Large language models (LLMs) have demonstrated strong persuasive capabilities comparable to those of humans, offering promising benefits while raising societal concerns about their deployment. However, systematically evaluating the persuasive capabilities of LLMs is inherently challenging, as the effectiveness of persuasion among humans varies significantly across different domains. In this paper, we take a theory-driven approach to provide a scalable and principled framework for measuring the persuasive capabilities of LLMs. Grounded in the Bayesian Persuasion (BP) framework, we repurpose existing human-human persuasion datasets to construct environments for evaluating and training LLMs in strategic persuasion. Our results reveal that frontier models can consistently achieve high persuasion gains and exhibit sophisticated persuasion strategies that align with theoretical predictions. Building on this, we use reinforcement learning to train LLMs for strategic persuasion in our environments. Our results also demonstrate that even small LLMs can obtain significantly higher persuasion gains through reinforcement learning.

Large Language Models (LLMs) vary widely in their capabilities, with larger models often having better performance but higher cost: choosing an LLM model often involves trading off performance and cost. The LLM Cascade is a paradigm that defers difficult queries from weak/cheap to strong/expensive models. This approach is nonadaptive: the deferral decision is trained offline. When confronted with similar or repeated queries, the LLM Cascade may then repeatedly consult the expensive model and incur higher cost. To improve the cascading efficiency, we propose Inter-Cascade, an online and interactive LLM Cascade that extends the role of strong model from a backup helper to a long-term teacher. In our system, when a strong model resolves a difficult query, it also distills its solution into a generalized, reusable problem-solving strategy that boosts the weak model on subsequent queries. Adding strategies to queries enables the weak model to dynamically improve its performance over time, avoiding computationally and time-intensive fine-tuning. Empirically, compared with standard LLM Cascade baselines across multiple benchmarks, the Inter-Cascade significantly improves the accuracy of the weak model (by up to 33.06 absolute percentage points) and the overall system (by up to 5.53 absolute percentage points), while reducing the calls to strong models (by up to 48.05% relative reduction) and saving the corresponding fees (by up to 49.63% relative reduction). Inter-Cascade demonstrates the effective in-context knowledge transfer between LLMs, and provides a general, scalable framework applicable to both open-source and API-based LLMs.

As Generative AI (GenAI) systems see growing adoption, a key concern involves the external validity of evaluations, or the extent to which they generalize from lab-based to real-world deployment conditions. Threats to the external validity of GenAI evaluations arise when the source sample of human raters and system outputs used to obtain a system quality estimate differs from the target distribution at deployment time. In this work, we propose a doubly-robust estimation framework designed to address this evaluation sampling bias. Key to our approach is the use of "persona" ratings produced by prompting an LLM evaluator (i.e., an LLM-as-a-judge) to behave as a human rater with specific sociodemographic characteristics. Our doubly-robust framework combines these informative yet imperfect persona ratings with human ratings obtained under evaluation sampling bias to produce statistically valid system quality estimates. In particular, we show that our approach yields valid system quality estimates when either (i) a model trained to predict human ratings using persona ratings and source data observed under sampling bias, or (ii) a reweighting model that corrects for sampling bias is of sufficient quality. We validate our framework theoretically and via a novel Persona Simulation Framework (PSF) designed to systematically manipulate persona quality and the degree of evaluation sampling bias present in source data. Our work provides a principled foundation for combining imperfect persona ratings with human ratings observed under sampling bias to obtain valid system quality estimates.

The QMoE model provides a practical approach for compression of massive Mixture-of-Experts (MoE) models. QMoE offers a solution geared towards memory limitations that often reach terabyte scales, and it has the advantage of working with high sparsity models which implicitly lend themselves to compression techniques. QMoE also has the advantage of only taking MoE models into account and does not evaluate its use with non mixture of expert systems. Although this prior attempt focuses on the limitations of large servers with the latest NVIDIA hardware which in the case of the H100 and V100 which have 80 GB of HBM (High Bandwidth Memory), what is not being considered is a significantly more constrained environment, such as in the case of mobile devices which may have in the case of the iPhone anywhere from 4 to 8 GB of unified memory which also needs to be shared with the operating system and additional processes. Although edge devices such as phones and laptops are becoming increasingly more computationally powerful, they are still not close to the level of advanced server machines such as NVIDIA. An additional constraint that we must consider is that of latency. The communication time of sending a request to an LLM server and then getting it back is an additional waiting time that can be removed. We may also want to use LLM technology in environments where there is no reliable network connection.

