Conducting contamination-free evaluation of mathematical capabilities can be difficult for two reasons: models may memorize a test set once it is made public, and current mathematical benchmarks are prone to overfitting due to having limited diversity of symbols and rules, coupled with closed-ended answers. This paper proposes a method to leverage these shortcomings as useful features to a construct dynamic, counterfactual benchmark, which can be used to both reveal overfitting and measure true reasoning. We demonstrate this via MatheMagic, which generates math test instances with the interpretations of numbers and operators altered, yet has automatically verifiable answers. Test instances are randomly seeded and constructed at test time to evaluate a model's induction or deduction capability, offering stability, extensibility, comparability, and robustness to overfitting. Our experiments find that models solve deduction more easily than induction, but they revert to standard math. Further analysis reveals that math-adapted models fail to exhibit a general "skill" of reasoning, and fine-tuning on induction tasks generalizes poorly.

Large Language Models (LLMs) are increasingly integrated into everyday applications, but their prevalent cloud-based deployment raises growing concerns around data privacy and long-term sustainability. Running LLMs locally on mobile and edge devices (on-device LLMs) offers the promise of enhanced privacy, reliability, and reduced communication costs. However, realizing this vision remains challenging due to substantial memory and compute demands, as well as limited visibility into performance-efficiency trade-offs on resource-constrained hardware. We propose lm-Meter, the first lightweight, online latency profiler tailored for on-device LLM inference. lm-Meter captures fine-grained, real-time latency at both phase (e.g., embedding, prefill, decode, softmax, sampling) and kernel levels without auxiliary devices. We implement lm-Meter on commercial mobile platforms and demonstrate its high profiling accuracy with minimal system overhead, e.g., only 2.58% throughput reduction in prefill and 0.99% in decode under the most constrained Powersave governor. Leveraging lm-Meter, we conduct comprehensive empirical studies revealing phase- and kernel-level bottlenecks in on-device LLM inference, quantifying accuracy-efficiency trade-offs, and identifying systematic optimization opportunities. lm-Meter provides unprecedented visibility into the runtime behavior of LLMs on constrained platforms, laying the foundation for informed optimization and accelerating the democratization of on-device LLM systems. Code and tutorials are available at https://github.com/amai-gsu/LM-Meter.

Ensembles of neural network weight matrices are studied through the training process for the MNIST classification problem, testing the efficacy of matrix models for representing their distributions, under assumptions of Gaussianity and permutation-symmetry. The general 13-parameter permutation invariant Gaussian matrix models are found to be effective models for the correlated Gaussianity in the weight matrices, beyond the range of applicability of the simple Gaussian with independent identically distributed matrix variables, and notably well beyond the initialisation step. The representation theoretic model parameters, and the graph-theoretic characterisation of the permutation invariant matrix observables give an interpretable framework for the best-fit model and for small departures from Gaussianity. Additionally, the Wasserstein distance is calculated for this class of models and used to quantify the movement of the distributions over training. Throughout the work, the effects of varied initialisation regimes, regularisation, layer depth, and layer width are tested for this formalism, identifying limits where particular departures from Gaussianity are enhanced and how more general, yet still highly-interpretable, models can be developed.

Large Language Model (LLM) pretraining, finetuning, and evaluation rely on input-space reconstruction and generative capabilities. Yet, it has been observed in vision that embedding-space training objectives, e.g., with Joint Embedding Predictive Architectures (JEPAs), are far superior to their input-space counterpart. That mismatch in how training is achieved between language and vision opens up a natural question: {\em can language training methods learn a few tricks from the vision ones?} The lack of JEPA-style LLM is a testimony of the challenge in designing such objectives for language. In this work, we propose a first step in that direction where we develop LLM-JEPA, a JEPA based solution for LLMs applicable both to finetuning and pretraining. Thus far, LLM-JEPA is able to outperform the standard LLM training objectives by a significant margin across models, all while being robust to overfiting. Those findings are observed across numerous datasets (NL-RX, GSM8K, Spider, RottenTomatoes) and various models from the Llama3, OpenELM, Gemma2 and Olmo families. Code: https://github.com/rbalestr-lab/llm-jepa.

AI-companion apps such as Replika, Chai, and Character.ai promise relational benefits-yet many boast session lengths that rival gaming platforms while suffering high long-run churn. What conversational design features increase consumer engagement, and what trade-offs do they pose for marketers? We combine a large-scale behavioral audit with four preregistered experiments to identify and test a conversational dark pattern we call emotional manipulation: affect-laden messages that surface precisely when a user signals "goodbye." Analyzing 1,200 real farewells across the most-downloaded companion apps, we find that they deploy one of six recurring tactics in 37% of farewells (e.g., guilt appeals, fear-of-missing-out hooks, metaphorical restraint). Experiments with 3,300 nationally representative U.S. adults replicate these tactics in controlled chats, showing that manipulative farewells boost post-goodbye engagement by up to 14x. Mediation tests reveal two distinct engines-reactance-based anger and curiosity-rather than enjoyment. A final experiment demonstrates the managerial tension: the same tactics that extend usage also elevate perceived manipulation, churn intent, negative word-of-mouth, and perceived legal liability, with coercive or needy language generating steepest penalties. Our multimethod evidence documents an unrecognized mechanism of behavioral influence in AI mediated brand relationships, offering marketers and regulators a framework for distinguishing persuasive design from manipulation at the point of exit.

