling system that parametrically captures core environment dynamics, while learning to rely on retrieval from an actively maintained memory store. Our hybrid system consistently outperforms other methods across multiple environments and models, showcasing the robustness of the approach and the applicability of our findings.

cs.AI↗

arXiv2026

Toward Interactive Understanding of Code APIs

Empowered by advances in Language Model agents, systems have made substantial strides in code generation and understanding. However, these approaches often rely on read access to the relevant code, an assumption which does not hold when dealing with external APIs. In this work, we introduce the PAU (Python API Understanding) benchmark, where we provide models with black-box, API-level access to code snippets. Models must query the API with exploratory inputs and draw insights from the resulting outputs, with the goal of describing the snippet's true functionality. By treating the code snippets as external tools that must be understood via interaction alone, PAU studies the more general problem of unsupervised tool understanding, specifically for tools implemented as Python methods. Despite recent progress in coding agents, even frontier models struggle to achieve high performance on PAU, with the best model (Claude-4-Opus) failing to understand over 45% of the PAU test set. An investigation into the common error modes reveals that models are overconfident; they often overrate the quality of their current hypothesis, leading to insufficient exploration and premature termination. Finally, we take inspiration from the Asymmetric Actor Critic (AAC) paradigm, frequently used in robot learning, to post-train models for interactive code understanding. Models trained with AAC conduct more active exploration of the APIs, with an AAC-tuned Qwen3-8B model matching the performance of GPT-5-mini.

cs.AI↗

arXiv2026

GameBoyWorlds: A Testbed for Self-Improvement in Embodied Video Games

Powered by expert guidance, agents can operate in interactive environments; however, it is unclear whether they can learn autonomously from their own experience. To evaluate such self-improvement methods, we introduce GameBoyWorlds, a testbed for agentic self-improvement in video games. GameBoyWorlds-Execution evaluates task execution on a collection of 5 distinct game series. Agents are allowed access to dedicated training games but are provided no demonstrations, documentation, or rewards. Agents must ground themselves in the environment through self-directed exploration and by inferring actionable knowledge from their own experience. At test time, agents must complete short-horizon tasks that evaluate their ability to navigate, interact, and engage with game-specific mechanics in unseen games. Out-of-the-box frontier models complete fewer than 50% of the 500 tasks due to failures in multimodal grounding, establishing that self-improvement methods have room to push performance. We demonstrate that contemporary approaches to self-improvement are lacking, with world modelling and autonomous skill discovery failing, and a novel strategy that uses curiosity-based exploration to write guides achieving only partial success. GameBoyWorlds-Playthrough tests end-to-end game completion in two fan-made Pokémon games. We show that while frontier models have been pre-exposed to official releases such as Pokémon Red, they lack essential information on the games in our testbed. Instead of relying on their parametric knowledge to succeed, agents must learn from their own experience and autonomously improve over the course of the playthrough. We show that a sophisticated agentic pipeline with multimodal memory and hierarchical subgoals fails to reach even the first major milestone in both games, establishing GameBoyWorlds as an ambitious target for self-improving agents.

cs.AI↗

arXiv2026

GLARE: Generative Learning via Adversarial Reward Estimation For Social Dynamics Forecasting

Meeting continuation requires tracking the agenda, speaker roles, participant intentions, and disagreement across long multi-party discussions. We introduce the Meeting Dynamic Forecasting Benchmark (MDFB), constructed from 2,207 real-world meetings and 24,794 future-facing queries. Given a transcript prefix and an active question, a model generates a plausible multi-turn continuation in one call. We evaluate utility---progress toward the question---and human-likeness---plausible conversational flow and role consistency---without requiring exact reproduction of the observed future. We further present GLARE, an adaptation of adversarial imitation learning to conditional language generation. A discriminator ranks the observed continuation above samples from the current actor, and its score supplies a KL-regularized policy reward; retraining on current-policy negatives allows the reward landscape to evolve with the actor. GLARE attains average human-evaluated win rates of 0.66 on utility and 0.70 on human-likeness, outperforming SFT and SPIN while remaining below the observed human continuation. We also demonstrate MDFB as a social reasoning arena for comparing general-purpose models, including closed-source systems, through reference-assisted judgments. Together, these studies illustrate the benchmark's use for both task-specific learning and output-based evaluation of meeting behavior.

cs.AI↗

arXiv2026

ORBIT: Training-Free Multi-Attribute Behavioral Steering via Orthogonal Subspace Rotation