Does explicitly exercising the induction circuit during pretraining improve in-context learning (ICL), or is natural text sufficient when compute is held constant (iso-FLOPs)? To test whether targeted synthetic data can accelerate induction-head emergence and enhance ICL, we introduce Bi-Induct, a lightweight curriculum that injects forward-copy (Induction), backward-copy (Anti), or a balanced mix into the pretraining stream. We train models from 0.13B to 1B parameters under iso-FLOPs, evaluating (i) few-shot ICL benchmarks, (ii) head-level telemetry, and (iii) held-out language modeling perplexity. Our findings challenge the assumption that early induction circuit activation directly improves ICL. While Bi-Induct accelerates induction-head emergence at small scales, this does not consistently yield stronger generalization. On standard LM benchmarks, Bi-Induct matches natural-only training; on function-style ICL probes, the 1B natural-only performs best. Stress tests (e.g., label permutation, HITS@1 vs. HITS@3, 1 vs. 10 shots) preserve these trends. Telemetry shows larger natural-only models develop broader, earlier induction heads without explicit induction patterns. Anti-induction data fails to elicit meaningful activation. Perplexity penalties from synthetic data shrink with scale, suggesting larger models can absorb non-natural patterns with minimal cost. Crucially, ablating the top 2% of induction heads degrades ICL more than random ablations, especially for natural-only models, indicating more centralized, load-bearing circuits. Bi-Induct variants exhibit more redundant induction activity, implying different circuit utilization. Overall, inducing activation is not sufficient: ICL gains depend on these circuits becoming functionally necessary. These results underscore mechanism-aware pretraining diagnostics and data mixtures that foster load-bearing, not merely present, structure.

Generative AI has established the opportunity to readily transform content from one medium to another. This capability is especially powerful for storytelling, where visual illustrations can illuminate a story originally expressed in text. In this paper, we focus on the task of narrative scene illustration, which involves automatically generating an image depicting a scene in a story. Motivated by recent progress on text-to-image models, we consider a pipeline that uses LLMs as an interface for prompting text-to-image models to generate scene illustrations given raw story text. We apply variations of this pipeline to a prominent story corpus in order to synthesize illustrations for scenes in these stories. We conduct a human annotation task to obtain pairwise quality judgments for these illustrations. The outcome of this process is the SceneIllustrations dataset, which we release as a new resource for future work on cross-modal narrative transformation. Through our analysis of this dataset and experiments modeling illustration quality, we demonstrate that LLMs can effectively verbalize scene knowledge implicitly evoked by story text. Moreover, this capability is impactful for generating and evaluating illustrations.

Quantization-aware training (QAT) is a leading technique for improving the accuracy of quantized neural networks. Previous work has shown that decomposing training into a full-precision (FP) phase followed by a QAT phase yields superior accuracy compared to QAT alone. However, the optimal allocation of compute between the FP and QAT phases remains unclear. We conduct extensive experiments with various compute budgets, QAT bit widths, and model sizes from 86.0M to 2.2B to investigate how different QAT durations impact final performance. We demonstrate that, contrary to previous findings, the loss-optimal ratio of QAT to FP training increases with the total amount of compute. Moreover, the optimal fraction can be accurately predicted for a wide range of model sizes and quantization widths using the tokens-per-parameter-byte statistic. From experimental data, we derive a loss scaling law that predicts both optimal QAT ratios and final model performance across different QAT/FP compute allocation strategies and QAT bit widths. We use the scaling law to make further predictions, which we verify experimentally, including which QAT bit width is optimal under a given memory constraint and how QAT accuracy with different bit widths compares to full-precision model accuracy. Additionally, we propose a novel cooldown and QAT fusion approach that performs learning rate decay jointly with quantization-aware training, eliminating redundant full-precision model updates and achieving significant compute savings. These findings provide practical insights into efficient QAT planning and enable the training of higher-quality quantized models with the same compute budget.