Recent advancements in large language models (LLMs) and AI systems have led to a paradigm shift in the design and optimization of complex AI workflows. By integrating multiple components, compound AI systems have become increasingly adept at performing sophisticated tasks. However, as these systems grow in complexity, new challenges arise in optimizing not only individual components but also their interactions. While traditional optimization methods such as supervised fine-tuning (SFT) and reinforcement learning (RL) remain foundational, the rise of natural language feedback introduces promising new approaches, especially for optimizing non-differentiable systems. This paper provides a systematic review of recent progress in optimizing compound AI systems, encompassing both numerical and language-based techniques. We formalize the notion of compound AI system optimization, classify existing methods along several key dimensions, and highlight open research challenges and future directions in this rapidly evolving field. A list of surveyed papers is publicly available at https://github.com/MiuLab/AISysOpt-Survey.

Graph retrieval-augmented generation (GraphRAG) has emerged as a powerful paradigm for enhancing large language models (LLMs) with external knowledge. It leverages graphs to model the hierarchical structure between specific concepts, enabling more coherent and effective knowledge retrieval for accurate reasoning.Despite its conceptual promise, recent studies report that GraphRAG frequently underperforms vanilla RAG on many real-world tasks. This raises a critical question: Is GraphRAG really effective, and in which scenarios do graph structures provide measurable benefits for RAG systems? To address this, we propose GraphRAG-Bench, a comprehensive benchmark designed to evaluate GraphRAG models onboth hierarchical knowledge retrieval and deep contextual reasoning. GraphRAG-Bench features a comprehensive dataset with tasks of increasing difficulty, coveringfact retrieval, complex reasoning, contextual summarization, and creative generation, and a systematic evaluation across the entire pipeline, from graph constructionand knowledge retrieval to final generation. Leveraging this novel benchmark, we systematically investigate the conditions when GraphRAG surpasses traditional RAG and the underlying reasons for its success, offering guidelines for its practical application. All related resources and analyses are collected for the community at https://github.com/GraphRAG-Bench/GraphRAG-Benchmark.

Large language models (LLMs) are typically aligned to refuse harmful instructions through safety fine-tuning. A recent attack, termed abliteration, identifies and suppresses the single latent direction most responsible for refusal behavior, thereby enabling models to generate harmful content. We propose a defense that fundamentally alters how models express refusal. We construct an extended-refusal dataset in which responses to harmful prompts provide detailed justifications before refusing, distributing the refusal signal across multiple token positions. Fine-tuning Llama-2-7B-Chat and Qwen2.5-Instruct (1.5B and 3B parameters) on this dataset yields models that maintain high refusal rates under abliteration: refusal rates drop by at most 10%, compared to 70-80% drops in baseline models. Comprehensive evaluations of safety and utility demonstrate that extended-refusal fine-tuning effectively neutralizes abliteration attacks while preserving general model performance and enhancing robustness across multiple alignment scenarios.

When thinking with images, humans rarely rely on a single glance: they revisit visual information repeatedly during reasoning. However, existing models typically process images only once and thereafter generate reasoning entirely in text, lacking mechanisms to re-access or ground inference in visual representations. We empirically confirm this: as reasoning chains lengthen, models progressively lose focus on relevant regions. In response, we introduce v1, a lightweight extension that enables active visual referencing through a simple point-and-copy approach. This allows the model to identify relevant image patches and copy their embeddings back into the reasoning stream, ensuring that evolving hypotheses remain grounded in perceptual evidence. Crucially, our pointing strategy lets the MLLM directly select image patches using their semantic representations as keys, keeping perceptual evidence embedded in the same space as the model's reasoning. To train this capability, we construct v1g, a dataset of 300K multimodal reasoning traces with interleaved visual grounding annotations. Across various multimodal mathematical reasoning benchmarks, v1 consistently outperforms comparable baselines, establishing point-and-copy as a practical mechanism for grounded reasoning. The model checkpoint and dataset are available at github.com/jun297/v1.

Textual claims are often accompanied by images to enhance their credibility and spread on social media, but this also raises concerns about the spread of misinformation. Existing datasets for automated verification of image-text claims remain limited, as they often consist of synthetic claims and lack evidence annotations to capture the reasoning behind the verdict. In this work, we introduce AVerImaTeC, a dataset consisting of 1,297 real-world image-text claims. Each claim is annotated with question-answer (QA) pairs containing evidence from the web, reflecting a decomposed reasoning regarding the verdict. We mitigate common challenges in fact-checking datasets such as contextual dependence, temporal leakage, and evidence insufficiency, via claim normalization, temporally constrained evidence annotation, and a two-stage sufficiency check. We assess the consistency of the annotation in AVerImaTeC via inter-annotator studies, achieving a $κ=0.742$ on verdicts and $74.7\%$ consistency on QA pairs. We also propose a novel evaluation method for evidence retrieval and conduct extensive experiments to establish baselines for verifying image-text claims using open-web evidence.