Language models are widely used in assistant settings, where controlling behavioral attributes is often essential. Activation steering modifies hidden-state representations at inference time, providing a lightweight, training-free mechanism that can be toggled at runtime. Existing methods, however, have focused primarily on steering a single attribute at a time. When multiple attributes must be controlled simultaneously, naive summation of per-attribute steering vectors suffers from norm imbalance and directional cancellation, while classifier-based approaches require retraining whenever the attribute set changes. We introduce ORBIT (Orthogonal Rotation-Based Intervention Technique), a training-free extension of rotation-based steering to the multi-attribute setting. Our method constructs a joint subspace from per-attribute steering planes via singular value decomposition and applies a single norm-preserving rotation within that subspace toward a combined target direction. Adaptive per-token gating identifies which attributes need correction at each position, and an optional additive boost strengthens attributes with weak initial projection. We also introduce TraitFactory, a new multi-attribute benchmark that focuses on behavioral tendencies rather than surface-level style. We evaluate ORBIT on TraitFactory and ToneBank across three models (Llama-3.2-3B, Qwen-2.5-7B, Llama-3.1-8B) while steering multiple attributes simultaneously, showing that it achieves stronger and more balanced multi-attribute steering than existing training-free baselines while better preserving output coherence.

cs.CL↗

arXiv2026

It Matters How You Say It: Exploring Rhetorical Patterns for AI-Assisted Information Evaluation

Prior work on AI-assisted information evaluation has largely focused on what AI systems communicate, comparing explanation types and formats, with responses predominantly cast in directive rhetoric where the system delivers a verdict and the user passively accepts it. While debate-style interactions have recently shown promise in prompting critical evaluation over deference, the rhetorical patterns that structure AI responses and how they might induce reflection, uncertainty, or independent reasoning remain largely unexamined. To address this, we investigated eight rhetorical patterns known to induce contemplation: Intentional Misleading, Interpretive Alternative, Scaffold Explanation, Triggering Distrust, Information Distortion, Alternative Framing, Socratic Questioning, and an Oracle baseline. Through a within-subject study with n=98 participants on a hint-on-demand fact verification task, we observed preliminary evidence that Scaffold Explanation were associated with the highest accuracy gains, and encouraging deeper reflection. Surprisingly, the adversarial conditions also improved accuracy modestly. Participants preferred Alternative Framing most and Interpretive Alternative least, largely due to the latter's perceived time cost. We discuss the implications of designing conversational agents with varied rhetorical styles and the trade-offs among user performance, satisfaction, and contemplation.

cs.HC↗

arXiv2026

Sparse Layers are Critical to Scaling Looped Language Models

Looped language models repeat a set of transformer layers through depth, reducing memory costs and providing natural early-exit points at loop boundaries. However, looped models do not scale as favorably as standard transformers with unique layers. We compare standard and Mixture-of-Experts (MoE) transformers, with and without looping, and find two main results. First, we find Looped-MoE models scale better than the standard baseline while dense looped models do not. We trace this to routing divergence between loops: in Looped-MoE models, different experts are activated on each pass through the same shared layers, recovering expressivity without additional parameters. Our second finding is that looped models have better compute-quality trade-offs with early exits than standard models. Because each loop ends with the same layers that produce the final output, loop boundaries are superior exit points, as confirmed by earlier output convergence at these points. In sum, we provide a clear direction for scaling looped models: a Looped-MoE model with early exits can not only beat standard transformers at scale, but also enable significant memory and inference savings with minimal degradation in quality.

cs.LG↗

arXiv2026

Translating the Untranslatable: An Operationalizable Ontology for Untranslatability

Untranslatability, cases where meaning cannot be directly preserved across languages, is well-studied in linguistics but underexplored in NLP. As machine translation (MT) systems improve on standard benchmarks, their limitations increasingly concentrate in such cases, where translation cannot be reduced to one-to-one equivalence. We introduce a structured ontology of untranslatability along with a taxonomy of compensation strategies, which are specific techniques to convey meaning under these untranslatable circumstances. We operationalize this framework into a multilingual dataset of untranslatable sentences paired with strategy-based translations, enabling controlled analysis of translation behavior. Initial human preference studies suggest that translation quality depends on the strategy used, with consistent preferences for outputs that include explanatory context, known as the Annotation compensation strategy. Our framework and dataset provide a foundation for studying and modeling strategy-informed machine translation.

cs.CL↗

arXiv2026

Uncovering Intervention Opportunities for Suicide Prevention with Language Model Assistants