We introduce a novel approach to large language model (LLM) distillation by formulating it as a constrained reinforcement learning problem. While recent work has begun exploring the integration of task-specific rewards into distillation processes, existing methods typically rely on ad-hoc reward weighting. We propose a principled optimization framework that maximizes task-specific rewards while constraining the divergence from the teacher model to remain below a specified threshold. Our approach adapts constrained state augmented reinforcement learning to the distillation setting, introducing a modified reward function that maintains theoretical guarantees of constraint satisfaction without requiring state augmentation or teacher model access during deployment and without the computational overhead of the dual Lagrangian methods. Through extensive experiments on mathematical reasoning tasks, we demonstrate that our method achieves better constraint satisfaction rates and better reasoning compared to the soft Lagrangian relaxation baselines while maintaining competitive task performance. Our framework provides a theoretically grounded and practically efficient solution for reward-aware distillation in resource-constrained settings. Large Language Models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks (V aswani et al., 2017; Trinh et al., 2024; Chervonyi et al., 2025; Guo et al., 2025; Christianos et al., 2023), but their size and complexity make them impractical for deployment in resource-constrained environments. Distillation (Hinton et al., 2015; Czarnecki et al., 2019), a technique where a smaller student model learns from a larger teacher model, has been widely used to transfer knowledge while reducing computational costs. Conventional distillation methods (Sanh et al., 2020; Gu et al., 2024; Ko et al., 2024) typically focus on minimizing the divergence between the student and teacher models, often using metrics such as Kullback-Leibler (KL) divergence. However, these methods do not fully leverage additional reward signals that can provide valuable guidance, particularly in tasks requiring complex reasoning.

We present and test the largest benchmark for vericoding, LLM-generation of formally verified code from formal specifications -- in contrast to vibe coding, which generates potentially buggy code from a natural language description. Our benchmark contains 12,504 formal specifications, with 3,029 in Dafny, 2,334 in V erus/Rust and 7,141 in Lean. Of these, 6,174 are new unseen problems. We find vericoding success rates of 27% in Lean, 44% in V erus/Rust and 82% in Dafny using off-the-shelf LLMs. Adding natural-language descriptions does not significantly improve performance. We also find that LLM progress has improved progress on pure Dafny verification from 68% to 96% over the past year. The benchmark and vericoding results are shared in this GitHub repo. Rapid AI progress has popularized vibe coding, which generates computer programs from natural language descriptions. For example, Google has reported that over 30% of its software is created this way (Google Earnings Call). Unfortunately, the resulting code can be buggy, and traditional bug hunting with test cases can typically only demonstrate the presence and not the absence of bugs, since there are too many test cases to try them all. For example, major code-testing efforts failed to prevent bugs causing an Ariane-V rocket explosion (Ariane 5 Failure) and an embarrassing security vulnerability in the Bash shell (Shellshock Bug) that was built into the Unix operating system for 25 years before being discovered. The 2024 CrowdStrike outage disrupted 8.5 million devices globally, harming airlines, hospitals, banking, broadcasting, emergency services (CrowdStrike Outage). Fortunately, rigorous correctness guarantees can be created via formal verification, by generating a machine-checkable proof that code meets its human-written specifications.

Standard LLM evaluation practices compress diverse abilities into single scores, obscuring their inherently multidimensional nature. We present JE-IRT, a geometric item-response framework that embeds both LLMs and questions in a shared space. For question embeddings, the direction encodes semantics and the norm encodes difficulty, while correctness on each question is determined by the geometric interaction between the model and question embeddings. This geometry replaces a global ranking of LLMs with topical specialization and enables smooth variation across related questions. Building on this framework, our experimental results reveal that out-of-distribution behavior can be explained through directional alignment, and that larger norms consistently indicate harder questions. Moreover, JE-IRT naturally supports generalization: once the space is learned, new LLMs are added by fitting a single embedding. The learned space further reveals an LLM-internal taxonomy that only partially aligns with human-defined subject categories. JE-IRT thus establishes a unified and interpretable geometric lens that connects LLM abilities with the structure of questions, offering a distinctive perspective on model evaluation and generalization.