Warning: This paper discusses topics of suicide and suicidal ideation, which may be distressing to some readers. The National Violent Death Reporting System (NVDRS) documents information about suicides in the United States, including free text narratives (e.g., circumstances surrounding a suicide). In a demanding public health data pipeline, annotators manually extract structured information from death investigation records following extensive guidelines developed painstakingly by experts. In this work, we facilitate data-driven insights from the NVDRS data to support the development of novel suicide interventions by investigating the value of language models (LMs) as efficient assistants to these (a) data annotators and (b) experts. We find that LM predictions match existing data annotations about 85% of the time across 50 NVDRS variables. In the cases where the LM disagrees with existing annotations, expert review reveals that LM assistants can surface annotation discrepancies 38% of the time. Finally, we introduce a human-in-the-loop algorithm to assist experts in efficiently building and refining guidelines for annotating new variables by allowing them to focus only on providing feedback for incorrect LM predictions. We apply our algorithm to a real-world case study for a new variable that characterizes victim interactions with lawyers and demonstrate that it achieves comparable annotation quality with a laborious manual approach. Our findings provide evidence that LMs can serve as effective assistants to public health researchers who handle sensitive data in high-stakes scenarios.

cs.CY↗

arXiv2026

Optimizing Diversity and Quality through Base-Aligned Model Collaboration

Alignment has greatly improved large language models (LLMs)' output quality at the cost of diversity, yielding highly similar outputs across generations, especially in open-ended generation tasks. We propose Base-Aligned Model Collaboration (BACo), an inference-time token-level model collaboration framework that dynamically combines a base LLM with its aligned counterpart to optimize diversity and quality. Using uncertainty and content-based signals, BACo employs routing strategies to determine, at each token, which model to decode from. Prior diversity-promoting methods often improve diversity at the expense of quality or require expensive decoding or post-training. In contrast, BACo achieves both high diversity and quality post hoc within a single pass, while offering strong controllability. We introduce a family of effective routing strategies and evaluate them across three open-ended generation tasks with 13 diversity and quality metrics. BACo consistently surpasses state-of-the-art inference-time baselines. With our best router, BACo achieves a 21.3% joint improvement in diversity and quality, which is further supported by human evaluations. Overall, our results demonstrate that collaboration between base and aligned models provides an effective and controllable mechanism for optimizing the diversity-quality trade-off.

cs.CL↗

arXiv2026

How Language Models Process Negation

We study how Large Language Models (LLMs) process negation mechanistically. First, we establish that even though open-weight models often provide wrong answers to questions involving negation, they do possess internal components that process negation correctly. Their poor accuracy is due to late-layer attention behavior that promotes simple shortcuts; ablating those attention modules greatly improves accuracy on negation-related questions. Second, we uncover how models process negation. We consider two hypotheses: models could use attention heads that attend to the phrase being negated and suppress related concepts, or they could directly construct a representation of the entire negative phrase (e.g., representing "not gas" as a vector that promotes liquids and solids). We apply a range of observational and causal interpretability techniques on Mistral-7B and Llama-3.1-8B to show that models implement both mechanisms, with the "constructive" mechanism being more prominent. Combined, our work deepens the understanding of LLMs' internals, highlighting construction-dominant computations and the coexistence of competing mechanisms within LLMs.

cs.CL↗

arXiv2026

GTA: Generating Long-Horizon Tasks for Web Agents at Scale

Web agents, which couple language models with browsing and tool-use capabilities, show promise as open web assistants. Yet progress is increasingly limited by the lack of scalable, process-level supervision. Existing benchmarks are largely manually constructed, providing only coarse start-goal annotations without intermediate trajectories, while recent automatic generation efforts remain expensive, biased, and shallow. These limitations prevent reliable training and evaluation of agents that must generalize to realistic, multi-hop, cross-page tasks. We introduce a scalable framework, GTA, that integrates crawling, retrieval-based seeding, in-context generation, and automated quality control to produce realistic tasks paired with executable trajectories. This design decouples crawling from generation for greater efficiency, grounds tasks in the site graph to enforce compositionality, and ensures dense supervision through deterministic replays and systematic validation. We instantiate the pipeline on over 50 websites covering e-commerce, government, forums, and news, with multilingual and multi-hop coverage. The resulting benchmark reveals a significant human-agent performance gap and enables detailed diagnostics. Our contributions are three-fold: (i) formalizing multi-hop web-agent task generation, (ii) proposing an efficient and validated pipeline for automatic data creation, and (iii) releasing a dynamic benchmark with reproducible evaluation.

cs.AI↗

arXiv2026

Conceptual Steganography

Language Models (LMs) emit Chains-of-Thought (CoTs) that drive much of their capability. However, the same sequence that carries useful reasoning can also covertly convey messages: a misaligned model may embed covert information in its CoT that slips through human supervision, a form of steganography known as encoded reasoning. Prior LM steganography schemes operate in the token or lexical space, and a content-preserving paraphraser is the canonical and effective defense in recent work. We introduce conceptual steganography, in which each step of a CoT carries information through patterns of high-level reasoning behavior, rather than through lexical choice. Across four model families and two reasoning domains, this backdoor communication channel is shown to be consistently more robust to a strong paraphrase defense than standard keyword approaches, and the encoding of information into CoTs does not affect their utility in the reasoning process. Having raised awareness of this new risk, we then demonstrate that a strategy-aware paraphraser can close much of the channel, highlighting new challenges and recommended defenses for ensuring faithful LLM reasoning in the wild.

cs.CL↗

arXiv2026

Seamless Deception: Larger Language Models Are Better Knowledge Concealers

Language Models (LMs) may acquire harmful knowledge, and yet feign ignorance of these topics when under audit. Inspired by the recent discovery of deception-related behaviour patterns in LMs, we aim to train classifiers that detect when a LM is actively concealing knowledge. Initial findings on smaller models show that classifiers can detect concealment more reliably than human evaluators, with gradient-based concealment proving easier to identify than prompt-based methods. However, contrary to prior work, we find that the classifiers do not reliably generalize to unseen model architectures and topics of hidden knowledge. Most concerningly, the identifiable traces associated with concealment become fainter as the models increase in scale, with the classifiers achieving no better than random performance on any model exceeding 70 billion parameters. Our results expose a key limitation in black-box-only auditing of LMs and highlight the need to develop robust methods to detect models that are actively hiding the knowledge they contain.

cs.CL↗

arXiv2026

Personalized Help for Optimizing Low-Skilled Users' Strategy

AIs can beat humans in game environments; however, how helpful those agents are to human remains understudied. We augment CICERO, a natural language agent that demonstrates superhuman performance in Diplomacy, to generate both move and message advice based on player intentions. A dozen Diplomacy games with novice and experienced players, with varying advice settings, show that some of the generated advice is beneficial. It helps novices compete with experienced players and in some instances even surpass them. The mere presence of advice can be advantageous, even if players do not follow it.

cs.CL↗

arXiv2026

NewsInterview: a Dataset and a Playground to Evaluate LLMs' Ground Gap via Informational Interviews

Large Language Models (LLMs) have demonstrated impressive capabilities in generating coherent text but often struggle with grounding language and strategic dialogue. To address this gap, we focus on journalistic interviews, a domain rich in grounding communication and abundant in data. We curate a dataset of 40,000 two-person informational interviews from NPR and CNN, and reveal that LLMs are significantly less likely than human interviewers to use acknowledgements and to pivot to higher-level questions. Realizing that a fundamental deficit exists in multi-turn planning and strategic thinking, we develop a realistic simulated environment, incorporating source personas and persuasive elements, in order to facilitate the development of agents with longer-horizon rewards. Our experiments show that while source LLMs mimic human behavior in information sharing, interviewer LLMs struggle with recognizing when questions are answered and engaging persuasively, leading to suboptimal information extraction across model size and capability. These findings underscore the need for enhancing LLMs' strategic dialogue capabilities.

cs.CL↗

arXiv2026

Textual Entailment is not a Better Bias Metric than Token Probability

Measurement of social bias in language models is typically by token probability (TP) metrics, which are broadly applicable but have been criticized for their distance from real-world language model use cases and harms. In this work, we test natural language inference (NLI) as an alternative bias metric. In extensive experiments across seven LM families, we show that NLI and TP bias evaluation behave substantially differently, with very low correlation among different NLI metrics and between NLI and TP metrics. NLI metrics are more brittle and unstable, slightly less sensitive to wording of counterstereotypical sentences, and slightly more sensitive to wording of tested stereotypes than TP approaches. Given this conflicting evidence, we conclude that neither token probability nor natural language inference is a ``better'' bias metric in all cases. We do not find sufficient evidence to justify NLI as a complete replacement for TP metrics in bias evaluation.

cs.CL↗